repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
interactive-image2video-synthesis
interactive-image2video-synthesis-main/models/discriminator.py
import torch from torch import nn from torch.optim import Adam import functools from torch.nn.utils import spectral_norm import math import numpy as np from utils.general import get_member from models.blocks import SPADE class GANTrainer(object): def __init__(self, config, load_fn,logger,spatial_size=128, parallel=False, devices=None, debug=False, temporal=False, sequence_length = None): self.config = config self.logger = logger # disc self.logger.info("Load discriminator model") self.temporal = temporal if self.temporal: assert sequence_length is not None self.key = "gan_temp" self.disc = resnet(config=config[self.key],spatial_size=spatial_size,sequence_length=sequence_length) self.load_key = "disc_temp" self.postfix = "temp" if self.disc.cond: self.logger.info(f"Using Conditional temporal discriminator.") else: self.key = "gan" self.disc = PatchDiscriminator(self.config[self.key]) self.load_key = "disc_patch" self.postfix = "patch" self.cond = self.config[self.key]["conditional"] if self.temporal and "conditional" in self.config[self.key] else False self.logger.info(f"Number of parameters in discriminator_{self.postfix} is {sum(p.numel() for p in self.disc.parameters())}.") self.parallel = parallel self.devices = devices if self.parallel: assert self.devices is not None # load checkpoint if there's any and it is required disc_ckpt = disc_op_ckpt = None if self.config["general"]["restart"] and not debug: disc_ckpt, disc_op_ckpt = load_fn(key=self.load_key) if disc_ckpt is not None: self.logger.info(f"Resuming training of discriminator...loading weights.") self.disc.load_state_dict(disc_ckpt) if self.parallel: self.disc = nn.DataParallel(self.disc,device_ids=self.devices) self.disc.cuda(self.devices[0]) else: self.disc.cuda() self.logger.info("Discriminator on gpu!") # disc optimizer self.disc_opt = Adam(self.disc.parameters(), lr=self.config["training"]["lr"]) if self.config["general"]["restart"] and disc_op_ckpt is not None: self.disc_opt.load_state_dict(disc_op_ckpt) # scheduler for disc optimizer milestones = [int(self.config["training"]["n_epochs"] * t) for t in self.config["training"]["tau"]] self.disc_scheduler = torch.optim.lr_scheduler.MultiStepLR(self.disc_opt, milestones=milestones, gamma=self.config["training"]["lr_reduce"]) def train_step(self, x_in_true, x_in_fake, cond=None): # predict cond = cond if self.cond else None self.disc.train() # if self.parallel: # x_in_true = x_in_true.cuda(self.devices[0]) # x_in_fake = x_in_fake.cuda(self.devices[0]) # set gradient to zero self.disc_opt.zero_grad() # real examples x_in_true.requires_grad_() pred_true, _ = self.disc(x_in_true, cond) loss_real = get_member(self.disc,"loss")(pred_true, real=True) if self.config[self.key]["gp_weight"] > 0: loss_real.backward(retain_graph=True) # gradient penalty loss_gp = get_member(self.disc,"gp")(pred_true, x_in_true).mean() gp_weighted = self.config[self.key]["gp_weight"] * loss_gp gp_weighted.backward() else: loss_real.backward() # fake examples pred_fake, _ = self.disc(x_in_fake.detach(),cond) loss_fake = get_member(self.disc,"loss")(pred_fake, real=False) loss_fake.backward() # optmize parameters self.disc_opt.step() loss_disc = ((loss_real + loss_fake) / 2.).item() out_dict = {f"loss_disc_{self.postfix}": loss_disc, f"p_true_{self.postfix}": torch.sigmoid(pred_true).mean().item(), f"p_fake_{self.postfix}": torch.sigmoid(pred_fake).mean().item(), f"loss_gp_{self.postfix}": loss_gp.item() if self.config[self.key]["gp_weight"] > 0 else 0 } # train generator pred_fake, fmap_fake = self.disc(x_in_fake,cond) _, fmap_true = self.disc(x_in_true,cond) if get_member(self.disc,"bce_loss"): loss_gen = get_member(self.disc,"bce")(pred_fake, torch.ones_like(pred_fake)) else: loss_gen = -torch.mean(pred_fake) loss_fmap = get_member(self.disc,"fmap_loss")(fmap_fake, fmap_true) # if self.parallel: # loss_fmap = loss_fmap.cuda(self.devices[0]) # loss_gen = loss_gen.cuda(self.devices[0]) return out_dict, loss_gen, loss_fmap # code taken from https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py class PatchDiscriminator(nn.Module): """Defines a PatchGAN discriminator""" def __init__(self, config, norm_layer=nn.InstanceNorm2d): super().__init__() if type(norm_layer) == functools.partial: # no need to use bias as BatchNorm2d has affine parameters use_bias = norm_layer.func == nn.InstanceNorm2d else: use_bias = norm_layer == nn.InstanceNorm2d deep_disc = config["deep_disc"] if "deep_disc" in config else False input_nc = 6 if config["pixel_dynamics"] else 3 n_deep_layers = config["deep_layers"] ndf = 64 n_layers = config["n_layers"] self.bce_loss = config["bce_loss"] if self.bce_loss: self.bce = nn.BCEWithLogitsLoss() kw = 4 padw = 1 self.layers = nn.ModuleList() self.norms = nn.ModuleList() self.in_conv = nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw) nf_mult = 1 nf_mult_prev = 1 self.act_fn = nn.LeakyReLU(0.2, True) for n in range(1, n_layers): # gradually increase the number of filters nf_mult_prev = nf_mult nf_mult = min(2 ** n, 8) self.layers.append(nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=2, padding=padw, bias=use_bias)) self.norms.append(norm_layer(ndf * nf_mult)) nf_mult_prev = nf_mult nf_mult = min(2 ** n_layers, 8) self.layers.append(nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias)) self.norms.append(norm_layer(ndf * nf_mult)) n_d = ndf * nf_mult if deep_disc: n_max = 1024 for i in range(n_deep_layers): # add one layer to the original patch discrminator to make it more powerful self.layers.append(nn.Conv2d(n_d, min(n_max, n_d*2), kernel_size=kw, stride=1, padding=padw, bias=use_bias)) self.norms.append(norm_layer(min(n_max, n_d*2))) n_d = min(n_max, n_d*2) self.out_conv = nn.Conv2d(n_d, 1, kernel_size=kw, stride=1, padding=padw) # output 1 channel prediction map def forward(self, input,cond=None): """Standard forward.""" x = self.act_fn(self.in_conv(input)) fmap = [] for i in range(len(self.layers)): x = self.layers[i](x) x = self.act_fn(self.norms[i](x)) fmap.append(x) x = self.out_conv(x) return x, fmap def loss(self, pred, real): if self.bce_loss: # vanilla gan loss return self.bce(pred, torch.ones_like(pred) if real else torch.zeros_like(pred)) else: # hinge loss if real: l = torch.mean(torch.nn.ReLU()(1.0 - pred)) else: l = torch.mean(torch.nn.ReLU()(1.0 + pred)) return l def gp(self, pred_fake, x_fake): batch_size = x_fake.size(0) grad_dout = torch.autograd.grad( outputs=pred_fake.sum(), inputs=x_fake, create_graph=True, retain_graph=True, only_inputs=True )[0] grad_dout2 = grad_dout.pow(2) assert (grad_dout2.size() == x_fake.size()) reg = grad_dout2.view(batch_size, -1).sum(1) return reg def fmap_loss(self, fmap1, fmap2, loss="l1"): recp_loss = 0 for idx in range(len(fmap1)): if loss == "l1": recp_loss += torch.mean(torch.abs((fmap1[idx] - fmap2[idx]))) if loss == "l2": recp_loss += torch.mean((fmap1[idx] - fmap2[idx]) ** 2) return recp_loss / len(fmap1) ###################################################################################################### ###3D-ConvNet Implementation from https://github.com/tomrunia/PyTorchConv3D ########################## def resnet10(**kwargs): """Constructs a ResNet-10 model. """ model = ResNet(BasicBlock, [1, 1, 1, 1], **kwargs) return model def resnet(**kwargs): """Constructs a ResNet-18 model. """ model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs) return model def resnet34(**kwargs): """Constructs a ResNet-34 model. """ model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs) return model def conv3x3x3(in_planes, out_planes, stride=1, stride_t=1): # 3x3x3 convolution with padding return spectral_norm(nn.Conv3d( in_planes, out_planes, kernel_size=(3, 3, 3), stride=[stride_t, stride, stride], padding=[1, 1, 1], bias=False)) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, stride_t=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3x3(inplanes, planes, stride, stride_t) self.bn1 = nn.GroupNorm(num_groups=16, num_channels=planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3x3(planes, planes) self.bn2 = nn.GroupNorm(num_groups=16, num_channels=planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers, spatial_size, sequence_length, config): super(ResNet, self).__init__() # spatial_size = config["spatial_size"] self.inplanes = 64 self.bce_loss = config["bce_loss"] min_spatial_size = int(spatial_size / 8) #sample_duration = dic.Network['sequence_length']-1 self.max_channels = config["max_channels"] if "max_channels" in config else 256 self.conv1 = spectral_norm(nn.Conv3d( 3, 64, kernel_size=(3, 7, 7), stride=(1, 2, 2), padding=(1, 3, 3), bias=False)) self.gn1 = nn.GroupNorm(num_groups=16, num_channels=64) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool3d(kernel_size=(3, 3, 3), stride=(1, 2, 2), padding=1) self.layers = nn.ModuleList() self.patch_temp = config["patch_temp_disc"] self.spatio_temporal = config["spatio_temporal"] if"spatio_temporal" in config else False if self.patch_temp: self.layer1 = self._make_layer(block, 64, layers[0], stride=1) self.layers.append(self._make_layer(block, 128, layers[1], stride=1, stride_t=1)) self.layers.append(self._make_layer(block, 128, layers[2], stride=2, stride_t=1)) self.layers.append(self._make_layer(block, 256, layers[3], stride=2, stride_t=1)) last_size = int(math.ceil(spatial_size / 16)) last_duration = 1 self.avgpool = nn.AvgPool3d((last_duration, last_size, last_size), stride=1) self.cond = config["conditional"] if "conditional" in config else False if self.cond: self.spade_emb = SPADE(norm_nc=block.expansion * 256, label_nc=2, config=config) self.fc = nn.Linear(256 * block.expansion, config["num_classes"], bias=False) else: spatial_size /= 2 self.layer1 = self._make_layer(block, 32, layers[0], stride=1) n_channels = 64 if "conditional" in config and config["conditional"]: raise ValueError("If non-patch-gan temporal discriminator is used, conditional must not be True!") self.cond = False n = 0 while sequence_length > 1: blocks = layers[n] if n<sequence_length-1 else layers[-1] n_channels = min(2*n_channels,self.max_channels) stride = 1 if spatial_size <= min_spatial_size else 2 spatial_size = int(spatial_size / stride) stride_t = 1 if self.spatio_temporal else (2 if sequence_length > 1 else 1) self.layers.append(self._make_layer(block,n_channels,blocks,stride=stride,stride_t=stride_t)) sequence_length = int(math.ceil(sequence_length / 2)) n += 1 self.final = nn.Conv2d(n_channels,1,3,padding=1) print(f"Temporal discriminator has {len(self.layers)} layers") for m in self.modules(): if isinstance(m, nn.Conv3d): m.weight = nn.init.orthogonal_(m.weight) def _make_layer(self, block, planes, blocks, stride=1, stride_t=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion or stride_t != 1: downsample = nn.Sequential( spectral_norm(nn.Conv3d( self.inplanes, planes * block.expansion, kernel_size=[3, 3, 3], stride=[stride_t, stride, stride], padding=[1, 1, 1], bias=False)), nn.GroupNorm(num_channels=planes * block.expansion, num_groups=16)) layers = [] layers.append(block(self.inplanes, planes, stride, stride_t, downsample)) self.inplanes = planes * block.expansion for _ in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x, cond=None): out = [] x = self.conv1(x) x = self.gn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) out.append(x) for n in range(len(self.layers)): x = self.layers[n](x) out.append(x) if self.patch_temp: if self.cond: x_norm = [] for i in range(x.size(2)): x_norm.append(self.spade_emb(x[:,:,i],cond)) x_norm = torch.stack(x_norm,2) else: x_norm = x x1 = self.avgpool(x_norm) output = [] for i in range(x1.size(2)): output.append(self.fc(x1[:,:,i].reshape(x1.size(0), -1))) return torch.cat(output, dim=1), out else: output = self.final(x.squeeze(2)) return output, out def loss(self, pred, real): if self.bce_loss: # vanilla gan loss return self.bce(pred, torch.ones_like(pred) if real else torch.zeros_like(pred)) else: # hinge loss if real: l = torch.mean(torch.nn.ReLU()(1.0 - pred)) else: l = torch.mean(torch.nn.ReLU()(1.0 + pred)) return l def gp(self, pred_fake, x_fake): batch_size = x_fake.size(0) grad_dout = torch.autograd.grad( outputs=pred_fake.sum(), inputs=x_fake, create_graph=True, retain_graph=True, only_inputs=True )[0] grad_dout2 = grad_dout.pow(2) assert (grad_dout2.size() == x_fake.size()) reg = grad_dout2.view(batch_size, -1).sum(1) return reg def fmap_loss(self, fmap1, fmap2, loss="l1"): recp_loss = 0 for idx in range(len(fmap1)): if loss == "l1": recp_loss += torch.mean(torch.abs((fmap1[idx] - fmap2[idx]))) if loss == "l2": recp_loss += torch.mean((fmap1[idx] - fmap2[idx]) ** 2) return recp_loss / len(fmap1) # return output, out, mu if __name__ == '__main__': ## Test 3dconvnet with dummy input import os os.environ["CUDA_VISIBLE_DEVICES"] = '7' config = {"num_classes": 1, "patch_temp_disc": True,"spatial_size": 128, "bce_loss": False, "conditional": True} dummy = torch.rand((2, 3, 6, 128, 128)).cuda() dummy_cond = torch.rand((2, 2, 128, 128)).cuda() model = resnet(config=config,spatial_size=128, sequence_length=dummy.shape[2]).cuda() print("Number of parameters in generator", sum(p.numel() for p in model.parameters())) if config["conditional"]: out, out2 = model(dummy,dummy_cond) else: out, out2,= model(dummy) test = 1
17,349
37.988764
191
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/models/latent_flow_net.py
import torch from torch import nn from torch.nn import functional as F import numpy as np import math from models.blocks import Conv2dBlock, ResBlock, AdaINLinear, NormConv2d,ConvGRU class OscillatorModel(nn.Module): def __init__(self,spatial_size,config,n_no_motion=2, logger=None): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.reparamterize = config["reparameterize_poke"] if "reparameterize_poke" in config else False self.norm_layer = config["norm_layer"] if "norm_layer" in config else "in" self.layers = config["layers"] self.n_gru_layers = config["n_gru_layers"] if "n_gru_layers" in config else 3 self.n_no_motion = n_no_motion self.n_stages = len(self.layers) assert self.n_no_motion is not None nf_first_shape_enc = int(max(32, config["nf_deep"] / (2 ** self.n_stages))) self.shape_enc = SkipConnectionEncoder(nf_in=3, nf_max=self.layers[-1], n_stages=self.n_stages, n_skip_stages=self.n_stages ,nf_first=nf_first_shape_enc, norm_layer=self.norm_layer,layers=self.layers) self.dynamics_enc = Encoder(2, nf_max=self.layers[-1], n_stages=self.n_stages, variational=self.reparamterize, norm_layer=self.norm_layer, layers=self.layers) ups = [False] * self.n_gru_layers #input_sizes = [self.layers[-1]*2] + (len(self.layers)-1) * [self.layers[-1]] self.fusion_block = ConvGRU(input_size=self.layers[-1] * 2, hidden_sizes=self.layers[-1], kernel_sizes=3, n_layers=self.n_gru_layers, upsampling=ups,) self.dec = SkipConnectionDecoder(nf_in=self.layers[-1], in_channels=self.shape_enc.depths, n_skip_stages=len(self.layers), disentanglement=False, norm_layer=self.norm_layer, layers=self.layers) if logger is not None: logger.info("Constructed OscillatorModel") logger.info(f"Layers of OscillatorModel is {self.layers}") logger.info(f"Encoder channels of oscillator model is {self.layers}") def forward(self,input_img, poke, len,n_ref,target_img=None): imgs = [] sigmas_hat_out = [] if target_img==None: target_img = input_img if self.reparamterize: delta, mu, _ = self.dynamics_enc(poke) else: delta = self.dynamics_enc(poke)[0] # only first time shape encoding # if self.poke_scale_mode and not poke_linear: sigmas = self.shape_enc(input_img) sigma_tgt = self.shape_enc(target_img)[-1] sigma_dyn = sigmas.pop() pred = [sigma_dyn] * self.n_gru_layers pred_out = pred[-1] for n in range(len): # apply fusion block: input is delta, hidden states are the sigma_n # get inputs for network delta_in = delta * (1. - float(n)/(n_ref-1)) if n <= n_ref else torch.zeros_like(delta) if self.training: sigma_diff = pred_out - sigma_tgt if n < len - self.n_no_motion else torch.zeros_like(pred_out) else: sigma_diff = pred_out - sigma_tgt delta_in = torch.cat([delta_in,sigma_diff],1) # predict object encoding at next time step pred = self.fusion_block(delta_in, pred) pred_out = pred[-1] sigmas.append(pred_out) # decode x = self.dec(sigmas, [], del_shape=False) imgs.append(x) sigmas_hat_out.append(pred_out) #remove pred sigmas.pop() imgs = torch.stack(imgs, dim=1) #sigmas_hat_out[-1].reverse() return imgs, sigmas_hat_out class SAVPArchModel(nn.Module): def __init__(self, spatial_size, config): super().__init__() self.n_stages = int(np.log2(spatial_size[0] // 16)) self.poke_every_t = config["poke_every_t"] if "poke_every_t" in config else True self.dynamics_enc = Encoder(nf_in=2, nf_max=64, n_stages=self.n_stages) self.gen = SAVPGenerator(self.poke_every_t) def forward(self,img,poke,len): # encode dynamics delta = self.dynamics_enc(poke)[0] out = self.gen(img,delta,len) return out class SAVPGenerator(nn.Module): def __init__(self, poke_every): super().__init__() self.poke_every_t = poke_every # encoder stuff self.conv_e1 = Conv2dBlock(3, 32, 3, 2, norm="in", padding=1, activation="relu") # ssize 32 self.rnn_e1 = ConvGRU(32, 32, 3, 1) self.conv_e2 = Conv2dBlock(32, 64, 3, 2, norm="in", padding=1, activation="relu") # ssize 16 self.rnn_e2 = ConvGRU(64, 64, 3, 1) # bottleneck self.conv_bn = Conv2dBlock(128, 128, 3, 2, norm="in", padding=1, activation="relu") # ssize 8 self.rnn_bn = ConvGRU(128, 64, 3, 1) # decoder stuff self.up1 = nn.Upsample((16, 16), mode="bilinear") # ssize 16 self.conv_d1 = Conv2dBlock(128, 64, 3, 1, norm="in", padding=1, activation="relu") self.rnn_d1 = ConvGRU(64, 32, 3, 1) self.up2 = nn.Upsample((32, 32), mode="bilinear") # ssize 32 self.conv_d2 = Conv2dBlock(64, 32, 3, 1, norm="in", padding=1, activation="relu") self.rnn_d2 = ConvGRU(32, 32, 3, 1) self.up3 = nn.Upsample((64, 64), mode="bilinear") self.conv_out = Conv2dBlock(32, 3, 3, 1, 1, norm="none", activation="tanh") def forward(self,img,delta,len): x = img out_imgs = [] for t in range(len): x1e = self.conv_e1(x) x1er = self.rnn_e1(x1e,[x1er] if t>0 else None)[0] x2e = self.conv_e2(x1er) x2er = self.rnn_e2(x2e,[x2er] if t>0 else None)[0] if t > 0 and not self.poke_every_t: delta = np.zeros_like(delta) xbn = torch.cat([x2er,delta],dim=1) xbn = self.conv_bn(xbn) xbnr = self.rnn_bn(xbn,[xbnr] if t>0 else None)[0] x1d = self.up1(xbnr) x1d = self.conv_d1(torch.cat([x1d,x2er],1)) x1dr = self.rnn_d1(x1d,[x1dr] if t>0 else None)[0] x2d = self.up2(x1dr) x2d = self.conv_d2(torch.cat([x2d,x1er],1)) x2dr = self.rnn_d2(x2d,[x2dr] if t > 0 else None)[0] x = self.conv_out(self.up3(x2dr)) out_imgs.append(x) out_imgs = torch.stack(out_imgs,1) return out_imgs class ForegroundBackgroundModel(nn.Module): def __init__(self, spatial_size,config): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.cat_poke_img = config["poke_and_img"] self.zeroflow_baseline = config["zeroflow_baseline"] self.variational = config["variational"] if "variational" in config else False foreground_background_div = config["foreground_background_div"] assert foreground_background_div >= 1. nf_first_shape_enc = int(max(32, config["nf_deep"] / (2 ** self.n_stages))) if self.variational: self.shape_enc = VariationalSkipConnectionEncoderFGBG(nf_in=3,nf_max=config["nf_deep"],n_stages=self.n_stages, n_skip_stages=self.n_stages, nf_first=nf_first_shape_enc) else: self.shape_enc = SkipConnectionEncoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, n_skip_stages=self.n_stages, nf_first=nf_first_shape_enc, fg_bg=True, div=foreground_background_div) self.dynamics_enc = Encoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages) hidden_sizes = [int(config["nf_deep"]/foreground_background_div)] + [int(d/foreground_background_div) for d in self.shape_enc.depths] ups = [False] + self.shape_enc.downs self.fusion_block = ConvGRU(input_size=config["nf_deep"], hidden_sizes=hidden_sizes, kernel_sizes=3, n_layers=self.n_stages + 1, upsampling=ups) self.dec = SkipConnectionDecoder(nf_in=config["nf_deep"], in_channels=self.shape_enc.depths, n_skip_stages=self.n_stages, disentanglement=False) def forward(self,fg_img,bg_img,poke,len): x = fg_img mus = logstds = None if len > 0: if self.zeroflow_baseline: poke = torch.zeros_like(poke) if self.cat_poke_img: poke = torch.cat([poke, x], dim=1) imgs = [] sigmas_fg = [] sigmas_bg = [] # infer dynamics input delta = self.dynamics_enc(poke)[0] # only first time shape encoding sigma_n = self.shape_enc(x)[0] sigma_bg = self.shape_enc(bg_img)[1] for n in range(len): # apply fusion block: input is delta, hidden states are the sigma_n sigma_n.reverse() sigma_n = self.fusion_block(delta, sigma_n) sigma_n.reverse() sigma_cat = [torch.cat([sfg,sbg],dim=1) for sfg,sbg in zip(sigma_n,sigma_bg)] x = self.dec(sigma_cat, None, del_shape=True) imgs.append(x) # output foreground representation sigmas_fg.append(sigma_n) # out sigmas_bg.append(sigma_bg) imgs = torch.stack(imgs, dim=1) sigmas_fg[-1].reverse() else: if self.variational: sigmas_fg, sigmas_bg1, mus, logstds = self.shape_enc(x) _, sigmas_bg2, *_ = self.shape_enc(bg_img) else: sigmas_fg, sigmas_bg1 = self.shape_enc(x) _, sigmas_bg2 = self.shape_enc(bg_img) sigmas_bg = (sigmas_bg1,sigmas_bg2) sigmas = [torch.cat([sfg,sbg],dim=1) for sfg,sbg in zip(sigmas_fg,sigmas_bg2)] imgs = self.dec(sigmas, None, del_shape=True) sigmas_fg.reverse() return imgs, sigmas_fg, sigmas_bg, mus, logstds class SkipSequenceModel(nn.Module): def __init__(self,spatial_size,config,n_no_motion=None): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.min_spatial_size = config["min_spatial_size"] if "min_spatial_size" in config else 16 self.n_stages = int(np.log2(spatial_size[0] // self.min_spatial_size)) print(f"number of stages in model is {self.n_stages}") self.disentanglement = config["disentanglement"] if "disentanglement" in config else False self.cat_poke_img = config["poke_and_img"] self.zeroflow_baseline = config["zeroflow_baseline"] self.poke_every_t = config["poke_every_t"] if "poke_every_t" in config else True use_spectral_norm = config["spectnorm_decoder"] if "spectnorm_decoder" in config else False self.reparamterize = config["reparameterize_poke"] if "reparameterize_poke" in config else False self.norm_layer = config["norm_layer"] if "norm_layer" in config else "in" self.layers = config["layers"] if "layers" in config and len(config["layers"])>0 else None self.poke_scale_mode = config["poke_scale"] if "poke_scale" in config else False # self.n_no_motion = n_no_motion # if self.poke_scale_mode: # assert self.n_no_motion is not None # self.multiscale_fusion_block = config["multiscale_dynamics"] # default is dynamics model # if self.disentanglement: # self.appearance_enc = Encoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, prepare_adain=True, # resnet_down=config["resnet_down"] if "resnet_down" in config else False) # n_skip_stages = min(config["n_skip_stages"], self.n_stages) if "n_skip_stages" in config else self.n_stages nf_first_shape_enc = int(max(32, config["nf_deep"] / (2 ** self.n_stages))) self.shape_enc = SkipConnectionEncoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, n_skip_stages=self.n_stages ,nf_first=nf_first_shape_enc, norm_layer=self.norm_layer,layers=self.layers) self.dynamics_enc = Encoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages, variational=self.reparamterize, norm_layer=self.norm_layer, layers=self.layers) hidden_sizes = [config["nf_deep"]]+self.shape_enc.depths ups = [False] + self.shape_enc.downs self.fusion_block = ConvGRU(input_size=config["nf_deep"], hidden_sizes=hidden_sizes, kernel_sizes=3, n_layers=self.n_stages+1 if self.layers is None else len(self.layers)+1, upsampling=ups,) self.dec = SkipConnectionDecoder(nf_in=config["nf_deep"], in_channels=self.shape_enc.depths, n_skip_stages=self.n_stages if self.layers is None else len(self.layers), disentanglement=False, spectral_norm=use_spectral_norm, norm_layer=self.norm_layer, layers=self.layers) def forward(self, app_img, shape_img, poke, len, poke_linear=False,delta_scaling = None, n_zero_frames=0, invert_poke=False, poke_jump=False): # if self.disentanglement: # alpha, *_ = self.appearance_enc(app_img) # else: # alpha = None # sigma = self.shape_enc(shape_img) x = shape_img if len > 0: if self.zeroflow_baseline: poke = torch.zeros_like(poke) if self.cat_poke_img: poke = torch.cat([poke, app_img], dim=1) imgs = [] sigmas_hat_out = [] sigmas_out = [] # infer dynamics input if self.reparamterize: delta, mu, _ = self.dynamics_enc(poke) else: delta = self.dynamics_enc(poke)[0] sigmas_out = delta # only first time shape encoding #if self.poke_scale_mode and not poke_linear: sigma_n = self.shape_enc(x) for n in range(len): # apply fusion block: input is delta, hidden states are the sigma_n sigma_n.reverse() if self.poke_scale_mode: if poke_linear: if invert_poke: delta_in = delta * (1 - float(n) / int(len/2)) if n < int(len/2) else delta * (float(n - int(len/2)) / int(math.ceil(float(len)/2)) - 1) else: delta_in = (1 - float(n) / (len-n_zero_frames)) * delta if n <= len - n_zero_frames else torch.zeros_like(delta) else: delta_in = delta_scaling[n] * delta_in else: if poke_jump: delta_in = delta if n < len -n_zero_frames else torch.zeros_like(delta) else: delta_in = delta if self.poke_every_t else (delta if n == 0 else torch.zeros_like(delta)) sigma_n = self.fusion_block(delta_in, sigma_n) sigma_n.reverse() x = self.dec(sigma_n, [], del_shape=False) imgs.append(x) sigmas_hat_out.append(sigma_n) imgs = torch.stack(imgs, dim=1) sigmas_hat_out[-1].reverse() else: sigmas = self.shape_enc(x) sigmas_out = sigmas sigmas_hat_out = None imgs = self.dec(sigmas, [], del_shape=False) sigmas_out.reverse() return imgs, sigmas_out, sigmas_hat_out, [] class SingleScaleBaseline(nn.Module): def __init__(self, spatial_size, config, n_no_motion=None): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.disentanglement = config["disentanglement"] if "disentanglement" in config else False self.cat_poke_img = config["poke_and_img"] self.zeroflow_baseline = config["zeroflow_baseline"] self.poke_scale_mode = config["poke_scale"] if "poke_scale" in config else False self.poke_every_t = config["poke_every_t"] if "poke_every_t" in config else True self.n_no_motion = n_no_motion if self.poke_scale_mode: assert self.n_no_motion is not None print("Initialize SingleScaleBaseline") # self.multiscale_fusion_block = config["multiscale_dynamics"] # default is dynamics model if self.disentanglement: self.appearance_enc = Encoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, prepare_adain=True, resnet_down=config["resnet_down"] if "resnet_down" in config else False) self.shape_enc = SkipConnectionEncoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, n_skip_stages=0) self.dynamics_enc = Encoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages) self.n_gru_layers = 3 self.fusion_block = ConvGRU(input_size=config["nf_deep"], hidden_sizes=config["nf_deep"], kernel_sizes=3, n_layers=config["n_gru_layers"]) self.dec = SkipConnectionDecoder(nf_in=config["nf_deep"], in_channels=self.shape_enc.depths, n_skip_stages=0, disentanglement=self.disentanglement) def forward(self, app_img, shape_img, poke, len, poke_linear=False,delta_scaling = None, n_zero_frames=0, invert_poke=False,poke_jump=False): if self.disentanglement: alpha, *_ = self.appearance_enc(app_img) else: alpha = None # sigma = self.shape_enc(shape_img) if self.zeroflow_baseline: poke = torch.zeros_like(poke) if self.cat_poke_img: poke = torch.cat([poke, app_img], dim=1) x = shape_img if len > 0: imgs = [] sigmas_hat_out = [] sigmas_out = [] # infer dynamics input delta = self.dynamics_enc(poke)[0] sigma_n = self.shape_enc(x)[0] sigma_n = torch.stack([sigma_n] * self.n_gru_layers) for n in range(len): # delta scaling delta_in = delta if self.poke_every_t else (delta if n == 0 else torch.zeros_like(delta)) if self.poke_scale_mode: if poke_linear: if invert_poke: delta_in = delta * (1 - float(n) / int(len / 2)) if n < int(len / 2) else delta * (float(n - int(len / 2)) / int(math.ceil(float(len) / 2)) - 1) else: delta_in = (1 - float(n) / (len - n_zero_frames)) * delta if n <= len - n_zero_frames else torch.zeros_like(delta) else: delta_in = delta_scaling[n] * delta_in # apply fusion block sigma_n = self.fusion_block(delta_in, sigma_n) # residual connection sigma_n1 = sigma_n[-1] x = self.dec([sigma_n1], alpha, del_shape=False) imgs.append(x) sigmas_hat_out.append(sigma_n1) #sigmas_hat_out = torch.stack(sigmas_hat_out) imgs = torch.stack(imgs, dim=1) else: sigmas = self.shape_enc(x) sigmas_out = sigmas[-1] sigmas_hat_out = None imgs = self.dec(sigmas, alpha, del_shape=False) return imgs, sigmas_out, sigmas_hat_out, alpha class ResidualSequenceBaseline(nn.Module): def __init__(self,spatial_size,config): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.disentanglement = config["disentanglement"] if "disentanglement" in config else False self.cat_poke_img = config["poke_and_img"] self.zeroflow_baseline = config["zeroflow_baseline"] #self.multiscale_fusion_block = config["multiscale_dynamics"] # default is dynamics model if self.disentanglement: self.appearance_enc = Encoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, prepare_adain=True, resnet_down=config["resnet_down"] if "resnet_down" in config else False) self.shape_enc = SkipConnectionEncoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, n_skip_stages=0) self.dynamics_enc = Encoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages) self.n_gru_layers = 3 self.fusion_block = ConvGRU(input_size=config["nf_deep"], hidden_sizes=config["nf_deep"],kernel_sizes=3, n_layers=3) self.dec = SkipConnectionDecoder(nf_in=config["nf_deep"], in_channels=self.shape_enc.depths, n_skip_stages=0, disentanglement=self.disentanglement) def forward(self,app_img,shape_img,poke,len): if self.disentanglement: alpha, *_ = self.appearance_enc(app_img) else: alpha = None #sigma = self.shape_enc(shape_img) if self.zeroflow_baseline: poke = torch.zeros_like(poke) if self.cat_poke_img: poke = torch.cat([poke,app_img],dim=1) x = shape_img if len>0: imgs = [] sigmas_hat_out = [] sigmas_out = [] # infer dynamics input delta = self.dynamics_enc(poke)[0] delta = torch.stack([delta]*self.n_gru_layers) for n in range(len): # shape encoding sigma_n = self.shape_enc(x)[0] # apply fusion block delta = self.fusion_block(sigma_n,delta) # residual connection sigma_n1 = sigma_n + delta[-1] x = self.dec([sigma_n1],alpha) imgs.append(x) sigmas_hat_out.append(sigma_n1) sigmas_hat_out = torch.stack(sigmas_hat_out,) imgs = torch.stack(imgs,dim=1) else: sigmas = self.shape_enc(x) sigmas_out = sigmas[-1] sigmas_hat_out = None imgs = self.dec(sigmas,alpha,del_shape=False) return imgs, sigmas_out, sigmas_hat_out, alpha class DynamicSkipModel(nn.Module): def __init__(self, spatial_size,config): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.disentanglement = config["disentanglement"] if "disentanglement" in config else False self.cat_poke_img = config["poke_and_img"] self.zeroflow_baseline = config["zeroflow_baseline"] self.multiscale_fusion_block = config["multiscale_dynamics"] # default is dynamics model if self.disentanglement: self.appearance_enc = Encoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, prepare_adain=True, resnet_down=config["resnet_down"] if "resnet_down" in config else False) n_skip_stages = min(config["n_skip_stages"],self.n_stages) if "n_skip_stages" in config else self.n_stages self.shape_enc = SkipConnectionEncoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages,n_skip_stages=n_skip_stages) if config["multiscale_dynamics"]: self.dynamics_enc = SkipConnectionEncoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages, n_skip_stages=n_skip_stages) self.fusion_block = FusionBlockMultiscale(nf_in=config["nf_deep"],nfs=self.shape_enc.depths,n_blocks=config["n_blocks"]) else: self.dynamics_enc = Encoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages) self.fusion_block = LearnedFusionBlock(nf=config["nf_deep"], n_blocks=config["n_blocks"]) self.dec = SkipConnectionDecoder(nf_in=config["nf_deep"],in_channels=self.shape_enc.depths, n_skip_stages=n_skip_stages,disentanglement=self.disentanglement) def forward(self,app_img,shape_img,poke, apply_dynamics = False): if self.disentanglement: alpha, *_ = self.appearance_enc(app_img) else: alpha = None sigma = self.shape_enc(shape_img) if self.zeroflow_baseline: poke = torch.zeros_like(poke) if self.cat_poke_img: poke = torch.cat([poke,app_img],dim=1) delta = self.dynamics_enc(poke) if self.multiscale_fusion_block else self.dynamics_enc(poke)[0] if apply_dynamics: sigma_hat = self.fusion_block(sigma if self.multiscale_fusion_block else sigma.pop(), delta) if self.multiscale_fusion_block: sigma_in_dec = sigma_hat else: sigma_in_dec = sigma + [sigma_hat] sigma_out = sigma_hat else: sigma_out = sigma if self.multiscale_fusion_block else sigma[-1] img = self.dec(sigma_in_dec if apply_dynamics else sigma,alpha,del_shape=False) return img, sigma_out, alpha class DisentangledModelWithoutDynamics(nn.Module): def __init__(self, spatial_size,config): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.adain = config["adain"] # default is dynamics model self.latent_fusion = config["latent_fusion"] if "latent_fusion" in config else None self.appearance_enc = Encoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, prepare_adain=self.adain, resnet_down=config["resnet_down"] if "resnet_down" in config else False) self.shape_enc = Encoder(nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages, variational=config["ib_shape"]) if self.adain: self.dec = AdaINDecoderDisentangled(nf_in=config["nf_deep"], n_stages=self.n_stages, latent_fusion=self.latent_fusion, nf_in_bn=self.appearance_enc.nf_in_bn) else: self.dec = DecoderEntangled(nf_in=2 * config["nf_deep"],n_stages=self.n_stages) def forward(self,app_img,shape_img): # appearance representation alpha, alpha_spatial, *_ = self.appearance_enc(app_img) # shape representation sigma, shape_mean, shape_logstd = self.shape_enc(shape_img) # decode img = self.dec(alpha, sigma, alpha_spatial) return img, alpha, sigma, shape_mean,shape_logstd class BasicDisentangledModel(nn.Module): def __init__(self, spatial_size,config): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.zero_flow_baseline = config["zero_flow_baseline"] self.adain = config["adain"] self.cat_poke_img = config["poke_and_img"] # default is dynamics model self.latent_fusion = config["latent_fusion"] if "latent_fusion" in config else None self.appearance_enc = Encoder(nf_in=3,nf_max = config["nf_deep"], n_stages=self.n_stages, prepare_adain=self.adain, resnet_down=config["resnet_down"] if "resnet_down" in config else False) self.shape_enc = Encoder(nf_in=3, nf_max = config["nf_deep"],n_stages=self.n_stages, variational=config["ib_shape"]) self.dynamics_enc = Encoder(nf_in=5 if self.cat_poke_img else 2, nf_max=config["nf_deep"], n_stages=self.n_stages) self.fusion_block = LearnedFusionBlock(nf=config["nf_deep"],n_blocks=config["n_blocks"]) if self.adain: self.dec = AdaINDecoderDisentangled(nf_in=config["nf_deep"],n_stages=self.n_stages,latent_fusion=self.latent_fusion,nf_in_bn=self.appearance_enc.nf_in_bn) else: self.dec = DecoderEntangled(nf_in=2 * config["nf_deep"],n_stages=self.n_stages) def forward(self, app_img, shape_img, poke, apply_dynamics = False): # appearance representation alpha, alpha_spatial, *_ = self.appearance_enc(app_img) # shape representation sigma, shape_mean, shape_logstd = self.shape_enc(shape_img) # dynamics representation if self.zero_flow_baseline: poke = torch.zeros_like(poke) if self.cat_poke_img: poke = torch.cat([poke,app_img],dim=1) delta, *_ = self.dynamics_enc(poke) if self.zero_flow_baseline: delta = torch.zeros_like(delta) # apply dynamics to shape represenation sigma_hat = self.fusion_block(sigma,delta) # decode if apply_dynamics: img = self.dec(alpha,sigma_hat,alpha_spatial) else: img = self.dec(alpha,sigma,alpha_spatial) return img, sigma, sigma_hat, alpha, delta, shape_mean, shape_logstd class LearnedFusionBlock(nn.Module): def __init__(self,nf,n_blocks): super().__init__() assert n_blocks >= 1 blocks = [ResBlock(2*nf,nf)] for i in range(1,n_blocks): blocks.append(ResBlock(nf,nf)) self.model = nn.Sequential(*blocks) def forward(self,sigma,delta): x = torch.cat([sigma,delta],dim=1) x = self.model(x) return x class BasicModel(nn.Module): def __init__(self, spatial_size, config): super().__init__() # number of downsampling layers; always such that spatial bottleneck size is 16x16 self.n_stages = int(np.log2(spatial_size[0] // 16)) self.zero_flow_baseline = config["zero_flow_baseline"] self.adain = config["adain"] self.obj_enc = Encoder( nf_in=3, nf_max=config["nf_deep"], n_stages=self.n_stages,variational=config["variational"]) self.flow_enc = Encoder( nf_in=2, nf_max=config["nf_deep"], n_stages=self.n_stages, prepare_adain=self.adain ) if self.adain: self.dec = AdaINDecoderEntangled( nf_in=config["nf_deep"], n_stages=self.n_stages,latent_fusion=config["latent_fusion"] ) else: self.dec = DecoderEntangled(nf_in=2*config["nf_deep"],n_stages=self.n_stages) def forward(self, image, flow,sample_prior=False): # get object code and variational paramezers if model is variational object_code, mean, logstd = self.obj_enc(image,sample_prior) # get dynamics codes dynamics_code1, dynamics_code2, _ = self.flow_enc(flow) if self.zero_flow_baseline: # this is without flow usage, to measure the impacts of the flow as adain input dynamics_code1 = torch.zeros_like(dynamics_code1) dynamics_code2 = torch.zeros_like(dynamics_code2) # decode if self.adain: img = self.dec(object_code, dynamics_code1,dynamics_code2) else: img = self.dec(object_code,dynamics_code1) return img, object_code, dynamics_code1, mean, logstd class VariationalSkipConnectionEncoderFGBG(nn.Module): def __init__(self,nf_in,nf_max, n_stages, n_skip_stages, act = "relu", nf_first=None): super().__init__() self.blocks = nn.ModuleList() self.n_stages = n_stages self.depths = [] self.downs = [] if nf_first is None: nf = 64 else: nf = nf_first # required self.blocks.append( NormConv2d( nf_in, int(1.5 * nf), 3, 2, padding=1 ) ) self.n_skip_stages = n_skip_stages self.depths.append(nf) for n in range(self.n_stages - 1): self.blocks.append( NormConv2d( nf, min(nf * 3, int(1.5*nf_max)), 3, 2, padding=1, ) ) nf = min(nf * 2, nf_max) self.depths.insert(0, nf) self.downs.insert(0, True) self.bottleneck = ResBlock(nf, int(1.5 * nf_max), activation=act, stride=1) self.downs.insert(0, False) self.squash = nn.Sigmoid() def _reparameterize(self,codes): mu = codes[:,:int(codes.shape[1]/2)] logstd = codes[:,int(codes.shape[1]/2):] logstd = self.squash(logstd) std = torch.exp(logstd) eps = torch.randn_like(std) return eps.mul(std) + mu, mu, logstd def forward(self,x): out_fg = [] out_bg = [] out_logstd = [] out_mu = [] for i in range(self.n_stages): x = self.blocks[i](x) if i >= self.n_stages - self.n_skip_stages: act_div = int(x.shape[1] * 2. / 3.) sample, mu, logstd = self._reparameterize(x[:,:act_div]) out_fg.append(sample) bg = x[:,act_div:] out_bg.append(bg) out_mu.append(mu) out_logstd.append(logstd) x = torch.cat([mu,bg], dim=1) x = self.bottleneck(x) act_div = int(x.shape[1] * 2. / 3.) sample, mu, logstd = self._reparameterize(x[:, :act_div]) out_fg.append(sample) bg = x[:, act_div:] out_bg.append(bg) out_mu.append(mu) out_logstd.append(logstd) return out_fg, out_bg, out_mu, out_logstd class SkipConnectionEncoder(nn.Module): def __init__(self,nf_in,nf_max, n_stages, n_skip_stages, act = "elu", nf_first=None, fg_bg = False, div= None, norm_layer="in", layers=None): super().__init__() self.blocks = nn.ModuleList() self.n_stages = n_stages if layers is None else len(layers) self.depths = [] self.downs = [] if nf_first is None: nf = 32 else: nf = nf_first if layers is not None: nf = layers[0] self.fg_bg = fg_bg if self.fg_bg: assert div is not None self.div = div self.blocks.append( Conv2dBlock( nf_in, nf, 3, 2, norm=norm_layer, activation=act, padding=1 ) ) self.n_skip_stages = n_skip_stages if layers is None else len(layers) self.depths.append(nf) for n in range(self.n_stages - 1): self.blocks.append( Conv2dBlock( nf, min(nf * 2, nf_max) if layers is None else layers[n+1], 3, 2, norm=norm_layer, activation=act, padding=1, ) ) nf = min(nf * 2, nf_max) if layers is None else layers[n+1] self.depths.insert(0,nf) self.downs.insert(0,True) self.bottleneck = ResBlock(nf, nf_max, activation=act, stride=1,norm=norm_layer) self.downs.insert(0,False) def forward(self,x): if self.fg_bg: out_fg = [] out_bg = [] else: out = [] for i in range(self.n_stages): x = self.blocks[i](x) if i >= self.n_stages - self.n_skip_stages: if self.fg_bg: act_div = int(x.shape[1] / self.div) out_fg.append(x[:,:act_div]) out_bg.append(x[:,act_div:]) else: out.append(x) x = self.bottleneck(x) if self.fg_bg: act_div = int(x.shape[1] / self.div) out_fg.append(x[:,:act_div]) out_bg.append(x[:,act_div:]) return out_fg, out_bg else: out.append(x) return out class Encoder(nn.Module): def __init__(self, nf_in, nf_max, n_stages, prepare_adain=False, variational=False, resnet_down=False, norm_layer = "in", layers=None): super().__init__() self.prepare_adain = prepare_adain self.variational = variational if self.prepare_adain: assert not self.variational, "Encoder should not be variational if adain is prepared" if self.prepare_adain: self.final_linear = nn.Linear(nf_max, nf_max) act = "elu" #if self.variational else "relu" blocks = [] bottleneck = [] nf = 32 if layers is None else layers[0] blocks.append( Conv2dBlock( nf_in, nf, 3, 2, norm=norm_layer, activation=act, padding=1 ) ) n_stages = n_stages if layers is None else len(layers) for n in range(n_stages - 1): blocks.append( Conv2dBlock( nf, min(nf * 2, nf_max) if layers is None else layers[n+1], 3, 2, norm=norm_layer, activation=act, padding=1, ) ) nf = min(nf * 2, nf_max) if layers is None else layers[n+1] self.resnet_down = resnet_down and self.prepare_adain self.nf_in_bn = nf bottleneck.append(ResBlock(nf, nf_max,activation=act, stride=2 if self.resnet_down else 1, norm=norm_layer)) if layers is None: bottleneck.append(ResBlock(nf_max, nf_max,activation=act, stride=2 if self.resnet_down else 1, norm=norm_layer)) if self.resnet_down: self.make_vector = Conv2dBlock(nf_max,nf_max,4,1,0) if self.variational: self.make_mu = NormConv2d(nf_max,nf_max,3, padding=1) self.make_sigma = NormConv2d(nf_max,nf_max,3, padding=1) self.squash = nn.Sigmoid() self.model = nn.Sequential(*blocks) self.bottleneck = nn.Sequential(*bottleneck) def forward(self, input, sample_prior=False): out = self.model(input) mean = out out = self.bottleneck(out) logstd = None if self.prepare_adain: # mean is a false name here, this is the raw channels of the conv model # mean = out if self.resnet_down: # in this case, mean has spatial_size 4x4 out = self.make_vector(out).squeeze(-1).squeeze(-1) else: out = F.avg_pool2d(out, out.size(2), padding=0) out = out.squeeze(-1).squeeze(-1) # no activation for the first trial, as relu would not allow for values < 0 out = self.final_linear(out) elif self.variational: mean = self.make_mu(out) # normalize sigma in between logstd = self.squash(self.make_sigma(out)) if sample_prior: out = torch.randn_like(mean) else: out = self.reparametrize(mean,logstd) return out, mean, logstd def reparametrize(self,mean,logstd): std = torch.exp(logstd) eps = torch.randn_like(std) return eps.mul(std) + mean class AdaINDecoderEntangled(nn.Module): """ We sample up from spatial resolution 16x16, given quadratic images """ def __init__(self, nf_in, n_stages,latent_fusion): super().__init__() self.blocks = nn.ModuleList() self.affines = nn.ModuleList() self.n_stages = n_stages self.latent_fusion = latent_fusion # results latent fusion results in a deeper model nf = nf_in * 2 if self.latent_fusion else nf_in self.in_block = ResBlock(nf, nf) for n in range(self.n_stages): self.affines.append(AdaINLinear(nf_in, int(nf // 2))) # upsampling adain layers self.blocks.append( ResBlock(nf, int(nf // 2), norm="adain", upsampling=True) ) nf = int(nf // 2) self.out_conv = Conv2dBlock( nf, 3, 3, 1, padding=1, norm="none", activation="tanh" ) def forward(self, object_code, dynamics_linear,dynamics_spatial): if self.latent_fusion: in_code = torch.cat([object_code,dynamics_spatial],dim=1) else: in_code = object_code x = self.in_block(in_code) for n in range(self.n_stages): adain_params = self.affines[n](dynamics_linear) x = self.blocks[n](x, adain_params) x = self.out_conv(x) return x class AdaINDecoderDisentangled(nn.Module): def __init__(self,nf_in, n_stages, latent_fusion = None, nf_in_bn = 0): super().__init__() self.blocks = nn.ModuleList() self.affines = nn.ModuleList() self.n_stages = n_stages self.latent_fusion = False if latent_fusion is None else latent_fusion if self.latent_fusion: assert nf_in_bn > 0 # results latent fusion results in a deeper model nf = nf_in + nf_in_bn if self.latent_fusion else nf_in # self.bottleneck_adain = bottleneck_adain self.in_block = ResBlock(nf, nf,) for n in range(self.n_stages): self.affines.append(AdaINLinear(nf_in, int(nf // 2))) # upsampling adain layers self.blocks.append( ResBlock(nf, int(nf // 2), norm="adain", upsampling=True) ) nf = int(nf // 2) self.out_conv = Conv2dBlock( nf, 3, 3, 1, padding=1, norm="none", activation="tanh" ) def forward(self,alpha,sigma,alpha_spatial=None): if self.latent_fusion: assert alpha_spatial is not None in_code = torch.cat([sigma, alpha_spatial],dim=1) else: in_code = sigma x = self.in_block(in_code) for n in range(self.n_stages): adain_params = self.affines[n](alpha) x = self.blocks[n](x, adain_params) x = self.out_conv(x) return x class SkipConnectionDecoder(nn.Module): def __init__(self,nf_in, in_channels, n_skip_stages, disentanglement=False, spectral_norm=False, norm_layer="in",layers=None): super().__init__() self.n_stages = len(in_channels) self.disentanglement = disentanglement self.n_skip_stages = n_skip_stages self.blocks = nn.ModuleList() if self.disentanglement: self.affines = nn.ModuleList() nf = nf_in self.in_block = ResBlock(nf,in_channels[0], snorm=spectral_norm, norm=norm_layer) for i,nf in enumerate(in_channels): if layers is None: n_out = int(nf // 2) if i < len(in_channels) - 1 else nf if self.disentanglement: self.affines.append(AdaINLinear(nf_in,n_out)) nf_in_dec = 2 * nf if i < self.n_skip_stages else nf if layers is not None: nf_in_dec = 2 * nf n_out = in_channels[i+1] if i < len(in_channels) -1 else nf self.blocks.append(ResBlock(nf_in_dec, n_out , norm="adain" if self.disentanglement else norm_layer, upsampling=True,snorm=spectral_norm)) self.out_conv = Conv2dBlock(nf,3,3,1,1,norm="none",activation="tanh") def forward(self,shape, appearance = None, del_shape=True): x = self.in_block(shape.pop() if del_shape else shape[-1]) for n in range(self.n_stages): if n < self.n_skip_stages: x = torch.cat([x,shape.pop() if del_shape else shape[self.n_skip_stages-1-n]],1) if self.disentanglement: adain_params = self.affines[n](appearance) x = self.blocks[n](x,adain_params) else: x = self.blocks[n](x) if del_shape: assert not shape out = self.out_conv(x) return out class DecoderEntangled(nn.Module): """ We sample up from spatial resolution 16x16, given quadratic images """ def __init__(self, nf_in, n_stages): super().__init__() self.blocks = nn.ModuleList() self.n_stages = n_stages nf = nf_in self.in_block = ResBlock(nf, nf) for n in range(self.n_stages): self.blocks.append( ResBlock(nf, int(nf // 2), norm="in", upsampling=True) ) nf = int(nf // 2) self.out_conv = Conv2dBlock( nf, 3, 3, 1, padding=1, norm="none", activation="tanh" ) def forward(self, object_code, dynamics_code,*args): in_code = torch.cat([object_code,dynamics_code],dim=1) x = self.in_block(in_code) for n in range(self.n_stages): x = self.blocks[n](x) x = self.out_conv(x) return x class FusionBlockMultiscale(nn.Module): def __init__(self,nf_in,nfs,n_blocks): super().__init__() self.blocks = nn.ModuleList() self.n_stages = len(nfs) + 1 nf = nf_in for n in range(self.n_stages): self.blocks.append(LearnedFusionBlock(nf,n_blocks)) if n < len(nfs): nf = nfs[n] def forward(self,sigmas,deltas): out = [] for i,n in enumerate(range(len(sigmas)-1,-1,-1)) : out.insert(0,self.blocks[i](sigmas[n],deltas[n])) return out
45,992
39.274081
181
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/models/blocks.py
import torch from torch import nn from torch.nn import functional as F from torch.nn.utils import weight_norm, spectral_norm from torch.nn import init class ResBlock(nn.Module): def __init__( self, dim_in, dim_out, norm="in", activation="elu", pad_type="zero", upsampling=False, stride = 1, snorm=False ): super(ResBlock, self).__init__() self.norm = norm self.model = nn.ModuleList() if upsampling: self.conv1 = Conv2dTransposeBlock( dim_in, dim_out, 3, 2, 1, norm=self.norm, activation=activation, snorm= snorm ) self.conv2 = Conv2dBlock( dim_out, dim_out, 3, 1, 1, norm=self.norm, activation="none", pad_type=pad_type, snorm=snorm ) else: self.conv1 = Conv2dBlock( dim_in, dim_out, 3, stride, 1, norm=self.norm, activation=activation, pad_type=pad_type, snorm=snorm ) self.conv2 = Conv2dBlock( dim_out, dim_out, 3, 1, 1, norm=self.norm, activation="none", pad_type=pad_type, snorm=snorm ) self.convolve_res = dim_in != dim_out or upsampling or stride != 1 if self.convolve_res: if not upsampling: self.res_conv = Conv2dBlock(dim_in,dim_out,3,stride,1, norm="in", activation=activation, pad_type=pad_type, snorm=snorm) else: self.res_conv = Conv2dTransposeBlock(dim_in,dim_out,3,2,1, norm="in", activation=activation, snorm=snorm) def forward(self, x,adain_params=None): residual = x if self.convolve_res: residual = self.res_conv(residual) out = self.conv1(x,adain_params) out = self.conv2(out,adain_params) out += residual return out class Conv2dBlock(nn.Module): def __init__( self, in_dim, out_dim, ks, st, padding=0, norm="none", activation="elu", pad_type="zero", use_bias=True, activation_first=False, snorm=False ): super().__init__() self.use_bias = use_bias self.activation_first = activation_first # initialize padding if pad_type == "reflect": self.pad = nn.ReflectionPad2d(padding) elif pad_type == "replicate": self.pad = nn.ReplicationPad2d(padding) elif pad_type == "zero": self.pad = nn.ZeroPad2d(padding) else: assert 0, "Unsupported padding type: {}".format(pad_type) # initialize normalization norm_dim = out_dim if norm == "bn": self.norm = nn.BatchNorm2d(norm_dim) elif norm == "in": self.norm = nn.InstanceNorm2d(norm_dim) elif norm == "group": self.norm = nn.GroupNorm(num_channels=norm_dim,num_groups=16) elif norm == "adain": self.norm = AdaptiveInstanceNorm2d(norm_dim) elif norm == "none": self.norm = None else: assert 0, "Unsupported normalization: {}".format(norm) # initialize activation if activation == "relu": self.activation = nn.ReLU(inplace=True) elif activation == "lrelu": self.activation = nn.LeakyReLU(0.2, inplace=True) elif activation == "tanh": self.activation = nn.Tanh() elif activation == "elu": self.activation = nn.ELU() elif activation == "none": self.activation = None else: assert 0, "Unsupported activation: {}".format(activation) if snorm: self.conv = spectral_norm(nn.Conv2d(in_dim, out_dim, ks, st, bias=self.use_bias)) else: self.conv = nn.Conv2d(in_dim, out_dim, ks, st, bias=self.use_bias) def forward(self, x, adain_params=None): if self.activation_first: if self.activation: x = self.activation(x) x = self.conv(self.pad(x)) if self.norm and not isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x) elif isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x, adain_params) else: x = self.conv(self.pad(x)) if self.norm and not isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x) elif isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x, adain_params) if self.activation: x = self.activation(x) return x class NormConv2d(nn.Module): """ Convolutional layer with l2 weight normalization and learned scaling parameters """ def __init__( self, in_channels, out_channels, kernel_size, stride=1, padding=0 ): super().__init__() self.beta = nn.Parameter( torch.zeros([1, out_channels, 1, 1], dtype=torch.float32) ) self.gamma = nn.Parameter( torch.ones([1, out_channels, 1, 1], dtype=torch.float32) ) self.conv = weight_norm( nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding), name="weight", ) def forward(self, x): # weight normalization # self.conv.weight = normalize(self.conv.weight., dim=[0, 2, 3]) out = self.conv(x) out = self.gamma * out + self.beta return out class Conv2dTransposeBlock(nn.Module): def __init__( self, in_dim, out_dim, ks, st, padding=0, norm="none", activation="elu", use_bias=True, activation_first=False, snorm=False ): super().__init__() self.use_bias = use_bias self.activation_first = activation_first # initialize normalization norm_dim = out_dim if norm == "bn": self.norm = nn.BatchNorm2d(norm_dim) elif norm == "in": self.norm = nn.InstanceNorm2d(norm_dim) elif norm == "group": self.norm = nn.GroupNorm(num_channels=norm_dim,num_groups=16) elif norm == "adain": self.norm = AdaptiveInstanceNorm2d(norm_dim) elif norm == "none": self.norm = None else: assert 0, "Unsupported normalization: {}".format(norm) # initialize activation if activation == "elu": self.activation = nn.ReLU(inplace=True) elif activation == "lrelu": self.activation = nn.LeakyReLU(0.2, inplace=True) elif activation == "tanh": self.activation = nn.Tanh() elif activation == "none": self.activation = None else: assert 0, "Unsupported activation: {}".format(activation) if snorm: self.conv = spectral_norm(nn.ConvTranspose2d(in_dim, out_dim, ks, st, bias=self.use_bias, padding=padding, output_padding=padding)) else: self.conv = nn.ConvTranspose2d(in_dim, out_dim, ks, st, bias=self.use_bias, padding=padding,output_padding=padding) def forward(self, x, adain_params=None): if self.activation_first: if self.activation: x = self.activation(x) x = self.conv(x) if self.norm and not isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x) elif isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x, adain_params) else: x = self.conv(x) if self.norm and not isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x) elif isinstance(self.norm,AdaptiveInstanceNorm2d): x = self.norm(x, adain_params) if self.activation: x = self.activation(x) return x class AdaptiveInstanceNorm2d(nn.Module): def __init__(self, num_features, eps=1e-5, momentum=0.1): super().__init__() self.num_features = num_features self.eps = eps self.momentum = momentum self.register_buffer("running_mean", torch.zeros(num_features)) self.register_buffer("running_var", torch.ones(num_features)) def forward(self, x, adain_params): b, c = x.size(0), x.size(1) running_mean = self.running_mean.repeat(b) running_var = self.running_var.repeat(b) x_reshaped = x.contiguous().view(1, b * c, *x.size()[2:]) out = F.batch_norm( x_reshaped, running_mean, running_var, adain_params["weight"], adain_params["bias"], True, self.momentum, self.eps, ) return out.view(b, c, *x.size()[2:]) def __repr__(self): return self.__class__.__name__ + "(" + str(self.num_features) + ")" class AdaINLinear(nn.Module): def __init__(self, in_units, target_units, use_bias=True, actfn=nn.ReLU): super().__init__() self.linear = nn.Linear(in_units, 2 * target_units, bias=use_bias) self.act_fn = actfn() def forward(self, x): out = self.act_fn(self.linear(x)) out = { "weight": out[:, : out.size(1) // 2], "bias": out[:, out.size(1) // 2 :], } return out class ConvGRUCell(nn.Module): """ Generate a convolutional GRU cell """ def __init__(self, input_size, hidden_size, kernel_size,upsample=False): super().__init__() padding = kernel_size // 2 self.input_size = input_size self.upsample = upsample self.hidden_size = hidden_size self.reset_gate = nn.Conv2d(input_size + hidden_size, hidden_size, kernel_size, padding=padding) self.update_gate = nn.Conv2d(input_size + hidden_size, hidden_size, kernel_size, padding=padding) self.out_gate = nn.Conv2d(input_size + hidden_size, hidden_size, kernel_size, padding=padding) if self.upsample: self.up_gate = nn.ConvTranspose2d(input_size,input_size,kernel_size,2,padding=padding, output_padding=padding) init.orthogonal_(self.reset_gate.weight) init.orthogonal_(self.update_gate.weight) init.orthogonal_(self.out_gate.weight) init.constant_(self.reset_gate.bias, 0.) init.constant_(self.update_gate.bias, 0.) init.constant_(self.out_gate.bias, 0.) if self.upsample: init.orthogonal_(self.up_gate.weight) init.constant_(self.up_gate.bias, 0.) def forward(self, input_, prev_state): if self.upsample: input_ = self.up_gate(input_) # get batch and spatial sizes batch_size = input_.data.size()[0] spatial_size = input_.data.size()[2:] # generate empty prev_state, if None is provided if prev_state is None: state_size = [batch_size, self.hidden_size] + list(spatial_size) if torch.cuda.is_available(): prev_state = torch.zeros(state_size).cuda() else: prev_state = torch.zeros(state_size) # data size is [batch, channel, height, width] stacked_inputs = torch.cat([input_, prev_state], dim=1) update = torch.sigmoid(self.update_gate(stacked_inputs)) reset = torch.sigmoid(self.reset_gate(stacked_inputs)) out_inputs = torch.tanh(self.out_gate(torch.cat([input_, prev_state * reset], dim=1))) new_state = prev_state * (1 - update) + out_inputs * update return new_state class ConvGRU(nn.Module): def __init__(self, input_size, hidden_sizes, kernel_sizes, n_layers, upsampling:list=None): ''' Generates a multi-layer convolutional GRU. Preserves spatial dimensions across cells, only altering depth. Parameters ---------- input_size : integer. depth dimension of input tensors. hidden_sizes : integer or list. depth dimensions of hidden state. if integer, the same hidden size is used for all cells. kernel_sizes : integer or list. sizes of Conv2d gate kernels. if integer, the same kernel size is used for all cells. n_layers : integer. number of chained `ConvGRUCell`. ''' super(ConvGRU, self).__init__() if upsampling is None: upsampling = [False]*n_layers self.input_size = input_size if type(hidden_sizes) != list: self.hidden_sizes = [hidden_sizes]*n_layers else: assert len(hidden_sizes) == n_layers, '`hidden_sizes` must have the same length as n_layers' self.hidden_sizes = hidden_sizes if type(kernel_sizes) != list: self.kernel_sizes = [kernel_sizes]*n_layers else: assert len(kernel_sizes) == n_layers, '`kernel_sizes` must have the same length as n_layers' self.kernel_sizes = kernel_sizes self.n_layers = n_layers self.cells = [] for i in range(self.n_layers): if i == 0: input_dim = self.input_size else: input_dim = self.hidden_sizes[i - 1] self.cells.append(ConvGRUCell(input_dim, self.hidden_sizes[i], self.kernel_sizes[i],upsample=upsampling[i])) self.cells = nn.Sequential(*self.cells) def forward(self, x, hidden=None): ''' Parameters ---------- x : 4D input tensor. (batch, channels, height, width). hidden : list of 4D hidden state representations. (layer, batch, channels, height, width). Returns ------- upd_hidden : 5D hidden representation. (layer, batch, channels, height, width). ''' if hidden is None: hidden = [None]*self.n_layers input_ = x upd_hidden = [] for layer_idx in range(self.n_layers): cell = self.cells[layer_idx] cell_hidden = hidden[layer_idx] # pass through layer upd_cell_hidden = cell(input_, cell_hidden) upd_hidden.append(upd_cell_hidden) # update input_ to the last updated hidden layer for next pass input_ = upd_cell_hidden # retain tensors in list to allow different hidden sizes return upd_hidden # taken from official NVLabs implementation # Creates SPADE normalization layer based on the given configuration # SPADE consists of two steps. First, it normalizes the activations using # your favorite normalization method, such as Batch Norm or Instance Norm. # Second, it applies scale and bias to the normalized output, conditioned on # the segmentation map. # The format of |config_text| is spade(norm)(ks), where # (norm) specifies the type of parameter-free normalization. # (e.g. syncbatch, batch, instance) # (ks) specifies the size of kernel in the SPADE module (e.g. 3x3) # Example |config_text| will be spadesyncbatch3x3, or spadeinstance5x5. # Also, the other arguments are # |norm_nc|: the #channels of the normalized activations, hence the output dim of SPADE # |label_nc|: the #channels of the input semantic map, hence the input dim of SPADE class SPADE(nn.Module): def __init__(self, norm_nc, label_nc, config): super().__init__() param_free_norm_type = config["base_norm_spade"] if "base_norm_spade" in config else "instance" ks = 3 if param_free_norm_type == 'instance': self.param_free_norm = nn.InstanceNorm2d(norm_nc, affine=False) elif param_free_norm_type == 'batch': self.param_free_norm = nn.BatchNorm2d(norm_nc, affine=False) else: raise ValueError('%s is not a recognized param-free norm type in SPADE' % param_free_norm_type) # The dimension of the intermediate embedding space. Yes, hardcoded. nhidden = 128 pw = ks // 2 self.mlp_shared = nn.Sequential( nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw), nn.ReLU() ) self.mlp_gamma = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw) self.mlp_beta = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw) def forward(self, x, segmap): # Part 1. generate parameter-free normalized activations normalized = self.param_free_norm(x) # Part 2. produce scaling and bias conditioned on semantic map segmap = F.interpolate(segmap, size=x.size()[2:], mode='nearest') actv = self.mlp_shared(segmap) gamma = self.mlp_gamma(actv) beta = self.mlp_beta(actv) # apply scale and bias out = normalized * (1 + gamma) + beta return out
17,622
33.690945
143
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/experiments/experiment.py
from abc import abstractmethod import torch import wandb import os from os import path from glob import glob import numpy as np from utils.general import get_logger WANDB_DISABLE_CODE = True class Experiment: def __init__(self, config:dict, dirs: dict, device): self.parallel = isinstance(device, list) self.config = config self.logger = get_logger(self.config["general"]["project_name"]) self.is_debug = self.config["general"]["debug"] if self.is_debug: self.logger.info("Running in debug mode") if self.parallel: self.device = torch.device( f"cuda:{device[0]}" if torch.cuda.is_available() else "cpu" ) self.all_devices = device self.logger.info("Running experiment on multiple gpus!") else: self.device = device self.all_devices = [device] self.dirs = dirs if torch.cuda.is_available(): os.environ["CUDA_VISIBLE_DEVICES"] = ",".join([str(dev.index if self.parallel else dev) for dev in self.all_devices]) if self.config["general"]["restart"]: self.logger.info(f'Resume training run with name "{self.config["general"]["project_name"]}" on device(s) {self.all_devices}') else: self.logger.info(f'Start new training run with name "{self.config["general"]["project_name"]}" on device(s) {self.all_devices}') ########## seed setting ########## torch.manual_seed(self.config["general"]["seed"]) torch.cuda.manual_seed(self.config["general"]["seed"]) np.random.seed(self.config["general"]["seed"]) # random.seed(opt.seed) torch.backends.cudnn.deterministic = True torch.manual_seed(self.config["general"]["seed"]) rng = np.random.RandomState(self.config["general"]["seed"]) if self.config["general"]["mode"] == "train": project = "visual_poking_unsupervised" wandb.init( dir=self.dirs["log"], project=project, name=self.config["general"]["project_name"], group=self.config["general"]["experiment"], ) # log paramaters self.logger.info("Training parameters:") for key in self.config: if key != "testing": self.logger.info(f"{key}: {self.config[key]}") # print to console wandb.config.update({key: self.config[key]}) # update wandb config def _load_ckpt(self, key, dir=None,name=None, single_opt = True, use_best=False, load_name = "model"): if dir is None: dir = self.dirs["ckpt"] if name is None: if len(os.listdir(dir)) > 0: ckpts = glob(path.join(dir,"*.pt")) # load latest stored checkpoint ckpts = [ckpt for ckpt in ckpts if key in ckpt.split("/")[-1]] if len(ckpts) == 0: self.logger.info(f"*************No ckpt found****************") op_ckpt = mod_ckpt = None return mod_ckpt, op_ckpt if use_best: ckpts = [x for x in glob(path.join(dir,"*.pt")) if "=" in x.split("/")[-1]] ckpts = {float(x.split("=")[-1].split(".")[0]): x for x in ckpts} ckpt = torch.load( ckpts[max(list(ckpts.keys()))], map_location="cpu" ) else: ckpts = {float(x.split("_")[-1].split(".")[0]): x for x in ckpts} ckpt = torch.load( ckpts[max(list(ckpts.keys()))], map_location="cpu" ) mod_ckpt = ckpt[load_name] if load_name in ckpt else None if single_opt: key = [key for key in ckpt if key.startswith("optimizer")] assert len(key) == 1 key = key[0] op_ckpt = ckpt[key] else: op_ckpt = {key: ckpt[key] for key in ckpt if "optimizer" in key} msg = "best model" if use_best else "model" if mod_ckpt is not None: self.logger.info(f"*************Restored {msg} with key {key} from checkpoint****************") else: self.logger.info(f"*************No ckpt for {msg} with key {key} found, not restoring...****************") if op_ckpt is not None: self.logger.info(f"*************Restored optimizer with key {key} from checkpoint****************") else: self.logger.info(f"*************No ckpt for optimizer with key {key} found, not restoring...****************") else: mod_ckpt = op_ckpt = None return mod_ckpt, op_ckpt else: # fixme add checkpoint loading for best performing models ckpt_path = path.join(dir,name) if not path.isfile(ckpt_path): self.logger.info(f"*************No ckpt for model and optimizer found under {ckpt_path}, not restoring...****************") mod_ckpt = op_ckpt = None else: if "epoch_ckpts" in ckpt_path: mod_ckpt = torch.load( ckpt_path, map_location="cpu" ) op_path = ckpt_path.replace("model@","opt@") op_ckpt = torch.load(op_path,map_location="cpu") return mod_ckpt,op_ckpt ckpt = torch.load(ckpt_path, map_location="cpu") mod_ckpt = ckpt[load_name] if load_name in ckpt else None op_ckpt = ckpt["optimizer"] if "optimizer" in ckpt else None if mod_ckpt is not None: self.logger.info(f"*************Restored model under {ckpt_path} ****************") else: self.logger.info(f"*************No ckpt for model found under {ckpt_path}, not restoring...****************") if op_ckpt is not None: self.logger.info(f"*************Restored optimizer under {ckpt_path}****************") else: self.logger.info(f"*************No ckpt for optimizer found under {ckpt_path}, not restoring...****************") return mod_ckpt,op_ckpt @abstractmethod def train(self): """ Here, the experiment shall be run :return: """ pass @abstractmethod def test(self): """ Here the prediction shall be run :param ckpt_path: The path where the checkpoint file to load can be found :return: """ pass
6,865
40.361446
140
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/experiments/fixed_length_model.py
import torch from torch.utils.data import DataLoader, RandomSampler, SequentialSampler from torch.optim import Adam from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from ignite.contrib.handlers import ProgressBar from ignite.metrics import Average, MetricUsage import numpy as np import wandb from functools import partial from lpips import LPIPS from tqdm import tqdm from experiments.experiment import Experiment from data import get_dataset from data.samplers import FixedLengthSampler from models.latent_flow_net import SingleScaleBaseline,SkipSequenceModel from models.discriminator import GANTrainer from utils.losses import PerceptualVGG,vgg_loss_agg,DynamicsLoss, style_loss from utils.testing import make_flow_grid, make_img_grid, make_video, make_plot from utils.metrics import metric_fid, FIDInceptionModel, metric_lpips, psnr_lightning, ssim_lightning from utils.general import linear_var, get_member, get_patches class FixedLengthModel(Experiment): def __init__(self, config, dirs, device): super().__init__(config, dirs, device) self.datakeys = ["images","poke"] if self.config["architecture"]["disentanglement"]: self.datakeys.append("img_aT") self.datakeys.append("app_img_random") # used for efficient metrics computation self.fid_feats_real_per_frame = {} self.fid_feats_fake_per_frame = {} self.psnrs = {"t": [], "tk": [], "pl" : []} self.ssims = {"t": [], "tk": [], "pl" : []} self.lpips = {"t": [], "tk": []} self.use_gan = self.config["gan"]["use"] self.use_temp_disc = self.config["gan_temp"]["use"] if self.use_temp_disc: if not self.config["gan_temp"]["patch_temp_disc"]: assert not self.config["gan_temp"]["conditional"] #self.pixel_decoder_loss = self.config["training"]["pixel_dynamics_weight"] > 0 self.lr_dec_t = 0 self.target_dev = None # metrics for each frame self.ssims_per_frame = {} self.lpips_per_frame = {} self.psnrs_per_frame = {} # self.ssims_per_frame_pl = {} # self.psnrs_per_frame_pl = {} self.lpips_avg = None self.custom_sampler = self.config["training"]["custom_sampler"] if "custom_sampler" in self.config["training"] else False self.poke_jump = self.config["training"]["poke_jump"] if "poke_jump" in self.config["training"] else False self.poke_scale_mode = self.config["architecture"]["poke_scale"] if "poke_scale" in self.config["architecture"] else False if self.poke_jump: assert not self.poke_scale_mode def __clear_metric_arrs(self): [self.psnrs[key].clear() for key in self.psnrs] [self.ssims[key].clear() for key in self.ssims] [self.lpips[key].clear() for key in self.lpips] self.lpips_per_frame = {} self.psnrs_per_frame = {} self.ssims_per_frame = {} self.fid_feats_real_per_frame = {} self.fid_feats_fake_per_frame = {} # self.ssims_per_frame_pl = {} # self.psnrs_per_frame_pl = {} def train(self): ########## checkpoints ########## if self.config["general"]["restart"] and not self.is_debug: mod_ckpt, op_ckpts = self._load_ckpt("reg_ckpt", single_opt=False) op_ckpt_dis = op_ckpts["optimizer_dis"] op_ckpt_dyn = op_ckpts["optimizer_dyn"] else: mod_ckpt = op_ckpt_dis = op_ckpt_dyn = None # get datasets for training and testing def w_init_fn(worker_id): return np.random.seed(np.random.get_state()[1][0] + worker_id) if not self.poke_scale_mode: del self.config["data"]["n_ref_frames"] dataset, transforms = get_dataset(config=self.config["data"]) train_dataset = dataset(transforms, self.datakeys, self.config["data"], train=True) test_datakeys = self.datakeys + ["app_img_random"] if self.config["testing"]["eval_app_transfer"] and "app_img_random" not in self.datakeys else self.datakeys test_datakeys.append("flow") test_dataset = dataset(transforms, test_datakeys, self.config["data"], train=False) if self.custom_sampler: train_sampler = FixedLengthSampler(train_dataset, self.config["training"]["batch_size"],shuffle=True, weighting=train_dataset.obj_weighting,drop_last=True,zero_poke=True, zero_poke_amount=self.config["training"]["zeropoke_amount"]) train_loader = DataLoader(train_dataset, batch_sampler=train_sampler,num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn,) test_sampler = FixedLengthSampler(test_dataset, batch_size=self.config["training"]["batch_size"], shuffle=True, drop_last=True, weighting=test_dataset.obj_weighting,zero_poke=True,zero_poke_amount=self.config["training"]["zeropoke_amount"]) test_loader = DataLoader( test_dataset, batch_sampler=test_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], # worker_init_fn=w_init_fn, ) eval_sampler = FixedLengthSampler(test_dataset,batch_size=self.config["testing"]["test_batch_size"],shuffle=True, drop_last=True,weighting=test_dataset.obj_weighting,zero_poke=False) eval_loader = DataLoader(test_dataset, batch_sampler=eval_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn,) self.logger.info("Using custom fixed length sampler.") else: self.logger.info("Using standard pytorch random sampler") train_sampler = RandomSampler(train_dataset) train_loader = DataLoader( train_dataset, sampler=train_sampler, batch_size=self.config["training"]["batch_size"], num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn, drop_last=True ) test_sampler = RandomSampler(test_dataset,) test_loader = DataLoader( test_dataset, batch_size=self.config["training"]["batch_size"], sampler=test_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn, drop_last=True ) # no zeropoke for evaluation as zeropoke is only to ensure no reaction when poking outside eval_sampler = SequentialSampler(test_dataset,) eval_loader = DataLoader(test_dataset, sampler=eval_sampler, batch_size=self.config["testing"]["test_batch_size"], num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn, drop_last=True) # define model self.logger.info(f"Load model...") #net_model = SkipSequenceModel if self.config["architecture"]["use_skip_model"] else ResidualSequenceBaseline net = SkipSequenceModel(spatial_size=self.config["data"]["spatial_size"],config=self.config["architecture"]) if self.config["architecture"]["use_skip_model"] else \ SingleScaleBaseline(spatial_size=self.config["data"]["spatial_size"], config=self.config["architecture"], ) self.logger.info( f"Number of trainable parameters in model is {sum(p.numel() for p in net.parameters())}" ) if self.config["general"]["restart"] and mod_ckpt is not None: self.logger.info("Load pretrained paramaters and resume training.") net.load_state_dict(mod_ckpt) if self.parallel: net = torch.nn.DataParallel(net, device_ids=self.all_devices) net.cuda(self.all_devices[0]) self.logger.info("Model on gpu!") # log weights and gradients wandb.watch(net, log="all") # define optimizers # appearance and shape disentanglement dis_params = [{"params": get_member(net,"shape_enc").parameters(), "name": "shape_encoder"}, {"params": get_member(net,"dec").parameters(), "name": "decoder"} ] optimizer_dis = Adam(dis_params, lr=self.config["training"]["lr"]) if self.config["general"]["restart"] and op_ckpt_dis is not None: self.logger.info("Load state_dict of optimizer.") optimizer_dis.load_state_dict(op_ckpt_dis) milestones = [int(self.config["training"]["n_epochs"] * t) for t in self.config["training"]["tau"]] scheduler_dis = torch.optim.lr_scheduler.MultiStepLR(optimizer_dis, milestones=milestones, gamma=self.config["training"]["lr_reduce"]) # dynamics dyn_params = [{"params": get_member(net,"dynamics_enc").parameters(), "name": "dynamics_encoder", }, {"params": get_member(net,"fusion_block").parameters(), "name": "fusion_block",},] if self.config["training"]["decoder_update_tk"]: dyn_params.append({"params": get_member(net,"dec").parameters(), "name": "decoder"}) if "singlestage" in self.config["training"] and self.config["training"]["singlestage"]: dyn_params.append({"params": get_member(net, "shape_enc").parameters(), "name": "shape_encoder"}) optimizer_dyn = Adam(dyn_params, lr = self.config["training"]["lr"]) if self.config["general"]["restart"] and op_ckpt_dyn is not None: self.logger.info("Load state_dict of optimizer.") optimizer_dyn.load_state_dict(op_ckpt_dyn) milestones = [int(self.config["training"]["n_epochs"] * t) for t in self.config["training"]["tau"]] scheduler_dyn = torch.optim.lr_scheduler.MultiStepLR(optimizer_dyn, milestones=milestones, gamma=self.config["training"]["lr_reduce"]) # initialize disc if gan mode is enabled if self.use_gan: gan_trainer = GANTrainer(self.config, self._load_ckpt, self.logger,spatial_size=self.config["data"]["spatial_size"][0] , parallel=self.parallel, devices=self.all_devices, debug=self.is_debug) if self.use_temp_disc: gan_trainer_temp = GANTrainer(self.config, self._load_ckpt,self.logger,spatial_size=self.config["data"]["spatial_size"][0], parallel=self.parallel,devices=self.all_devices, debug=self.is_debug,temporal=True, sequence_length=train_dataset.max_frames) # set start iteration and epoch in case model training is resumed start_it = 0 start_epoch = 0 n_epoch_train = self.config["training"]["n_epochs"] if self.config["general"]["restart"] and op_ckpts is not None: start_it = list(optimizer_dis.state_dict()["state"].values())[-1]["step"] start_epoch = int(np.floor(start_it / len(train_loader))) assert self.config["training"]["n_epochs"] > start_epoch n_epoch_train = self.config["training"]["n_epochs"] - start_epoch # lr_dec_rec = partial(linear_var,start_it=0, end_it=self.config["training"]["lr_dec_end_it"], start_val=self.config["training"]["lr"], end_val=self.config["training"]["lr_dec_end_val"], clip_min=0, clip_max=self.config["training"]["lr"],) self.lr_dec_t = lr_dec_rec(start_it) # losses self.logger.info("Load VGG") self.vgg = PerceptualVGG() if self.parallel: self.vgg = torch.nn.DataParallel(self.vgg,device_ids=self.all_devices) self.vgg.cuda(self.all_devices[0]) self.logger.info("VGG on gpu") # from torchsummary import summary # summary(vgg.vgg,(3,224,224)) self.logger.info("Initialize persistent losses") latent_dynamics_loss = DynamicsLoss(config=self.config["training"]) self.logger.info("Finished initializing persistent losses.") def train_step(engine,batch): net.train() # prepare data weights=None loss_dis = 0 out_dict = {} if train_dataset.flow_weights: poke = batch["poke"][0].cuda(self.all_devices[0]) weights = batch["poke"][1].cuda(self.all_devices[0]) else: poke = batch["poke"].cuda(self.all_devices[0]) x_t = batch["images"][:, 0].cuda(self.all_devices[0]) x_seq = batch["images"][:, 1:].cuda(self.all_devices[0]) if self.config["architecture"]["disentanglement"]: shape_img = batch["img_aT"].cuda(self.all_devices[0]) # apply style loss app_img_tr = batch["app_img_random"].cuda(self.all_devices[0]) x_trans, *_ = net(app_img_tr,x_t,poke,len=0) loss_style = style_loss(self.vgg,app_img_tr,x_trans) loss_dis = self.config["training"]["style_loss_weight"] * loss_style out_dict.update({"style_loss": loss_style.item()}) else: shape_img = x_t x_t_hat_i, sigma_t, _ , alpha = net(x_seq[:,-1],shape_img,poke,len=0) n_ref_frames = self.config["data"]["n_ref_frames"] - 1 if self.poke_scale_mode else train_dataset.max_frames -1 # static loss to obtain fixed image state space if "singlestage" not in self.config["training"] or not self.config["training"]["singlestage"]: loss_dis = loss_dis + vgg_loss_agg(self.vgg, x_t, x_t_hat_i) #optimize parameter of appearance, shape encoders and decoder optimizer_dis.zero_grad() loss_dis.backward() optimizer_dis.step() out_dict.update({"loss_dis" : loss_dis.item()}) #optimize in alternating gradient descent as this results in equal results than training the static/dynamic model in two completely seperate stages # however, it performs significantly better than training both models jointly with a single optimizer step (see ablations or run the model with 'singlestage' set to true) # forward pass for training of dynamics part of the model # dynamics losses seq_len = x_seq.shape[1] seq_rec, mu_delta, sigmas_hat, logstd_delta = net(x_t,shape_img,poke,len=seq_len, poke_linear=self.poke_scale_mode, n_zero_frames=seq_len-n_ref_frames-1, poke_jump=self.poke_jump) sigmas_gt = [] ll_loss_dyn = [] rec_imgs = [] if weights is not None: seq_rec = get_patches(seq_rec,weights,self.config["data"],train_dataset.weight_value_flow, logger=self.logger) x_seq = get_patches(x_seq,weights,self.config["data"],train_dataset.weight_value_flow, logger=self.logger) for n in range(seq_len): x_hat_tn,s_tn,*_ = net(x_seq[:,n],x_seq[:,n],poke,len=0) sigmas_gt.append(s_tn) rec_imgs.append(x_hat_tn) w = 1. if n != n_ref_frames else self.config["training"]["target_weight"] ll_dyn_n =w * vgg_loss_agg(self.vgg,x_seq[:,n],seq_rec[:,n]) ll_loss_dyn.append(ll_dyn_n) ll_loss_dyn = torch.stack(ll_loss_dyn,dim=0).mean() rec_imgs = torch.stack(rec_imgs,1) #latent dynamics dyn_losses = [] for s_tk,s_hat_tk in zip(sigmas_gt,sigmas_hat): dyn_losses.append(latent_dynamics_loss(s_hat_tk,s_tk,[])) latent_loss_dyn = torch.stack(dyn_losses).mean() loss_dyn = self.config["training"]["vgg_dyn_weight"] * ll_loss_dyn + self.config["training"]["latent_dynamics_weight"] * latent_loss_dyn if self.use_gan and engine.state.iteration >= self.config["gan"]["start_iteration"]: if self.config["gan"]["pixel_dynamics"]: offsets = np.random.choice(np.arange(max(1,x_seq.shape[1]-train_dataset.max_frames)),size=x_seq.shape[0]) true_exmpls = torch.stack([seq[o:o+train_dataset.max_frames] for seq, o in zip(x_seq,offsets)],dim=0) fake_exmpls = torch.stack([seq[o:o+train_dataset.max_frames] for seq, o in zip(seq_rec, offsets)], dim=0) x_true = torch.cat([true_exmpls[:,1:],true_exmpls[:,:-1]],dim=2).reshape(-1,2*true_exmpls.shape[2],*true_exmpls.shape[3:]) x_fake = torch.cat([fake_exmpls[:, 1:], true_exmpls[:, :-1]], dim=2).reshape(-1, 2 * fake_exmpls.shape[2], *fake_exmpls.shape[3:]) else: true_exmpls = np.random.choice(np.arange(x_seq.shape[0]*x_seq.shape[1]),self.config["gan"]["n_examples"]) fake_exmpls = np.random.choice(np.arange(seq_rec.shape[0]*seq_rec.shape[1]), self.config["gan"]["n_examples"]) x_true = x_seq.view(-1,*x_seq.shape[2:])[true_exmpls] x_fake = seq_rec.view(-1,*seq_rec.shape[2:])[fake_exmpls] disc_dict, loss_gen, loss_fmap = gan_trainer.train_step(x_true, x_fake) loss_dyn = loss_dyn + self.config["gan"]["gen_weight"] * loss_gen + self.config["gan"]["fmap_weight"] * loss_fmap if self.use_temp_disc and engine.state.iteration >= self.config["gan_temp"]["start_iteration"]: seq_len_act = x_seq.shape[1] offset = int(np.random.choice(np.arange(max(1,seq_len_act-train_dataset.max_frames)),1)) # offset_fake = int(np.random.choice(np.arange(max(1,seq_len_act-seq_len_temp_disc)), 1)) x_fake_tmp = seq_rec[:,offset:offset+train_dataset.max_frames].permute(0,2,1,3,4) x_true_tmp = x_seq[:, offset:offset+train_dataset.max_frames].permute(0,2,1,3,4) if self.config["gan_temp"]["conditional"]: cond = get_patches(poke,weights,self.config["data"],test_dataset.weight_value_flow,self.logger) if test_dataset.flow_weights else poke else: cond = None disc_dict_temp, loss_gen_temp, loss_fmap_temp = gan_trainer_temp.train_step(x_true_tmp,x_fake_tmp,cond) loss_dyn = loss_dyn + self.config["gan_temp"]["gen_weight"] * loss_gen_temp + self.config["gan_temp"]["fmap_weight"] * loss_fmap_temp # optimize parameters of dynamics part optimizer_dyn.zero_grad() loss_dyn.backward() optimizer_dyn.step() out_dict.update({"loss_dyn":loss_dyn.item() ,"vgg_loss_dyn" : ll_loss_dyn.item(), "latent_loss_dyn": latent_loss_dyn.item(), "lr_dec_t": self.lr_dec_t}) if self.use_gan and engine.state.iteration >= self.config["gan"]["start_iteration"]: out_dict.update(disc_dict) out_dict.update({"loss_gen_patch" :loss_gen.item(), "loss_fmap_patch": loss_fmap.item()}) if self.use_temp_disc and engine.state.iteration >= self.config["gan_temp"]["start_iteration"]: out_dict.update(disc_dict_temp) out_dict.update({"loss_gen_temp" :loss_gen_temp.item(), "loss_fmap_temp": loss_fmap_temp.item()}) return out_dict self.logger.info("Initialize inception model...") self.inception_model = FIDInceptionModel() self.logger.info("Finished initialization of inception model...") # note that lpips is exactly vgg-cosine similarity as proposed in the google papers and savp self.lpips_fn = LPIPS(net="vgg") def eval_step(engine, eval_batch): net.eval() out_dict = {} with torch.no_grad(): # prepare data weights = None if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) weights = eval_batch["poke"][1].cuda(self.all_devices[0]) else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:,0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:,1:].cuda(self.all_devices[0]) if self.config["architecture"]["disentanglement"]: app_img_tr = eval_batch["app_img_random"].cuda(self.all_devices[0]) x_trans, *_ = net(app_img_tr, x_t, poke,len=0) loss_style = style_loss(self.vgg, app_img_tr, x_trans) out_dict.update({"style_loss_eval": loss_style.item()}) n_ref_frames = self.config["data"]["n_ref_frames"] - 1 if self.poke_scale_mode else train_dataset.max_frames -1 # eval forward passes seq_len = x_seq_gt.shape[1] x_t_hat, sigma_t, _, alpha = net(x_t,x_t,poke,len=0) x_seq_hat, _, sigmas_hat,_ = net(x_t, x_t, poke,len=seq_len,poke_linear=self.poke_scale_mode, n_zero_frames=seq_len-n_ref_frames-1,poke_jump=self.poke_jump) if weights is not None and self.config["testing"]["metrics_on_patches"]: x_seq_hat = get_patches(x_seq_hat,weights,self.config["data"],test_dataset.weight_value_flow, logger=self.logger) x_seq_gt = get_patches(x_seq_gt, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) sigmas_gt = [] ll_loss_dyn = [] rec_imgs = [] for n in range(seq_len): x_hat_tn, s_tn, *_ = net(x_seq_gt[:, n], x_seq_gt[:, n], poke, len=0) sigmas_gt.append(s_tn) rec_imgs.append(x_hat_tn) ll_dyn_n = vgg_loss_agg(self.vgg, x_seq_gt[:, n], x_seq_hat[:, n]) ll_loss_dyn.append(ll_dyn_n) ll_loss_tk_eval = torch.stack(ll_loss_dyn,dim=0).mean() rec_imgs = torch.stack(rec_imgs,1) if weights is not None and self.config["testing"]["metrics_on_patches"]: rec_imgs = get_patches(rec_imgs, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) # apply inception model for fid calculation at all timesteps for t in range(x_seq_gt.shape[1]): real_features_t = self.inception_model(x_seq_gt[:, t]).cpu().numpy() fake_features_t = self.inception_model(x_seq_hat[:, t]).cpu().numpy() if t not in self.fid_feats_fake_per_frame: self.fid_feats_fake_per_frame.update({t: fake_features_t}) self.fid_feats_real_per_frame.update({t: real_features_t}) else: self.fid_feats_fake_per_frame[t] = np.concatenate([self.fid_feats_fake_per_frame[t], fake_features_t], axis=0) self.fid_feats_real_per_frame[t] = np.concatenate([self.fid_feats_real_per_frame[t], real_features_t], axis=0) # evaluate training losses # ll_loss_tk_eval = vgg_loss_agg(self.vgg, x_tk, x_tk_hat) ll_loss_t_i_eval = vgg_loss_agg(self.vgg, x_t, x_t_hat) dyn_losses = [] for s_tk, s_hat_tk in zip(sigmas_gt, sigmas_hat): dyn_losses.append(latent_dynamics_loss(s_hat_tk, s_tk, [])) latent_loss_dyn_eval = torch.stack(dyn_losses).mean() out_dict.update({"vgg_loss_dyn_eval": ll_loss_tk_eval.item(), "loss_dis_i_eval": ll_loss_t_i_eval.item(), "latent_loss_dyn_eval": latent_loss_dyn_eval.item()}) # compute metrics ssim_t = ssim_lightning(x_t, x_t_hat) psnr_t = psnr_lightning(x_t, x_t_hat) lpips_t = metric_lpips(x_t,x_t_hat, self.lpips_fn, reduce=False) ssim_tk, ssim_per_frame = ssim_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) psnr_tk, psnr_per_frame = psnr_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) lpips_avg, lpips_per_frame = metric_lpips(x_seq_gt, x_seq_hat,self.lpips_fn,reduce=False,return_per_frame=True) # ssim_pl, ssim_pl_per_frame = ssim_lightning(x_seq_gt,x_seq_hat,return_per_frame=True) # psnr_pl, psnr_pl_per_frame = psnr_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) # append to arrays self.lpips["t"].append(lpips_t) self.psnrs["t"].append(psnr_t) self.ssims["t"].append(ssim_t) self.psnrs["tk"].append(psnr_tk) self.ssims["tk"].append(ssim_tk) self.lpips["tk"].append(lpips_avg) #self.ssims["pl"].append(ssim_pl) #self.psnrs["pl"].append(psnr_pl) # append the values of the respective sequence length [self.ssims_per_frame[key].append(ssim_per_frame[key]) if key in self.ssims_per_frame else self.ssims_per_frame.update({key:[ssim_per_frame[key]]}) for key in ssim_per_frame] [self.psnrs_per_frame[key].append(psnr_per_frame[key]) if key in self.psnrs_per_frame else self.psnrs_per_frame.update({key:[psnr_per_frame[key]]}) for key in psnr_per_frame] [self.lpips_per_frame[key].append(lpips_per_frame[key]) if key in self.lpips_per_frame else self.lpips_per_frame.update({key:[lpips_per_frame[key]]}) for key in lpips_per_frame] #[self.ssims_per_frame_pl[key].append(ssim_pl_per_frame[key]) if key in self.ssims_per_frame_pl else self.ssims_per_frame_pl.update({key: [ssim_pl_per_frame[key]]}) for key in ssim_pl_per_frame] #[self.psnrs_per_frame_pl[key].append(psnr_pl_per_frame[key]) if key in self.psnrs_per_frame_pl else self.psnrs_per_frame_pl.update({key: [psnr_pl_per_frame[key]]}) for key in psnr_pl_per_frame] return out_dict # test_it steps are performed while generating test_imgs, there n_test_img is overall number divided by number of test iterations n_test_img = int(self.config["testing"]["n_test_img"] // self.config["testing"]["test_it"]) def eval_visual(engine, eval_batch): net.eval() with torch.no_grad(): # prepare data if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) weights = eval_batch["poke"][1] else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:, 1:].cuda(self.all_devices[0]) flow = eval_batch["flow"] if self.config["architecture"]["disentanglement"]: shape_img = eval_batch["img_aT"].cuda(self.all_devices[0]) else: shape_img = x_t n_ref_frames = self.config["data"]["n_ref_frames"] - 1 if self.poke_scale_mode else train_dataset.max_frames -1 seq_len = x_seq_gt.shape[1] x_seq_hat, *_ = net(x_t,x_t, poke, len=seq_len,poke_linear=self.poke_scale_mode,n_zero_frames=seq_len-n_ref_frames-1, poke_jump = self.poke_jump) x_t_hat , *_ = net(x_seq_gt[:,-1],shape_img,poke,len=0) grid_dis = make_img_grid(x_seq_gt[:,-1],shape_img, x_t_hat,x_t, n_logged=n_test_img) grid_dyn = make_flow_grid(x_t, poke, x_seq_hat[:,-1], x_seq_gt[:,-1], n_logged=n_test_img, flow=flow) seq_vis_hat = torch.cat([x_t.unsqueeze(1), x_seq_hat], 1) seq_vis_gt = torch.cat([x_t.unsqueeze(1), x_seq_gt], 1) grid_anim = make_video(x_t,poke,seq_vis_hat,seq_vis_gt,n_logged=n_test_img,flow=flow, display_frame_nr=True) it = engine.state.iteration log_dict = {"Last Frame Comparison Test data": wandb.Image(grid_dyn, caption=f"Last frames test grid #{it}."), "Disentanglement Grid Test Data": wandb.Image(grid_dis, caption=f"Test grid disentanglement #{it}."), "Video Grid Test Data": wandb.Video(grid_anim,caption=f"Test Video Grid #{it}.",fps=5)} if self.config["testing"]["eval_app_transfer"]: app_img_unrelated = eval_batch["app_img_random"].cuda(self.all_devices[0]) x_transferred, *_ = net(app_img_unrelated,x_t, poke,len=0) transfer_grid = make_img_grid(app_img_unrelated,x_t,x_transferred) log_dict.update({"Appearance transfer grid Test Data": wandb.Image(transfer_grid, caption=f"Test_grid appearance transfer #{it}")}) wandb.log(log_dict) return None self.logger.info("Initialize engines...") trainer = Engine(train_step) evaluator = Engine(eval_step) test_img_generator = Engine(eval_visual) self.logger.info("Finish engine initialization...") # checkpointing self.logger.info("Add checkpointing and pbar...") n_saved = 10 self.logger.info(f"Checkpoint saving window is {n_saved}") ckpt_handler = ModelCheckpoint(self.dirs["ckpt"], "reg_ckpt", n_saved=n_saved, require_empty=False) save_dict = {"model": net, "optimizer_dis": optimizer_dis, "optimizer_dyn": optimizer_dyn} trainer.add_event_handler(Events.ITERATION_COMPLETED(every=self.config["testing"]["ckpt_intervall"]), ckpt_handler, save_dict) if self.use_gan: ckpt_handler_disc = ModelCheckpoint(self.dirs["ckpt"], gan_trainer.load_key, n_saved=10, require_empty=False) save_dict_disc = {"model": gan_trainer.disc, "optimizer": gan_trainer.disc_opt} trainer.add_event_handler(Events.ITERATION_COMPLETED(every=self.config["testing"]["ckpt_intervall"]), ckpt_handler_disc, save_dict_disc) if self.use_temp_disc: ckpt_handler_disc_temp = ModelCheckpoint(self.dirs["ckpt"], gan_trainer_temp.load_key, n_saved=10, require_empty=False) save_dict_disc_temp = {"model": gan_trainer_temp.disc, "optimizer": gan_trainer_temp.disc_opt} trainer.add_event_handler(Events.ITERATION_COMPLETED(every=self.config["testing"]["ckpt_intervall"]), ckpt_handler_disc_temp, save_dict_disc_temp) pbar = ProgressBar(ascii=True) pbar.attach(trainer, output_transform=lambda x: x) pbar.attach(evaluator, output_transform=lambda x: x) #reduce the learning rate of the decoder for the image reconstruction task, such that the model focusses more on t --> tk @trainer.on(Events.ITERATION_COMPLETED) def update_lr(engine): self.lr_dec_t = lr_dec_rec(engine.state.iteration) for g in optimizer_dis.param_groups: if g["name"] == "decoder": g["lr"] = self.lr_dec_t @trainer.on(Events.ITERATION_COMPLETED(every=self.config["testing"]["log_intervall"])) def log(engine): it = engine.state.iteration wandb.log({"iteration": it}) # log losses for key in engine.state.output: wandb.log({key: engine.state.output[key]}) data = engine.state.batch if test_dataset.flow_weights: poke = data["poke"][0].cuda(self.all_devices[0]) else: poke = data["poke"].cuda(self.all_devices[0]) x_t = data["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = data["images"][:, 1:].cuda(self.all_devices[0]) if self.config["architecture"]["disentanglement"]: shape_img = data["img_aT"].cuda(self.all_devices[0]) else: shape_img = x_t n_ref_frames = self.config["data"]["n_ref_frames"] - 1 if self.poke_scale_mode else train_dataset.max_frames -1 net.eval() seq_len = x_seq_gt.shape[1] with torch.no_grad(): x_seq_hat, *_ = net(x_t, x_t, poke, len=seq_len, poke_linear=self.poke_scale_mode, n_zero_frames=seq_len-n_ref_frames-1, poke_jump=self.poke_jump) x_t_hat, *_ = net(x_seq_gt[:,-1], shape_img, poke,len=0) #x_t_hat_e, *_ = net(img_aT, img_sT, poke) grid_dis_i = make_img_grid(x_seq_gt[:,-1], shape_img, x_t_hat, x_t, n_logged=n_test_img) grid_dyn = make_flow_grid(x_t, poke, x_seq_hat[:,-1], x_seq_gt[:,-1], n_logged=n_test_img) seq_vis_hat = torch.cat([x_t.unsqueeze(1),x_seq_hat],1) seq_vis_gt = torch.cat([x_t.unsqueeze(1), x_seq_gt], 1) grid_anim = make_video(x_t,poke,seq_vis_hat,seq_vis_gt,n_logged=n_test_img, display_frame_nr=True) wandb.log({"Last Frame Comparison Train Data": wandb.Image(grid_dyn, caption=f"Last frames train grid after {it} train steps."), "Disentanglement Grid Invariance Train Data": wandb.Image(grid_dis_i, caption=f"Invariance Disentanglement Grid on train set after {it} train steps."), "Video Grid Train Data": wandb.Video(grid_anim, caption=f"Train Video Grid after {it} train steps",fps=5)}) #"Disentanglement Grid Equivariance Train Data": wandb.Image(grid_dis_e, caption=f"Eqiuvariance Disentanglement Grid on train set after {it} train steps.") self.logger.info("Initialize metrics...") # compute loss average over epochs # Average(output_transform=lambda x: x["loss_dis"]).attach(trainer, "loss_dis-epoch_avg") if "singlestage" not in self.config["training"] or not self.config["training"]["singlestage"]: Average(output_transform=lambda x: x["loss_dis"]).attach(trainer, "loss_dis-epoch_avg") Average(output_transform=lambda x: x["loss_dis_i_eval"]).attach(evaluator, "loss_dis_i_eval") Average(output_transform=lambda x: x["vgg_loss_dyn"]).attach(trainer, "vgg_loss_dyn-epoch_avg") Average(output_transform=lambda x: x["latent_loss_dyn"]).attach(trainer, "latent_loss_dyn-epoch_avg") if "disentanglement" in self.config["architecture"] and self.config["architecture"]["disentanglement"]: Average(output_transform=lambda x: x["style_loss"]).attach(trainer, "style_loss-epoch_avg") Average(output_transform=lambda x: x["style_loss_eval"]).attach(evaluator, "style_loss_eval") if self.use_temp_disc or self.use_gan: def gan_training_started(engine,epoch, key="gan"): return engine.state.iteration >= self.config[key]["start_iteration"] if self.use_gan: use_patchgan_metrics = MetricUsage(started=Events.EPOCH_STARTED(event_filter=gan_training_started), completed=Events.EPOCH_COMPLETED(event_filter=gan_training_started), iteration_completed=Events.ITERATION_COMPLETED(event_filter=gan_training_started)) # gan losses Average(output_transform=lambda x: x["loss_gen_patch"]).attach(trainer, "loss_gen_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["loss_fmap_patch"]).attach(trainer, "loss_fmap_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["loss_disc_patch"]).attach(trainer, "loss_disc_patch-epoch_avg",usage=use_patchgan_metrics) #if self.config["gan"]["gp_weighflow_video_generatort"] > 0: Average(output_transform=lambda x: x["loss_gp_patch"]).attach(trainer, "loss_gp_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["p_""true_patch"]).attach(trainer, "p_true_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["p_fake_patch"]).attach(trainer, "p_fake_patch-epoch_avg",usage=use_patchgan_metrics) @trainer.on(Events.EPOCH_COMPLETED(event_filter=gan_training_started)) def gan_stuff(engine): gan_trainer.disc_scheduler.step() if self.use_temp_disc: use_tmpgan_metrics = MetricUsage(started=Events.EPOCH_STARTED(event_filter=partial(gan_training_started,key="gan_temp")), completed=Events.EPOCH_COMPLETED(event_filter=partial(gan_training_started,key="gan_temp")), iteration_completed=Events.ITERATION_COMPLETED(event_filter=partial(gan_training_started,key="gan_temp"))) # gan losses Average(output_transform=lambda x: x["loss_gen_temp"]).attach(trainer, "loss_gen_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["loss_fmap_temp"]).attach(trainer, "loss_fmap_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["loss_disc_temp"]).attach(trainer, "loss_disc_temp-epoch_avg",usage=use_tmpgan_metrics) #if self.config["gan"]["gp_weight"] > 0: Average(output_transform=lambda x: x["loss_gp_temp"]).attach(trainer, "loss_gp_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["p_true_temp"]).attach(trainer, "p_true_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["p_fake_temp"]).attach(trainer, "p_fake_temp-epoch_avg",usage=use_tmpgan_metrics) @trainer.on(Events.EPOCH_COMPLETED(event_filter=gan_training_started)) def temp_disc_stuff(engine): gan_trainer_temp.disc_scheduler.step() # evaluation losses Average(output_transform=lambda x: x["vgg_loss_dyn_eval"]).attach(evaluator, "vgg_loss_dyn_eval") Average(output_transform=lambda x: x["latent_loss_dyn_eval"]).attach(evaluator, "latent_loss_dyn_eval") self.logger.info("Finish metric initialization.") @trainer.on(Events.EPOCH_COMPLETED(every=self.config["testing"]["n_epoch_metrics"])) def metrics(engine): # set incpetion model to cpu self.inception_model.eval() self.inception_model.cuda(self.all_devices[0]) self.lpips_fn.cuda(self.all_devices[0]) self.lpips_fn.eval() if self.config["gan_temp"]["use"]: gan_trainer_temp.disc.cpu() if self.config["gan"]["use"]: gan_trainer.disc.cpu() # compute metrics over an epoch self.logger.info(f"Computing metrics after epoch #{engine.state.epoch}") batch_size = eval_sampler.batch_size if self.config["training"]["custom_sampler"] else eval_loader.batch_size bs = 20 if self.is_debug else (int(8000 / batch_size) if len(test_dataset) > 8000 else len(eval_loader)) evaluator.run(eval_loader, max_epochs=1, epoch_length=bs) [wandb.log({key: evaluator.state.metrics[key]}) for key in evaluator.state.metrics] # compute metrics test = np.stack(self.ssims["t"], axis=0) ssim_t = np.mean(np.stack(self.ssims["t"], axis=0)) psnr_t = np.mean(np.stack(self.psnrs["t"], axis=0)) lpips_t = np.mean(np.concatenate(self.lpips["t"], axis=0)) ssim_tk = np.mean(np.stack(self.ssims["tk"], axis=0)) psnr_tk = np.mean(np.stack(self.psnrs["tk"], axis=0)) lpips_avg = np.mean(np.concatenate(self.lpips["tk"], axis=0)) self.lpips_avg = lpips_avg fid_per_frame = {} for key in tqdm(self.fid_feats_real_per_frame, desc="Computing FID per frame"): fid_per_frame[key] = metric_fid(self.fid_feats_real_per_frame[key], self.fid_feats_fake_per_frame[key]) #fid_tk = metric_fid(self.features_real_fid["tk"], self.features_fake_fid["tk"]) fid_avg = np.mean([fid_per_frame[key] for key in fid_per_frame]) log_dict = {"ssim-t": ssim_t, "psnr-t": psnr_t, "lpips-t": lpips_t,"ssim-tk": ssim_tk, "psnr-tk": psnr_tk, "fid-tk": fid_avg, "lpips-avg": lpips_avg} # add histograms for per-frame-metrics self.lpips_per_frame = {key: np.concatenate(self.lpips_per_frame[key], axis=0).mean() for key in self.lpips_per_frame} self.ssims_per_frame = {key: np.stack(self.ssims_per_frame[key], axis=0).mean() for key in self.ssims_per_frame} self.psnrs_per_frame = {key: np.stack(self.psnrs_per_frame[key], axis=0).mean() for key in self.psnrs_per_frame} # self.ssims_per_frame_pl = {key: np.stack(self.ssims_per_frame_pl[key], axis=0).mean() for key in self.ssims_per_frame_pl} # self.psnrs_per_frame_pl = {key: np.stack(self.psnrs_per_frame_pl[key], axis=0).mean() for key in self.psnrs_per_frame_pl} x = [k + 1 for k in self.lpips_per_frame] make_plot(x, list(self.lpips_per_frame.values()), "LPIPS of predicted frames", ylabel="Average LPIPS") make_plot(x, list(self.ssims_per_frame.values()), "SSIM of predicted frames", ylabel="Average SSIM") make_plot(x, list(self.psnrs_per_frame.values()), "PSNR of predicted frames", ylabel="Average PSNR") make_plot(x, list(fid_per_frame.values()), "FIDs of predicted frames", ylabel="FID") wandb.log(log_dict) # clear collection arrays self.__clear_metric_arrs() self.inception_model.cpu() self.lpips_fn.cpu() if self.config["gan_temp"]["use"]: gan_trainer_temp.disc.cuda(self.all_devices[0]) if self.config["gan"]["use"]: gan_trainer.disc.cuda(self.all_devices[0]) # set #toggle_gpu(True) @trainer.on(Events.ITERATION_COMPLETED(every=self.config["testing"]["test_img_intervall"])) def make_test_grid(engine): test_img_generator.run(test_loader, max_epochs=1, epoch_length=self.config["testing"]["test_it"]) @trainer.on(Events.EPOCH_COMPLETED) def log_train_avg(engine): wandb.log({"epoch": engine.state.epoch}) [wandb.log({key: engine.state.metrics[key]}) for key in engine.state.metrics] # also perform scheduler step scheduler_dis.step() scheduler_dyn.step() def score_fn(engine): assert self.lpips_avg is not None return -self.lpips_avg # define best ckpt best_ckpt_handler = ModelCheckpoint(self.dirs["ckpt"],filename_prefix="ckpt_metric" ,score_function=score_fn,score_name="lpips",n_saved=5,require_empty=False) trainer.add_event_handler(Events.EPOCH_COMPLETED(every=self.config["testing"]["n_epoch_metrics"]),best_ckpt_handler,save_dict) @trainer.on(Events.STARTED) def set_start_it(engine): self.logger.info(f'Engine starting from iteration {start_it}, epoch {start_epoch}') engine.state.iteration = start_it engine.state.epoch = start_epoch # run everything n_step_per_epoch = 10 if self.is_debug else len(train_loader) self.logger.info("Start training...") trainer.run(train_loader, max_epochs=n_epoch_train, epoch_length=n_step_per_epoch) self.logger.info("End training.") def test(self): from tqdm import tqdm import cv2 from os import makedirs,path mod_ckpt, _ = self._load_ckpt("reg_ckpt", single_opt=False) dataset, transforms = get_dataset(config=self.config["data"]) test_dataset = dataset(transforms, self.datakeys, self.config["data"], train=False) # get datasets for training and testing def w_init_fn(worker_id): return np.random.seed(np.random.get_state()[1][0] + worker_id) if self.custom_sampler: test_sampler = FixedLengthSampler(test_dataset, batch_size=self.config["testing"]["test_batch_size"], shuffle=True, drop_last=True, weighting=test_dataset.obj_weighting, zero_poke=False) test_loader = DataLoader( test_dataset, batch_sampler=test_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], # worker_init_fn=w_init_fn, ) self.logger.info("Using custom data sampler") else: test_sampler = RandomSampler(test_dataset, ) test_loader = DataLoader(test_dataset, sampler=test_sampler, batch_size=16, num_workers=self.config["data"]["num_workers"], worker_init_fn=w_init_fn, drop_last=True) self.logger.info("Using common torch sampler") # define model self.logger.info(f"Sequence length is {test_dataset.max_frames}") self.logger.info(f"Load model...") net_model = SkipSequenceModel if self.config["architecture"]["use_skip_model"] else SingleScaleBaseline net = net_model(spatial_size=self.config["data"]["spatial_size"], config=self.config["architecture"], ) weights = [5, 5, 0] self.logger.info( f"Number of trainable parameters in model is {sum(p.numel() for p in net.parameters())}" ) net.load_state_dict(mod_ckpt) net.cuda(self.all_devices[0]) self.logger.info("Model on gpu!") net.eval() if self.config["testing"]["mode"] == "metrics": fid_feats_real_per_frame = {} fid_feats_fake_per_frame = {} def metric_step(engine, eval_batch): net.eval() out_dict = {} with torch.no_grad(): # prepare data weights = None if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) weights = eval_batch["poke"][1].cuda(self.all_devices[0]) else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:, 1:].cuda(self.all_devices[0]) n_ref_frames = self.config["data"]["n_ref_frames"] -1 if "n_ref_frames" in self.config["data"] else self.config["data"]["max_frames"] # eval forward passes seq_len = x_seq_gt.shape[1] x_t_hat, sigma_t, _, alpha = net(x_t, x_t, poke, len=0) x_seq_hat, _, sigmas_hat, _ = net(x_t, x_t, poke, len=seq_len, poke_linear=self.poke_scale_mode, n_zero_frames=seq_len-n_ref_frames-1, poke_jump=self.poke_jump) if weights is not None and self.config["testing"]["metrics_on_patches"]: x_seq_hat = get_patches(x_seq_hat, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) x_seq_gt = get_patches(x_seq_gt, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) # apply inception model for fid calculation at time t+k for t in range(x_seq_gt.shape[1]): real_features_t = self.inception_model(x_seq_gt[:, t]).cpu().numpy() fake_features_t = self.inception_model(x_seq_hat[:, t]).cpu().numpy() if t not in fid_feats_fake_per_frame: fid_feats_fake_per_frame.update({t: fake_features_t}) fid_feats_real_per_frame.update({t: real_features_t}) else: fid_feats_fake_per_frame[t] = np.concatenate([fid_feats_fake_per_frame[t], fake_features_t], axis=0) fid_feats_real_per_frame[t] = np.concatenate([fid_feats_real_per_frame[t], real_features_t], axis=0) ssim_tk, ssim_per_frame = ssim_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) psnr_tk, psnr_per_frame = psnr_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) lpips_avg, lpips_per_frame = metric_lpips(x_seq_gt, x_seq_hat, self.lpips_fn, reduce=False, return_per_frame=True) # append to arrays self.psnrs["tk"].append(psnr_tk) self.ssims["tk"].append(ssim_tk) self.lpips["tk"].append(lpips_avg) # append the values of the respective sequence length [self.ssims_per_frame[key].append(ssim_per_frame[key]) if key in self.ssims_per_frame else self.ssims_per_frame.update({key: [ssim_per_frame[key]]}) for key in ssim_per_frame] [self.psnrs_per_frame[key].append(psnr_per_frame[key]) if key in self.psnrs_per_frame else self.psnrs_per_frame.update({key: [psnr_per_frame[key]]}) for key in psnr_per_frame] [self.lpips_per_frame[key].append(lpips_per_frame[key]) if key in self.lpips_per_frame else self.lpips_per_frame.update({key: [lpips_per_frame[key]]}) for key in lpips_per_frame] return out_dict evaluator = Engine(metric_step) self.logger.info("Initialize inception model...") self.inception_model = FIDInceptionModel() self.logger.info("Finished initialization of inception model...") # note that lpips is exactly vgg-cosine similarity as proposed in the google papers and savp self.lpips_fn = LPIPS(net="vgg") pbar = ProgressBar(ascii=True) pbar.attach(evaluator, output_transform=lambda x: x) # set incpetion model to cpu self.inception_model.eval() self.inception_model.cuda(self.all_devices[0]) self.lpips_fn.cuda(self.all_devices[0]) self.lpips_fn.eval() # compute metrics over an epoch self.logger.info(f"Start metrics computation.") batch_size = test_sampler.batch_size if self.custom_sampler else test_loader.batch_size el = (int(8000 / batch_size) if len(test_dataset) > 8000 else len(test_loader)) evaluator.run(test_loader, max_epochs=1, epoch_length=el) # [wandb.log({key: evaluator.state.metrics[key]}) for key in evaluator.state.metrics] # compute metrics ssim_tk = np.mean(np.stack(self.ssims["tk"], axis=0)) psnr_tk = np.mean(np.stack(self.psnrs["tk"], axis=0)) lpips_avg = np.mean(np.concatenate(self.lpips["tk"], axis=0)) assert list(fid_feats_real_per_frame.keys()) == list(fid_feats_fake_per_frame.keys()) fid_per_frame = {} for key in tqdm(fid_feats_real_per_frame, desc="Computing FID per frame"): fid_per_frame[key] = metric_fid(fid_feats_real_per_frame[key], fid_feats_fake_per_frame[key]) # fid_tk = metric_fid(self.features_real_fid["tk"], self.features_fake_fid["tk"]) fid_avg = np.mean([fid_per_frame[key] for key in fid_per_frame]) log_dict = {"ssim-avg-temp": ssim_tk, "psnr-avg_temp": psnr_tk, "fid-avg_temp": fid_avg, "lpips-avg-temp": lpips_avg} # add histograms for per-frame-metrics self.lpips_per_frame = {key: np.concatenate(self.lpips_per_frame[key], axis=0).mean() for key in self.lpips_per_frame} self.ssims_per_frame = {key: np.stack(self.ssims_per_frame[key], axis=0).mean() for key in self.ssims_per_frame} self.psnrs_per_frame = {key: np.stack(self.psnrs_per_frame[key], axis=0).mean() for key in self.psnrs_per_frame} savedir = path.join(self.dirs["generated"], "metric_summaries") makedirs(savedir, exist_ok=True) x = [k + 1 for k in self.lpips_per_frame] make_plot(x, list(self.lpips_per_frame.values()), "LPIPS of predicted frames", ylabel="Average LPIPS", savename=path.join(savedir, "lpips.svg")) make_plot(x, list(self.ssims_per_frame.values()), "SSIM of predicted frames", ylabel="Average SSIM", savename=path.join(savedir, "ssim.svg")) make_plot(x, list(self.psnrs_per_frame.values()), "PSNR of predicted frames", ylabel="Average PSNR", savename=path.join(savedir, "psnr.svg")) make_plot(x, list(fid_per_frame.values()), "FIDs of predicted frames", ylabel="FID", savename=path.join(savedir, "fid.svg")) self.logger.info("Averaged metrics: ") for key in log_dict: self.logger.info(f'{key}: {log_dict[key]}') elif self.config["testing"]["mode"] == "fvd": batch_size = test_sampler.batch_size if self.custom_sampler else test_loader.batch_size el = (int(1000 / batch_size) if len(test_dataset) > 1000 else len(test_loader)) real_samples = [] fake_samples = [] real_samples_out = [] def generate_vids(engine, eval_batch): net.eval() with torch.no_grad(): # prepare data if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) else: poke = eval_batch["poke"].cuda(self.all_devices[0]) if engine.state.iteration < el: x_t = eval_batch["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:, 1:].cuda(self.all_devices[0]) n_ref_frames = self.config["data"]["n_ref_frames"] -1 if "n_ref_frames" in self.config["data"] else self.config["data"]["max_frames"] # eval forward passes seq_len = x_seq_gt.shape[1] x_seq_hat, *_ = net(x_t, x_t, poke, len=seq_len,poke_linear=self.poke_scale_mode, n_zero_frames=seq_len-n_ref_frames-1, poke_jump=self.poke_jump) real_batch = ((x_seq_gt + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy().astype(np.uint8) fake_batch = ((x_seq_hat + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy().astype(np.uint8) real_samples.append(real_batch) fake_samples.append(fake_batch) else: real_batch = ((eval_batch["images"][:, 1:] + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy().astype(np.uint8) real_samples_out.append(real_batch) generator = Engine(generate_vids) pbar = ProgressBar(ascii=True) pbar.attach(generator, output_transform=lambda x: x) self.logger.info(f"Start collecting sequences for fvd computation...") generator.run(test_loader, max_epochs=1, epoch_length=el) savedir = path.join(self.dirs["generated"], "samples_fvd") savedir_exmpls = path.join(savedir,"vid_examples") makedirs(savedir, exist_ok=True) makedirs(savedir_exmpls, exist_ok=True) real_samples = np.stack(real_samples, axis=0) fake_samples = np.stack(fake_samples, axis=0) real_samples_out = np.stack(real_samples_out, axis=0) n_ex = 0 self.logger.info(f"Generating example videos") for i,(r,f) in enumerate(zip(real_samples,fake_samples)): savename = path.join(savedir_exmpls,f"sample{i}.mp4") r = np.concatenate([v for v in r],axis=2) f = np.concatenate([v for v in f],axis=2) all = np.concatenate([r,f],axis=1) writer = cv2.VideoWriter( savename, cv2.VideoWriter_fourcc(*"MP4V"), 5, (all.shape[2], all.shape[1]), ) # writer = vio.FFmpegWriter(savename,inputdict=inputdict,outputdict=outputdict) for frame in all: frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) writer.write(frame) writer.release() n_ex+=1 if n_ex > 20: break self.logger.info(f"Saving samples to {savedir}") np.save(path.join(savedir, "real_samples.npy"), real_samples) np.save(path.join(savedir, "fake_samples.npy"), fake_samples) self.logger.info(f'Finish generation of vid samples.')
57,839
54.776278
210
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/experiments/sequence_model.py
import torch from torch.utils.data import DataLoader from torch.optim import Adam from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from ignite.contrib.handlers import ProgressBar from ignite.metrics import Average, MetricUsage import numpy as np import wandb from functools import partial from lpips import LPIPS from tqdm import tqdm from experiments.experiment import Experiment from data import get_dataset from data.samplers import SequenceLengthSampler from models.latent_flow_net import ResidualSequenceBaseline,SkipSequenceModel from models.discriminator import GANTrainer from utils.losses import PerceptualVGG,vgg_loss_agg,DynamicsLoss, pixel_triplet_loss, style_loss, PixelDynamicsLoss, kl_loss from utils.testing import make_flow_grid, make_img_grid, make_video, make_hist, make_plot from utils.metrics import metric_fid, FIDInceptionModel, metric_lpips, psnr_lightning, ssim_lightning from utils.general import linear_var, get_member, get_patches class SequencePokeModel(Experiment): def __init__(self, config, dirs, device): super().__init__(config, dirs, device) self.datakeys = ["images","poke"] if self.config["architecture"]["disentanglement"]: self.datakeys.append("img_aT") self.datakeys.append("app_img_random") if self.config["gan_temp"]["conditional"]: self.datakeys.append("flow") # used for efficient metrics computation self.fid_feats_fake_per_frame = {} self.fid_feats_real_per_frame = {} self.psnrs = {"t": [], "tk": [], "pl" : []} self.ssims = {"t": [], "tk": [], "pl" : []} self.lpips = {"t": [], "tk": []} self.fvd_vids_real = [] self.fvd_vids_fake = [] self.use_gan = self.config["gan"]["use"] self.use_temp_disc = self.config["gan_temp"]["use"] #self.pixel_decoder_loss = self.config["training"]["pixel_dynamics_weight"] > 0 self.lr_dec_t = 0 self.target_dev = None # metrics for each frame self.ssims_per_frame = {} self.lpips_per_frame = {} self.psnrs_per_frame = {} # self.ssims_per_frame_pl = {} # self.psnrs_per_frame_pl = {} self.lpips_avg = None self.use_norm_loss = self.config["training"]["norm_loss_weight"] > 0 if "norm_loss_weight" in self.config["training"] else False if self.use_norm_loss: assert not self.config["architecture"]["dynamics_var"] def __clear_metric_arrs(self): [self.psnrs[key].clear() for key in self.psnrs] [self.ssims[key].clear() for key in self.ssims] [self.lpips[key].clear() for key in self.lpips] self.fvd_vids_real.clear() self.fvd_vids_fake.clear() self.lpips_per_frame = {} self.psnrs_per_frame = {} self.ssims_per_frame = {} self.fid_feats_fake_per_frame = {} self.fid_feats_real_per_frame = {} # self.ssims_per_frame_pl = {} # self.psnrs_per_frame_pl = {} def train(self): ########## checkpoints ########## if self.config["general"]["restart"] and not self.is_debug: mod_ckpt, op_ckpts = self._load_ckpt("reg_ckpt", single_opt=not self.config["training"]["two_stage"]) if self.config["training"]["two_stage"]: op_ckpt_dis = op_ckpts["optimizer_dis"] op_ckpt_dyn = op_ckpts["optimizer_dyn"] else: op_ckpt_dyn = op_ckpts else: mod_ckpt = op_ckpt_dis = op_ckpt_dyn = None # get datasets for training and testing def w_init_fn(worker_id): return np.random.seed(np.random.get_state()[1][0] + worker_id) dataset, transforms = get_dataset(config=self.config["data"]) train_dataset = dataset(transforms, self.datakeys, self.config["data"], train=True) test_datakeys = self.datakeys + ["app_img_random"] if self.config["testing"]["eval_app_transfer"] and "app_img_random" not in self.datakeys else self.datakeys test_datakeys.append("flow") test_dataset = dataset(transforms, test_datakeys, self.config["data"], train=False) train_sampler = SequenceLengthSampler(train_dataset, batch_size=self.config["training"]["batch_size"], shuffle=True, drop_last=True, zero_poke=self.config["data"]["include_zeropoke"]) train_loader = DataLoader( train_dataset, batch_sampler=train_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn ) test_sampler = SequenceLengthSampler(test_dataset, batch_size=self.config["training"]["batch_size"], shuffle=True, drop_last=True, zero_poke=self.config["data"]["include_zeropoke"]) test_loader = DataLoader( test_dataset, batch_sampler=test_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn ) # n_eval_frames = int(test_dataset.max_frames/2) # if int(test_dataset.min_frames + n_eval_frames) % 2 != 0: # n_eval_frames+=1 # no zeropoke for evaluation as zeropoke is only to ensure no reaction when poking outside eval_sampler = SequenceLengthSampler(test_dataset, batch_size=self.config["testing"]["test_batch_size"], shuffle=False, drop_last=True, zero_poke = False ) eval_loader = DataLoader(test_dataset, batch_sampler=eval_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn ) # define model self.logger.info(f"Load model...") net_model = SkipSequenceModel if self.config["architecture"]["use_skip_model"] else ResidualSequenceBaseline net = net_model(spatial_size=self.config["data"]["spatial_size"], config=self.config["architecture"],) self.logger.info( f"Number of trainable parameters in model is {sum(p.numel() for p in net.parameters())}" ) if self.config["general"]["restart"] and mod_ckpt is not None: self.logger.info("Load pretrained paramaters and resume training.") net.load_state_dict(mod_ckpt) if self.parallel: net = torch.nn.DataParallel(net, device_ids=self.all_devices) net.cuda(self.all_devices[0]) self.logger.info("Model on gpu!") # log weights and gradients wandb.watch(net, log="all") # define optimizers # appearance and shape disentanglement if self.config["training"]["two_stage"]: dis_params = [{"params": get_member(net,"shape_enc").parameters(), "name": "shape_encoder"}, {"params": get_member(net,"dec").parameters(), "name": "decoder"} ] if self.config["architecture"]["disentanglement"]: dis_params.append({"params": net.appearance_enc.parameters(), "name": "appearance_encoder", },) optimizer_dis = Adam(dis_params, lr=self.config["training"]["lr"]) if self.config["general"]["restart"] and op_ckpt_dis is not None: self.logger.info("Load state_dict of optimizer.") optimizer_dis.load_state_dict(op_ckpt_dis) milestones = [int(self.config["training"]["n_epochs"] * t) for t in self.config["training"]["tau"]] scheduler_dis = torch.optim.lr_scheduler.MultiStepLR(optimizer_dis, milestones=milestones, gamma=self.config["training"]["lr_reduce"]) # dynamics dyn_params = [{"params": get_member(net,"dynamics_enc").parameters(), "name": "dynamics_encoder", }, {"params": get_member(net,"fusion_block").parameters(), "name": "fusion_block",},] if self.config["training"]["decoder_update_tk"] or not self.config["training"]["two_stage"]: dyn_params.append({"params": get_member(net,"dec").parameters(), "name": "decoder"}) if not self.config["training"]["two_stage"]: dyn_params.append({"params": get_member(net,"shape_enc").parameters(), "name": "shape_encoder"}) optimizer_dyn = Adam(dyn_params, lr = self.config["training"]["lr"]) if self.config["general"]["restart"] and op_ckpt_dyn is not None: self.logger.info("Load state_dict of optimizer.") optimizer_dyn.load_state_dict(op_ckpt_dyn) milestones = [int(self.config["training"]["n_epochs"] * t) for t in self.config["training"]["tau"]] scheduler_dyn = torch.optim.lr_scheduler.MultiStepLR(optimizer_dyn, milestones=milestones, gamma=self.config["training"]["lr_reduce"]) # initialize disc if gan mode is enabled if self.use_gan: gan_trainer = GANTrainer(self.config, self._load_ckpt, self.logger,spatial_size=self.config["data"]["spatial_size"][0] , parallel=self.parallel, devices=self.all_devices, debug=self.is_debug) if self.use_temp_disc: gan_trainer_temp = GANTrainer(self.config, self._load_ckpt,self.logger,spatial_size=self.config["data"]["spatial_size"][0], parallel=self.parallel,devices=self.all_devices, debug=self.is_debug,temporal=True, sequence_length=train_dataset.min_frames) # set start iteration and epoch in case model training is resumed start_it = 0 start_epoch = 0 n_epoch_train = self.config["training"]["n_epochs"] n_epoch_overall = self.config["training"]["n_epochs"] if self.config["general"]["restart"] and op_ckpts is not None: start_it = list(optimizer_dyn.state_dict()["state"].values())[-1]["step"] start_epoch = int(np.floor(start_it / len(train_loader))) assert self.config["training"]["n_epochs"] > start_epoch n_epoch_train = self.config["training"]["n_epochs"] - start_epoch # lr_dec_rec = partial(linear_var,start_it=0, end_it=self.config["training"]["lr_dec_end_it"], start_val=self.config["training"]["lr"], end_val=self.config["training"]["lr_dec_end_val"], clip_min=0, clip_max=self.config["training"]["lr"],) self.lr_dec_t = lr_dec_rec(start_it) # losses self.logger.info("Load VGG") self.vgg = PerceptualVGG() if self.parallel: self.vgg = torch.nn.DataParallel(self.vgg,device_ids=self.all_devices) self.vgg.cuda(self.all_devices[0]) self.logger.info("VGG on gpu") # from torchsummary import summary # summary(vgg.vgg,(3,224,224)) self.logger.info("Initialize persistent losses") latent_dynamics_loss = DynamicsLoss(config=self.config["training"]) pixel_dynamics_loss = partial(pixel_triplet_loss,vgg = self.vgg, diff_pp=self.config["training"]["pixel_dyn_spatial"]) if self.config["training"]["pixel_dynamics_vgg"] else PixelDynamicsLoss() self.logger.info("Finished initializing persistent losses.") def train_step(engine,batch): net.train() # prepare data weights=None loss_dis = 0 out_dict = {} if train_dataset.flow_weights: poke = batch["poke"][0].cuda(self.all_devices[0]) weights = batch["poke"][1].cuda(self.all_devices[0]) else: poke = batch["poke"].cuda(self.all_devices[0]) x_t = batch["images"][:, 0].cuda(self.all_devices[0]) x_seq = batch["images"][:, 1:].cuda(self.all_devices[0]) if self.config["architecture"]["disentanglement"]: shape_img = batch["img_aT"].cuda(self.all_devices[0]) # apply style loss app_img_tr = batch["app_img_random"].cuda(self.all_devices[0]) x_trans, *_ = net(app_img_tr,x_t,poke,len=0) loss_style = style_loss(self.vgg,app_img_tr,x_trans) loss_dis = self.config["training"]["style_loss_weight"] * loss_style out_dict.update({"style_loss": loss_style.item()}) else: shape_img = x_t x_t_hat_i, sigma_t, _, alpha = net(x_seq[:, -1], shape_img, poke, len=0) # disentanglement loss loss_dis = loss_dis + vgg_loss_agg(self.vgg, x_t, x_t_hat_i) if engine.state.epoch <= self.config["training"]["stop_seq_stat"] and self.config["training"]["two_stage"]: #optimize parameter of appearance, shape encoders and decoder optimizer_dis.zero_grad() loss_dis.backward() optimizer_dis.step() out_dict.update({"loss_dis" : loss_dis.item()}) # forward pass for training of dynamics part of the model # dynamics losses seq_len = x_seq.shape[1] seq_rec, mu_delta, sigmas_hat, logstd_delta = net(x_t,shape_img,poke,len=seq_len) sigmas_gt = [] ll_loss_dyn = [] rec_imgs = [] if weights is not None: seq_rec = get_patches(seq_rec,weights,self.config["data"],train_dataset.weight_value_flow, logger=self.logger) x_seq = get_patches(x_seq,weights,self.config["data"],train_dataset.weight_value_flow, logger=self.logger) for n in range(seq_len): x_hat_tn,s_tn,*_ = net(x_seq[:,n],x_seq[:,n],poke,len=0) sigmas_gt.append(s_tn) rec_imgs.append(x_hat_tn) ll_dyn_n = vgg_loss_agg(self.vgg,x_seq[:,n],seq_rec[:,n]) ll_loss_dyn.append(ll_dyn_n) ll_loss_dyn = torch.stack(ll_loss_dyn,dim=0).mean() rec_imgs = torch.stack(rec_imgs,1) if weights is not None: rec_imgs = get_patches(rec_imgs,weights,self.config["data"],train_dataset.weight_value_flow, logger=self.logger) #latent dynamics dyn_losses = [] for s_tk,s_hat_tk in zip(sigmas_gt,sigmas_hat): dyn_losses.append(latent_dynamics_loss(s_hat_tk,s_tk,[])) latent_loss_dyn = torch.stack(dyn_losses).mean() loss_dyn = self.config["training"]["vgg_dyn_weight"] * ll_loss_dyn + self.config["training"]["latent_dynamics_weight"] * latent_loss_dyn if self.use_norm_loss: poke_norms = [] for p in poke: magns = p.norm(dim=0) ids = magns.nonzero(as_tuple=True) if ids[0].shape[0] > 0: poke_norms.append(magns[ids].mean().unsqueeze(0)) else: poke_norms.append(torch.zeros(1).cuda(self.all_devices[0])) poke_norms = torch.cat(poke_norms, 0) norm_loss = ((poke_norms - mu_delta.reshape(poke_norms.shape[0], -1).norm(dim=-1)) ** 2).mean() loss_dyn = loss_dyn + self.config["training"]["norm_loss_weight"] * norm_loss out_dict.update({"norm_loss": norm_loss.item()}) # kl loss for if self.config["architecture"]["dynamics_var"]: kl_dyn = kl_loss(mu_delta,logstd_delta) loss_dyn = self.config["training"]["kl_weight"] * kl_dyn out_dict.update({"kl_dyn": kl_dyn.item()}) # pixel dynamics loss_dec_dyn = [] for n in range(seq_len-1): loss_dec_dyn_tn = pixel_dynamics_loss(x_seq[:,n],x_seq[:,n+1],rec_imgs[:,n],seq_rec[:,n+1]) loss_dec_dyn.append(loss_dec_dyn_tn) loss_dec_dyn = torch.stack(loss_dec_dyn,dim=0).mean() loss_dyn = loss_dyn + self.config["training"]["pixel_dynamics_weight"] * loss_dec_dyn if self.use_gan and engine.state.iteration >= self.config["gan"]["start_iteration"]: if self.config["gan"]["pixel_dynamics"]: offsets = np.random.choice(np.arange(max(1,x_seq.shape[1]-train_dataset.min_frames)),size=x_seq.shape[0]) true_exmpls = torch.stack([seq[o:o+train_dataset.min_frames] for seq, o in zip(x_seq,offsets)],dim=0) fake_exmpls = torch.stack([seq[o:o+train_dataset.min_frames] for seq, o in zip(seq_rec, offsets)], dim=0) x_true = torch.cat([true_exmpls[:,1:],true_exmpls[:,:-1]],dim=2).reshape(-1,2*true_exmpls.shape[2],*true_exmpls.shape[3:]) x_fake = torch.cat([fake_exmpls[:, 1:], true_exmpls[:, :-1]], dim=2).reshape(-1, 2 * fake_exmpls.shape[2], *fake_exmpls.shape[3:]) else: true_exmpls = np.random.choice(np.arange(x_seq.shape[0]*x_seq.shape[1]),self.config["gan"]["n_examples"]) fake_exmpls = np.random.choice(np.arange(seq_rec.shape[0]*seq_rec.shape[1]), self.config["gan"]["n_examples"]) x_true = x_seq.view(-1,*x_seq.shape[2:])[true_exmpls] x_fake = seq_rec.view(-1,*seq_rec.shape[2:])[fake_exmpls] disc_dict, loss_gen, loss_fmap = gan_trainer.train_step(x_true, x_fake) loss_dyn = loss_dyn + self.config["gan"]["gen_weight"] * loss_gen + self.config["gan"]["fmap_weight"] * loss_fmap if self.use_temp_disc and engine.state.iteration >= self.config["gan_temp"]["start_iteration"]: seq_len_act = x_seq.shape[1] offset = int(np.random.choice(np.arange(max(1,seq_len_act-train_dataset.min_frames)),1)) # offset_fake = int(np.random.choice(np.arange(max(1,seq_len_act-seq_len_temp_disc)), 1)) x_fake_tmp = seq_rec[:,offset:offset+train_dataset.min_frames].permute(0,2,1,3,4) x_true_tmp = x_seq[:, offset:offset+train_dataset.min_frames].permute(0,2,1,3,4) if self.config["gan_temp"]["conditional"]: flow = batch["flow"].cuda(self.all_devices[0]) cond = get_patches(flow,weights,self.config["data"],test_dataset.weight_value_flow,self.logger) if test_dataset.flow_weights else flow else: cond = None disc_dict_temp, loss_gen_temp, loss_fmap_temp = gan_trainer_temp.train_step(x_true_tmp,x_fake_tmp,cond) loss_dyn = loss_dyn + self.config["gan_temp"]["gen_weight"] * loss_gen_temp + self.config["gan_temp"]["fmap_weight"] * loss_fmap_temp # optimize parameters of dynamics part optimizer_dyn.zero_grad() loss_dyn.backward() optimizer_dyn.step() out_dict.update({"loss_dyn":loss_dyn.item() ,"vgg_loss_dyn" : ll_loss_dyn.item(), "latent_loss_dyn": latent_loss_dyn.item(), "lr_dec_t": self.lr_dec_t}) if self.use_gan and engine.state.iteration >= self.config["gan"]["start_iteration"]: out_dict.update(disc_dict) out_dict.update({"loss_gen_patch" :loss_gen.item(), "loss_fmap_patch": loss_fmap.item()}) if self.use_temp_disc and engine.state.iteration >= self.config["gan_temp"]["start_iteration"]: out_dict.update(disc_dict_temp) out_dict.update({"loss_gen_temp" :loss_gen_temp.item(), "loss_fmap_temp": loss_fmap_temp.item()}) #if self.pixel_decoder_loss: out_dict.update({"pixel_loss_dec": loss_dec_dyn.item()}) return out_dict self.logger.info("Initialize inception model...") self.inception_model = FIDInceptionModel() self.logger.info("Finished initialization of inception model...") # note that lpips is exactly vgg-cosine similarity as proposed in the google papers and savp self.lpips_fn = LPIPS(net="vgg") def eval_step(engine, eval_batch): net.eval() out_dict = {} with torch.no_grad(): # prepare data weights = None if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) weights = eval_batch["poke"][1].cuda(self.all_devices[0]) else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:,0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:,1:].cuda(self.all_devices[0]) if self.config["architecture"]["disentanglement"]: app_img_tr = eval_batch["app_img_random"].cuda(self.all_devices[0]) x_trans, *_ = net(app_img_tr, x_t, poke,len=0) loss_style = style_loss(self.vgg, app_img_tr, x_trans) out_dict.update({"style_loss_eval": loss_style.item()}) # eval forward passes seq_len = x_seq_gt.shape[1] x_t_hat, sigma_t, _, alpha = net(x_t,x_t,poke,len=0) x_seq_hat, mu_delta, sigmas_hat,_ = net(x_t, x_t, poke,len=seq_len) if weights is not None and self.config["testing"]["metrics_on_patches"]: x_seq_hat = get_patches(x_seq_hat,weights,self.config["data"],test_dataset.weight_value_flow, logger=self.logger) x_seq_gt = get_patches(x_seq_gt, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) sigmas_gt = [] ll_loss_dyn = [] rec_imgs = [] for n in range(seq_len): x_hat_tn, s_tn, *_ = net(x_seq_gt[:, n], x_seq_gt[:, n], poke, len=0) sigmas_gt.append(s_tn) rec_imgs.append(x_hat_tn) ll_dyn_n = vgg_loss_agg(self.vgg, x_seq_gt[:, n], x_seq_hat[:, n]) ll_loss_dyn.append(ll_dyn_n) ll_loss_tk_eval = torch.stack(ll_loss_dyn,dim=0).mean() rec_imgs = torch.stack(rec_imgs,1) if self.use_norm_loss: poke_norms = [] for p in poke: magns = p.norm(dim=0) ids = magns.nonzero(as_tuple=True) if ids[0].shape[0] > 0: poke_norms.append(magns[ids].mean().unsqueeze(0)) else: poke_norms.append(torch.zeros(1).cuda(self.all_devices[0])) poke_norms = torch.cat(poke_norms, 0) norm_loss = ((poke_norms - mu_delta.reshape(poke_norms.shape[0], -1).norm(dim=-1)) ** 2).mean() out_dict.update({"norm_loss": norm_loss.item()}) if weights is not None and self.config["testing"]["metrics_on_patches"]: rec_imgs = get_patches(rec_imgs, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) # apply inception model for fid calculation at time t+k for t in range(x_seq_gt.shape[1]): real_features_t = self.inception_model(x_seq_gt[:, t]).cpu().numpy() fake_features_t = self.inception_model(x_seq_hat[:, t]).cpu().numpy() if t not in self.fid_feats_fake_per_frame: self.fid_feats_fake_per_frame.update({t: fake_features_t}) self.fid_feats_real_per_frame.update({t: real_features_t}) else: self.fid_feats_fake_per_frame[t] = np.concatenate([self.fid_feats_fake_per_frame[t], fake_features_t], axis=0) self.fid_feats_real_per_frame[t] = np.concatenate([self.fid_feats_real_per_frame[t], real_features_t], axis=0) # evaluate training losses # ll_loss_tk_eval = vgg_loss_agg(self.vgg, x_tk, x_tk_hat) ll_loss_t_i_eval = vgg_loss_agg(self.vgg, x_t, x_t_hat) dyn_losses = [] for s_tk, s_hat_tk in zip(sigmas_gt, sigmas_hat): dyn_losses.append(latent_dynamics_loss(s_hat_tk, s_tk, [])) latent_loss_dyn_eval = torch.stack(dyn_losses).mean() out_dict.update({"vgg_loss_dyn_eval": ll_loss_tk_eval.item(), "loss_dis_i_eval": ll_loss_t_i_eval.item(), "latent_loss_dyn_eval": latent_loss_dyn_eval.item()}) #if self.pixel_decoder_loss: #x_t_hat_dec = net.dec(sigma_t, alpha) #loss_dec_dyn = (vgg_loss_agg(self.vgg, x_t_hat_dec, x_tk_hat) - vgg_loss_agg(self.vgg, x_t, x_tk)) ** 2 loss_dec_dyn = [] for n in range(seq_len - 1): loss_dec_dyn_tn = pixel_dynamics_loss(x_seq_gt[:, n], x_seq_gt[:, n + 1], rec_imgs[:,n], x_seq_hat[:, n + 1]) loss_dec_dyn.append(loss_dec_dyn_tn) loss_dec_dyn = torch.stack(loss_dec_dyn, dim=0).mean() out_dict.update({"pixel_loss_dec_eval": loss_dec_dyn.item()}) # compute metrics ssim_t = ssim_lightning(x_t, x_t_hat) psnr_t = psnr_lightning(x_t, x_t_hat) lpips_t = metric_lpips(x_t,x_t_hat, self.lpips_fn, reduce=False) ssim_tk, ssim_per_frame = ssim_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) psnr_tk, psnr_per_frame = psnr_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) lpips_avg, lpips_per_frame = metric_lpips(x_seq_gt, x_seq_hat,self.lpips_fn,reduce=False,return_per_frame=True) # ssim_pl, ssim_pl_per_frame = ssim_lightning(x_seq_gt,x_seq_hat,return_per_frame=True) # psnr_pl, psnr_pl_per_frame = psnr_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) # append to arrays self.lpips["t"].append(lpips_t) self.psnrs["t"].append(psnr_t) self.ssims["t"].append(ssim_t) self.psnrs["tk"].append(psnr_tk) self.ssims["tk"].append(ssim_tk) self.lpips["tk"].append(lpips_avg) #self.ssims["pl"].append(ssim_pl) #self.psnrs["pl"].append(psnr_pl) # append the values of the respective sequence length [self.ssims_per_frame[key].append(ssim_per_frame[key]) if key in self.ssims_per_frame else self.ssims_per_frame.update({key:[ssim_per_frame[key]]}) for key in ssim_per_frame] [self.psnrs_per_frame[key].append(psnr_per_frame[key]) if key in self.psnrs_per_frame else self.psnrs_per_frame.update({key:[psnr_per_frame[key]]}) for key in psnr_per_frame] [self.lpips_per_frame[key].append(lpips_per_frame[key]) if key in self.lpips_per_frame else self.lpips_per_frame.update({key:[lpips_per_frame[key]]}) for key in lpips_per_frame] #[self.ssims_per_frame_pl[key].append(ssim_pl_per_frame[key]) if key in self.ssims_per_frame_pl else self.ssims_per_frame_pl.update({key: [ssim_pl_per_frame[key]]}) for key in ssim_pl_per_frame] #[self.psnrs_per_frame_pl[key].append(psnr_pl_per_frame[key]) if key in self.psnrs_per_frame_pl else self.psnrs_per_frame_pl.update({key: [psnr_pl_per_frame[key]]}) for key in psnr_pl_per_frame] return out_dict # test_it steps are performed while generating test_imgs, there n_test_img is overall number divided by number of test iterations n_test_img = int(self.config["testing"]["n_test_img"] // self.config["testing"]["test_it"]) def eval_visual(engine, eval_batch): net.eval() with torch.no_grad(): # prepare data if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) weights = eval_batch["poke"][1] else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:, 1:].cuda(self.all_devices[0]) flow = eval_batch["flow"] if self.config["architecture"]["disentanglement"]: shape_img = eval_batch["img_aT"].cuda(self.all_devices[0]) else: shape_img = x_t seq_len = x_seq_gt.shape[1] x_seq_hat, *_ = net(x_t,x_t, poke, len=seq_len) x_t_hat , *_ = net(x_seq_gt[:,-1],shape_img,poke,len=0) grid_dis = make_img_grid(x_seq_gt[:,-1],shape_img, x_t_hat,x_t, n_logged=n_test_img) grid_dyn = make_flow_grid(x_t, poke, x_seq_hat[:,-1], x_seq_gt[:,-1], n_logged=n_test_img, flow=flow) seq_vis_hat = torch.cat([x_t.unsqueeze(1), x_seq_hat], 1) seq_vis_gt = torch.cat([x_t.unsqueeze(1), x_seq_gt], 1) grid_anim = make_video(x_t,poke,seq_vis_hat,seq_vis_gt,n_logged=n_test_img,flow=flow) it = engine.state.iteration log_dict = {"Last Frame Comparison Test data": wandb.Image(grid_dyn, caption=f"Last frames test grid #{it}."), "Disentanglement Grid Test Data": wandb.Image(grid_dis, caption=f"Test grid disentanglement #{it}."), "Video Grid Test Data": wandb.Video(grid_anim,caption=f"Test Video Grid #{it}.",fps=5)} if self.config["testing"]["eval_app_transfer"]: app_img_unrelated = eval_batch["app_img_random"].cuda(self.all_devices[0]) x_transferred, *_ = net(app_img_unrelated,x_t, poke,len=0) transfer_grid = make_img_grid(app_img_unrelated,x_t,x_transferred) log_dict.update({"Appearance transfer grid Test Data": wandb.Image(transfer_grid, caption=f"Test_grid appearance transfer #{it}")}) wandb.log(log_dict) return None self.logger.info("Initialize engines...") trainer = Engine(train_step) evaluator = Engine(eval_step) test_img_generator = Engine(eval_visual) self.logger.info("Finish engine initialization...") # checkpointing ckpt_handler = ModelCheckpoint(self.dirs["ckpt"], "reg_ckpt", n_saved=10, require_empty=False) if self.config["training"]["two_stage"]: save_dict = {"model": net, "optimizer_dis": optimizer_dis, "optimizer_dyn": optimizer_dyn} else: save_dict = {"model": net, "optimizer_dyn": optimizer_dyn} trainer.add_event_handler(Events.ITERATION_COMPLETED(every=self.config["testing"]["ckpt_intervall"]), ckpt_handler, save_dict) if self.use_gan: ckpt_handler_disc = ModelCheckpoint(self.dirs["ckpt"], gan_trainer.load_key, n_saved=10, require_empty=False) save_dict_disc = {"model": gan_trainer.disc, "optimizer": gan_trainer.disc_opt} trainer.add_event_handler(Events.ITERATION_COMPLETED(every=self.config["testing"]["ckpt_intervall"]), ckpt_handler_disc, save_dict_disc) if self.use_temp_disc: ckpt_handler_disc_temp = ModelCheckpoint(self.dirs["ckpt"], gan_trainer_temp.load_key, n_saved=10, require_empty=False) save_dict_disc_temp = {"model": gan_trainer_temp.disc, "optimizer": gan_trainer_temp.disc_opt} trainer.add_event_handler(Events.ITERATION_COMPLETED(every=self.config["testing"]["ckpt_intervall"]), ckpt_handler_disc_temp, save_dict_disc_temp) pbar = ProgressBar(ascii=True) pbar.attach(trainer, output_transform=lambda x: x) pbar.attach(evaluator, output_transform=lambda x: x) #reduce the learning rate of the decoder for the image reconstruction task, such that the model focusses more on t --> tk if self.config["training"]["two_stage"]: @trainer.on(Events.ITERATION_COMPLETED) def update_lr(engine): self.lr_dec_t = lr_dec_rec(engine.state.iteration) for g in optimizer_dis.param_groups: if g["name"] == "decoder": g["lr"] = self.lr_dec_t @trainer.on(Events.ITERATION_COMPLETED(every=self.config["testing"]["log_intervall"])) def log(engine): it = engine.state.iteration wandb.log({"iteration": it}) # log losses for key in engine.state.output: wandb.log({key: engine.state.output[key]}) data = engine.state.batch if test_dataset.flow_weights: poke = data["poke"][0].cuda(self.all_devices[0]) else: poke = data["poke"].cuda(self.all_devices[0]) x_t = data["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = data["images"][:, 1:].cuda(self.all_devices[0]) if self.config["architecture"]["disentanglement"]: shape_img = data["img_aT"].cuda(self.all_devices[0]) else: shape_img = x_t net.eval() seq_len = x_seq_gt.shape[1] with torch.no_grad(): x_seq_hat, *_ = net(x_t, x_t, poke, len=seq_len) x_t_hat, *_ = net(x_seq_gt[:,-1], shape_img, poke,len=0) #x_t_hat_e, *_ = net(img_aT, img_sT, poke) grid_dis_i = make_img_grid(x_seq_gt[:,-1], shape_img, x_t_hat, x_t, n_logged=n_test_img) grid_dyn = make_flow_grid(x_t, poke, x_seq_hat[:,-1], x_seq_gt[:,-1], n_logged=n_test_img) seq_vis_hat = torch.cat([x_t.unsqueeze(1),x_seq_hat],1) seq_vis_gt = torch.cat([x_t.unsqueeze(1), x_seq_gt], 1) grid_anim = make_video(x_t,poke,seq_vis_hat,seq_vis_gt,n_logged=n_test_img) wandb.log({"Last Frame Comparison Train Data": wandb.Image(grid_dyn, caption=f"Last frames train grid after {it} train steps."), "Disentanglement Grid Invariance Train Data": wandb.Image(grid_dis_i, caption=f"Invariance Disentanglement Grid on train set after {it} train steps."), "Video Grid Train Data": wandb.Video(grid_anim, caption=f"Train Video Grid after {it} train steps",fps=5)}) #"Disentanglement Grid Equivariance Train Data": wandb.Image(grid_dis_e, caption=f"Eqiuvariance Disentanglement Grid on train set after {it} train steps.") self.logger.info("Initialize metrics...") # compute loss average over epochs # Average(output_transform=lambda x: x["loss_dis"]).attach(trainer, "loss_dis-epoch_avg") Average(output_transform=lambda x: x["loss_dis"]).attach(trainer, "loss_dis-epoch_avg") Average(output_transform=lambda x: x["vgg_loss_dyn"]).attach(trainer, "vgg_loss_dyn-epoch_avg") Average(output_transform=lambda x: x["latent_loss_dyn"]).attach(trainer, "latent_loss_dyn-epoch_avg") if self.config["architecture"]["disentanglement"]: Average(output_transform=lambda x: x["style_loss"]).attach(trainer, "style_loss-epoch_avg") Average(output_transform=lambda x: x["style_loss_eval"]).attach(evaluator, "style_loss_eval") if self.use_norm_loss: Average(output_transform=lambda x: x["norm_loss"]).attach(trainer, "norm_loss-epoch_avg") Average(output_transform=lambda x: x["norm_loss"]).attach(evaluator, "norm_loss_eval") if self.config["architecture"]["dynamics_var"]: Average(output_transform=lambda x: x["kl_dyn"]).attach(trainer, "kl_dyn_loss-epoch_avg") if self.use_temp_disc or self.use_gan: def gan_training_started(engine,epoch, key="gan"): return engine.state.iteration >= self.config[key]["start_iteration"] if self.use_gan: use_patchgan_metrics = MetricUsage(started=Events.EPOCH_STARTED(event_filter=gan_training_started), completed=Events.EPOCH_COMPLETED(event_filter=gan_training_started), iteration_completed=Events.ITERATION_COMPLETED(event_filter=gan_training_started)) # gan losses Average(output_transform=lambda x: x["loss_gen_patch"]).attach(trainer, "loss_gen_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["loss_fmap_patch"]).attach(trainer, "loss_fmap_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["loss_disc_patch"]).attach(trainer, "loss_disc_patch-epoch_avg",usage=use_patchgan_metrics) #if self.config["gan"]["gp_weighflow_video_generatort"] > 0: Average(output_transform=lambda x: x["loss_gp_patch"]).attach(trainer, "loss_gp_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["p_""true_patch"]).attach(trainer, "p_true_patch-epoch_avg",usage=use_patchgan_metrics) Average(output_transform=lambda x: x["p_fake_patch"]).attach(trainer, "p_fake_patch-epoch_avg",usage=use_patchgan_metrics) @trainer.on(Events.EPOCH_COMPLETED(event_filter=gan_training_started)) def gan_stuff(engine): gan_trainer.disc_scheduler.step() if self.use_temp_disc: use_tmpgan_metrics = MetricUsage(started=Events.EPOCH_STARTED(event_filter=partial(gan_training_started,key="gan_temp")), completed=Events.EPOCH_COMPLETED(event_filter=partial(gan_training_started,key="gan_temp")), iteration_completed=Events.ITERATION_COMPLETED(event_filter=partial(gan_training_started,key="gan_temp"))) # gan losses Average(output_transform=lambda x: x["loss_gen_temp"]).attach(trainer, "loss_gen_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["loss_fmap_temp"]).attach(trainer, "loss_fmap_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["loss_disc_temp"]).attach(trainer, "loss_disc_temp-epoch_avg",usage=use_tmpgan_metrics) #if self.config["gan"]["gp_weight"] > 0: Average(output_transform=lambda x: x["loss_gp_temp"]).attach(trainer, "loss_gp_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["p_true_temp"]).attach(trainer, "p_true_temp-epoch_avg",usage=use_tmpgan_metrics) Average(output_transform=lambda x: x["p_fake_temp"]).attach(trainer, "p_fake_temp-epoch_avg",usage=use_tmpgan_metrics) @trainer.on(Events.EPOCH_COMPLETED(event_filter=gan_training_started)) def temp_disc_stuff(engine): gan_trainer_temp.disc_scheduler.step() # if self.pixel_decoder_loss: Average(output_transform=lambda x: x["pixel_loss_dec"]).attach(trainer, "pixel_loss_dec-epoch_avg") Average(output_transform=lambda x: x["pixel_loss_dec_eval"]).attach(evaluator, "pixel_loss_dec_eval") # evaluation losses Average(output_transform=lambda x: x["vgg_loss_dyn_eval"]).attach(evaluator, "vgg_loss_dyn_eval") Average(output_transform=lambda x: x["loss_dis_i_eval"]).attach(evaluator, "loss_dis_i_eval") Average(output_transform=lambda x: x["latent_loss_dyn_eval"]).attach(evaluator, "latent_loss_dyn_eval") self.logger.info("Finish metric initialization.") @trainer.on(Events.EPOCH_COMPLETED(every=self.config["testing"]["n_epoch_metrics"])) def metrics(engine): # set incpetion model to cpu self.inception_model.eval() self.inception_model.cuda(self.all_devices[0]) self.lpips_fn.cuda(self.all_devices[0]) self.lpips_fn.eval() if self.use_temp_disc: gan_trainer_temp.disc.cpu() if self.use_gan: gan_trainer.disc.cpu() # compute metrics over an epoch self.logger.info(f"Computing metrics after epoch #{engine.state.epoch}") bs = 20 if self.is_debug else (int(8000 / eval_sampler.batch_size) if len(test_dataset) > 8000 else len(eval_loader)) evaluator.run(eval_loader, max_epochs=1, epoch_length=bs) [wandb.log({key: evaluator.state.metrics[key]}) for key in evaluator.state.metrics] # compute metrics ssim_t = np.mean(np.stack(self.ssims["t"], axis=0)) psnr_t = np.mean(np.stack(self.psnrs["t"], axis=0)) lpips_t = np.mean(np.concatenate(self.lpips["t"], axis=0)) ssim_tk = np.mean(np.stack(self.ssims["tk"], axis=0)) psnr_tk = np.mean(np.stack(self.psnrs["tk"], axis=0)) lpips_avg = np.mean(np.concatenate(self.lpips["tk"], axis=0)) self.lpips_avg = lpips_avg fid_per_frame = {} for key in tqdm(self.fid_feats_real_per_frame, desc="Computing FID per frame"): fid_per_frame[key] = metric_fid(self.fid_feats_real_per_frame[key], self.fid_feats_fake_per_frame[key]) # fid_tk = metric_fid(self.features_real_fid["tk"], self.features_fake_fid["tk"]) fid_avg = np.mean([fid_per_frame[key] for key in fid_per_frame]) log_dict = {"ssim-t": ssim_t, "psnr-t": psnr_t, "fid-avg": fid_avg, "lpips-t": lpips_t,"ssim-tk": ssim_tk, "psnr-tk": psnr_tk, "lpips-avg": lpips_avg} # add histograms for per-frame-metrics self.lpips_per_frame = {key: np.concatenate(self.lpips_per_frame[key], axis=0).mean() for key in self.lpips_per_frame} self.ssims_per_frame = {key: np.stack(self.ssims_per_frame[key], axis=0).mean() for key in self.ssims_per_frame} self.psnrs_per_frame = {key: np.stack(self.psnrs_per_frame[key], axis=0).mean() for key in self.psnrs_per_frame} # self.ssims_per_frame_pl = {key: np.stack(self.ssims_per_frame_pl[key], axis=0).mean() for key in self.ssims_per_frame_pl} # self.psnrs_per_frame_pl = {key: np.stack(self.psnrs_per_frame_pl[key], axis=0).mean() for key in self.psnrs_per_frame_pl} x = [k+1 for k in self.lpips_per_frame] make_plot(x,list(self.lpips_per_frame.values()),"LPIPS of predicted frames", ylabel="Average LPIPS",) make_plot(x, list(self.ssims_per_frame.values()), "SSIM of predicted frames", ylabel="Average SSIM",) make_plot(x, list(self.psnrs_per_frame.values()), "PSNR of predicted frames", ylabel="Average PSNR",) make_plot(x, list(fid_per_frame.values()), "FIDs of predicted frames", ylabel="FID") wandb.log(log_dict) # clear collection arrays self.__clear_metric_arrs() self.inception_model.cpu() self.lpips_fn.cpu() if self.use_temp_disc: gan_trainer_temp.disc.cuda(self.all_devices[0]) if self.use_gan: gan_trainer.disc.cuda(self.all_devices[0]) @trainer.on(Events.ITERATION_COMPLETED(every=self.config["testing"]["test_img_intervall"])) def make_test_grid(engine): test_img_generator.run(test_loader, max_epochs=1, epoch_length=self.config["testing"]["test_it"]) @trainer.on(Events.EPOCH_COMPLETED) def log_train_avg(engine): wandb.log({"epoch": engine.state.epoch}) [wandb.log({key: engine.state.metrics[key]}) for key in engine.state.metrics] # also perform scheduler step if self.config["training"]["two_stage"]: scheduler_dis.step() scheduler_dyn.step() def score_fn(engine): assert self.lpips_avg is not None return -self.lpips_avg # define best ckpt best_ckpt_handler = ModelCheckpoint(self.dirs["ckpt"],filename_prefix="ckpt_metric" ,score_function=score_fn,score_name="lpips",n_saved=5,require_empty=False) trainer.add_event_handler(Events.EPOCH_COMPLETED(every=self.config["testing"]["n_epoch_metrics"]),best_ckpt_handler,save_dict) @trainer.on(Events.STARTED) def set_start_it(engine): self.logger.info(f'Engine starting from iteration {start_it}, epoch {start_epoch}') engine.state.iteration = start_it engine.state.epoch = start_epoch # run everything n_step_per_epoch = 10 if self.is_debug else len(train_loader) self.logger.info("Start training...") trainer.run(train_loader, max_epochs=n_epoch_overall, epoch_length=n_step_per_epoch) self.logger.info("End training.") def test(self): from tqdm import tqdm import cv2 from os import makedirs,path # load checkpoint mod_ckpt, _ = self._load_ckpt("reg_ckpt", single_opt=False,use_best=self.config["testing"]["best_ckpt"],) #dir="/export/data/ablattma/visual_poking/final_models/final_models/var_length/iper/",name="baseline_wo_upsampling.pt") dataset, transforms = get_dataset(config=self.config["data"]) test_dataset = dataset(transforms, self.datakeys, self.config["data"], train=False) # get datasets for training and testing def w_init_fn(worker_id): return np.random.seed(np.random.get_state()[1][0] + worker_id) test_sampler = SequenceLengthSampler(test_dataset, batch_size=self.config["testing"]["test_batch_size"] if self.config["testing"]["test_batch_size"] > 16 else 16, shuffle=False, drop_last=True, zero_poke=False) test_loader = DataLoader( test_dataset, batch_sampler=test_sampler, num_workers=0 if self.is_debug else self.config["data"]["num_workers"], worker_init_fn=w_init_fn ) # define model self.logger.info(f"Load model...") net_model = SkipSequenceModel if self.config["architecture"]["use_skip_model"] else ResidualSequenceBaseline net = net_model(spatial_size=self.config["data"]["spatial_size"], config=self.config["architecture"],) weights = [5,5,0] self.logger.info( f"Number of trainable parameters in model is {sum(p.numel() for p in net.parameters())}" ) net.load_state_dict(mod_ckpt) net.cuda(self.all_devices[0]) self.logger.info("Model on gpu!") net.eval() if self.config["testing"]["mode"] == "noise_comp": target_dir = path.join(self.dirs["generated"], "test_decoder_noise") makedirs(target_dir, exist_ok=True) n_gen = 0 for i,batch in enumerate(test_loader): if n_gen > 100: break if test_dataset.flow_weights: poke = batch["poke"][0].cuda(self.all_devices[0]) else: poke = batch["poke"].cuda(self.all_devices[0]) img = batch["images"][:,0].cuda(self.all_devices[0]) with torch.no_grad(): x_rec, sigmas, *_ = net(img,img,poke,len=0) # add noise to sigmas for n in tqdm(range(x_rec.shape[0]),desc=f'Generating noise xs for batch #{i}'): simgs_noise = [weights[k] * torch.randn((self.config["testing"]["n_examples_noise"],*sigm[n].shape)).cuda(self.all_devices[0])+sigm[n].unsqueeze(0) for k,sigm in enumerate(sigmas)] # xs = torch.stack([img[n]]*self.config["testing"]["n_examples"]) simgs_noise.reverse() xs_noise = net.dec(simgs_noise,None,del_shape=True) xs_disp = torch.cat([x_rec[n].unsqueeze(0),xs_noise],dim=0) xs_disp = np.concatenate([((x.permute(1,2,0).cpu().numpy()+ 1.)*127.5).astype(np.uint8) for x in xs_disp],axis=1) n_gen += 1 xs_disp = cv2.cvtColor(xs_disp,cv2.COLOR_RGB2BGR) cv2.imwrite(path.join(target_dir,f"noise_imgs_{n_gen}.png"),xs_disp) elif self.config["testing"]["mode"] == "metrics": fid_feats_real_per_frame = {} fid_feats_fake_per_frame = {} def metric_step(engine, eval_batch): net.eval() out_dict = {} with torch.no_grad(): # prepare data weights = None if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) weights = eval_batch["poke"][1].cuda(self.all_devices[0]) else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:, 1:].cuda(self.all_devices[0]) # eval forward passes seq_len = x_seq_gt.shape[1] x_t_hat, sigma_t, _, alpha = net(x_t, x_t, poke, len=0) x_seq_hat, _, sigmas_hat, _ = net(x_t, x_t, poke, len=seq_len) if weights is not None and self.config["testing"]["metrics_on_patches"]: x_seq_hat = get_patches(x_seq_hat, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) x_seq_gt = get_patches(x_seq_gt, weights, self.config["data"], test_dataset.weight_value_flow, logger=self.logger) # apply inception model for fid calculation at time t+k for t in range(x_seq_gt.shape[1]): real_features_t = self.inception_model(x_seq_gt[:, t]).cpu().numpy() fake_features_t = self.inception_model(x_seq_hat[:, t]).cpu().numpy() if t not in fid_feats_fake_per_frame: fid_feats_fake_per_frame.update({t:fake_features_t}) fid_feats_real_per_frame.update({t: real_features_t}) else: fid_feats_fake_per_frame[t] = np.concatenate([fid_feats_fake_per_frame[t],fake_features_t], axis=0) fid_feats_real_per_frame[t] = np.concatenate([fid_feats_real_per_frame[t], real_features_t], axis=0) #self.features_real_fid["tk"].append(real_features_tk) #self.features_fake_fid["tk"].append(fake_features_tk) self.fvd_vids_real.append(x_seq_gt.cpu()) self.fvd_vids_fake.append(x_seq_hat.cpu()) ssim_tk, ssim_per_frame = ssim_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) psnr_tk, psnr_per_frame = psnr_lightning(x_seq_gt, x_seq_hat, return_per_frame=True) lpips_avg, lpips_per_frame = metric_lpips(x_seq_gt, x_seq_hat, self.lpips_fn, reduce=False, return_per_frame=True) # append to arrays self.psnrs["tk"].append(psnr_tk) self.ssims["tk"].append(ssim_tk) self.lpips["tk"].append(lpips_avg) # append the values of the respective sequence length [self.ssims_per_frame[key].append(ssim_per_frame[key]) if key in self.ssims_per_frame else self.ssims_per_frame.update({key: [ssim_per_frame[key]]}) for key in ssim_per_frame] [self.psnrs_per_frame[key].append(psnr_per_frame[key]) if key in self.psnrs_per_frame else self.psnrs_per_frame.update({key: [psnr_per_frame[key]]}) for key in psnr_per_frame] [self.lpips_per_frame[key].append(lpips_per_frame[key]) if key in self.lpips_per_frame else self.lpips_per_frame.update({key: [lpips_per_frame[key]]}) for key in lpips_per_frame] return out_dict evaluator = Engine(metric_step) self.logger.info("Initialize inception model...") self.inception_model = FIDInceptionModel() self.logger.info("Finished initialization of inception model...") # note that lpips is exactly vgg-cosine similarity as proposed in the google papers and savp self.lpips_fn = LPIPS(net="vgg") pbar = ProgressBar(ascii=True) pbar.attach(evaluator, output_transform=lambda x: x) # set incpetion model to cpu self.inception_model.eval() self.inception_model.cuda(self.all_devices[0]) self.lpips_fn.cuda(self.all_devices[0]) self.lpips_fn.eval() # compute metrics over an epoch self.logger.info(f"Start metrics computation.") el = (int(8000 / test_sampler.batch_size) if len(test_dataset) > 8000 else len(test_loader)) evaluator.run(test_loader, max_epochs=1, epoch_length=el) # [wandb.log({key: evaluator.state.metrics[key]}) for key in evaluator.state.metrics] # compute metrics ssim_tk = np.mean(np.stack(self.ssims["tk"], axis=0)) psnr_tk = np.mean(np.stack(self.psnrs["tk"], axis=0)) lpips_avg = np.mean(np.concatenate(self.lpips["tk"], axis=0)) assert list(fid_feats_real_per_frame.keys()) == list(fid_feats_fake_per_frame.keys()) fid_per_frame = {} for key in tqdm(fid_feats_real_per_frame, desc="Computing FID per frame"): fid_per_frame[key] = metric_fid(fid_feats_real_per_frame[key], fid_feats_fake_per_frame[key]) #fid_tk = metric_fid(self.features_real_fid["tk"], self.features_fake_fid["tk"]) fid_avg = np.mean([fid_per_frame[key] for key in fid_per_frame]) log_dict = {"ssim-avg-temp": ssim_tk, "psnr-avg_temp": psnr_tk, "fid-avg_temp": fid_avg, "lpips-avg-temp": lpips_avg} # add histograms for per-frame-metrics self.lpips_per_frame = {key: np.concatenate(self.lpips_per_frame[key], axis=0).mean() for key in self.lpips_per_frame} self.ssims_per_frame = {key: np.stack(self.ssims_per_frame[key], axis=0).mean() for key in self.ssims_per_frame} self.psnrs_per_frame = {key: np.stack(self.psnrs_per_frame[key], axis=0).mean() for key in self.psnrs_per_frame} savedir = path.join(self.dirs["generated"],"metric_summaries") makedirs(savedir, exist_ok=True) x = [k+1 for k in self.lpips_per_frame] make_plot(x,list(self.lpips_per_frame.values()),"LPIPS of predicted frames", ylabel="Average LPIPS",savename=path.join(savedir,"lpips.svg")) make_plot(x, list(self.ssims_per_frame.values()), "SSIM of predicted frames", ylabel="Average SSIM", savename=path.join(savedir, "ssim.svg")) make_plot(x, list(self.psnrs_per_frame.values()), "PSNR of predicted frames", ylabel="Average PSNR", savename=path.join(savedir, "psnr.svg")) make_plot(x,list(fid_per_frame.values()), "FIDs of predicted frames", ylabel="FID", savename=path.join(savedir, "fid.svg")) self.logger.info("Averaged metrics: ") for key in log_dict: self.logger.info(f'{key}: {log_dict[key]}') elif self.config["testing"]["mode"] == "fvd": test_sampler = SequenceLengthSampler(test_dataset,16,shuffle=True, drop_last=True, n_frames=10,zero_poke=False) test_loader = DataLoader(test_dataset,batch_sampler=test_sampler,num_workers=self.config["data"]["num_workers"],worker_init_fn=w_init_fn) real_samples = [] fake_samples = [] def generate_vids(engine,eval_batch): net.eval() with torch.no_grad(): # prepare data if test_dataset.flow_weights: poke = eval_batch["poke"][0].cuda(self.all_devices[0]) else: poke = eval_batch["poke"].cuda(self.all_devices[0]) x_t = eval_batch["images"][:, 0].cuda(self.all_devices[0]) x_seq_gt = eval_batch["images"][:, 1:].cuda(self.all_devices[0]) # eval forward passes seq_len = x_seq_gt.shape[1] x_seq_hat, *_ = net(x_t, x_t, poke, len=seq_len) real_batch = ((x_seq_gt + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy().astype(np.uint8) fake_batch = ((x_seq_hat + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy().astype(np.uint8) real_samples.append(real_batch) fake_samples.append(fake_batch) generator = Engine(generate_vids) pbar = ProgressBar(ascii=True) pbar.attach(generator, output_transform=lambda x: x) self.logger.info(f"Start collecting sequences for fvd computation...") el = (int(1000 / test_sampler.batch_size) if len(test_dataset) > 1000 else len(test_loader)) generator.run(test_loader,max_epochs=1,epoch_length=el) savedir = path.join(self.dirs["generated"],"samples_fvd") makedirs(savedir,exist_ok=True) real_samples = np.stack(real_samples,axis=0) fake_samples = np.stack(fake_samples, axis=0) np.save(path.join(savedir,"real_samples.npy"),real_samples) np.save(path.join(savedir,"fake_samples.npy"),fake_samples) self.logger.info(f'Finish generation of vid samples.') else: raise ValueError(f'Specified testing mode "{self.config["testing"]["mode"]}" does not exist.')
58,565
53.581547
210
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/utils/losses.py
import torch from torch import nn from torchvision.models import vgg19 from collections import namedtuple from operator import mul from functools import reduce from utils.general import get_member VGGOutput = namedtuple( "VGGOutput", ["input", "relu1_2", "relu2_2", "relu3_2", "relu4_2", "relu5_2"], ) StyleLayers = namedtuple("StyleLayers",["relu1_2","relu2_2","relu3_3", "relu4_3"]) class PerceptualVGG(nn.Module): def __init__(self, weights=None): super().__init__() self.vgg = vgg19(pretrained=True) self.vgg.eval() self.vgg_layers = self.vgg.features self.register_buffer( "mean", torch.tensor([0.485, 0.456, 0.406], dtype=torch.float) .unsqueeze(dim=0) .unsqueeze(dim=-1) .unsqueeze(dim=-1), ) self.register_buffer( "std", torch.tensor([0.229, 0.224, 0.225], dtype=torch.float) .unsqueeze(dim=0) .unsqueeze(dim=-1) .unsqueeze(dim=-1), ) self.target_layers = { "3": "relu1_2", "8": "relu2_2", "13": "relu3_2", "15" : "relu3_3", "22": "relu4_2", "24" : "relu4_3", "31": "relu5_2", } if weights is None: self.loss_weights = {"input":1., "relu1_2": 1.,"relu2_2": 1.,"relu3_2": 1.,"relu3_3": 1.,"relu4_2": 1.,"relu4_3": 1.,"relu5_2": 1. } else: assert isinstance(weights, dict) and list(weights.keys()) == list(self.target_layers.keys()), f"The weights passed to PerceptualVGG have to be a dict with the keys {list(self.target_layers.keys())}" self.loss_weights = weights def forward(self, x): # IMPORTANT: Input is assumed to be in range [-1,1] here. x = (x + 1.0) / 2.0 x = (x - self.mean) / self.std # add also common reconstruction loss in pixel space out = {"input": x} for name, submodule in self.vgg_layers._modules.items(): # x = submodule(x) if name in self.target_layers: x = submodule(x) out[self.target_layers[name]] = x else: x = submodule(x) return out def vgg_loss(custom_vgg:PerceptualVGG, target, pred, weights=None): """ Implements a vgg based perceptual loss, as extensively used for image/video generation tasks :param custom_vgg: The vgg feature extractor for the perceptual loss, definition see above :param target: :param pred: :return: """ target_feats = custom_vgg(target) pred_feats = custom_vgg(pred) target_feats = VGGOutput(**{key: target_feats[key] for key in VGGOutput._fields}) pred_feats = VGGOutput(**{key: pred_feats[key] for key in VGGOutput._fields}) names = list(pred_feats._asdict().keys()) if weights is None: losses = {} for i, (tf, pf) in enumerate(zip(target_feats, pred_feats)): loss = get_member(custom_vgg,"loss_weights")[VGGOutput._fields[i]] * torch.mean( torch.abs(tf - pf) ).unsqueeze(dim=-1) losses.update({names[i]: loss}) else: losses = { names[0]: get_member(custom_vgg,"loss_weights")[VGGOutput._fields[0]] * torch.mean(weights * torch.abs(target_feats[0] - pred_feats[0])) .unsqueeze(dim=-1) .to(torch.float) } for i, (tf, pf) in enumerate(zip(target_feats[1:], pred_feats[1:])): loss = get_member(custom_vgg,"loss_weights")[i + 1] * torch.mean( torch.abs(tf - pf) ).unsqueeze(dim=-1) losses.update({names[i + 1]: loss}) return losses def vgg_loss_agg(vgg, target, pred, weights=None): """ To aggreagate the vgg losses :param vgg: :param target: :param pred: :param weights: :return: """ # basic_device = target.get_device() # net_device = list(vgg.parameters())[0].get_device() # pred = pred.cuda(net_device) # target = target.cuda(net_device) loss_list = vgg_loss(vgg,target,pred,weights) loss_tensor = torch.stack([loss_list[key] for key in loss_list],dim=0,) return loss_tensor.sum()#.cuda(basic_device) class PixelDynamicsLoss(nn.Module): def __init__(self, diff_pp=False): super().__init__() self.diff_pp = diff_pp def forward(self,target_t,target_tk,pred_t,pred_tk): if self.diff_pp: loss = (((target_t-target_tk).abs()-(pred_t.detach()-pred_tk).abs()).mean())**2 else: loss = ((target_t-target_tk).abs().mean()-(pred_t.detach()-pred_tk).abs().mean())**2 return loss def pixel_triplet_loss(target_t,target_tk,pred_t, pred_tk,vgg:PerceptualVGG,layerwise = True, detach=True, diff_pp=False): """ :param vgg: :param target_t: :param target_tk: :param pred_t: :param pred_tk: :param layerwise: :param detach: whether or not to detach the predicted feats at time t :param diff_pp: whether to consider differences for each spatial location in each channel or average over all (default average) :return: """ if layerwise: losses = {} # old_device = target_tk.get_device() # new_device = list(vgg.parameters())[0].get_device() # timestep t # target_t = target_t.cuda(new_device) # pred_t = pred_t.cuda(new_device) target_feats_t = vgg(target_t.cuda()) pred_feats_t = vgg(pred_t.detach() if detach else pred_t) target_feats_t = VGGOutput(**{key: target_feats_t[key] for key in VGGOutput._fields}) pred_feats_t = VGGOutput(**{key: pred_feats_t[key] for key in VGGOutput._fields}) # timestep tk # target_tk = target_tk.cuda(new_device) # pred_tk = pred_tk.cuda(new_device) target_feats_tk = vgg(target_tk) pred_feats_tk = vgg(pred_tk) target_feats_tk = VGGOutput(**{key: target_feats_tk[key] for key in VGGOutput._fields}) pred_feats_tk = VGGOutput(**{key: pred_feats_tk[key] for key in VGGOutput._fields}) names = list(pred_feats_t._asdict().keys()) for i, (tft, pft, tftk, pftk) in enumerate(zip(target_feats_t, pred_feats_t,target_feats_tk, pred_feats_tk)): if diff_pp: loss = get_member(vgg,"loss_weights")[VGGOutput._fields[i]] * torch.mean((torch.abs(tft - tftk) - torch.abs(pft - pftk)) ** 2).unsqueeze(dim=-1) else: loss = get_member(vgg,"loss_weights")[VGGOutput._fields[i]] * (torch.mean(torch.abs(tft - tftk)).unsqueeze(dim=-1) - torch.mean(torch.abs(pft - pftk)).unsqueeze(dim=-1))**2 losses.update({names[i]: loss}) loss_tensor = torch.stack([losses[key] for key in losses], dim=0, ) ptl = loss_tensor.sum() #.cuda(old_device) else: ptl = (vgg_loss_agg(vgg, pred_t.detach(), pred_tk) - vgg_loss_agg(vgg, target_t, target_tk)) ** 2 return ptl def style_loss(vgg,style_target, pred): target_feats = vgg(style_target) pred_feats = vgg(pred) target_feats = StyleLayers(**{key: target_feats[key] for key in StyleLayers._fields}) pred_feats = StyleLayers(**{key: pred_feats[key] for key in StyleLayers._fields}) names = list(pred_feats._asdict().keys()) style_outs = {} # compute gram matrices and take squared frobenius norm for i, (tf,pf) in enumerate(zip(target_feats,pred_feats)): shape = pf.shape pf = pf.reshape(*shape[:2],-1) tf = tf.reshape(*shape[:2],-1) gram_diff = 1. / (shape[1]*shape[2]*shape[3]) * (torch.matmul(pf,pf.permute(0,2,1)) - torch.matmul(tf,tf.permute(0,2,1))) loss = (torch.norm(gram_diff, p="fro",dim=[1,2])**2).mean() style_outs.update({names[i]:loss}) style_outs = torch.stack([style_outs[key] for key in style_outs]) return style_outs.sum() class DynamicsLoss(nn.Module): """ Triplet loss Takes embeddings of an anchor sample, a positive sample and a negative sample """ def __init__(self, config): super().__init__() self.mse = nn.MSELoss() def forward(self, anchor, positive, negative, ): if isinstance(anchor,list) and isinstance(positive,list): losses = [] for a,p in zip(anchor,positive): losses.append(self.mse(a,p)) return torch.stack(losses).mean() else: return self.mse(anchor,positive) def kl_loss_check(latents): """ Estimates a gaussian from the latents and returns the kl_divergence between this gaussian and the standard normal :param latents: :return: """ mu = latents[:,:int(latents.shape[1] / 2)] sigma = latents[:,int(latents.shape[1] / 2):] # reparameterize logstd = nn.Sigmoid()(sigma) return kl_loss(mu,logstd) def kl_loss(mu, logstd): if len(mu.shape) != 2: mu = mu.reshape(mu.shape[0],-1) logstd = logstd.reshape(mu.shape[0],-1) dim = mu.shape[1] std = torch.exp(logstd) kl = torch.sum(-logstd + 0.5 * (std ** 2 + mu ** 2), dim=-1) - (0.5 * dim) return kl.mean()
9,190
33.423221
210
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/utils/metric_fvd.py
import numpy as np import argparse from os import path import torch import ssl from glob import glob from natsort import natsorted ssl._create_default_https_context = ssl._create_unverified_context import cv2 from utils.metrics import compute_fvd from utils.general import get_logger if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("-s", "--source", type=str, required=True, help="Source directory where the data is stored.") parser.add_argument("--gpu",type=int, required=True, help="The target device.") parser.add_argument("-v","--visualize",default=False,action="store_true") args = parser.parse_args() if not path.isdir(args.source): raise NotADirectoryError(f'The specified, data-holding directory {args.source} is not existing...') file = path.basename(__file__) logger = get_logger(file) logger.info("Read in data...") real_samples_list = natsorted(glob(path.join(args.source, "real_samples_*.npy"))) fake_samples_list = natsorted(glob(path.join(args.source, "fake_samples_*.npy"))) if len(real_samples_list) == 0: fake_samples_list = [path.join(args.source, "fake_samples.npy")] real_samples_list = [path.join(args.source, "real_samples.npy")] for i,(real_samples, fake_samples) in enumerate(zip(real_samples_list,fake_samples_list)): try: length = int(real_samples.split("/")[-1].split(".")[0].split("_")[2]) context = int(real_samples.split("/")[-1].split(".npy")[0].split("_")[-1]) logger.info(f"processing samples of length {length} with {context} context frames.") except: logger.info(f"Processing standard samples") real_samples = np.load(real_samples) fake_samples = np.load(fake_samples) if args.visualize: vis_real = real_samples[0,0] vis_fake = fake_samples[0,0] # visualize writer = cv2.VideoWriter( path.join(args.source, "test_vid_fake.mp4"), cv2.VideoWriter_fourcc(*"MP4V"), 5, (vis_fake.shape[2], vis_fake.shape[1]), ) # writer = vio.FFmpegWriter(savename,inputdict=inputdict,outputdict=outputdict) for frame in vis_fake: frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) writer.write(frame) writer.release() writer = cv2.VideoWriter( path.join(args.source, "test_vid_real.mp4"), cv2.VideoWriter_fourcc(*"MP4V"), 5, (vis_real.shape[2], vis_real.shape[1]), ) # writer = vio.FFmpegWriter(savename,inputdict=inputdict,outputdict=outputdict) for frame in vis_real: frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) writer.write(frame) writer.release() real_samples = real_samples[:62] if real_samples.shape[0] > 62 else real_samples fake_samples = fake_samples[:62] if fake_samples.shape[0] > 62 else fake_samples logger.info(f'Number of samples: {len(fake_samples)}') target_device = args.gpu real_samples = list(real_samples) fake_samples = list(fake_samples) real_samples = [torch.from_numpy(r) for r in real_samples] fake_samples = [torch.from_numpy(r) for r in fake_samples] fvd_val = compute_fvd(real_samples,fake_samples,target_device,logger)
3,569
30.59292
107
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/utils/testing.py
import numpy as np import torch from skimage.metrics import structural_similarity as ssim import cv2 import math import imutils import matplotlib.pyplot as plt import wandb from os import path import math def make_flow_grid(src, poke, pred, tgt, n_logged, flow=None): """ :param src: :param poke: :param pred: :param tgt: :param n_logged: :return: """ src = ((src.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] # poke = poke.permute(0, 2, 3, 1).cpu().numpy()[:n_logged] # poke -= poke.min() # poke /= poke.max() # poke = (poke * 255.0).astype(np.uint8) # poke = np.concatenate([poke, np.expand_dims(np.zeros_like(poke).sum(-1), axis=-1)], axis=-1).astype(np.uint8) poke = vis_flow(poke[:n_logged]) poke = np.concatenate(poke,axis=1) # if prediction is image, just take the premuted image if pred.shape[1] == 3: pred = ((pred.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype( np.uint8)[:n_logged] else: # if prediction is flow, you need treat it like that pred = pred.permute(0, 2, 3, 1).cpu().numpy()[:n_logged] pred -= pred.min() pred /= pred.max() pred = (pred * 255.0).astype(np.uint8) pred = np.concatenate([pred, np.expand_dims(np.zeros_like(pred).sum(-1), axis=-1)], axis=-1).astype(np.uint8) tgt = ((tgt.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype( np.uint8)[:n_logged] tgt_gr = [cv2.cvtColor(t,cv2.COLOR_RGB2GRAY) for t in tgt] pred_gr = [cv2.cvtColor(t,cv2.COLOR_RGB2GRAY) for t in pred] ssim_imgs = [ssim(rimg, fimg, multichannel=True, data_range=255, gaussian_weights=True, use_sample_covariance=False, full=True)[1] for rimg, fimg in zip(tgt_gr, pred_gr)] additional = [np.concatenate([cv2.cvtColor((s * 255.).astype(np.uint8),cv2.COLOR_GRAY2RGB) for s in ssim_imgs],axis=1)] if flow is not None: # if provided, use additional flow information (in case of poking, that's the entire flow src --> tgt # add = flow.permute(0, 2, 3, 1).cpu().numpy()[:n_logged] # add -= add.min() # add /= add.max() # add = (add * 255.0).astype(np.uint8) # add = np.concatenate([add, np.expand_dims(np.zeros_like(add).sum(-1), axis=-1)], axis=-1).astype(np.uint8) # additional = additional + [np.concatenate([a for a in add], axis=1)] add = vis_flow(flow[:n_logged]) add = np.concatenate(add,axis=1) additional = additional + [add] # compute ssim_img in grayscale src = np.concatenate([s for s in src], axis=1) #poke = np.concatenate([f for f in poke], axis=1) pred = np.concatenate([p for p in pred], axis=1) tgt = np.concatenate([t for t in tgt], axis=1) grid = np.concatenate([src,poke,pred,tgt,*additional],axis=0) return grid def vis_flow(flow_map, normalize=False): if isinstance(flow_map,torch.Tensor): flow_map = flow_map.cpu().numpy() flows_vis = [] for flow in flow_map: hsv = np.zeros((*flow.shape[1:],3),dtype=np.uint8) hsv[...,1] = 255 mag, ang = cv2.cartToPolar(flow[0], flow[1]) # since 360 is not valid for uint8, 180° corresponds to 360° for opencv hsv representation. Therefore, we're dividing the angle by 2 after conversion to degrees hsv[...,0] = ang * 180 / np.pi / 2 hsv[...,2] = cv2.normalize(mag,None,alpha=0,beta=255, norm_type=cv2.NORM_MINMAX) as_rgb = cv2.cvtColor(hsv,cv2.COLOR_HSV2RGB) if normalize: as_rgb = as_rgb.astype(np.float) - as_rgb.min(axis=(0,1),keepdims=True) as_rgb = (as_rgb / as_rgb.max(axis=(0,1),keepdims=True)*255.).astype(np.uint8) flows_vis.append(as_rgb) return flows_vis def vis_flow_dense(flow_map,**kwargs): if isinstance(flow_map,torch.Tensor): flow_map = flow_map.cpu().numpy() flows_vis = [] for flow in flow_map: h, w = flow.shape[1:] fx, fy = flow[0], flow[1] ang = np.arctan2(fy, fx) + np.pi v = np.sqrt(fx * fx + fy * fy) hsv = np.zeros((h, w, 3), np.uint8) hsv[..., 0] = ang * (180 / np.pi / 2) hsv[..., 1] = 255 hsv[..., 2] = cv2.normalize(v,None,alpha=0,beta=255, norm_type=cv2.NORM_MINMAX) bgr = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR) flows_vis.append(bgr) return flows_vis def make_trf_video(img1,img2,v12,v21,poke,n_logged,logwandb=True,length_divisor=5): """ :param img1: :param img2: :param v12: :param v21: :param poke: :param n_logged: :param lomake_flow_grid()gwandb: :param length_divisor: :return: """ seq_len = v12.shape[1] pokes = vis_flow(poke[:n_logged]) img1 = ((img1.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] img2 = ((img2.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] img1_with_arrow = [] img2_with_arrow = [] eps = 1e-6 for i, (poke_p, img1_i, img2_i) in enumerate(zip(poke[:n_logged], img1, img2)): poke_points = np.nonzero(pokes[i].any(-1) > 0) if poke_points[0].size == 0: img1_with_arrow.append(img1_i) img2_with_arrow.append(img2_i) else: min_y = np.amin(poke_points[0]) max_y = np.amax(poke_points[0]) min_x = np.amin(poke_points[1]) max_x = np.amax(poke_points[1]) # plot mean direction of flow in poke region avg_flow = np.mean(poke_p[:, min_y:max_y, min_x:max_x].cpu().numpy(), axis=(1, 2)) arrow_dir = avg_flow / (np.linalg.norm(avg_flow) + eps) * (poke_p.shape[1] / length_divisor) if not math.isnan(arrow_dir[0]) or not math.isnan(arrow_dir[1]): arrow_start = (int((min_x + max_x) / 2), int((min_y + max_y) / 2)) arrow_end = (arrow_start[0] + int(arrow_dir[0]), arrow_start[1] + int(arrow_dir[1])) img1_with_arrow.append(cv2.UMat.get(cv2.arrowedLine(cv2.UMat(img1_i), arrow_start, arrow_end, (255, 0, 0), max(int(img1_i.shape[0] / 64), 1)))) img2_with_arrow.append(cv2.UMat.get(cv2.arrowedLine(cv2.UMat(img2_i), arrow_start, arrow_end, (255, 0, 0), max(int(img2_i.shape[0] / 64), 1)))) else: img1.append(img1_i) img2.append(img2_i) vid_st1 = np.concatenate(img1_with_arrow, axis=1) vid_st2 = np.concatenate(img2_with_arrow, axis=1) vid_st1 = put_text_to_video_row(np.stack([vid_st1] * seq_len, axis=0), "Image 1", color=(255, 0, 0)) vid_st2 = put_text_to_video_row(np.stack([vid_st2] * seq_len, axis=0), "Image 2", color=(255, 0, 0)) v12 = ((v12.permute(0, 1, 3, 4, 2).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] v12 = np.concatenate(list(v12), axis=2) v12 = put_text_to_video_row(v12, "Vid: FG1-BG2") v21 = ((v21.permute(0, 1, 3, 4, 2).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] v21 = np.concatenate(list(v21), axis=2) v21 = put_text_to_video_row(v21, "Vid: FG2-BG1") full = np.concatenate([vid_st1, vid_st2, v12, v21], axis=1) if logwandb: full = np.moveaxis(full, [0, 1, 2, 3], [0, 2, 3, 1]) return full def draw_arrow(traj): arrow_imgs = [] for c,t in enumerate(traj): active_points = np.nonzero(t.astype(np.uint8).any(0) > 0) img = np.zeros((*t.shape[1:],3),dtype=np.uint8) if active_points[0].size>0: for i in range(active_points[0].shape[0]): y =active_points[0][i] x = active_points[1][i] arrow_dir = t[:,y,x] if not math.isnan(arrow_dir[0]) or not math.isnan(arrow_dir[1]): arrow_start = (x, y) arrow_end = (int(np.clip(x + int(arrow_dir[0]),0,img.shape[0])), int(np.clip(y + int(arrow_dir[1]),0,img.shape[0]))) img = cv2.arrowedLine(img, arrow_start, arrow_end, (255, 0, 0), max(int(traj.shape[1] / 64), 1)) arrow_imgs.append(img) arrow_imgs = np.concatenate(arrow_imgs,axis=1) return arrow_imgs def img_grid_ci(src,traj,pred,tgt,n_logged): src = ((src.permute(0, 2, 3, 1).cpu().numpy()) * 255.).astype(np.uint8)[:n_logged] # if prediction is image, just take the premuted image pred = ((pred.permute(0, 2, 3, 1).cpu().numpy()) * 255).astype( np.uint8)[:n_logged] tgt = ((tgt.permute(0, 2, 3, 1).cpu().numpy()) * 255).astype( np.uint8)[:n_logged] src = np.concatenate([s for s in src], axis=1) # poke = np.concatenate([f for f in poke], axis=1) pred = np.concatenate([p for p in pred], axis=1) tgt = np.concatenate([t for t in tgt], axis=1) arrows = draw_arrow(traj[:n_logged].cpu().numpy()) grid = np.concatenate([src, arrows, pred, tgt], axis=0) return grid def make_video_ci(src,traj,pred,tgt,n_logged,logwandb=True, display_frame_nr=True): seq_len = tgt.shape[1] srcs = np.concatenate([s for s in ((src.permute(0, 2, 3, 1).cpu().numpy()) * 255.).astype(np.uint8)[:n_logged]],axis=1) traj_vis = [] for t in range(traj.shape[1]): arrows = draw_arrow(traj[:n_logged,t].cpu().numpy()) traj_vis.append(arrows) traj_vis = np.stack(traj_vis,axis=0) traj_vis = put_text_to_video_row(traj_vis, "Flow Vectors", display_frame_nr=display_frame_nr) srcs = cv2.UMat.get(cv2.putText(cv2.UMat(srcs), f"Sequence length {seq_len}", (int(srcs.shape[1] // 3), int(srcs.shape[0] / 6)), cv2.FONT_HERSHEY_SIMPLEX, float(srcs.shape[0] / 256), (255, 0, 0), int(srcs.shape[0] / 128))) srcs = np.stack([srcs] * seq_len, axis=0) srcs = put_text_to_video_row(srcs, "Input Image", display_frame_nr=display_frame_nr) pred = ((pred.permute(0, 1, 3, 4, 2).cpu().numpy()) * 255).astype(np.uint8)[:n_logged] pred = np.concatenate(list(pred), axis=2) pred = put_text_to_video_row(pred, "Predicted Video", display_frame_nr=display_frame_nr) tgt = ((tgt.permute(0, 1, 3, 4, 2).cpu().numpy()) * 255).astype(np.uint8)[:n_logged] tgt = np.concatenate(list(tgt), axis=2) tgt = put_text_to_video_row(tgt, "Groundtruth Video", display_frame_nr=display_frame_nr) full = np.concatenate([srcs, pred, tgt, traj_vis], axis=1) if logwandb: full = np.moveaxis(full, [0, 1, 2, 3], [0, 2, 3, 1]) return full def make_video(src,poke,pred,tgt,n_logged,flow=None,length_divisor=5,logwandb=True,flow_weights= None, display_frame_nr=False,invert_poke = False): """ :param src: src image :param poke: poke, also input to the network :param pred: predicted video of the network :param tgt: target video the network was trained to reconstruct :param n_logged: numvber of logged examples :param flow: src flow from which the poke is originating :param length_divisor: divisor for the length of the arrow, that's drawn ti visualize the mean direction of the flow within the poke patch :param logwandb: whether the output video grid is intended to be logged with wandb or not (in this case the grid channels have to be changed) :param flow_weights: Optional weights for the flow which are also displayed if they are not None. :return: """ seq_len = tgt.shape[1] src = ((src.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] pokes = vis_flow(poke[:n_logged]) flows_vis = None if flow is not None: flows = vis_flow(flow[:n_logged]) flows_with_rect = [] for i,(poke_p,flow) in enumerate(zip(pokes,flows)): poke_points = np.nonzero(poke_p.any(-1) > 0) if poke_points[0].size == 0: flows_with_rect.append(np.zeros_like(flow)) else: min_y = np.amin(poke_points[0]) max_y = np.amax(poke_points[0]) min_x = np.amin(poke_points[1]) max_x = np.amax(poke_points[1]) # draw rect flow_with_rect = cv2.rectangle(flow,(min_x,min_y),(max_x,max_y),(255,255,255),max(1,int(flow.shape[0]//64))) # flow_with_rect = cv2.UMat.get(cv2.putText(cv2.UMat(flow_with_rect), f"Flow Complete",(int(flow_with_rect.shape[1] // 3), int(5 * flow_with_rect.shape[0] / 6) ), cv2.FONT_HERSHEY_SIMPLEX, # float(flow_with_rect.shape[0] / 256), (255, 255, 255), int(flow_with_rect.shape[0] / 128))) flows_with_rect.append(flow_with_rect) flow_cat = np.concatenate(flows_with_rect,axis=1) flows_vis= [np.stack([flow_cat]*seq_len,axis=0)] flows_vis[0] = put_text_to_video_row(flows_vis[0], "Flow Complete", color=(255, 255, 255)) if flow_weights is not None: flow_weights = flow_weights.cpu().numpy() heatmaps = [] for i, weight in enumerate(flow_weights): weight_map = ((weight - weight.min()) / weight.max() * 255.).astype(np.uint8) heatmap = cv2.applyColorMap(weight_map, cv2.COLORMAP_HOT) heatmap = cv2.cvtColor(heatmap, cv2.COLOR_RGB2BGR) heatmaps.append(heatmap) heatmaps = np.concatenate(heatmaps, axis=1) heatmaps = np.stack([heatmaps]*seq_len, axis=0) heatmaps = put_text_to_video_row(heatmaps, "Flow Weights", color=(255,255,255)) if flows_vis is None: flows_vis = [heatmaps] else: flows_vis.insert(0,heatmaps) srcs_with_arrow = [] pokes_with_arrow = [] if invert_poke: srcs_with_arrow_inv = [] pokes_with_arrow_inv = [] eps = 1e-6 for i, (poke_p,src_i) in enumerate(zip(poke[:n_logged],src)): poke_points = np.nonzero(pokes[i].any(-1) > 0) if poke_points[0].size==0: pokes_with_arrow.append(np.zeros_like(pokes[i])) srcs_with_arrow.append(src_i) else: min_y = np.amin(poke_points[0]) max_y = np.amax(poke_points[0]) min_x = np.amin(poke_points[1]) max_x = np.amax(poke_points[1]) # plot mean direction of flow in poke region avg_flow = np.mean(poke_p[:, min_y:max_y, min_x:max_x].cpu().numpy(), axis=(1, 2)) arrow_dir = avg_flow / (np.linalg.norm(avg_flow) + eps) * (poke_p.shape[1] / length_divisor) if not math.isnan(arrow_dir[0]) or not math.isnan(arrow_dir[1]): arrow_start = (int((min_x + max_x) / 2), int((min_y + max_y) / 2)) arrow_end = (arrow_start[0] + int(arrow_dir[0]), arrow_start[1] + int(arrow_dir[1])) test = pokes[i] # test = cv2.UMat.get(cv2.putText(cv2.UMat(test), f"Poke", (int(test.shape[1] // 3), int(5 * test .shape[0] / 6)), cv2.FONT_HERSHEY_SIMPLEX, # float(test.shape[0] / 256), (255, 255, 255), int(test.shape[0] / 128))) pokes_with_arrow.append(cv2.arrowedLine(test, arrow_start, arrow_end, (255, 0, 0), max(int(src_i.shape[0] / 64),1))) srcs_with_arrow.append(cv2.UMat.get(cv2.arrowedLine(cv2.UMat(src_i), arrow_start, arrow_end, (255, 0, 0), max(int(src_i.shape[0] / 64),1)))) if invert_poke: arrow_end_inv = (arrow_start[0] - int(arrow_dir[0]), arrow_start[1] - int(arrow_dir[1])) pokes_with_arrow_inv.append(cv2.arrowedLine(test, arrow_start, arrow_end_inv, (0, 255, 0), max(int(src_i.shape[0] / 64), 1))) srcs_with_arrow_inv.append(cv2.UMat.get(cv2.arrowedLine(cv2.UMat(src_i), arrow_start, arrow_end, (0, 255, 0), max(int(src_i.shape[0] / 64), 1)))) else: pokes_with_arrow.append(np.zeros_like(pokes[i])) srcs_with_arrow.append(src_i) poke = np.concatenate(pokes_with_arrow, axis=1) if invert_poke: poke_inv = np.concatenate(pokes_with_arrow_inv, axis=1) poke = put_text_to_video_row(np.stack([*[poke] * int(math.ceil(float(seq_len)/2)),*[poke_inv]*int(seq_len/2)], axis=0),"Pokes",color=(255,255,255)) else: poke = put_text_to_video_row(np.stack([poke] * seq_len, axis=0),"Poke",color=(255,255,255)) if flows_vis is None: flows_vis = [poke] else: flows_vis.append(poke) srcs = np.concatenate(srcs_with_arrow,axis=1) srcs = cv2.UMat.get(cv2.putText(cv2.UMat(srcs), f"Sequence length {seq_len}", (int(srcs.shape[1] // 3), int(srcs.shape[0]/6)), cv2.FONT_HERSHEY_SIMPLEX, float(srcs.shape[0] / 256), (255, 0, 0), int(srcs.shape[0] / 128))) if invert_poke: srcs_inv = np.concatenate(srcs_with_arrow_inv, axis=1) srcs_inv = cv2.UMat.get(cv2.putText(cv2.UMat(srcs_inv), f"Sequence length {seq_len}", (int(srcs_inv.shape[1] // 3), int(srcs_inv.shape[0] / 6)), cv2.FONT_HERSHEY_SIMPLEX, float(srcs_inv.shape[0] / 256), (255, 0, 0), int(srcs_inv.shape[0] / 128))) srcs = np.stack([*[srcs] * int(math.ceil(float(seq_len)/2)),*[srcs_inv]*int(seq_len/2)],axis=0) else: srcs = np.stack([srcs]*seq_len,axis=0) srcs = put_text_to_video_row(srcs,"Input Image",display_frame_nr=display_frame_nr) pred = ((pred.permute(0, 1, 3, 4, 2).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] pred = np.concatenate(list(pred),axis=2) pred = put_text_to_video_row(pred, "Predicted Video",display_frame_nr=display_frame_nr) tgt = ((tgt.permute(0, 1, 3, 4, 2).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] tgt = np.concatenate(list(tgt),axis=2) tgt = put_text_to_video_row(tgt,"Groundtruth Video",display_frame_nr=display_frame_nr) full = np.concatenate([srcs,pred,tgt,*flows_vis],axis=1) if logwandb: full = np.moveaxis(full,[0,1,2,3],[0,2,3,1]) return full def put_text_to_video_row(video_row,text, color = None,display_frame_nr=False): written = [] for i,frame in enumerate(video_row): current = cv2.UMat.get(cv2.putText(cv2.UMat(frame), text, (int(frame.shape[1] // 3), frame.shape[0] - int(frame.shape[0]/6)), cv2.FONT_HERSHEY_SIMPLEX, float(frame.shape[0] / 256), (255, 0, 0) if color is None else color, int(frame.shape[0] / 128))) if display_frame_nr: current = cv2.UMat.get(cv2.putText(cv2.UMat(current), str(i+1), (int(frame.shape[1] / 32), frame.shape[0] - int(frame.shape[0] / 6)), cv2.FONT_HERSHEY_SIMPLEX, float(frame.shape[0] / 256), (255, 0, 0) if color is None else color, int(frame.shape[0] / 128))) written.append(current) return np.stack(written) def make_animated_grid(src, poke, pred, tgt, n_logged, flow=None, length_divisor=5,logwandb=True): # visualize flows pokes = vis_flow(poke[:n_logged]) pokes_with_arrow = [] for i,poke_p in enumerate(poke[:n_logged]): poke_points = np.nonzero(pokes[i].any(-1) > 0) min_y = np.amin(poke_points[0]) max_y = np.amax(poke_points[0]) min_x = np.amin(poke_points[1]) max_x = np.amax(poke_points[1]) # plot mean direction of flow in poke region avg_flow = np.mean(poke_p[:,min_y:max_y,min_x:max_x].cpu().numpy(), axis=(1, 2)) arrow_dir = avg_flow / np.linalg.norm(avg_flow) * (poke_p.shape[1] / length_divisor) arrow_start = (int((min_x+max_x)/2),int((min_y+max_y)/2)) arrow_end = (arrow_start[0]+int(arrow_dir[0]),arrow_start[1]+int(arrow_dir[1])) test = pokes[i] pokes_with_arrow.append(cv2.arrowedLine(test,arrow_start,arrow_end,(0,0,255),2)) poke = np.concatenate(pokes_with_arrow, axis=1) flows_vis = [np.stack([poke]*3,axis=0)] if flow is not None: flows = vis_flow(flow[:n_logged]) flows_with_rect = [] for i,(poke_p,flow) in enumerate(zip(pokes,flows)): poke_points = np.nonzero(poke_p.any(-1) > 0) min_y = np.amin(poke_points[0]) max_y = np.amax(poke_points[0]) min_x = np.amin(poke_points[1]) max_x = np.amax(poke_points[1]) # draw rect flow_with_rect = cv2.rectangle(flow,(min_x,min_y),(max_x,max_y),(255,255,255),max(1,int(flow.shape[0]//64))) flows_with_rect.append(flow_with_rect) flow_cat = np.concatenate(flows_with_rect,axis=1) flows_vis.insert(0,np.stack([flow_cat]*3,axis=0)) # visualize images src = ((src.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] pred = ((pred.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] tgt = ((tgt.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] src = np.concatenate(list(src),axis=1) pred = np.concatenate(list(pred), axis=1) tgt = np.concatenate(list(tgt), axis=1) src = cv2.UMat.get(cv2.putText(cv2.UMat(src), "Source", (int(src.shape[0] // 4), 30), cv2.FONT_HERSHEY_SIMPLEX, float(src.shape[0] / 256), (0, 0, 0), int(src.shape[0] / 128))) pred = cv2.UMat.get(cv2.putText(cv2.UMat(pred), "Predicted", (int(pred.shape[0] // 4), 30), cv2.FONT_HERSHEY_SIMPLEX, float(pred.shape[0] / 256), (0, 0, 0), int(pred.shape[0] / 128))) tgt = cv2.UMat.get(cv2.putText(cv2.UMat(tgt), "Target", (int(tgt.shape[0] // 4), 30), cv2.FONT_HERSHEY_SIMPLEX, float(tgt.shape[0] / 256), (0, 0, 0), int(tgt.shape[0] / 128))) animation = np.stack([src,pred,tgt],axis=0) # this generates a video grid which can be used by wandb.Video() full = np.concatenate([animation,*flows_vis],axis=1) # wandb requires video to have shape (time, channels, height, width) if logwandb: full = np.moveaxis(full,[0,1,2,3],[0,2,3,1]) return full def make_generic_grid(data, dtype, n_logged): from utils.visualizer import FlowVisualizer visualizer = FlowVisualizer() final_data = [] assert(len(data)==len(dtype)) for i, batch in enumerate(data): if dtype[i] == "flow": add = batch.permute(0, 2, 3, 1).cpu().numpy()[:n_logged] add -= add.min() add /= add.max() add = (add * 255.0).astype(np.uint8) image = np.concatenate( [add, np.expand_dims(np.zeros_like(add).sum(-1), axis=-1)], axis=-1).astype(np.uint8) elif dtype[i] == "flow_3D": add = batch.permute(0, 2, 3, 1).cpu().numpy()[:n_logged] add -= add.min() add /= add.max() add = (add * 255.0).astype(np.uint8) image = add elif dtype[i] == "img": image = ((batch.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] elif dtype[i] == "diff_flow_amplitude": generated = batch[0][:n_logged].detach().cpu() target = batch[1][:n_logged].detach().cpu() image = visualizer.create_diff_amplitude( visualizer.make_3d_to_2d(generated), visualizer.make_3d_to_2d(target)) image = (image*255).astype(np.uint8)[:, None].repeat(3, axis=1).transpose(0, 2, 3, 1) elif dtype[i] == "diff_flow_direction": generated = batch[0][:n_logged].detach().cpu() target = batch[1][:n_logged].detach().cpu() image = visualizer.create_diff_direction( visualizer.make_3d_to_2d(generated), visualizer.make_3d_to_2d(target)) image = (image * 255).astype(np.uint8)[:, None].repeat(3, axis=1).transpose(0, 2, 3, 1) elif dtype[i] == "diff_flow_clipped": generated = batch[0][:n_logged].permute(0, 2, 3, 1).cpu().numpy() target = batch[1][:n_logged].permute(0, 2, 3, 1).cpu().numpy() image = np.sum(np.abs(generated-target), axis=-1) image = (image[:, :, :, None]).astype(np.uint8) image = np.clip(image, 0, 255) image = np.repeat(image, 3, axis=-1) elif dtype[i] == "diff_scaled": generated = batch[0][:n_logged].permute(0, 2, 3, 1).cpu().numpy() target = batch[1][:n_logged].permute(0, 2, 3, 1).cpu().numpy() image = np.sum(np.abs(generated-target), axis=-1) image /= image.max(axis=0) image = (image[:, :, :, None]*255.0).astype(np.uint8) image = np.repeat(image, 3, axis=-1) image = np.concatenate([s for s in image], axis=1) final_data.append(image) grid = np.concatenate(final_data, axis=0) return grid def make_img_grid(appearance, shape, pred, tgt= None, n_logged=4, target_label="Target Images", label_app = "Appearance Images", label_gen = "Generated Images", label_shape = "Shape Images"): """ :param appearance: :param shape: :param pred: :param tgt: :param n_logged: :return: """ appearance = ((appearance.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] shape = ((shape.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype(np.uint8)[:n_logged] pred = ((pred.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype( np.uint8)[:n_logged] if tgt is not None: tgt = ((tgt.permute(0, 2, 3, 1).cpu().numpy() + 1.) * 127.5).astype( np.uint8)[:n_logged] tgt = np.concatenate([t for t in tgt], axis=1) tgt = cv2.UMat.get(cv2.putText(cv2.UMat(tgt), target_label , (int(tgt.shape[1] // 3), tgt.shape[0] - int(tgt.shape[0]/6)), cv2.FONT_HERSHEY_SIMPLEX, float(tgt.shape[0] / 256), (255, 0, 0), int(tgt.shape[0] / 128))) appearance = np.concatenate([s for s in appearance], axis=1) appearance = cv2.UMat.get(cv2.putText(cv2.UMat(appearance),label_app, (int(appearance.shape[1] // 3), appearance.shape[0] - int(appearance.shape[0]/6)), cv2.FONT_HERSHEY_SIMPLEX, float(appearance.shape[0] / 256), (255, 0, 0), int(appearance.shape[0] / 128))) shape = np.concatenate([f for f in shape], axis=1) shape = cv2.UMat.get(cv2.putText(cv2.UMat(shape), label_shape, (int(shape.shape[1] // 3), shape.shape[0] - int(shape.shape[0]/6)), cv2.FONT_HERSHEY_SIMPLEX, float(shape.shape[0] / 256), (255, 0, 0), int(shape.shape[0] / 128))) pred = np.concatenate([p for p in pred], axis=1) pred = cv2.UMat.get(cv2.putText(cv2.UMat(pred), label_gen, (int(pred.shape[1] // 3), pred.shape[0] - int(pred.shape[0]/6)), cv2.FONT_HERSHEY_SIMPLEX, float(pred.shape[0] / 256), (255, 0, 0), int(pred.shape[0] / 128))) if tgt is None: grid = np.concatenate([appearance, shape, pred], axis=0) else: grid = np.concatenate([appearance, shape, pred, tgt], axis=0) return grid def scale_img(img): """ Rescales an image to the actual pixel domain in [0,255] and converts it to the dtype :param img: :return: """ img = (img + 1.) * 127.5 if isinstance(img, torch.Tensor): img = img.to(torch.uint8) else: # assumed to be numpy array img = img.astype(np.uint8) return img def human_graph_cut_map(img, poke_size): import cv2 import matplotlib.pyplot as plt # make backgound foreground segementation mask = np.zeros(img.shape[:2], np.uint8) rect = (int(img.shape[1] / 5), poke_size, int(3. * img.shape[1] / 5), int(img.shape[0] - 2 * poke_size)) fgm = np.zeros((1, 65), dtype=np.float64) bgm = np.zeros((1, 65), dtype=np.float64) mask2, fgm, bgm = cv2.grabCut(img, mask, rect, fgm, bgm, 5, cv2.GC_INIT_WITH_RECT) mask3 = np.where((mask2 == 2) | (mask2 == 0), 0, 1).astype(np.bool) tuples = np.where(mask3[:, :]) return tuples new_img = np.zeros_like(img) for t in tuples: for i in range(3): new_img[t[0], t[1], i] = 255 # show the output frame plt.imshow(new_img) plt.show() # human_NN_map_weights = "/export/home/jsieber/poking/models/mask-rcnn-coco/frozen_inference_graph.pb" # human_NN_map_classes = "/export/home/jsieber/poking/models/mask-rcnn-coco/object_detection_classes_coco.txt" # human_NN_map_config = "/export/home/jsieber/poking/models/mask-rcnn-coco/mask_rcnn_inception_v2_coco_2018_01_28.pbtxt" # human_NN_map_LABELS = open(human_NN_map_classes).read().strip().split("\n") # # human_NN_map_net = cv2.dnn.readNetFromTensorflow(human_NN_map_weights, human_NN_map_config) # def human_NN_map(frame, conf=0.5, threshold=0.3): # # # # modified from https://www.pyimagesearch.com/2018/11/26/instance-segmentation-with-opencv/ # before_width = frame.shape[1] # frame = imutils.resize(frame, width=600) # (H, W) = frame.shape[:2] # # # construct a blob from the input image and then perform a # # forward pass of the Mask R-CNN, giving us (1) the bounding # # box coordinates of the objects in the image along with (2) # # the pixel-wise segmentation for each specific object # blob = cv2.dnn.blobFromImage(frame, swapRB=True, crop=False) # human_NN_map_net.setInput(blob) # (boxes, masks) = human_NN_map_net.forward(["detection_out_final", # "detection_masks"]) # # # sort the indexes of the bounding boxes in by their corresponding # # prediction probability (in descending order) # idxs = np.argsort(boxes[0, 0, :, 2])[::-1] # # # initialize the mask, ROI, and coordinates of the person for the # # current frame # mask = None # roi = None # coords = None # # # loop over the indexes # for i in idxs: # # extract the class ID of the detection along with the # # confidence (i.e., probability) associated with the # # prediction # classID = int(boxes[0, 0, i, 1]) # confidence = boxes[0, 0, i, 2] # # # if the detection is not the 'person' class, ignore it # if human_NN_map_LABELS[classID] != "person": # continue # # filter out weak predictions by ensuring the detected # # probability is greater than the minimum probability # if confidence > conf: # # scale the bounding box coordinates back relative to the # # size of the image and then compute the width and the # # height of the bounding box # box = boxes[0, 0, i, 3:7] * np.array([W, H, W, H]) # (startX, startY, endX, endY) = box.astype("int") # coords = (startX, startY, endX, endY) # boxW = endX - startX # boxH = endY - startY # # # extract the pixel-wise segmentation for the object, # # resize the mask such that it's the same dimensions of # # the bounding box, and then finally threshold to create # # a *binary* mask # mask = masks[i, classID] # mask = cv2.resize(mask, (boxW, boxH), # interpolation=cv2.INTER_NEAREST) # mask = (mask > threshold) # # # extract the ROI and break from the loop (since we make # # the assumption there is only *one* person in the frame # # who is also the person with the highest prediction # # confidence) # roi = frame[startY:endY, startX:endX][mask] # break # # # initialize our output frame # output = frame.copy() # # # if the mask is not None *and* we are in privacy mode, then we # # know we can apply the mask and ROI to the output image # if mask is not None: # # blur the output frame # output = np.zeros_like(output) # # # add the ROI to the output frame for only the masked region # (startX, startY, endX, endY) = coords # roi = np.ones_like(roi)*255 # output[startY:endY, startX:endX][mask] = roi # output = imutils.resize(output, width=before_width) # # tuples = np.where(output[:, :, 0] > 0) # return tuples # # new_img = np.zeros_like(output) # for t in tuples: # for i in range(3): # new_img[t[0], t[1], i] = 255 # # show the output frame # plt.imshow(new_img) # plt.show() def make_hist(hist, title, ylabel, xlabel="Frame number", bins_edges = None): plt.ioff() if bins_edges is None: bins_edges = np.arange(0, len(hist) + 1).astype(np.float) else: assert len(list(bins_edges)) == len(list(hist)) + 1 centroids = (bins_edges[1:] + bins_edges[:-1]) / 2 hist_, bins_, _ = plt.hist( centroids, bins=len(hist), weights=np.asarray(hist), range=(min(bins_edges), max(bins_edges)), ) plt.xlabel(xlabel) plt.ylabel(ylabel) wandb.log({title: wandb.Image(plt)}) plt.close() def make_plot(x,y,title,ylabel,xlabel="frame idx", savename=None): plt.ioff() fig, ax = plt.subplots() ax.plot(x,y,'rv-') ax.set(xlabel=xlabel, ylabel=ylabel, title = title) ax.grid() if savename is None: wandb.log({title:wandb.Image(plt)}) else: fig.savefig(savename) plt.close() if __name__=="__main__": frame_path = "/export/data/ablattma/Datasets/iPER/processed/001_10_1/frame_277.png" frame = cv2.imread(frame_path) frame = imutils.resize(frame, width=128) import time for i in range(3): start_time = time.time() human_graph_cut_map(frame, 15) print("--- %s seconds ---" % (time.time() - start_time)) for i in range(3): start_time = time.time() human_NN_map(frame) print("--- %s secondss ---" % (time.time() - start_time))
33,806
43.424442
204
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/utils/flownet_loader.py
import torch from torch.nn import functional as F from PIL import Image from models.flownet2.models import * from torchvision import transforms import matplotlib.pyplot as plt import argparse from utils.general import get_gpu_id_with_lowest_memory class FlownetPipeline: def __init__(self): super(FlownetPipeline, self).__init__() def load_flownet(self, args, device): """ :param args: args from argparser :return: The flownet pytorch model """ # load model savefile save = torch.load( "/export/scratch/compvis/datasets/plants/pretrained_models/FlowNet2_checkpoint.pth.tar") model = FlowNet2(args, batchNorm=False) untrained_statedict = model.state_dict() # load it into proper clean model model.load_state_dict(save["state_dict"]) model.eval() return model.to(device) def preprocess_image(self, img, img2, channelOrder="RGB",spatial_size= None): """ This preprocesses the images for FlowNet input. Preserves the height and width order! :param channelOrder: RGB(A) or BGR :param img: The first image in form of (W x H x RGBA) or (H x W x RGBA) :param img2: The first image in form of (W x H x RGBA) or (H x W x RGBA) :return: The preprocessed input for the prediction (BGR x Img# x W x H) or (BGR x Img# x H x W) """ # ToTensor transforms from (H x W x C) => (C x H x W) # also automatically casts into range [0, 1] if spatial_size is None: img, img2 = transforms.ToTensor()(img)[:3], transforms.ToTensor()(img2)[:3] else: ts = transforms.Compose([transforms.ToPILImage(), transforms.Resize(size=spatial_size,interpolation=Image.BILINEAR),transforms.ToTensor()]) img, img2 = ts(img)[:3],ts(img2)[:3] if channelOrder == "RGB": img, img2 = img[[2, 1, 0]], img2[[2, 1, 0]] # Cast to proper shape (Batch x BGR x #Img x H x W) s = img.shape img, img2 = img[:, :int(s[1] / 64) * 64, :int(s[2] / 64) * 64], \ img2[:, :int(s[1] / 64) * 64,:int(s[2] / 64) * 64] stacked = torch.cat([img[:, None], img2[:, None]], dim=1) return stacked def predict(self, model, stacked, spatial_size=None): """ :param stacked: The two input images. (Batch x BGR x Img# x H x W) :return: The flow result (2 x W x H) """ # predict model.eval() prediction = model(stacked) out_size = float(prediction.shape[-1]) if spatial_size is not None: prediction = F.interpolate( prediction.cpu(), size=(spatial_size,spatial_size), mode="bilinear" ) # rescale to make it fit to new shape (not grave, if this is skipped as flow is normalized anyways later) prediction = prediction / (out_size / spatial_size) flow = prediction[0] return flow def show_results(self, prediction, with_ampl=False): """ prediction (Tensor): The predicted flow (2 x W x H) :return: plots """ zeros = torch.zeros((1, prediction.shape[1], prediction.shape[2])) if with_ampl: ampl = torch.sum(prediction * prediction, dim=0) ampl = ampl.squeeze() else: ampl = torch.cat([prediction, zeros], dim=0) ampl -= ampl.min() ampl /= ampl.max() # show image im = transforms.ToPILImage()(ampl) if with_ampl: plt.imshow(im, cmap='gray') else: plt.imshow(im) if __name__ == "__main__": # parse args parser = argparse.ArgumentParser(description='Process some integers.') # always 1.0, because pytorch toTensor automatically converts into range [0.0, 1.0] parser.add_argument("--rgb_max", type=float, default=1.) parser.add_argument('--fp16', action='store_true', help='Run model in pseudo-fp16 mode (fp16 storage fp32 math).') parser.add_argument('--fp16_scale', type=float, default=1024., help='Loss scaling, positive power of 2 values can improve fp16 convergence.') args = parser.parse_args() # load test images in BGR mode img, img2 = np.asarray(Image.open(f"/export/data/ablattma/Datasets/plants/processed/hoch_misc1/frame_0.png")), \ np.asarray(Image.open(f"/export/data/ablattma/Datasets/plants/processed/hoch_misc1/frame_100.png")) # load Flownet pipeline = FlownetPipeline() flownet_device = get_gpu_id_with_lowest_memory() flownet = pipeline.load_flownet(args, flownet_device) # process to show flow stacked = pipeline.preprocess_image(img, img2).to(flownet_device) prediction = pipeline.predict(flownet, stacked[None]).cpu() pipeline.show_results(prediction) plt.show()
4,882
37.753968
151
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/utils/metrics.py
import torch from torch import nn from torch.nn import functional as F from torchvision.models import inception_v3 import numpy as np from scipy import linalg from skimage.metrics import peak_signal_noise_ratio as compare_psnr from skimage.metrics import structural_similarity as ssim from pytorch_lightning.metrics import functional as PF from tqdm import tqdm class FIDInceptionModel(nn.Module): def __init__(self, normalize_range=True): super().__init__() self.v3 = inception_v3(pretrained=True,aux_logits=False) # self.v3.aux_logits = False self.register_buffer( "mean", torch.tensor([0.485, 0.456, 0.406], dtype=torch.float) .unsqueeze(dim=0) .unsqueeze(dim=-1) .unsqueeze(dim=-1), ) self.register_buffer( "std", torch.tensor([0.229, 0.224, 0.225], dtype=torch.float) .unsqueeze(dim=0) .unsqueeze(dim=-1) .unsqueeze(dim=-1), ) self.resize = nn.Upsample(size=(299,299),mode="bilinear") self.normalize_range = normalize_range def forward(self, x): x = self.resize(x) if self.normalize_range: # normalize in between 0 and 1 x = (x + 1.) / 2. else: x = x.to(torch.float) / 255. # normalize to demanded values x = (x - self.mean) / self.std # this reimpleents the respective layers of the inception model, see model definition for name, submodule in self.v3._modules.items(): x = submodule(x) if name == "Mixed_7c": break elif name == "Conv2d_4a_3x3" or name == "Conv2d_2b_3x3": x = F.avg_pool2d(x, kernel_size=3, stride=2) out = F.adaptive_avg_pool2d(x, (1, 1)) out = torch.flatten(out, 1) return out def metrcis_MAE(tensor1, tensor2, mean=True): value = torch.sum(torch.abs(tensor1 - tensor2)) if mean: value /= tensor1.view((-1)).shape[0] return value.data.cpu().numpy() def metrcis_MSE(tensor1, tensor2, mean=True): diff = tensor1 - tensor2 value = torch.sum(diff * diff) if mean: value /= tensor1.view((-1)).shape[0] return value.data.cpu().numpy() def metrcis_l1(tensor1, tensor2, mean=False): return metrcis_MAE(tensor1, tensor2, mean) def metrcis_l2(tensor1, tensor2, mean=False): diff = tensor1 - tensor2 value = torch.sum(diff * diff) value = torch.sqrt(value) if mean: value /= tensor1.view((-1)).shape[0] return value.data.cpu().numpy() def metric_ssim(real, fake, reduce = True, return_per_frame=False): if real.dim() == 3: real = real[None,None] fake = fake[None,None] elif real.dim() == 4: real = real[None] fake = fake[None] # rescale to valid range real = ((real + 1.) / 2.).permute(0, 1, 3, 4, 2).cpu().numpy() fake = ((fake + 1.) / 2.).permute(0, 1, 3, 4, 2).cpu().numpy() ssim_batch = np.asarray([ssim(rimg, fimg, multichannel=True, data_range=1.0, gaussian_weights=True,use_sample_covariance=False, ) for rimg, fimg in zip(real.reshape(-1,*real.shape[2:]), fake.reshape(-1,*fake.shape[2:]))]) if return_per_frame: ssim_per_frame = {} for i in range(real.shape[1]): real_test = real[:,i] fake_test = fake[:,i] ssim_per_frame[i] = np.asarray([ssim(real_test, fake_test, multichannel=True, data_range=1., gaussian_weights=True, use_sample_covariance=False)]) # ssim_per_frame = {i:np.asarray([ssim(real[:,i], fake[:,i], # multichannel=True, data_range=1., gaussian_weights=True, use_sample_covariance=False)]) for i in range(real.shape[1])} if reduce: if return_per_frame: ssim_pf_reduced = {key: ssim_per_frame[key] for key in ssim_per_frame} return np.mean(ssim_batch), ssim_pf_reduced else: return np.mean(ssim_batch) if return_per_frame: return ssim_batch, ssim_per_frame else: return ssim_batch def ssim_lightning(real, fake, return_per_frame=False, normalize_range=True): if real.dim() == 3: real = real[None, None] fake = fake[None, None] elif real.dim() == 4: real = real[None] fake = fake[None] if normalize_range: real = (real + 1.) /2. fake = (fake + 1.) / 2. ssim_batch = PF.ssim(fake.reshape(-1,*fake.shape[2:]),real.reshape(-1,*real.shape[2:])).cpu().numpy() if return_per_frame: ssim_per_frame = {i: PF.ssim(fake[:,i],real[:,i]).cpu().numpy() for i in range(real.shape[1])} return ssim_batch, ssim_per_frame return ssim_batch def psnr_lightning(real, fake, return_per_frame=False, normalize_range=True): if real.dim() == 3: real = real[None, None] fake = fake[None, None] elif real.dim() == 4: real = real[None] fake = fake[None] if normalize_range: real = (real + 1.) / 2. fake = (fake + 1.) / 2. psnr_batch = PF.psnr(fake.reshape(-1, *fake.shape[2:]), real.reshape(-1, *real.shape[2:])).cpu().numpy() if return_per_frame: psnr_per_frame = {i: PF.psnr(fake[:, i].contiguous(), real[:, i].contiguous()).cpu().numpy() for i in range(real.shape[1])} return psnr_batch, psnr_per_frame return psnr_batch def metric_psnr(im_true, im_test,reduce = True, return_per_frame=False): if im_true.dim() == 3: im_true, im_test = im_true[None,None], im_test[None,None] elif im_true.dim() == 4: im_true, im_test = im_true[None], im_test[None] # make channel last real = ((im_true + 1.) / 2.).permute(0, 1, 3, 4, 2).cpu().numpy() fake = ((im_test + 1.) / 2.).permute(0, 1, 3, 4, 2).cpu().numpy() psnr_batch = np.asarray([compare_psnr(r,f, data_range=1.) for r, f in zip(real.reshape(-1,*real.shape[2:]),fake.reshape(-1,*fake.shape[2:]))]) if return_per_frame: psnr_per_frame = {i: np.asarray([compare_psnr(real[:,i], fake[:,i], data_range=1.)]) for i in range(real.shape[1])} if reduce: if return_per_frame: psnr_pf_reduced = {key: psnr_per_frame[key] for key in psnr_per_frame} return np.mean(psnr_batch), psnr_pf_reduced else: return np.mean(psnr_batch) if return_per_frame: return psnr_batch, psnr_per_frame else: return psnr_batch def metric_lpips(real, fake, lpips_func, reduce=True, return_per_frame=False, normalize=False): if real.dim() == 3: real, fake = real[None,None], fake[None,None] elif real.dim() == 4: real, fake = real[None], fake[None] if normalize: if fake.max() > 1: fake = (fake.to(torch.float) / 127.5) - 1. real = (real.to(torch.float) / 127.5) -1. else: real = (real * 2.) - 1. fake = (fake * 2.) - 1. lpips_batch = lpips_func(real.reshape(-1,*real.shape[2:]),fake.reshape(-1,*fake.shape[2:])).squeeze().cpu().numpy() if return_per_frame: lpips_per_frame = {i: lpips_func(real[:,i],fake[:,i]).squeeze().cpu().numpy() for i in range(real.shape[1])} if reduce: if return_per_frame: lpips_pf_reduced = {key: lpips_per_frame[key].mean() for key in lpips_per_frame} return lpips_batch.mean(), lpips_pf_reduced else: return lpips_batch.mean() if return_per_frame: return lpips_batch, lpips_per_frame else: return lpips_batch def mean_cov(features): mu = np.mean(features, axis=0) cov = np.cov(features, rowvar=False) return mu,cov def metric_fid(real_features, fake_features, eps=1e-6): # Taken and adapted from https://github.com/mseitzer/pytorch-fid/blob/master/fid_score.py if not isinstance(real_features,np.ndarray): real_features = np.concatenate(real_features,axis=0) fake_features = np.concatenate(fake_features,axis=0) mu1, cov1 = mean_cov(real_features) mu2, cov2 = mean_cov(fake_features) mu1 = np.atleast_1d(mu1) mu2 = np.atleast_1d(mu2) sigma1 = np.atleast_2d(cov1) sigma2 = np.atleast_2d(cov2) assert ( mu1.shape == mu2.shape ), "Training and test mean vectors have different lengths" assert ( sigma1.shape == sigma2.shape ), "Training and test covariances have different dimensions" diff = mu1 - mu2 # Product might be almost singular covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False) if not np.isfinite(covmean).all(): msg = f"fid calculation produces singular product; adding {eps} to diagonal of cov estimates" print(msg) offset = np.eye(sigma1.shape[0]) * eps covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset)) # Numerical error might give slight imaginary component if np.iscomplexobj(covmean): if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3): m = np.max(np.abs(covmean.imag)) raise ValueError(f"Imaginary component {m}") covmean = covmean.real tr_covmean = np.trace(covmean) return diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean def compute_fvd(real_videos,fake_videos, device,logger): import silence_tensorflow.auto import tensorflow.compat.v1 as tf from utils.frechet_video_distance import preprocess,Embedder,calculate_fvd # required for fvd computation # config = tf.ConfigProto() # config.gpu_options.visible_device_list = f"{device}" devs = tf.config.experimental.get_visible_devices("GPU") target_dev = [d for d in devs if d.name.endswith(str(device))][0] tf.config.experimental.set_visible_devices(target_dev, 'GPU') logger.info("Compute fvd score.") #dev = f"/gpu:{device}" logger.info(f"using device {device}") with tf.device("/gpu:0"): with tf.Graph().as_default(): # construct graph sess = tf.Session() input_shape = real_videos[0].shape input_real = tf.placeholder(dtype=tf.uint8, shape=input_shape) input_fake = tf.placeholder(dtype=tf.uint8, shape=input_shape) real_pre = preprocess(input_real, (224, 224)) emb_real = Embedder(real_pre) embed_real = emb_real.create_id3_embedding(real_pre) fake_pre = preprocess(input_fake, (224, 224)) emb_fake = Embedder(fake_pre) embed_fake = emb_fake.create_id3_embedding(fake_pre) sess.run(tf.global_variables_initializer()) sess.run(tf.tables_initializer()) real, fake = [], [] for rv, fv in tqdm(zip(real_videos, fake_videos)): # real_batch = ((rv + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy() # fake_batch = ((fv + 1.) * 127.5).permute(0, 1, 3, 4, 2).cpu().numpy() # real_batch = ((rv + 1.) * 127.5).cpu().numpy() # fake_batch = ((fv + 1.) * 127.5).cpu().numpy() feed_dict = {input_real: rv, input_fake: fv} r, f = sess.run([embed_fake, embed_real], feed_dict) real.append(r) fake.append(f) print('Compute FVD score') real = np.concatenate(real, axis=0) fake = np.concatenate(fake, axis=0) embed_real = tf.placeholder(dtype=tf.float32, shape=(real.shape[0], 400)) embed_fake = tf.placeholder(dtype=tf.float32, shape=(real.shape[0], 400)) result = calculate_fvd(embed_real, embed_fake) feed_dict = {embed_real: real, embed_fake: fake} fvd_val = sess.run(result, feed_dict) sess.close() logger.info(f"Results of fvd computation: fvd={fvd_val}") # for being sure return fvd_val if __name__ == "__main__": z, o = torch.rand((1080, 720, 3)), torch.rand((1080, 720, 3)) o[0, 0, 0], o[1, 0, 0] = 0, 1 z[0, 0, 0], z[1, 0, 0] = 0, 1
12,279
33.985755
153
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/utils/general.py
import torch import os import subprocess import logging import yaml import logging.config import inspect from os import walk import numpy as np import coloredlogs import multiprocessing as mp from threading import Thread from queue import Queue from collections import abc import cv2 from torch import nn # import kornia def get_member(model, name): if isinstance(model, nn.DataParallel): module = model.module else: module = model return getattr(module, name) def convert_flow_2d_to_3d(flow): amplitude = torch.sqrt(torch.sum(flow * flow, dim=0, keepdim=True)) # fix division by zero scaler = amplitude.clone() scaler[scaler==0.0] = 1.0 flow = flow/scaler flow = torch.cat([flow, amplitude], dim=0) # n_flow = torch.sqrt(torch.sum(flow[:2] * flow[:2], dim=0)) # print(torch.max(n_flow.view((-1)))) return flow def convert_flow_2d_to_3d_batch(flows): final = [] for flow in flows: converted = convert_flow_2d_to_3d(flow) final.append(converted[None]) final = torch.cat(final, dim=0) return final def get_flow_gradients(flow, device=None): """torch in, torch out""" flow = flow[:, None] sobel = [[1, 2, 1], [0, 0, 0], [-1, -2, -1]] sobel_kernel_y = torch.tensor(sobel, dtype=torch.float32).unsqueeze(0) sobel_kernel_x = torch.transpose(sobel_kernel_y, 1, 2) sobel_kernel_x, sobel_kernel_y = sobel_kernel_x.expand((1, 1, 3, 3)), sobel_kernel_y.expand((1, 1, 3, 3)) if flow.is_cuda: sobel_kernel_x, sobel_kernel_y = sobel_kernel_x.to(flow.get_device()), sobel_kernel_y.to(flow.get_device()) gradient_d1_x = torch.nn.functional.conv2d(flow, sobel_kernel_x, stride=1, padding=1) gradient_d2_x = torch.nn.functional.conv2d(gradient_d1_x, sobel_kernel_x, stride=1, padding=1) gradient_d1_y = torch.nn.functional.conv2d(flow, sobel_kernel_y, stride=1, padding=1) gradient_d2_y = torch.nn.functional.conv2d(gradient_d1_y, sobel_kernel_y, stride=1, padding=1) gradient_d1_x, gradient_d2_x, gradient_d1_y, gradient_d2_y = gradient_d1_x.squeeze(),\ gradient_d2_x.squeeze(),\ gradient_d1_y.squeeze(),\ gradient_d2_y.squeeze() gradient_d1_x = torch.sqrt(torch.sum(gradient_d1_x ** 2, dim=0)) gradient_d1_y = torch.sqrt(torch.sum(gradient_d1_y ** 2, dim=0)) gradient_d2_x = torch.sqrt(torch.sum(gradient_d2_x ** 2, dim=0)) gradient_d2_y = torch.sqrt(torch.sum(gradient_d2_y ** 2, dim=0)) return gradient_d1_x, gradient_d1_y, gradient_d2_x, gradient_d2_y def get_flow_gradients_batch(flows): final = [] for flow in flows: gradient_d1_x, gradient_d1_y, gradient_d2_x, gradient_d2_y = get_flow_gradients(flow) all_gradients = [gradient_d1_x, gradient_d1_y, gradient_d2_x, gradient_d2_y] stacked = torch.stack(all_gradients, dim=0).squeeze(dim=0) final.append(stacked) final = torch.stack(final, dim=0).squeeze(dim=0) return final class LoggingParent: def __init__(self): super(LoggingParent, self).__init__() # find project root mypath = inspect.getfile(self.__class__) mypath = "/".join(mypath.split("/")[:-1]) found = False while mypath!="" and not found: f = [] for (dirpath, dirnames, filenames) in walk(mypath): f.extend(filenames) break if ".gitignore" in f: found = True continue mypath = "/".join(mypath.split("/")[:-1]) project_root = mypath+"/" # Put it together file = inspect.getfile(self.__class__).replace(project_root, "").replace("/", ".").split(".py")[0] cls = str(self.__class__)[8:-2] cls = str(cls).replace("__main__.", "").split(".")[-1] self.logger = get_logger(f"{file}.{cls}") def get_gpu_id_with_lowest_memory(index=0, target_gpus:list=None): # get info from nvidia-smi result = subprocess.check_output( [ 'nvidia-smi', '--query-gpu=memory.free', '--format=csv,nounits,noheader' ], encoding='utf-8') gpu_memory = [int(x) for x in result.strip().split('\n')] # get the one with the lowest usage if target_gpus is None: indices = np.argsort(gpu_memory) else: indices = [i for i in np.argsort(gpu_memory) if i in target_gpus] return torch.device(f"cuda:{indices[-index-1]}") iuhihfie_logger_loaded = False def get_logger(name): # setup logging global iuhihfie_logger_loaded if not iuhihfie_logger_loaded: with open(f'{os.path.dirname(os.path.abspath(__file__))}/logging.yaml', 'r') as f: log_cfg = yaml.load(f.read(), Loader=yaml.FullLoader) logging.config.dictConfig(log_cfg) iuhihfie_logger_loaded = True logger = logging.getLogger(name) coloredlogs.install(logger=logger, level="DEBUG") return logger def save_model_to_disk(path, models, epoch): for i, model in enumerate(models): tmp_path = path if not os.path.exists(path): os.makedirs(path) tmp_path = tmp_path + f"model_{i}-epoch{epoch}" torch.save(model.state_dict(), tmp_path) def _do_parallel_data_prefetch(func, Q, data, idx): # create dummy dataset instance # run prefetching res = func(data) Q.put([idx, res]) Q.put("Done") def parallel_data_prefetch( func: callable, data, n_proc, target_data_type="ndarray",cpu_intensive=True ): if target_data_type not in ["ndarray", "list"]: raise ValueError( "Data, which is passed to parallel_data_prefetch has to be either of type list or ndarray." ) if isinstance(data, np.ndarray) and target_data_type == "list": raise ValueError("list expected but function got ndarray.") elif isinstance(data, abc.Iterable): if isinstance(data, dict): print( f'WARNING:"data" argument passed to parallel_data_prefetch is a dict: Using only its values and disregarding keys.' ) data = list(data.values()) if target_data_type == "ndarray": data = np.asarray(data) else: data = list(data) else: raise TypeError( f"The data, that shall be processed parallel has to be either an np.ndarray or an Iterable, but is actually {type(data)}." ) if cpu_intensive: Q = mp.Queue(1000) proc = mp.Process else: Q = Queue(1000) proc = Thread # spawn processes if target_data_type == "ndarray": arguments = [ [func, Q, part, i] for i, part in enumerate(np.array_split(data, n_proc)) ] else: step = ( int(len(data) / n_proc + 1) if len(data) % n_proc != 0 else int(len(data) / n_proc) ) arguments = [ [func, Q, part, i] for i, part in enumerate( [data[i : i + step] for i in range(0, len(data), step)] ) ] processes = [] for i in range(n_proc): p = proc(target=_do_parallel_data_prefetch, args=arguments[i]) processes += [p] # start processes print(f"Start prefetching...") import time start = time.time() gather_res = [[] for _ in range(n_proc)] try: for p in processes: p.start() k = 0 while k < n_proc: # get result res = Q.get() if res == "Done": k += 1 else: gather_res[res[0]] = res[1] except Exception as e: print("Exception: ", e) for p in processes: p.terminate() raise e finally: for p in processes: p.join() print(f"Prefetching complete. [{time.time() - start} sec.]") if not isinstance(gather_res[0], np.ndarray): return np.concatenate([np.asarray(r) for r in gather_res], axis=0) # order outputs return np.concatenate(gather_res, axis=0) def linear_var( act_it, start_it, end_it, start_val, end_val, clip_min, clip_max ): act_val = ( float(end_val - start_val) / (end_it - start_it) * (act_it - start_it) + start_val ) return np.clip(act_val, a_min=clip_min, a_max=clip_max) def get_patches(seq_batch,weights,config,fg_value, logger = None): """ :param seq_batch: Batch of videos :param weights: batch of flow weights for the videos :param config: config, containing spatial_size :param fg_value: foreground value of the weight map :return: """ import kornia weights_as_bool = torch.eq(weights,fg_value) cropped = [] for vid,weight in zip(seq_batch,weights_as_bool): vid_old = vid weight_ids = torch.nonzero(weight,as_tuple=True) try: min_y = weight_ids[0].min() max_y = weight_ids[0].max() min_x = weight_ids[1].min() max_x = weight_ids[1].max() vid = vid[...,min_y:max_y,min_x:max_x] if len(vid.shape) < 4: data_4d = vid[None,...] vid = kornia.transform.resize(data_4d, config["spatial_size"]) cropped.append(vid.squeeze(0)) else: vid = kornia.transform.resize(vid,config["spatial_size"]) cropped.append(vid) except Exception as e: if logger is None: print(e) else: logger.warn(f'Catched the following exception in "get_patches": {e.__class__.__name__}: {e}. Skip patching this sample...') cropped.append(vid_old) return torch.stack(cropped,dim=0) if __name__ == "__main__": print(get_gpu_id_with_lowest_memory())
10,061
33.108475
139
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/data/flow_dataset.py
from os import path import numpy as np import pickle from copy import deepcopy import torch from torch.nn import functional as F from torch.utils.data import Dataset from torchvision import transforms as tt from tqdm import tqdm import cv2 from natsort import natsorted import os from glob import glob from utils.general import parallel_data_prefetch, LoggingParent from data.helper_functions import preprocess_image from data.base_dataset import BaseDataset class PlantDataset(BaseDataset): def __init__(self, transforms, datakeys, config, train=True, google_imgs=False, n_ref_frames=None): self.excluded_objects = config["excluded_objects"] if "excluded_objects" in config else [] super().__init__(transforms, datakeys, config,train=train) self.logger.info(f"Initializing {self.__class__.__name__}.") if self.config["spatial_size"][0] <= 256: self.flow_in_ram = self.config["flow_in_ram"] if "flow_in_ram" in self.config else False if self.config["spatial_size"][0] <= 256: self.imgs_in_ram = self.config["imgs_in_ram"] if "imgs_in_ram" in self.config else False # set instace specific fixed values which shall not be parameters from yaml self._set_instance_specific_values() self.subsample_step = config["subsample_step"] if "subsample_step" in config else self.subsample_step self.logger.info(f'Subsample step of {self.__class__.__name__} is {self.subsample_step}.') filt_msg = "enabled" if self.filter_flow else "disabled" self.logger.info(f"Flow filtering is {filt_msg} in {self.__class__.__name__}!") self.logger.info(f"Valid lag of {self.__class__.__name__} is {self.valid_lags[0]}") # load data metafile_path = path.join(self.datapath, f"{self.metafilename}.p") with open(metafile_path, "rb") as handle: self.data = pickle.load(handle) if path.isfile(path.join(self.datapath,"dataset_stats.p")) and self.normalize_flows: with open(path.join(self.datapath,"dataset_stats.p"),"rb") as norm_file: self.flow_norms = pickle.load(norm_file) # choose filter procedure available_frame_nrs = np.asarray([int(p.split("/")[-1].split(".")[0].split("_")[-1]) - int(p.split("/")[-1].split(".")[0].split("_")[-2]) for p in self.data["flow_paths"][0]]) # filter invalid flow_paths self.data["flow_paths"] = [p for p in self.data["flow_paths"] if len(p) == len(available_frame_nrs)] self.filter_proc = self.config["filter"] if "filter" in self.config else "all" # remove invalid video valid_ids = np.logical_not(np.char.startswith(self.data["img_path"],"VID_0_3_1024x1024")) # set flow paths in right order after reading in the data if "max_fid" not in self.data: self.data["flow_paths"] = [natsorted(d) for d in self.data["flow_paths"]] # make absolute image and flow paths self.data["img_path"] = [ path.join(self.datapath, p if not p.startswith("/") else p[1:]) for p in self.data["img_path"] ] self.data["flow_paths"] = [ [path.join(self.datapath, f if not f.startswith("/") else f[1:]) for f in fs] for fs in self.data["flow_paths"] ] # convert to numpy array self.data = {key: np.asarray(self.data[key])[valid_ids] for key in self.data} # if max fid is not predefined, the videos, the dataset consists of are sufficiently long, such that it doesn't make much of a difference, # if some frames at the end are skipped, therefore, we set the last valid fid (which is indicated by "max_fid") to the maximum fid # in the respective sequence if "max_fid" not in self.data: available_frame_nrs = np.asarray([int(p.split("/")[-1].split(".")[0].split("_")[-1]) - int(p.split("/")[-1].split(".")[0].split("_")[-2]) for p in self.data["flow_paths"][0]]) self.data.update({"max_fid": np.zeros((np.asarray(self.data["fid"]).shape[0],max(len(available_frame_nrs),self.valid_lags[0]+1)),dtype=np.int)}) for vid in np.unique(self.data["vid"]): self.data["max_fid"][self.data["vid"] == vid] = np.amax(self.data["fid"][self.data["vid"] == vid]) if not self.var_sequence_length and ("poke" in self.datakeys or "flow" in self.datakeys) and not self.normalize_flows: # reset valid_lags, such that always the right flow which corresponds to the respective sequence length, is chosen if not self.__class__.__name__ == "Human36mDataset": available_frame_nrs = np.asarray([int(p.split("/")[-1].split(".")[0].split("_")[-1]) - int(p.split("/")[-1].split(".")[0].split("_")[-2]) for p in self.data["flow_paths"][0]]) if "n_ref_frames" not in self.config: assert self.max_frames * self.subsample_step in available_frame_nrs right_lag = int(np.argwhere(available_frame_nrs == self.max_frames * self.subsample_step)) self.logger.info(f'Last frames of sequence serves as target frame.') else: self.logger.info(f'Number of frames in between target and start frames is {self.config["n_ref_frames"]}') assert self.config["n_ref_frames"]*self.subsample_step in available_frame_nrs right_lag = int(np.argwhere(available_frame_nrs==self.config["n_ref_frames"] * self.subsample_step)) self.valid_lags = [right_lag] else: assert self.max_frames == 10 assert self.subsample_step in [1,2] self.valid_lags = [0] if self.subsample_step == 1 else [1] self.logger.info(f"Dataset is run in fixed length mode, valid lags are {self.valid_lags}.") filt_msg = "enabled" if self.filter_flow else "disabled" self.logger.info(f"Flow filtering is {filt_msg} in {self.__class__.__name__}!") self.logger.info(f"Valid lag of {self.__class__.__name__} is {self.valid_lags[0]}") filt_msg = "enabled" if self.obj_weighting else "disabled" self.logger.info(f"Object weighting is {filt_msg} in {self.__class__.__name__}!") filt_msg = "enabled" if self.flow_weights else "disabled" self.logger.info(f"Patch weighting is {filt_msg} in {self.__class__.__name__}!") filt_msg = "enabled" if self.use_flow_for_weights else "disabled" self.logger.info(f"Flow patch extraction is {filt_msg} in {self.__class__.__name__}!") if self.filter_proc == "action": self.data = {key:self.data[key][self.data["action_id"]==2] for key in self.data} elif self.filter_proc == "pose": self.data = {key: self.data[key][self.data["action_id"] == 1] for key in self.data} # on this point, the raw data is parsed and can be processed further # exclude invalid object ids from data self.logger.info(f"Excluding the following, user-defined object ids: {self.excluded_objects} from dataloading.") kept_ids = np.nonzero(np.logical_not(np.isin(self.data["object_id"], self.excluded_objects)))[0] self.data = {key:self.data[key][kept_ids] for key in self.data} self.split = self.config["split"] split_data, train_indices, test_indices = self._make_split(self.data) self.datadict = ( split_data["train"] if self.train else split_data["test"] ) msg = "train" if self.train else "test" vids, start_ids = np.unique(self.datadict["vid"],return_index=True) # get start and end ids per sequence self.eids_per_seq = {vid: np.amax(np.flatnonzero(self.datadict["vid"] == vid)) for vid in vids} seids = np.asarray([self.eids_per_seq[self.datadict["vid"][i]] for i in range(self.datadict["img_path"].shape[0])],dtype=np.int) self.datadict.update({"seq_end_id": seids}) self.sids_per_seq = {vid:i for vid,i in zip(vids,start_ids)} self.seq_len_T_chunk = {l: c for l,c in enumerate(np.linspace(0,self.flow_cutoff,self.max_frames,endpoint=False))} # add last chunk self.seq_len_T_chunk.update({self.max_frames: self.flow_cutoff}) if self.var_sequence_length: if "flow_range" in self.datadict.keys(): self.ids_per_seq_len = {length: np.flatnonzero(np.logical_and(np.logical_and(self.datadict["flow_range"][:,1,self.valid_lags[0]]>self.seq_len_T_chunk[length], np.less_equal(np.arange(self.datadict["img_path"].shape[0]) + (self.min_frames + length)*self.subsample_step + 1, self.datadict["seq_end_id"])), np.less_equal(self.datadict["fid"],self.datadict["max_fid"][:,self.valid_lags[0]]))) for length in np.arange(self.max_frames)} else: self.ids_per_seq_len = {length: np.flatnonzero(np.less_equal(self.datadict["fid"],self.datadict["max_fid"][:,self.valid_lags[0]])) for length in np.arange(self.max_frames)} for length in self.ids_per_seq_len: actual_ids = self.ids_per_seq_len[length] oids, counts_per_obj = np.unique(self.datadict["object_id"][actual_ids],return_counts=True) weights = np.zeros_like(actual_ids,dtype=np.float) for oid,c in zip(oids,counts_per_obj): weights[self.datadict["object_id"][actual_ids]==oid] = 1. / (c * oids.shape[0]) self.object_weights_per_seq_len.update({length:weights}) obj_ids, obj_counts = np.unique(self.datadict["object_id"], return_counts=True) weights = np.zeros_like(self.datadict["object_id"], dtype=np.float) for (oid, c) in zip(obj_ids, obj_counts): weights[self.datadict["object_id"] == oid] = 1. / c weights = weights / obj_ids.shape[0] self.datadict.update({"weights": weights}) if self.flow_in_ram: self.logger.warn(f"Load flow maps in RAM... please make sure to have enough capacity there.") assert len(self.valid_lags) == 1 self.loaded_flows = parallel_data_prefetch(self._read_flows, self.datadict["flow_paths"][:,self.valid_lags[0]],n_proc=72,cpu_intensive=True) assert self.loaded_flows.shape[0] == self.datadict["img_path"].shape[0] if self.imgs_in_ram: self.logger.warn(f"Load images in RAM... please make sure to have enough capacity there.") self.loaded_imgs = parallel_data_prefetch(self._read_imgs, self.datadict["img_path"],n_proc=72,cpu_intensive=True) assert self.loaded_imgs.shape[0] == self.datadict["img_path"].shape[0] if google_imgs: img_paths = [p for p in glob(path.join(self.datapath,"google_images", "*")) if path.isfile(p) and any(map(lambda x: p.endswith(x), ["jpg", "jpeg", "png"]))] self.datadict["img_path"] = np.asarray(img_paths) self.logger.info("Use images from Google.") msg = "Flow normalization enabled!" if self.normalize_flows else "Flow normalization disabled!" self.logger.info( f'Initialized {self.__class__.__name__} in "{msg}"-mode. Dataset consists of {self.__len__()} samples. ' + msg ) def _set_instance_specific_values(self): # set flow cutoff to 0.2 as this seems to be a good heuristic for Plants self.valid_lags = [0] self.flow_cutoff = 0.4 self.extended_annotations = False self.subsample_step = 2 self.min_frames = 5 self.obj_weighting = True if not 8 in self.excluded_objects: self.excluded_objects.append(8) self.metafilename = "meta" # self.metafilename = 'test_codeprep_metadata' def _read_flows(self,data): read_flows = [] flow_paths = data def proc_flow(flow): org_shape = float(flow.shape[-1]) dsize = None if "spatial_size" in self.config: dsize = self.config["spatial_size"] elif "resize_factor" in self.config: dsize = ( int(float(flow.shape[1]) / self.config["resize_factor"]), int(float(flow.shape[2]) / self.config["resize_factor"]), ) flow = F.interpolate( torch.from_numpy(flow).unsqueeze(0), size=dsize, mode="bilinear", align_corners=True ).numpy() flow = flow / (org_shape / dsize[0]) return flow for i, flow_path in enumerate(tqdm(flow_paths)): try: f = np.load(flow_path) f = proc_flow(f) except ValueError: try: f = np.load(flow_path, allow_pickle=True) f = proc_flow(f) except Exception as ex: self.logger.error(ex) read_flows.append("None") continue except: self.logger.error("Fallback error ocurred. Append None and continue") read_flows.append("None") continue read_flows.append(f) return np.concatenate(read_flows,axis=0) def _read_imgs(self,imgs): read_imgs = [] for img_path in tqdm(imgs): img = cv2.imread(img_path) # image is read in BGR img = preprocess_image(img, swap_channels=True) img = cv2.resize( img, self.config["spatial_size"], cv2.INTER_LINEAR ) read_imgs.append(img) return read_imgs def _make_split(self,data): vids = np.unique(self.data["vid"]) split_data = {"train": {}, "test": {}} if self.split == "videos": # split such that some videos are held back for testing self.logger.info("Splitting data after videos") shuffled_vids = deepcopy(vids) np.random.shuffle(shuffled_vids) train_vids = shuffled_vids[: int(0.8 * shuffled_vids.shape[0])] train_indices = np.nonzero(np.isin(data["vid"], train_vids))[0] test_indices = np.nonzero(np.logical_not(train_indices))[0] split_data["train"] = { key: data[key][train_indices] for key in data } split_data["test"] = { key: data[key][test_indices] for key in data } else: self.logger.info(f"splitting data across_videos") train_indices = np.asarray([],dtype=np.int) test_indices = np.asarray([], dtype=np.int) for vid in vids: indices = np.nonzero(data["vid"] == vid)[0] # indices = np.arange(len(tdata["img_path"])) # np.random.shuffle(indices) train_indices = np.append(train_indices,indices[: int(0.8 * indices.shape[0])]) test_indices = np.append(test_indices,indices[int(0.8 * indices.shape[0]) :]) split_data["train"] = { key: data[key][train_indices] for key in data } split_data["test"] = { key: data[key][test_indices] for key in data } return split_data, train_indices, test_indices class VegetationDataset(PlantDataset): def _set_instance_specific_values(self): self.filter_flow = False self.valid_lags = [0] self.flow_cutoff = .3 self.min_frames = 5 self.subsample_step = 2 # self.datapath = "/export/data/ablattma/Datasets/vegetation_new/" self.metafilename = "meta" self.datadict.update({"train": []}) self.obj_weighting = True # set flow_weights to false self.flow_weights = False def _make_split(self,data): split_data = {"train":{},"test":{}} train_ids = np.flatnonzero(data["train"]) test_ids = np.flatnonzero(np.logical_not(data["train"])) assert np.intersect1d(train_ids,test_ids).size == 0 split_data["train"] = { key: data[key][train_ids] for key in data } split_data["test"] = { key: data[key][test_ids] for key in data } return split_data, train_ids, test_ids class TaichiDataset(VegetationDataset): def _set_instance_specific_values(self): self.filter_flow = True self.valid_lags = [1] self.flow_cutoff = .1 self.min_frames = 5 self.subsample_step = 2 # self.datapath = "/export/scratch/compvis/datasets/taichi/taichi/" self.metafilename = 'meta' self.datadict.update({"train": []}) self.obj_weighting = False # set flow_weights to false self.flow_weights = self.config["flow_weights"] if "flow_weights" in self.config else True self.flow_width_factor = 5 self.target_lags = [10,20] class LargeVegetationDataset(VegetationDataset): def _set_instance_specific_values(self): self.filter_flow = False self.valid_lags = [0] self.flow_cutoff = .1 self.min_frames = 5 self.subsample_step = 2 # self.datapath = "/export/scratch/compvis/datasets/plants/processed_256_resized/" self.metafilename = "meta" self.datadict.update({"train": []}) self.excluded_objects = [1,2,3] self.obj_weighting = True class IperDataset(PlantDataset): def _set_instance_specific_values(self): self.filter_flow = True self.flow_width_factor = 5 self.valid_lags = [0] # set flow cutoff to 0.45 as this seems to be a good heuristic for Iper self.flow_cutoff = 0.6 self.min_frames = 5 # self.datapath = "/export/scratch/compvis/datasets/iPER/processed_256_resized/" self.metafilename = 'meta' #"test_codeprep_metadata" self.datadict.update({"actor_id": [], "action_id": []}) # set object weighting always to false self.obj_weighting = False self.flow_weights = self.config["flow_weights"] if "flow_weights" in self.config else True self.use_flow_for_weights = False def _make_split(self,data): split_data = {"train": {}, "test": {}} if self.split == "videos": key = "vid" elif self.split == "objects": key = "object_id" elif self.split == "actions": key = "action_id" elif self.split == "actors": key = "actor_id" elif self.split == "official": # this is the official train test split as in the original paper with open(path.join("/".join(self.datapath.split("/")[:-1]),"train.txt"),"r") as f: train_names = f.readlines() train_indices = np.asarray([],dtype=np.int) for n in train_names: n = n.replace("/","_").rstrip() train_indices = np.append(train_indices,np.flatnonzero(np.char.find(data["img_path"],n) != -1)) train_indices = np.sort(train_indices) test_indices = np.flatnonzero(np.logical_not(np.isin(np.arange(data["img_path"].shape[0]),train_indices))) split_data["train"] = { key: data[key][train_indices] for key in data } split_data["test"] = { key: data[key][test_indices] for key in data } return split_data, train_indices, test_indices else: vids = np.unique(self.data["vid"]) self.logger.info(f"splitting data across_videos") train_indices = np.asarray([], dtype=np.int) test_indices = np.asarray([], dtype=np.int) for vid in vids: indices = np.nonzero(data["vid"] == vid)[0] # indices = np.arange(len(tdata["img_path"])) # np.random.shuffle(indices) train_indices = np.append(train_indices, indices[: int(0.8 * indices.shape[0])]) test_indices = np.append(test_indices, indices[int(0.8 * indices.shape[0]):]) split_data["train"] = { key: data[key][train_indices] for key in data } split_data["test"] = { key: data[key][test_indices] for key in data } return split_data, train_indices, test_indices # split such that some objects are held back for testing self.logger.info(f"Splitting data after {key}") ids = np.unique(data[key]) shuffled_ids = deepcopy(ids) np.random.shuffle(shuffled_ids) train_ids = shuffled_ids[: int(0.8 * shuffled_ids.shape[0])] train_indices = np.flatnonzero(np.isin(data[key], train_ids)) test_indices = np.flatnonzero(np.logical_not(np.isin(np.arange(self.data["img_path"].shape[0]),train_indices))) train_indices = np.sort(train_indices) test_indices = np.sort(test_indices) split_data["train"] = { key: data[key][train_indices] for key in data } split_data["test"] = { key: data[key][test_indices] for key in data } return split_data, train_indices, test_indices class Human36mDataset(PlantDataset): def _set_instance_specific_values(self): self.valid_lags = [1] self.flow_cutoff = 0.3 self.min_frames = 5 self.subsample_step = 2 # self.datapath = "/export/scratch/compvis/datasets/human3.6M/video_prediction" self.metafilename = "meta" self.datadict.update({"actor_id": [], "action_id": [], "train": []}) # set object weighting always to false self.obj_weighting = False self.filter_flow = False self.flow_width_factor = 5 self.flow_weights = True self.use_flow_for_weights = True self.use_lanczos = True def _make_split(self,data): split_data = {"train": {}, "test": {}} if self.split == "official": train_ids = np.flatnonzero(data["train"]) test_ids = np.flatnonzero(np.logical_not(data["train"])) assert np.intersect1d(train_ids, test_ids).size == 0 split_data["train"] = { key: data[key][train_ids] for key in data } split_data["test"] = { key: data[key][test_ids] for key in data } return split_data, train_ids, test_ids elif self.split == "gui": vids = np.unique(self.data["vid"]) self.logger.info(f"splitting data across_videos") train_indices = np.asarray([], dtype=np.int) test_indices = np.asarray([], dtype=np.int) for vid in vids: indices = np.nonzero(data["vid"] == vid)[0] # indices = np.arange(len(tdata["img_path"])) # np.random.shuffle(indices) train_indices = np.append(train_indices, indices[: int(0.8 * indices.shape[0])]) test_indices = np.append(test_indices, indices[int(0.8 * indices.shape[0]):]) split_data["train"] = { key: data[key][train_indices] for key in data } split_data["test"] = { key: data[key][test_indices] for key in data } return split_data, train_indices, test_indices else: raise ValueError(f'Specified split type "{self.split}" is not valid for Human36mDataset.') class GoogleImgDataset(Dataset, LoggingParent): def __init__(self, base_dir, config,): Dataset.__init__(self) LoggingParent.__init__(self) self.logger.info(f"Initialize GoogleImgDataset with basepath {base_dir}") self.config = config img_paths = [p for p in glob(path.join(base_dir,"*")) if path.isfile(p) and any(map(lambda x: p.endswith(x),["jpg","jpeg","png"]))] self.datadict = {"img_path": np.asarray(img_paths)} self.transforms = tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ]) self.logger.info(f"Initialized Dataset with {self.__len__()} images") def __getitem__(self, idx): return self.datadict["img_path"][idx] def __len__(self): return self.datadict["img_path"].shape[0] if __name__ == "__main__": import yaml import torch from torchvision import transforms as tt from torch.utils.data import DataLoader, RandomSampler import cv2 from os import makedirs from tqdm import tqdm from data import get_dataset from data.samplers import SequenceSampler, SequenceLengthSampler from utils.testing import make_video, make_flow_grid from utils.general import get_patches seed = 42 torch.manual_seed(42) torch.cuda.manual_seed(42) np.random.seed(42) # random.seed(opt.seed) torch.backends.cudnn.deterministic = True torch.manual_seed(42) rng = np.random.RandomState(42) # load config fpath = path.dirname(path.realpath(__file__)) configpath = path.abspath(path.join(fpath, "../config/test_config.yaml")) with open(configpath, "r") as f: config = yaml.load(f, Loader=yaml.FullLoader) transforms = tt.Compose( [tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0)] ) datakeys = ["images", "img_aT", "img_sT", "app_img_cmp", "app_img_random","flow", "poke"] make_overlay = config["general"]["overlay"] # generate dataset dset, transforms = get_dataset(config["data"],transforms) test_dataset = dset(transforms, datakeys, config["data"],train=True) save_dir = f"test_data/{test_dataset.__class__.__name__}" makedirs(save_dir, exist_ok=True) print(test_dataset.datapath) if test_dataset.yield_videos: def init_fn(worker_id): return np.random.seed(np.random.get_state()[1][0] + worker_id) if test_dataset.var_sequence_length: sampler = SequenceLengthSampler(test_dataset,shuffle=True,drop_last=False, batch_size=config["training"]["batch_size"],zero_poke=config["data"]["include_zeropoke"]) loader = DataLoader(test_dataset, batch_sampler=sampler, num_workers=config["data"]["num_workers"], worker_init_fn=init_fn) else: sampler = RandomSampler(test_dataset) loader = DataLoader(test_dataset,batch_size=config["training"]["batch_size"], sampler=sampler,num_workers=config["data"]["num_workers"], worker_init_fn=init_fn, drop_last= True) n_logged = config["testing"]["n_logged"] for i, batch in enumerate(tqdm(loader)): if i >200: break imgs = batch["images"][:n_logged] src_img = imgs[:,0] tgt_img = imgs[:,-1] flow = batch["flow"][:n_logged] poke = batch["poke"][:n_logged][0] if test_dataset.flow_weights else batch["poke"][:n_logged] weights = batch["poke"][:n_logged][1] if test_dataset.flow_weights else None postfix = "weighted" if config["data"]["object_weighting"] else "unweighted" if weights is not None: imgs = get_patches(imgs,weights,config["data"],test_dataset.weight_value_flow) postfix = postfix + "_patched" out_vid = make_video(imgs[:,0],poke,imgs,imgs,n_logged=min(n_logged,config["training"]["batch_size"]),flow=flow,logwandb=False, flow_weights=weights) warping_test = make_flow_grid(src_img,flow,tgt_img,tgt_img,n_logged=min(n_logged,config["training"]["batch_size"])) warping_test = cv2.cvtColor(warping_test,cv2.COLOR_RGB2BGR) cv2.imwrite(path.join(save_dir,f'warping_test-{i}.png'),warping_test) savename = path.join(save_dir,f"vid-grid-{i}-{postfix}.mp4") writer = cv2.VideoWriter( savename, cv2.VideoWriter_fourcc(*"MP4V"), 5, (out_vid.shape[2], out_vid.shape[1]), ) # writer = vio.FFmpegWriter(savename,inputdict=inputdict,outputdict=outputdict) for frame in out_vid: frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) writer.write(frame) writer.release() else: sampler = SequenceSampler(test_dataset, batch_size=config["training"]["batch_size"], shuffle=False, drop_last=False) loader = DataLoader(test_dataset, batch_sampler=sampler, num_workers=config["data"]["num_workers"]) #assert sampler.batch_size == 1 postfix = "filt" if test_dataset.filter_flow else "nofilt " for i, batch in enumerate(tqdm(loader)): if i > 200: break batch = {key: batch[key].squeeze(0) if not isinstance(batch[key],list) else [e.squeeze(0) for e in batch[key]] for key in batch} src_img = batch["images"][0] tgt_img = batch["images"][-1] # vis augmented images img_aT = batch["img_aT"][0] img_sT = batch["img_sT"] img_dis = batch["app_img_random"] img_cmp = batch["app_img_cmp"] # # vis flow flow_map = batch["flow"].permute(1, 2, 0).cpu().numpy() flow_map -= flow_map.min() flow_map /= flow_map.max() flow_map = (flow_map * 255.0).astype(np.uint8) # vis poke poke = batch["poke"][0].permute(1, 2, 0).cpu().numpy() if test_dataset.flow_weights else batch["poke"].permute(1, 2, 0).cpu().numpy() if test_dataset.flow_weights: weight_map = batch["poke"][1].cpu().numpy() weight_map = ((weight_map - weight_map.min()) / weight_map.max() * 255.).astype(np.uint8) heatmap = cv2.applyColorMap(weight_map, cv2.COLORMAP_HOT) heatmap = cv2.cvtColor(heatmap, cv2.COLOR_RGB2BGR) # visualize poke patch in flow map as white region flow_map = np.where((poke**2).sum(-1,keepdims=True)>0, np.full_like(flow_map, 255), flow_map) poke -= poke.min() poke /= poke.max() poke = (poke * 255.0).astype(np.uint8) # vis inverted flow # flow_map_inv = batch["flow_inv"].permute(1, 2, 0).cpu().numpy() # flow_map_inv -= flow_map_inv.min() # flow_map_inv /= flow_map_inv.max() # flow_map_inv = (flow_map_inv * 255.0).astype(np.uint8) # vis images src_img = ( ((src_img.permute(1, 2, 0).cpu() + 1) * 127.5) .numpy() .astype(np.uint8) ) tgt_img = ( ((tgt_img.permute(1, 2, 0).cpu() + 1) * 127.5) .numpy() .astype(np.uint8) ) img_aT = ((img_aT.permute(1, 2, 0).cpu() + 1) * 127.5).numpy().astype(np.uint8) img_sT = ((img_sT.permute(1, 2, 0).cpu() + 1) * 127.5).numpy().astype(np.uint8) img_dis = ((img_dis.permute(1, 2, 0).cpu() + 1) * 127.5).numpy().astype(np.uint8) img_cmp = ((img_cmp.permute(1, 2, 0).cpu() + 1) * 127.5).numpy().astype(np.uint8) if make_overlay: overlay = cv2.addWeighted(src_img,0.5,tgt_img,0.5,0) else: tgt_img = [tgt_img,heatmap] if test_dataset.flow_weights else [tgt_img] zeros = np.expand_dims(np.zeros_like(flow_map).sum(2), axis=2) flow_map = np.concatenate([flow_map, zeros], axis=2) poke = np.concatenate([poke, zeros], axis=2) # flow_map_inv = np.concatenate([flow_map_inv,zeros],axis=2) if make_overlay: grid = np.concatenate([src_img, *tgt_img,overlay, img_sT, img_aT, img_dis, img_cmp, flow_map, poke], axis=1).astype(np.uint8) else: grid = np.concatenate([src_img, *tgt_img, img_sT, img_aT, img_dis, img_cmp, flow_map, poke], axis=1).astype(np.uint8) grid = cv2.cvtColor(grid,cv2.COLOR_BGR2RGB) cv2.imwrite(path.join(save_dir, f"test_grid_{i}-{postfix}.png"), grid)
32,596
41.947299
192
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/data/base_dataset.py
from functools import partial from itertools import chain import torch from torch.nn import functional as F from torch.utils.data import Dataset from torchvision import transforms as T from torchvision.transforms import functional as FT from PIL import Image import numpy as np from abc import abstractmethod import cv2 from utils.general import convert_flow_2d_to_3d, get_flow_gradients from data.helper_functions import preprocess_image from utils.general import LoggingParent class FlowError(Exception): """Raises an exception when no valid flow file could be found """ def __init__(self, path, msg=None): if msg is None: message = f'Could not load flow file "{path}" neither with "allow_pickle=False" nor with "allow_pickle=True". Considering different sequence....' else: message = msg super().__init__(message) class BaseDataset(Dataset, LoggingParent): def __init__(self, transforms, datakeys: list, config: dict, train=True): Dataset.__init__(self) LoggingParent.__init__(self) # list of keys for the data that shall be retained assert len(datakeys) > 0 self.datakeys = datakeys # torchvision.transforms self.transforms = transforms # config: contains all relevant configuration parameters self.config = config self.train = train assert "spatial_size" in self.config self.datapath = self.config['datapath'] # self.valid_lags = np.unique(self.config["valid_lags"]) if "valid_lags" in self.config else list(range(6)) self.yield_videos = self.config["yield_videos"] if "yield_videos" in self.config else False # everything, which has to deal with variable sequence lengths self.var_sequence_length = self.config["var_sequence_length"] if "var_sequence_length" in self.config and self.yield_videos else False self.longest_seq_weight = self.config["longest_seq_weight"] if "longest_seq_weight" in self.config else None self.scale_poke_to_res = self.config["scale_poke_to_res"] if "scale_poke_to_res" in self.config else False if self.scale_poke_to_res: self.logger.info(f'Scaling flows and pokes to dataset resolution, which is {self.config["spatial_size"]}') self.logger.info(f'Dataset is yielding {"videos" if self.yield_videos else "images"}.') self.poke_size = self.config["poke_size"] if "poke_size" in self.config else self.config["spatial_size"][0] / 128 * 10 if "poke" in self.datakeys: self.logger.info(f"Poke size is {self.poke_size}.") # for flow filtering: default values are such that nothing changes self.filter_flow = False self.flow_width_factor = None # whether fancy appearance augmentation shall be used or not self.fancy_aug = self.config["fancy_aug"] if "fancy_aug" in self.config else False # flow weighting, if intended to be enabled self.flow_weights = self.config["flow_weights"] if "flow_weights" in self.config else False self.weight_value_flow = self.config["foreground_value"] if "foreground_value" in self.config else 1. self.weight_value_poke = self.config["poke_value"] if "poke_value" in self.config else 1. self.weight_value_bg = self.config["background_weight"] if "background_weight" in self.config else 1. # whether to use only one value in for poke or the complete flow field within that patch self.equal_poke_val = self.config["equal_poke_val"] if "equal_poke_val" in self.config else True # Whether or not to normalize the flow values self.normalize_flows = self.config["normalize_flows"] if "normalize_flows" in self.config else False # Whether to weight different objects (i.e. samples with different object_ids) the way that the should be yield equally often (recommended for imbalanced datasets) self.obj_weighting = self.config["object_weighting"] if "object_weighting" in self.config else False self.p_col= self.config["p_col"] if "p_col" in self.config else 0 self.p_geom = self.config["p_geom"] if "p_geom" in self.config else 0 self.ab = self.config["augment_b"] if "augment_b" in self.config else 0 self.ac = self.config["augment_c"] if "augment_c" in self.config else 0 self.ah = self.config["augment_h"] if "augment_h" in self.config else 0 self.a_s = self.config["augment_s"] if "augment_s" in self.config else 0 self.ad = self.config["aug_deg"] if "aug_deg" in self.config else 0 self.at = self.config["aug_trans"] if "aug_trans" in self.config else (0,0) self.use_lanczos = self.config["use_lanczos"] if "use_lanczos" in self.config else False self.pre_T = T.ToPILImage() self.z1_normalize = "01_normalize" in self.config and self.config["01_normalize"] if self.z1_normalize: self.post_T = T.Compose([T.ToTensor(),]) else: self.post_T = T.Compose([T.ToTensor(),T.Lambda(lambda x: (x * 2.0) - 1.0)]) self.post_edges = T.Compose([T.ToTensor()]) # key:value mappings for every datakey in self.datakeys self._output_dict = { "images": [partial(self._get_imgs)], "poke": [self._get_poke], "flow": [self._get_flow], "img_aT": [partial(self._get_imgs,use_fb_aug = self.fancy_aug), ["color"]], "img_sT": [partial(self._get_imgs,sample=True),["geometry"]], "app_img_random": [self._get_transfer_img], "app_img_dis": [partial(self._get_imgs, sample=True), ["color", "geometry"]], "app_img_cmp": [self._get_transfer_img], "flow_3D": [self._get_3d_flow], "poke_3D": [self._get_3d_poke], "edge_image": [self._get_edge_image], "edge_flow": [self._get_edge_flow], "flow_3D_series": [self._get_flow_series], "image_series": [self._get_image_series] } if self.fancy_aug: assert "app_img_dis" not in self.datakeys # the data that's held by the dataset self.datadict = { "img_path": [], "flow_paths": [], "img_size": [], "flow_size": [], "vid": [], "fid": [], "object_id": [], # "original_id": [], "flow_range": [] } self.max_frames = self.config["max_frames"] if "max_frames" in self.config else 1 self.augment = self.config["augment_wo_dis"] if ("augment_wo_dis" in self.config and self.train) else False self.color_transfs = None self.geom_transfs = None self.subsample_step = 1 self.min_frames = None # sequence start and end ids are related to the entire dataset and so is self.img_paths self.eids_per_seq = {} self.sids_per_seq = {} self.seq_len_T_chunk = {} self.max_trials_flow_load = 50 #self.img_paths = {} self.mask=None self.flow_norms = None self.flow_in_ram = False self.imgs_in_ram = False self.outside_length = None self.loaded_flows = [] self.loaded_imgs = [] self.valid_lags = None self.ids_per_seq_len = {} self.object_weights_per_seq_len = {} if "weight_zeropoke" in self.config and "include_zeropoke" in self.config: self.zeropoke_weight = max(1.,float(self.max_frames) / 5) if self.config["weight_zeropoke"] and self.config["include_zeropoke"] else 1. else: self.zeropoke_weight = 1. # this is the value, which will be the upper bound for all normalized optical flows, when training on variable sequence lengths # per default, set to 1 here (max) can be adapted, if necessary, in the subclass of base dataset self.flow_cutoff = 1. self.valid_h = [self.poke_size, self.config["spatial_size"][0] - self.poke_size] self.valid_w = [self.poke_size, self.config["spatial_size"][1] - self.poke_size] self.use_flow_for_weights = False def __getitem__(self, idx): """ :param idx: The idx is here a tuple, consisting of the actual id and the sampled lag for the flow in the respective iteration :return: """ # collect outputs data = {} transforms = {"color": self._get_color_transforms(), "geometry" : self._get_geometric_transforms()} self.color_transfs = self._get_color_transforms() if self.augment else None self.geom_transfs = self._get_geometric_transforms() if self.augment else None # sample id (in case, sample is enabled) if self.var_sequence_length: idx = self._get_valid_ids(*idx) else: idx = self._get_valid_ids(length=None,index=idx) sidx = int(np.random.choice(np.flatnonzero(self.datadict["vid"] == self.datadict["vid"][idx[0]]), 1)) tr_vid = int(np.random.choice(self.datadict["vid"][self.datadict["vid"] != self.datadict["vid"][idx[0]]], 1)) for i in range(self.max_trials_flow_load): self.mask = {} try: self._get_mask(idx) data = {key: self._output_dict[key][0](idx, sample_idx = sidx, transforms = chain.from_iterable([transforms[tkey] for tkey in self._output_dict[key][1]]) if len(self._output_dict[key])>1 else None, transfer_vid= tr_vid) for key in self.datakeys} break except FlowError as fe: self.logger.error(fe) # sample new id and try again img_id = int(np.random.choice(np.arange(self.datadict["img_path"].shape[0]),1)) # don't change lag idx = (img_id,idx[1]) if len(data) == 0: raise IOError(f"Errors in flow files loading...tried it {self.max_trials_flow_load} times consecutively without success.") return data def _get_valid_ids(self,length,index = None): """ :param length: The sequence length (or flow step, depending on whether var_sequence_length is True or False) :param index: The id correspinding to the :return: """ # we need to do the following things: # take care, that choose one start id from all samples, which have the appropriate flow_magnitude and result in sequences which are within the same video if self.var_sequence_length: #ids = np.flatnonzero(np.logical_and(self.datadict["flow_range"][:,1]>self.seq_len_T_chunk[length],np.less_equal(np.arange(self.datadict["img_path"].shape[0]) + self.min_seq_length[0] + length*self.subsample_step,self.datadict["seq_end_id"]))) if length == -1: # use maximum sequence length for such cases # length = int(np.random.choice(np.arange(self.max_frames),1)) # in case length == -1: index corresponds to actual sampled length for the regarded batch self.outside_length = index start_id = int(np.random.choice(self.ids_per_seq_len[self.outside_length], 1)) else: ids = self.ids_per_seq_len[length] if self.obj_weighting: start_id = int(np.random.choice(ids, 1, p=self.object_weights_per_seq_len[length])) else: start_id = int(np.random.choice(ids, 1)) else: if index == -1: length = -1 if self.obj_weighting: index = int(np.random.choice(np.arange(self.datadict["object_id"].shape[0]),p=self.datadict["weights"],size=1)) else: index = int(np.random.choice(np.arange(self.datadict["object_id"].shape[0]), p=self.datadict["weights"], size=1)) max_id_fid = self.sids_per_seq[self.datadict["vid"][index]] + self.datadict["max_fid"][index,self.valid_lags[0]] - 1 start_id = min(min(index,self.datadict["seq_end_id"][index]-(self.max_frames* self.subsample_step) - 1),max_id_fid) return (start_id,length) def _get_3d_flow(self, ids, **kwargs): flow = self._get_flow(ids) flow = convert_flow_2d_to_3d(flow) return flow def _get_3d_poke(self, ids, **kwargs): flow = self._get_poke(ids) flow = convert_flow_2d_to_3d(flow) return flow def _get_edge_image(self, ids, sample_idx, transforms=None, sample=False, use_fb_aug=False, **kwargs): imgs = [] if sample: yield_ids = [sample_idx] else: yield_ids = self._get_yield_ids(ids) for i,idx in enumerate(yield_ids): img_path = self.datadict["img_path"][idx] img = cv2.imread(img_path) # image is read in BGR img = preprocess_image(img, swap_channels=True) img = cv2.resize( img, self.config["spatial_size"], cv2.INTER_LINEAR ) # transformations img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) gradient = cv2.Sobel(img/255, cv2.CV_64F, 1, 0, ksize=3) gradient = self.post_edges(gradient)[0] imgs.append(gradient) gradient = cv2.Sobel(img/255, cv2.CV_64F, 0, 1, ksize=3) gradient = self.post_edges(gradient)[0] imgs.append(gradient) return torch.stack(imgs, dim=0).squeeze(dim=0) def _get_edge_flow(self, ids, **kwargs): flow_path = self.datadict["flow_paths"][ids[0], self.valid_lags[0]] # debug, this path seems to be erroneous # flow_path = "/export/data/ablattma/Datasets/plants/processed_crops/VID_0_3_1024x1024/prediction_3_28.flow.npy" try: flow = np.load(flow_path) except ValueError: try: flow = np.load(flow_path,allow_pickle=True) except Exception as ex: print(ex) raise FlowError(flow_path) except: raise FlowError(flow_path) dsize = None if "spatial_size" in self.config: dsize = self.config["spatial_size"] elif "resize_factor" in self.config: dsize = ( int(float(flow.shape[1]) / self.config["resize_factor"]), int(float(flow.shape[2]) / self.config["resize_factor"]), ) flow = F.interpolate( torch.from_numpy(flow).unsqueeze(0), size=dsize, mode="nearest" ).squeeze(0) if self.config["predict_3D"]: flow = convert_flow_2d_to_3d(flow) gradient_d1_x, gradient_d1_y, gradient_d2_x, gradient_d2_y = get_flow_gradients(flow) all_gradients = [gradient_d1_x, gradient_d1_y, gradient_d2_x, gradient_d2_y] return torch.stack(all_gradients, dim=0).squeeze(dim=0) def _get_transfer_img(self, ids, transfer_vid,**kwargs): imgs=[] yield_ids = [int(np.random.choice(np.flatnonzero(self.datadict["vid"] == transfer_vid), 1))] for idx in yield_ids: img_path = self.datadict["img_path"][idx] img = cv2.imread(img_path) # image is read in BGR img = preprocess_image(img, swap_channels=True) if "spatial_size" in self.config: img = cv2.resize( img, self.config["spatial_size"], cv2.INTER_LINEAR ) elif "resize_factor" in self.config: dsize = ( int(float(img.shape[1]) / self.config["resize_factor"]), int(float(img.shape[0]) / self.config["resize_factor"]), ) img = cv2.resize(img, dsize, interpolation=cv2.INTER_LINEAR) # transformations img = self.pre_T(img) img = self.post_T(img) imgs.append(img) return torch.stack(imgs, dim=0).squeeze(dim=0) def _compute_mask(self,target_id): img = self._get_imgs([], sample_idx=target_id, sample=True) if self.z1_normalize: img = (img.permute(1, 2, 0).numpy() * 255.).astype(np.uint8) else: img = ((img.permute(1, 2, 0).numpy() + 1.) * 127.5).astype(np.uint8) mask = np.zeros(img.shape[:2], np.uint8) # rect defines starting background area rect = (int(img.shape[1] / self.flow_width_factor), int(self.valid_h[0]), int((self.flow_width_factor - 2) / self.flow_width_factor * img.shape[1]), int(self.valid_h[1] - self.valid_h[0])) # initialize background and foreground models fgm = np.zeros((1, 65), dtype=np.float64) bgm = np.zeros((1, 65), dtype=np.float64) # apply grab cut algorithm mask2, fgm, bgm = cv2.grabCut(img, mask, rect, fgm, bgm, 5, cv2.GC_INIT_WITH_RECT) return mask2 def _compute_mask_with_flow(self,target_id): flow = self._get_flow([target_id]) amplitude = torch.norm(flow, 2, dim=0) amplitude -= amplitude.min() amplitude /= amplitude.max() # use only such regions where the amplitude is larger than mean + 1 * std mask = torch.where(torch.gt(amplitude,amplitude.mean()+amplitude.std()),torch.ones_like(amplitude),torch.zeros_like(amplitude)).numpy().astype(np.bool) return mask def _get_mask(self,ids): if self.filter_flow or self.fancy_aug or (self.flow_weights and self.yield_videos): if self.use_flow_for_weights: mask_src = self._compute_mask_with_flow(ids[0]) self.mask.update({"img_start": mask_src}) else: mask_src = self._compute_mask(ids[0]) self.mask.update({"img_start" : np.where((mask_src == 2) | (mask_src == 0), 0, 1).astype(np.bool)}) if self.flow_weights: yield_ids = self._get_yield_ids(ids) tgt_id = yield_ids[-1] if self.use_flow_for_weights: mask_tgt = self._compute_mask_with_flow(tgt_id) self.mask.update({"img_tgt": mask_tgt}) else: mask_tgt = self._compute_mask(tgt_id) self.mask.update({"img_tgt": np.where((mask_tgt == 2) | (mask_tgt == 0), 0, 1).astype(np.bool)}) if self.yield_videos: mid_id = int((len(list(yield_ids))+yield_ids[0]) / 2) if self.use_flow_for_weights: mask_mid = self._compute_mask_with_flow(mid_id) self.mask.update({"img_mid": mask_mid}) else: mask_mid = self._compute_mask(mid_id) self.mask.update({"img_mid": np.where((mask_mid == 2) | (mask_mid == 0), 0, 1).astype(np.bool)}) def _get_yield_ids(self,ids): start_id = ids[0] if self.yield_videos: if ids[-1] == -1: if self.var_sequence_length: n_frames = self.min_frames + self.outside_length yield_ids = np.stack([start_id]* n_frames,axis=0).tolist() else: yield_ids = np.stack([start_id]* (self.max_frames+1),axis=0).tolist() else: yield_ids = range(start_id, start_id + (self.min_frames + ids[-1]) * self.subsample_step + 1 ,self.subsample_step) \ if self.var_sequence_length else range(start_id, start_id + self.max_frames * self.subsample_step + 1, self.subsample_step) else: yield_ids = (start_id, start_id + (self.valid_lags[0] + 1) * 5) return yield_ids def _get_image_series(self, ids, step_width=10, **kwargs): all_imgs = [] for i in range(1, step_width+1): new_ids = (ids[0] + i * (1 + self.valid_lags[0]) * 5, ids[1]) flow = self._get_imgs(new_ids, None) all_imgs.append(flow) return torch.from_numpy(np.stack(all_imgs, axis=0)) # grabs a series of images def _get_imgs(self, ids, sample_idx, transforms=None, sample=False, use_fb_aug=False, **kwargs): imgs = [] if sample: yield_ids = [sample_idx] else: # avoid generating the entire sequence for the color transformed image if transforms is not None and self._get_color_transforms in transforms and not sample: yield_ids = [ids[0]] else: yield_ids = self._get_yield_ids(ids) for i,idx in enumerate(yield_ids): faug = use_fb_aug and (i == 0 or i == len(yield_ids) - 1) if self.imgs_in_ram: img = self.loaded_imgs[idx] else: img_path = self.datadict["img_path"][idx] img = cv2.imread(img_path) img = preprocess_image(img, swap_channels=True) # image is read in BGR if self.use_lanczos and self.config["spatial_size"] == 64: img = np.array(Image.fromarray(img).resize(self.config["spatial_size"], resample=Image.LANCZOS)) else: img = cv2.resize( img, self.config["spatial_size"], cv2.INTER_LINEAR ) # transformations img = self.pre_T(img) if transforms is not None: for t in transforms: img = t(img) if faug: bts = self._get_color_transforms() img_back = img for bt in bts: img_back = bt(img_back) img_back = self.post_T(img_back) else: if self.color_transfs is not None: for t in self.color_transfs: img = t(img) if self.geom_transfs is not None: for t in self.geom_transfs: img = t(img) img = self.post_T(img) if faug: img = torch.where(torch.from_numpy(self.mask["img_start"]).unsqueeze(0),img,img_back) imgs.append(img) return torch.stack(imgs, dim=0).squeeze(dim=0) # extracts pokes as flow patches def _get_poke(self, ids, **kwargs): seq_len_idx = ids[-1] if seq_len_idx == -1: # make fake ids to avoid returning zero flow for poke sampling fake_ids = (ids[0],10) flow = self._get_flow(fake_ids) else: flow = self._get_flow(ids) # compute amplitude amplitude = torch.norm(flow[:, self.valid_h[0]:self.valid_h[1], self.valid_w[0]:self.valid_w[1]], 2, dim=0) amplitude -= amplitude.min() amplitude /= amplitude.max() if seq_len_idx == -1: # use only very small poke values, this should indicate background values amplitude_filt = amplitude if self.filter_flow: # only consider the part of the mask which corresponds to the region considered in flow #amplitude_filt = torch.from_numpy(np.where(self.mask["img_start"][self.valid_h[0]:self.valid_h[1],self.valid_w[0]:self.valid_w[1]], amplitude, np.zeros_like(amplitude))) indices_pre = np.nonzero(np.logical_not(self.mask["img_start"][self.valid_h[0]:self.valid_h[1],self.valid_w[0]:self.valid_w[1]])) indices = torch.from_numpy(np.stack(indices_pre,axis=-1)) if indices.shape[0] == 0: indices = torch.lt(amplitude, np.percentile(amplitude.numpy(), 5)).nonzero(as_tuple=False) else: indices = torch.lt(amplitude, np.percentile(amplitude.numpy(), 5)).nonzero(as_tuple=False) #amplitude_filt = amplitude std = amplitude_filt.std() mean = torch.mean(amplitude_filt) indices_mgn = torch.gt(amplitude_filt, mean + (std)).nonzero(as_tuple=False) if indices_mgn.shape[0] == 0: # if flow is not entirely equally distributed, there should be at least 1 value which is above the mean # self.logger.warn("Fallback in Dataloading bacause no values remain after filtering.") indices_mgn = torch.gt(amplitude_filt, mean).nonzero(as_tuple=False) indices_mgn = indices_mgn + np.asarray([[self.valid_h[0], self.valid_w[0]]], dtype=np.int) indices_mgn = (indices_mgn[:, 0], indices_mgn[:, 1]) else: if self.filter_flow: # only consider the part of the mask which corresponds to the region considered in flow amplitude_filt = torch.from_numpy(np.where(self.mask["img_start"][self.valid_h[0]:self.valid_h[1],self.valid_w[0]:self.valid_w[1]], amplitude, np.zeros_like(amplitude))) else: amplitude_filt = amplitude std = amplitude_filt.std() mean = torch.mean(amplitude_filt) if self.var_sequence_length: amplitude_filt = torch.where(torch.from_numpy(np.logical_and((amplitude_filt > self.seq_len_T_chunk[ids[-1]]).numpy(),(amplitude_filt<self.seq_len_T_chunk[ids[-1]+1]).numpy())), amplitude_filt, torch.zeros_like(amplitude_filt)) # compute valid indices by thresholding indices = torch.gt(amplitude_filt, mean + (std * 2.0)).nonzero(as_tuple=False) if indices.shape[0] == 0: indices = torch.gt(amplitude, mean + std).nonzero(as_tuple=False) if indices.shape[0] == 0: # if flow is not entirely equally distributed, there should be at least 1 value which is above the mean #self.logger.warn("Fallback in Dataloading bacause no values remain after filtering.") indices = torch.gt(amplitude, mean).nonzero(as_tuple=False) indices = indices + np.asarray([[self.valid_h[0], self.valid_w[0]]], dtype=np.int) # check if indices is not empty, if so, sample another frame (error is catched in __getitem__()) if indices.shape[0] == 0: raise FlowError(path=[],msg=f"Empty indices array at index {ids[0]}....") # shift ids to match size of real flow patch indices = (indices[:, 0], indices[:, 1]) # generate number of pokes n_pokes = int( np.random.randint( 1, min(self.config["n_pokes"], int(indices[0].shape[0])) + 1 ) ) if seq_len_idx == -1: ids_mgn = np.random.randint(indices_mgn[0].shape[0], size=n_pokes) row_ids_mgn = indices_mgn[0][ids_mgn] col_ids_mgn = indices_mgn[1][ids_mgn] # and generate the actual pokes ids = np.random.randint(indices[0].shape[0], size=n_pokes) row_ids = indices[0][ids] col_ids = indices[1][ids] pokes = [] half_poke_size = int(self.poke_size / 2) zeros = torch.zeros_like(flow) poke_targets = [] for n,ids in enumerate(zip(row_ids, col_ids)): poke = zeros if seq_len_idx == -1: poke_target =flow[:,row_ids_mgn[n],col_ids_mgn[n]].unsqueeze(-1).unsqueeze(-1) if self.equal_poke_val else \ flow[:,row_ids_mgn[n] - half_poke_size:row_ids_mgn[n] + half_poke_size +1, col_ids_mgn[n] - half_poke_size:col_ids_mgn[n] + half_poke_size +1] else: poke_target = flow[:,ids[0],ids[1]].unsqueeze(-1).unsqueeze(-1) if self.equal_poke_val else flow[:, ids[0] - half_poke_size : ids[0] + half_poke_size + 1, ids[1] - half_poke_size : ids[1] + half_poke_size + 1,] poke[ :, ids[0] - half_poke_size: ids[0] + half_poke_size + 1, ids[1] - half_poke_size: ids[1] + half_poke_size + 1, ] = poke_target pokes.append(poke) loc_and_poke = (ids,poke_target) poke_targets.append(loc_and_poke) # unsqueeze in case of num_pokes = 1 if self.flow_weights: if self.yield_videos: if seq_len_idx == -1: complete_mask = np.ones(self.config["spatial_size"], dtype=np.bool) else: complete_mask = np.logical_or(np.logical_or(self.mask["img_tgt"],self.mask["img_start"]), self.mask["img_mid"]) mask_ids = np.nonzero(complete_mask) try: min_h = mask_ids[0].min() max_h = mask_ids[0].max() min_w = mask_ids[1].min() max_w = mask_ids[1].max() weights = np.full(self.mask["img_start"].shape,self.weight_value_bg) weights[min_h:max_h,min_w:max_w] = self.weight_value_flow except Exception as e: self.logger.warn(f'Catch exception in "dataset._get_poke()": {e.__class__.__name__}: "{e}". Using full image instead of patch....') weights = np.full(self.mask["img_start"].shape,self.weight_value_bg) weights[self.valid_h[0]:self.valid_h[1],self.valid_w[0]:self.valid_w[1]] = self.weight_value_flow #weights = np.where(complete_mask,np.full_like(complete_mask,self.weight_value_flow,dtype=np.float),np.full_like(complete_mask,self.weight_value_bg,dtype=np.float),) else: weights = np.where(self.mask["img_tgt"],np.full_like(self.mask["img_tgt"],self.weight_value_flow,dtype=np.float),np.full_like(self.mask["img_tgt"],self.weight_value_bg,dtype=np.float),) # poke regions get higher weights # for poke in pokes: # weights = np.where(((poke**2).sum(0)>0),np.full_like(weights,self.weight_value_poke),weights) weights = torch.from_numpy(weights) pokes = torch.stack(pokes, dim=0).squeeze(0) if "yield_poke_target" in kwargs: return pokes, weights, poke_targets return pokes, weights else: pokes = torch.stack(pokes, dim=0).squeeze(0) if "yield_poke_target" in kwargs: return pokes, poke_targets return pokes def _get_flow_series(self, ids, step_width=10, **kwargs): all_flows = [] for i in range(1, step_width+1): new_ids = (ids[0] + i * (1 + self.valid_lags[0]) * 5, self.valid_lags[0], ids[1]) flow = self._get_3d_flow(new_ids) all_flows.append(flow) return torch.from_numpy(np.stack(all_flows, axis=0)) # extracts entire flow def _get_flow(self, ids, **kwargs): if self.flow_in_ram: flow = torch.from_numpy(self.loaded_flows[ids[0]]) else: flow_path = self.datadict["flow_paths"][ids[0], self.valid_lags[0]] # debug, this path seems to be erroneous # flow_path = "/export/data/ablattma/Datasets/plants/processed_crops/VID_0_3_1024x1024/prediction_3_28.flow.npy" try: flow = np.load(flow_path) except ValueError: try: flow = np.load(flow_path,allow_pickle=True) except Exception as ex: print(ex) raise FlowError(flow_path) except: raise FlowError(flow_path) if self.normalize_flows: flow = flow / self.flow_norms["max_norm"][self.valid_lags[0]] elif not self.normalize_flows and self.scale_poke_to_res: # scaling of poke magnitudes to current resolution flow = flow / (flow.shape[1]/self.config["spatial_size"][0]) dsize = self.config["spatial_size"] flow = F.interpolate( torch.from_numpy(flow).unsqueeze(0), size=dsize, mode="bilinear",align_corners=True ).squeeze(0) if ids[-1] == -1: flow = torch.zeros_like(flow) if self.geom_transfs is not None: c1 = Image.fromarray(flow[0].numpy(),mode="F") c2 = Image.fromarray(flow[1].numpy(),mode="F") for tr in self.geom_transfs: c1 = tr(c1) c2 = tr(c2) flow = torch.from_numpy(np.stack([np.array(c1.getdata()).reshape(c1.size[0],c1.size[1]), np.array(c2.getdata()).reshape(c2.size[0],c2.size[1])],axis=0)).to(torch.float) return flow def _get_color_transforms(self): # to make sure, the transformations are always coherent within the same sample make_trans = bool(np.random.choice(np.arange(2), size=1, p=[1 - self.p_col ,self.p_col])) brightness_val = float(np.random.uniform(-self.ab,self.ab,1)) if self.ab > 0. and make_trans else 0. contrast_val = float(np.random.uniform(-self.ac, self.ac, 1)) if self.ac > 0. and make_trans else 0. hue_val = float(np.random.uniform(-self.ah, 2 * self.ah, 1)) if self.ah > 0. and make_trans else 0. saturation_val = 1. + (float(np.random.uniform(-self.a_s,self.a_s)) if self.a_s > 0. and make_trans else 0) b_T = partial(FT.adjust_brightness,brightness_factor=1. + brightness_val) c_T = partial(FT.adjust_contrast,contrast_factor=1. + contrast_val) h_T = partial(FT.adjust_hue, hue_factor=hue_val) s_T = partial(FT.adjust_saturation,saturation_factor =saturation_val) return [b_T,c_T,h_T,s_T] def _get_geometric_transforms(self): # to make sure, the transformations are always coherent within the same sample make_trans = bool(np.random.choice(np.arange(2),size=1,p=[1-self.p_geom,self.p_geom])) rval = float(np.random.uniform(-self.ad,self.ad,1)) if self.ad > 0. and make_trans else 0. tval_vert = int(np.random.randint(int(-self.at[0] * self.config["spatial_size"][1] / 2), int(self.at[0] * self.config["spatial_size"][1] / 2), 1)) if self.at[0] > 0 and make_trans else 0 tval_hor = int(np.random.randint(int(-self.at[1] * self.config["spatial_size"][0] / 2), int(self.at[1] * self.config["spatial_size"][0] / 2), 1)) if self.at[1] > 0 and make_trans else 0 a_T = partial(FT.affine,angle=rval,translate=(tval_hor,tval_vert),scale=1.0,shear=0) p = partial(FT.pad,padding=(int(self.config["spatial_size"][0] / 2), int(self.config["spatial_size"][1] / 2)),padding_mode="reflect") c = partial(FT.center_crop,output_size=self.config["spatial_size"]) return [p,a_T,c] def _get_flip_transform(self): flip = bool(np.random.choice([True,False],size=1)) if flip: return FT.vflip else: return None @abstractmethod def __len__(self): # as len at least once before dataloading, generic checks can be put here assert self.valid_lags is not None assert self.min_frames is not None if self.filter_flow: assert self.flow_width_factor is not None, f"If the dataset shall be filtered, the flow width factor has to be set in the constructor of the respective child class of BaseDataset" assert isinstance(self.flow_width_factor,int) if self.flow_weights: assert self.flow_width_factor is not None if self.normalize_flows: assert self.flow_norms is not None if self.flow_in_ram: assert len(self.loaded_flows) == self.datadict["flow_paths"].shape[0] if self.imgs_in_ram: assert len(self.loaded_imgs) == self.datadict["img_path"].shape[0] if self.var_sequence_length: assert self.normalize_flows assert self.yield_videos assert len(self.ids_per_seq_len) > 0 assert len(self.object_weights_per_seq_len) == len(self.ids_per_seq_len) return self.datadict["flow_paths"].shape[0] if isinstance(self.datadict["flow_paths"],np.ndarray) else len(self.datadict["flow_paths"]) @abstractmethod def _set_instance_specific_values(self): pass @abstractmethod def get_test_app_images(self) -> dict: pass
36,611
46.119691
255
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/data/__init__.py
from data.base_dataset import BaseDataset from torchvision import transforms as tt from data.flow_dataset import PlantDataset, IperDataset,Human36mDataset, VegetationDataset, LargeVegetationDataset, TaichiDataset # add key value pair for datasets here, all datasets should inherit from base_dataset __datasets__ = {"IperDataset": IperDataset, "PlantDataset": PlantDataset, "Human36mDataset": Human36mDataset, "VegetationDataset": VegetationDataset, "LargeVegetationDataset": LargeVegetationDataset, "TaichiDataset": TaichiDataset, } # returns only the class, not yet an instance def get_transforms(config): return { "PlantDataset": tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ] ), "IperDataset": tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ] ), "Human36mDataset": tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ] ), "VegetationDataset": tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ] ), "LargeVegetationDataset": tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ] ), "TaichiDataset": tt.Compose( [ tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0), ] ), } def get_dataset(config, custom_transforms=None): dataset = __datasets__[config["dataset"]] if custom_transforms is not None: print("Returning dataset with custom transform") transforms = custom_transforms else: transforms = get_transforms(config)[config["dataset"]] return dataset, transforms
2,041
29.029412
129
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/data/prepare_dataset.py
import os import cv2 import re import argparse import torch import numpy as np from os import path, makedirs import pickle from tqdm import tqdm from glob import glob from natsort import natsorted import yaml import multiprocessing as mp from multiprocessing import Process from functools import partial from dotmap import DotMap from torchvision import transforms as tt import configparser from utils.general import parallel_data_prefetch from data import get_dataset from data.helper_functions import preprocess_image h36m_aname2aid = {name: i for i, name in enumerate(["Directions","Discussion","Eating","Greeting","Phoning", "Posing","Purchases","Sitting","SittingDown","Smoking", "Photo","Waiting","Walking","WalkDog","WalkTogether"])} h36m_aname2aid.update({"WalkingTogether": h36m_aname2aid["WalkTogether"]}) h36m_aname2aid.update({"WalkingDog": h36m_aname2aid["WalkDog"]}) h36m_aname2aid.update({"TakingPhoto": h36m_aname2aid["Photo"]}) def _do_parallel_data_prefetch(func, Q, data, idx): # create dummy dataset instance # run prefetching res = func(data) Q.put([idx, res]) Q.put("Done") def get_image(vidcap, frame_number,spatial_size=None): vidcap.set(1, frame_number) _, img = vidcap.read() if spatial_size is not None and spatial_size != img.shape[0]: img=cv2.resize(img,(spatial_size,spatial_size),interpolation=cv2.INTER_LINEAR) return img def process_video(f_name, args): from utils.flownet_loader import FlownetPipeline from utils.general import get_gpu_id_with_lowest_memory, get_logger target_gpus = None if len(args.target_gpus) == 0 else args.target_gpus gpu_index = get_gpu_id_with_lowest_memory(target_gpus=target_gpus) torch.cuda.set_device(gpu_index) #f_name = vid_path.split(vid_path)[-1] logger = get_logger(f"{gpu_index}") extract_device = torch.device("cuda", gpu_index.index if isinstance(gpu_index,torch.device) else gpu_index) # load flownet pipeline = FlownetPipeline() flownet = pipeline.load_flownet(args, extract_device) # open video base_raw_dir = args.raw_dir.split("*")[0] if not isinstance(f_name,list): f_name = [f_name] logger.info(f"Iterating over {len(f_name)} files...") for fn in tqdm(f_name,): if fn.startswith('/'): fn = fn[1:] vid_path = path.join(base_raw_dir, fn) # vid_path = f"Code/input/train_data/movies/{fn}" vidcap = cv2.VideoCapture() vidcap.open(vid_path) counter = 0 while not vidcap.isOpened(): counter += 1 time.sleep(1) if counter > 10: raise Exception("Could not open movie") # get some metadata number_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT)) height = int(vidcap.get(cv2.CAP_PROP_FRAME_HEIGHT)) width = int(vidcap.get(cv2.CAP_PROP_FRAME_WIDTH)) #upright = height > widt # create target path if not existent if args.data.dataset == 'Human36mDataset': vid_name = fn.split('/')[-1] if 'ALL' in vid_name: continue action = vid_name.split(' ')[0] if ' ' in vid_name else vid_name.split('.')[0] same_action_videos =list(filter(lambda y : y.startswith(action) and re.search(r'\d+$', y.split('.')[0]) is not None, map(lambda x: x.split('/')[-1],f_name))) subject = fn.split('/')[-2] if re.search(r'\d+$', fn.split('.')[0]) is not None: subaction_id = int(fn[-1]) else: max_id = max(map(lambda z: int(z.split(' ')[-1].split('.')[0]), same_action_videos)) if max_id ==2: subaction_id = 1 else: subaction_id = 2 cam_id = vid_name.split('.')[1] base_path = path.join(args.processed_dir,subject,f'{action}-{subaction_id}',cam_id) else: base_path = path.join(args.processed_dir, fn.split(".")[0]) #.replace(str,str(args.spatial_size))) # base_path = f"Code/input/train_data/images/{f_name.split('.')[0]}/" makedirs(base_path, exist_ok=True) delta = args.flow_delta diff = args.flow_max # begin extraction for frame_number in range(0, number_frames, args.frames_discr): # break if not enough frames to properly extract sequence if frame_number >= number_frames - diff * args.frames_discr: break first_fidx, second_fidx = frame_number, frame_number + diff * args.frames_discr image_target_file = path.join(base_path, f"frame_{frame_number}.png") # image_target_file = f"{base_path}frame_{frame_number}.png" # FRAME if not path.exists(image_target_file): # write frame itself img = get_image(vidcap, frame_number) if img is None: continue # if upright: # img = cv2.transpose(img) try: if args.spatial_size is None: success = cv2.imwrite(image_target_file, img) else: img_res = cv2.resize(img,(args.spatial_size,args.spatial_size), interpolation=cv2.INTER_LINEAR) success = cv2.imwrite(image_target_file,img_res) except cv2.error as e: print(e) continue except Exception as ex: print(ex) continue # if success: # logger.info(f'wrote img with shape {img.shape} to "{image_target_file}".') # FLOW for d in range(0, diff*args.frames_discr, delta*args.frames_discr): if second_fidx - d < number_frames: flow_target_file = path.join( base_path, f"prediction_{first_fidx}_{second_fidx-d}.flow" ) if not os.path.exists(flow_target_file + ".npy"): # predict and write flow prediction img, img2 = ( get_image(vidcap, first_fidx), get_image(vidcap, second_fidx - d), ) image_target_file2 = path.join(base_path, f"frame_{second_fidx - d}.png") if not path.exists(image_target_file2): try: if args.spatial_size is None: success = cv2.imwrite(image_target_file2, img2) else: img_res2 = cv2.resize(img2, (args.spatial_size, args.spatial_size), interpolation=cv2.INTER_LINEAR) success = cv2.imwrite(image_target_file2, img_res2) except cv2.error as e: print(e) continue except Exception as ex: print(ex) continue sample = pipeline.preprocess_image(img, img2, "BGR",spatial_size=args.input_size).to( extract_device ) prediction = ( pipeline.predict(flownet, sample[None],spatial_size=args.spatial_size) .cpu() .detach() .numpy() ) np.save(flow_target_file, prediction) logger.info( f'Finish processing video sequence "{fn}".') return "Finish" def extract(args): # if args.process_vids: base_dir = args.raw_dir.split("*")[0] if not args.raw_dir.endswith('*'): args.raw_dir =path.join(args.raw_dir,'*') data_names = [p.split(base_dir)[-1] for p in glob(args.raw_dir) if p.endswith(args.video_format)] # data_names = [d for d in data_names if d in ['/VID_0_5.mkv','/VID_7_0.mkv']] fn_extract = partial(process_video, args=args) Q = mp.Queue(1000) # step = ( # int(len(data_names) / args.num_workers + 1) # if len(data_names) % args.num_workers != 0 # else int(len(data_names) / args.num_workers) # ) splits = np.array_split(np.arange(len(data_names)), args.num_workers) arguments = [ [fn_extract, Q, part, i] for i, part in enumerate( [data_names[s[0]:s[-1]+1] for s in splits] ) ] processes = [] for i in range(args.num_workers): p = Process(target=_do_parallel_data_prefetch, args=arguments[i]) processes += [p] start = time.time() gather_res = [[] for _ in range(args.num_workers)] try: for p in processes: p.start() time.sleep(20) k = 0 while k < args.num_workers: # get result res = Q.get() if res == "Done": k += 1 else: gather_res[res[0]] = res[1] except Exception as e: print("Exception: ", e) for p in processes: p.terminate() raise e finally: for p in processes: p.join() print(f"Prefetching complete. [{time.time() - start} sec.]") def prepare(args): logger = get_logger("dataset_preparation") datadict = { "img_path": [], "flow_paths": [], "fid": [], "vid": [], "img_size": [], "flow_size": [], "object_id":[], "max_fid": [] } if "iPER" in args.processed_dir.split("/") or "human36m" in args.processed_dir.split("/") or \ "human3.6M" in args.processed_dir.split("/") : datadict.update({"action_id": [], "actor_id": []}) train_test_split = args.data.dataset == 'Human36mDataset' or args.data.dataset == 'TaichiDataset' fmax = args.flow_max fdelta = args.flow_delta fd = args.frames_discr if train_test_split: datadict.update({"train": []}) if args.data.dataset == 'TaichiDataset': oname2oid = {} # logger.info(f'Metafile is stored as "{args.meta_file_name}.p".') # logger.info(f"args.check_imgs is {args.check_imgs}") max_flow_length = int(fmax / fdelta) # if args.process_vids: if train_test_split: if args.data.dataset == 'Human36mDataset': videos = [d for d in glob(path.join(args.processed_dir, "*", "*", '*')) if path.isdir(d)] else: videos = [d for d in glob(path.join(args.processed_dir, "*", "*")) if path.isdir(d)] else: videos = [d for d in glob(path.join(args.processed_dir, "*")) if path.isdir(d)] videos = natsorted(videos) actual_oid = 0 for vid, vid_name in enumerate(videos): images = glob(path.join(vid_name, "*.png")) images = natsorted(images) actor_id = action_id = train = None if args.data.dataset == 'PlantDataset': object_id = int(vid_name.split("/")[-1].split("_")[1]) elif args.data.dataset == 'IperDataset': object_id = 100 * int(vid_name.split("/")[-1].split("_")[0]) + int(vid_name.split("/")[-1].split("_")[1]) actor_id = int(vid_name.split("/")[-1].split("_")[0]) action_id = int(vid_name.split("/")[-1].split("_")[-1]) elif args.data.dataset == 'TaichiDataset': train = "train" == vid_name.split("/")[-2] msg = "train" if train else "test" print(f"Video in {msg}-split") obj_name = vid_name.split("/")[-1].split("#")[0] if obj_name in oname2oid.keys(): object_id = oname2oid[obj_name] else: object_id = actual_oid oname2oid.update({obj_name: actual_oid}) actual_oid += 1 elif args.data.dataset == 'Human36mDataset': actor_id = int(vid_name.split('/')[-3][1:]) object_id = actor_id action_name = vid_name.split('/')[-2].split('-')[0] action_id = h36m_aname2aid[action_name] train = actor_id not in [9,11] else: raise ValueError("invalid dataset....") # max_flow_id = [len(images) - flow_step -1 for flow_step in range(fdelta*fd,fmax*fd+1, fdelta*fd)] for i, img_path in enumerate( tqdm( images, desc=f'Extracting meta information of video "{vid_name.split(args.processed_dir)[-1]}"', ) ): fid = int(img_path.split("_")[-1].split(".")[0]) #search_pattern = f'[{",".join([str(fid + n) for n in range(args.flow_delta,args.flow_max + 1, args.flow_delta)])}]' flows = natsorted([s for s in glob(path.join(vid_name, f"prediction_{fid}_*.npy")) if (int(s.split("_")[-1].split(".")[0]) - int(s.split("_")[-2])) % (fdelta * fd) == 0 and int(s.split("_")[-1].split(".")[0]) - int(s.split("_")[-2]) <= fmax*fd]) # skip example if second image path does not exist if any(map(lambda p: not path.isfile(path.join(vid_name, f'frame_{p.split("_")[-1].split(".")[0]}.png')),flows)): logger.info(f'Breaking meta file information processing earlier for video "{vid_name.split("/")[-1]}", since not all image frames have been extracted.') break # make relative paths img_path_rel = img_path.split(args.processed_dir)[1] flows_rel = [f.split(args.processed_dir)[1] for f in flows] # filter flows flows_rel = [f for f in flows_rel if (int(f.split("/")[-1].split(".")[0].split("_")[-1]) - int(f.split("/")[-1].split(".")[0].split("_")[-2])) <= fmax*fd] if len(flows_rel) < max_flow_length: diff = max_flow_length-len(flows_rel) [flows_rel.insert(len(flows_rel),last_flow_paths[len(flows_rel)]) for _ in range(diff)] w_img = args.spatial_size h_img = args.spatial_size if len(flows) > 0: w_f = args.spatial_size h_f = args.spatial_size else: h_f = w_f = None assert len(flows_rel) == max_flow_length datadict["img_path"].append(img_path_rel) datadict["flow_paths"].append(flows_rel) datadict["fid"].append(fid) datadict["vid"].append(vid) # image size compliant with numpy and torch datadict["img_size"].append((h_img, w_img)) datadict["flow_size"].append((h_f, w_f)) datadict["object_id"].append(object_id) # datadict["max_fid"].append(max_flow_id) if action_id is not None: datadict["action_id"].append(action_id) if actor_id is not None: datadict["actor_id"].append(actor_id) if train is not None: datadict["train"].append(train) last_flow_paths = flows_rel logger.info(f'Prepared dataset consists of {len(datadict["img_path"])} samples.') # Store data (serialize) save_path = path.join( args.processed_dir, "meta.p" ) with open(save_path, "wb") as handle: pickle.dump(datadict, handle, protocol=pickle.HIGHEST_PROTOCOL) def load_flow(flow_paths): norms = [] for i, flow_path in enumerate(tqdm(flow_paths)): # debug, this path seems to be erroneous # flow_path = "/export/data/ablattma/Datasets/plants/processed_crops/VID_0_3_1024x1024/prediction_3_28.flow.npy" try: flow = np.load(flow_path) except Exception as e: print(e) continue n = np.linalg.norm(flow,2,0) min_norm = np.amin(n) max_norm = np.amax(n) norms.append(np.stack([max_norm,min_norm])) norms = np.stack(norms,0) return norms def norms(cfg_dict): cfg_dict['data']['normalize_flows'] = False transforms = tt.Compose( [tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0)] ) datakeys = ["flow", "images"] dataset, _ = get_dataset(config=cfg_dict["data"]) test_dataset = dataset(transforms, datakeys, cfg_dict["data"], train=True) flow_paths = test_dataset.data["flow_paths"] stats_dict = {"max_norm": [], "min_norm": [], "percentiles": []} for i in range(flow_paths.shape[-1]): test_dataset.logger.info(f'Computing mean of flow with lag {(i + 1) * cfg_dict["flow_delta"]}') norms = parallel_data_prefetch(load_flow, flow_paths[:, i], cfg_dict['data']['num_workers']) max_n = np.amax(norms[:, 0]) min_n = np.amin(norms[:, 1]) percs_at = list(range(10, 100, 10)) percs = np.percentile(norms[:, 0], percs_at) stats_dict["percentiles"].append({pa: p for pa, p in zip(percs_at, percs)}) stats_dict["max_norm"].append(float(max_n)) stats_dict["min_norm"].append(float(min_n)) # save savepath = path.join(test_dataset.datapath, "dataset_stats.p") with open(savepath, "wb") as handle: pickle.dump(stats_dict, handle, protocol=pickle.HIGHEST_PROTOCOL) def process_flows(flow_data): out = np.zeros((len(flow_data), 3)) for i, dp in enumerate(tqdm(flow_data)): flow = np.load(dp[0]) lag = dp[2] test_dataset = dp[3] # flow = flow - test_dataset.flow_norms["min_norm"][test_dataset.valid_lags[0]] flow = flow / test_dataset.flow_norms["max_norm"][lag] img = cv2.imread(dp[1]) # image is read in BGR img = preprocess_image(img, swap_channels=True) mask = np.zeros(img.shape[:2], np.uint8) # rect defines starting background area if test_dataset.filter_flow: rect = ( int(img.shape[1] / test_dataset.flow_width_factor), test_dataset.valid_h[0], int((test_dataset.flow_width_factor - 2) / test_dataset.flow_width_factor * img.shape[1]), test_dataset.valid_h[1] - test_dataset.valid_h[0]) # initialize background and foreground models fgm = np.zeros((1, 65), dtype=np.float64) bgm = np.zeros((1, 65), dtype=np.float64) # apply grab cut algorithm mask2, fgm, bgm = cv2.grabCut(img, mask, rect, fgm, bgm, 5, cv2.GC_INIT_WITH_RECT) amplitude = np.linalg.norm(flow[:, test_dataset.valid_h[0]:test_dataset.valid_h[1], test_dataset.valid_w[0]:test_dataset.valid_w[1]], 2, axis=0) if test_dataset.filter_flow: # only consider the part of the mask which corresponds to the region considered in flow amplitude_filt = np.where(mask2[test_dataset.valid_h[0]:test_dataset.valid_h[1], test_dataset.valid_w[0]:test_dataset.valid_w[1]], amplitude, np.zeros_like(amplitude)) else: amplitude_filt = amplitude std = amplitude_filt.std() mean = np.mean(amplitude_filt) indices = np.argwhere(np.greater(amplitude_filt, mean + (std * 2.0))) if indices.shape[0] == 0: indices = np.argwhere(np.greater(amplitude_filt, np.mean(amplitude_filt) + amplitude_filt.std())) if indices.shape[0] == 0: print("Fallback in Dataloading bacause no values remain after filtering.") # there should be at least one element that is above the mean if flows are not entirely equally distributed indices = np.argwhere(np.greater(amplitude_filt, mean)) if indices.shape[0] == 0: print("strange case, cannot occure, skip") out[i, -1] = 1 continue values = np.asarray([amplitude_filt[idx[0], idx[1]] for idx in indices]) out[i, 0] = values.min() out[i, 1] = values.max() return out def stats(cfg_dict): logger=get_logger("stats_calculation") cfg_dict['data']['normalize_flows'] = True transforms = tt.Compose( [tt.ToTensor(), tt.Lambda(lambda x: (x * 2.0) - 1.0)] ) datakeys = ["flow", "images"] dataset, _ = get_dataset(config=cfg_dict["data"]) test_dataset = dataset(transforms, datakeys, cfg_dict["data"], train=True) all_frange_data = [] for l in tqdm(range(test_dataset.data['flow_paths'].shape[-1])): logger.info(f'Calculating stats for lag of {(l+1) * cfg_dict["flow_delta"]} frames...') in_data = [(f, i, l, test_dataset) for f, i in zip(test_dataset.data["flow_paths"][:, l], test_dataset.data["img_path"])] out_data = parallel_data_prefetch(process_flows,in_data[:100], n_proc=20, cpu_intensive=True, target_data_type="list") all_frange_data.append(out_data) n_error = np.count_nonzero(out_data[:, 2]) logger.info(f"While loading the data, {n_error} errors occurred.") all_frange_data = np.stack(all_frange_data,axis=-1) assert all_frange_data.shape[-1] == test_dataset.datadict['flow_paths'].shape[-1] with open(path.join(test_dataset.datapath,f"{test_dataset.metafilename}.p"),"rb") as f: datadict = pickle.load(f) #assert out_data.shape[0] == len(datadict["img_path"]) key = "flow_range" name_key = "frange" datadict.update({key: all_frange_data}) with open(path.join(test_dataset.datapath, f"{test_dataset.metafilename}_{name_key}.p"), "wb") as f: pickle.dump(datadict, f, protocol=pickle.HIGHEST_PROTOCOL) if __name__ == "__main__": import time from utils.general import get_logger parser = argparse.ArgumentParser() parser.add_argument('-c', '--config',type=str,required=True,help='Config file containing all parameters.') config_args = parser.parse_args() fpath = path.dirname(path.realpath(__file__)) configfile = path.abspath(path.join(fpath,f'../{config_args.config}')) with open(configfile,'r') as f: args = yaml.load(f,Loader=yaml.FullLoader) cfg_dict = args args = DotMap(args) if args.data.dataset == 'Human3.6mDataset': h36config = configparser.ConfigParser() h36config.read(path.join(fpath, 'config.ini')) args.raw_dir = path.join(h36config['General']['TARGETDIR'], 'videos','*','*') cfg_dict['data']['datapath'] = args.processed_dir if args.raw_dir == '': raise ValueError(f'The data holding directory is currently not defined. please define the field "raw_dir" in "{config_args.config}"') if args.processed_dir == '': raise ValueError(f'The target directory for the extracted image frames and flow maps is currently undefined. Please define the field "processed_dir" in "{config_args.config}"') pool = [] torch.multiprocessing.set_start_method("spawn") if args.mode == "extract": extract(args) elif args.mode == "prepare": # in this case, it is prepare prepare(args) elif args.mode == 'stats': # stats(cfg_dict) raise NotImplementedError() elif args.mode == 'norms': # norms(cfg_dict) raise NotImplementedError() elif args.mode == 'all': extract(args) prepare(args) # norms(cfg_dict) # stats(cfg_dict) else: raise ValueError(f'The "mode"-parameter in config file "{configfile}" must be in [all, extract, prepare, norms, stats], but is actually "{args.mode}"...')
23,809
36.974482
185
py
interactive-image2video-synthesis
interactive-image2video-synthesis-main/data/samplers.py
import numpy as np from torch.utils.data import BatchSampler,RandomSampler,SequentialSampler, WeightedRandomSampler from data.base_dataset import BaseDataset from data.flow_dataset import PlantDataset class SequenceSampler(BatchSampler): def __init__(self, dataset:BaseDataset, batch_size, shuffle, drop_last): assert isinstance(dataset, BaseDataset), "The used dataset in Sequence Sampler must inherit from BaseDataset" if shuffle: sampler = RandomSampler(dataset) else: sampler = SequentialSampler(dataset) super().__init__(sampler, batch_size, drop_last) self.dataset = dataset #self.max_lag = self.dataset.datadict["flow_paths"].shape[1] def __iter__(self): batch = [] # sample sequence length lag = int(np.random.choice(self.dataset.valid_lags, 1)) for idx in self.sampler: batch.append((idx, lag)) if len(batch) == self.batch_size: yield batch batch = [] # sample sequence length lag = int(np.random.choice(self.dataset.valid_lags, 1)) if len(batch) > 0 and not self.drop_last: yield batch class FixedLengthSampler(BatchSampler): def __init__(self, dataset:PlantDataset,batch_size,shuffle,drop_last, weighting, zero_poke,zero_poke_amount=None): if shuffle: if weighting: sampler = WeightedRandomSampler(weights=dataset.datadict["weights"],num_samples=len(dataset)) else: sampler = RandomSampler(dataset) else: sampler = SequentialSampler(dataset) super().__init__(sampler, batch_size, drop_last) self.shuffle = shuffle self.dataset = dataset self.zero_poke = zero_poke self.zero_poke_amount = zero_poke_amount if self.zero_poke: assert self.zero_poke_amount is not None def __iter__(self): batch = [] if self.zero_poke: # sample a certain proportion to be zero pokes zero_poke_ids = np.random.choice(np.arange(self.dataset.__len__()),size=int(self.dataset.__len__()/ self.zero_poke_amount),replace=False).tolist() self.dataset.logger.info(f"Sampling {len(zero_poke_ids)} zeropokes for next epoch") else: zero_poke_ids = [] for idx in self.sampler: if idx in zero_poke_ids: batch.append(-1) else: batch.append(idx) if len(batch) == self.batch_size: yield batch batch = [] if len(batch) > 0 and not self.drop_last: yield batch class SequenceLengthSampler(BatchSampler): def __init__(self, dataset:BaseDataset, batch_size, shuffle, drop_last, n_frames=None, zero_poke = False,): assert isinstance(dataset, BaseDataset), "The used dataset in Sequence Sampler must inherit from BaseDataset" assert dataset.var_sequence_length and dataset.yield_videos, "The dataset has to be run in sequence mode and has to output variable sequence lengths" sampler = SequentialSampler(dataset) super().__init__(sampler, batch_size, drop_last) self.dataset = dataset self.shuffle = shuffle if n_frames is not None: assert n_frames >= self.dataset.min_frames and n_frames <=(self.dataset.min_frames + self.dataset.max_frames) self.n_frames = (n_frames-self.dataset.min_frames) else: self.n_frames = n_frames self.start_n_frames = -1 if zero_poke else 0 if zero_poke: if self.dataset.train: self.len_p = np.asarray([self.dataset.zeropoke_weight] + [1.] * self.dataset.max_frames) else: self.len_p = np.asarray([1.] * (self.dataset.max_frames + 1)) else: self.len_p = np.asarray([1.] * self.dataset.max_frames) if self.dataset.longest_seq_weight != None and self.dataset.train: self.len_p[-1] = self.dataset.longest_seq_weight if zero_poke: # to keep sufficient outside pokes for the model to learn foreground and background self.len_p[0] = self.dataset.longest_seq_weight / 2 self.len_p = self.len_p /self.len_p.sum() def __iter__(self): batch = [] # sample sequence length if self.shuffle: # -1 corresponds to n_frames = int(np.random.choice(np.arange(self.start_n_frames,self.dataset.max_frames), 1, p=self.len_p)) else: last_n = self.start_n_frames n_frames = last_n if n_frames == -1: n_frames_actual = int(np.random.choice(np.arange(self.dataset.max_frames), 1)) appended = (n_frames, n_frames_actual) else: appended = (n_frames, None) for idx in self.sampler: appended = (appended[0] if self.n_frames is None else self.n_frames,appended[1]) batch.append(appended) if len(batch) == self.batch_size: yield batch batch = [] # sample sequence length if self.shuffle: n_frames = int(np.random.choice(np.arange(self.start_n_frames,self.dataset.max_frames), 1,p=self.len_p)) else: n_frames = last_n+1 if last_n<self.dataset.max_frames-1 else self.start_n_frames last_n = n_frames if n_frames == -1: n_frames_actual = int(np.random.choice(np.arange(self.dataset.max_frames), 1)) appended = (n_frames, n_frames_actual) else: appended = (n_frames, None) if len(batch) > 0 and not self.drop_last: yield batch
5,888
38.26
158
py
TensorFlowTTS
TensorFlowTTS-master/examples/tacotron2/train_tacotron2.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train Tacotron2.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import yaml from tqdm import tqdm import tensorflow_tts from examples.tacotron2.tacotron_dataset import CharactorMelDataset from tensorflow_tts.configs.tacotron2 import Tacotron2Config from tensorflow_tts.models import TFTacotron2 from tensorflow_tts.optimizers import AdamWeightDecay, WarmUp from tensorflow_tts.trainers import Seq2SeqBasedTrainer from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class Tacotron2Trainer(Seq2SeqBasedTrainer): """Tacotron2 Trainer class based on Seq2SeqBasedTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_mixed_precision (bool): Use mixed precision or not. """ super(Tacotron2Trainer, self).__init__( steps=steps, epochs=epochs, config=config, strategy=strategy, is_mixed_precision=is_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = [ "stop_token_loss", "mel_loss_before", "mel_loss_after", "guided_attention_loss", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() self.config = config def compile(self, model, optimizer): super().compile(model, optimizer) self.binary_crossentropy = tf.keras.losses.BinaryCrossentropy( from_logits=True, reduction=tf.keras.losses.Reduction.NONE ) self.mse = tf.keras.losses.MeanSquaredError( reduction=tf.keras.losses.Reduction.NONE ) self.mae = tf.keras.losses.MeanAbsoluteError( reduction=tf.keras.losses.Reduction.NONE ) def _train_step(self, batch): """Here we re-define _train_step because apply input_signature make the training progress slower on my experiment. Note that input_signature is apply on based_trainer by default. """ if self._already_apply_input_signature is False: self.one_step_forward = tf.function( self._one_step_forward, experimental_relax_shapes=True ) self.one_step_evaluate = tf.function( self._one_step_evaluate, experimental_relax_shapes=True ) self.one_step_predict = tf.function( self._one_step_predict, experimental_relax_shapes=True ) self._already_apply_input_signature = True # run one_step_forward self.one_step_forward(batch) # update counts self.steps += 1 self.tqdm.update(1) self._check_train_finish() def _one_step_evaluate_per_replica(self, batch): """One step evaluate per GPU Tacotron-2 used teacher-forcing when training and evaluation. So we need pass `training=True` for inference step. """ outputs = self._model(**batch, training=True) _, dict_metrics_losses = self.compute_per_example_losses(batch, outputs) self.update_eval_metrics(dict_metrics_losses) def _one_step_predict_per_replica(self, batch): """One step predict per GPU Tacotron-2 used teacher-forcing when training and evaluation. So we need pass `training=True` for inference step. """ outputs = self._model(**batch, training=True) return outputs def compute_per_example_losses(self, batch, outputs): """Compute per example losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ ( decoder_output, post_mel_outputs, stop_token_predictions, alignment_historys, ) = outputs mel_loss_before = calculate_3d_loss( batch["mel_gts"], decoder_output, loss_fn=self.mae ) mel_loss_after = calculate_3d_loss( batch["mel_gts"], post_mel_outputs, loss_fn=self.mae ) # calculate stop_loss max_mel_length = ( tf.reduce_max(batch["mel_lengths"]) if self.config["use_fixed_shapes"] is False else [self.config["max_mel_length"]] ) stop_gts = tf.expand_dims( tf.range(tf.reduce_max(max_mel_length), dtype=tf.int32), 0 ) # [1, max_len] stop_gts = tf.tile( stop_gts, [tf.shape(batch["mel_lengths"])[0], 1] ) # [B, max_len] stop_gts = tf.cast( tf.math.greater_equal(stop_gts, tf.expand_dims(batch["mel_lengths"], 1)), tf.float32, ) stop_token_loss = calculate_2d_loss( stop_gts, stop_token_predictions, loss_fn=self.binary_crossentropy ) # calculate guided attention loss. attention_masks = tf.cast( tf.math.not_equal(batch["g_attentions"], -1.0), tf.float32 ) loss_att = tf.reduce_sum( tf.abs(alignment_historys * batch["g_attentions"]) * attention_masks, axis=[1, 2], ) loss_att /= tf.reduce_sum(attention_masks, axis=[1, 2]) per_example_losses = ( stop_token_loss + mel_loss_before + mel_loss_after + loss_att ) dict_metrics_losses = { "stop_token_loss": stop_token_loss, "mel_loss_before": mel_loss_before, "mel_loss_after": mel_loss_after, "guided_attention_loss": loss_att, } return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt # predict with tf.function for faster. outputs = self.one_step_predict(batch) ( decoder_output, mel_outputs, stop_token_predictions, alignment_historys, ) = outputs mel_gts = batch["mel_gts"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: mels_before = decoder_output.values[0].numpy() mels_after = mel_outputs.values[0].numpy() mel_gts = mel_gts.values[0].numpy() alignment_historys = alignment_historys.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: mels_before = decoder_output.numpy() mels_after = mel_outputs.numpy() mel_gts = mel_gts.numpy() alignment_historys = alignment_historys.numpy() utt_ids = utt_ids.numpy() # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (mel_gt, mel_before, mel_after, alignment_history) in enumerate( zip(mel_gts, mels_before, mels_after, alignment_historys), 0 ): mel_gt = tf.reshape(mel_gt, (-1, 80)).numpy() # [length, 80] mel_before = tf.reshape(mel_before, (-1, 80)).numpy() # [length, 80] mel_after = tf.reshape(mel_after, (-1, 80)).numpy() # [length, 80] # plot figure and save it utt_id = utt_ids[idx] figname = os.path.join(dirname, f"{utt_id}.png") fig = plt.figure(figsize=(10, 8)) ax1 = fig.add_subplot(311) ax2 = fig.add_subplot(312) ax3 = fig.add_subplot(313) im = ax1.imshow(np.rot90(mel_gt), aspect="auto", interpolation="none") ax1.set_title("Target Mel-Spectrogram") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax1) ax2.set_title(f"Predicted Mel-before-Spectrogram @ {self.steps} steps") im = ax2.imshow(np.rot90(mel_before), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax2) ax3.set_title(f"Predicted Mel-after-Spectrogram @ {self.steps} steps") im = ax3.imshow(np.rot90(mel_after), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax3) plt.tight_layout() plt.savefig(figname) plt.close() # plot alignment figname = os.path.join(dirname, f"{idx}_alignment.png") fig = plt.figure(figsize=(8, 6)) ax = fig.add_subplot(111) ax.set_title(f"Alignment @ {self.steps} steps") im = ax.imshow( alignment_history, aspect="auto", origin="lower", interpolation="none" ) fig.colorbar(im, ax=ax) xlabel = "Decoder timestep" plt.xlabel(xlabel) plt.ylabel("Encoder timestep") plt.tight_layout() plt.savefig(figname) plt.close() def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train FastSpeech (See detail in tensorflow_tts/bin/train-fastspeech.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="usr norm-mels for train or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="pretrained weights .h5 file to load weights from. Auto-skips non-matching layers", ) parser.add_argument( "--use-fal", default=0, type=int, help="Use forced alignment guided attention loss or regular", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.mixed_precision = bool(args.mixed_precision) args.use_norm = bool(args.use_norm) args.use_fal = bool(args.use_fal) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ # get dataset if config["remove_short_samples"]: mel_length_threshold = config["mel_length_threshold"] else: mel_length_threshold = 0 if config["format"] == "npy": charactor_query = "*-ids.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" align_query = "*-alignment.npy" if args.use_fal is True else "" charactor_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") train_dataset = CharactorMelDataset( dataset=config["tacotron2_params"]["dataset"], root_dir=args.train_dir, charactor_query=charactor_query, mel_query=mel_query, charactor_load_fn=charactor_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, reduction_factor=config["tacotron2_params"]["reduction_factor"], use_fixed_shapes=config["use_fixed_shapes"], align_query=align_query, ) # update max_mel_length and max_char_length to config config.update({"max_mel_length": int(train_dataset.max_mel_length)}) config.update({"max_char_length": int(train_dataset.max_char_length)}) with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") train_dataset = train_dataset.create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = CharactorMelDataset( dataset=config["tacotron2_params"]["dataset"], root_dir=args.dev_dir, charactor_query=charactor_query, mel_query=mel_query, charactor_load_fn=charactor_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, reduction_factor=config["tacotron2_params"]["reduction_factor"], use_fixed_shapes=False, # don't need apply fixed shape for evaluation. align_query=align_query, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = Tacotron2Trainer( config=config, strategy=STRATEGY, steps=0, epochs=0, is_mixed_precision=args.mixed_precision, ) with STRATEGY.scope(): # define model. tacotron_config = Tacotron2Config(**config["tacotron2_params"]) tacotron2 = TFTacotron2(config=tacotron_config, name="tacotron2") tacotron2._build() tacotron2.summary() if len(args.pretrained) > 1: tacotron2.load_weights(args.pretrained, by_name=True, skip_mismatch=True) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) # AdamW for tacotron2 learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_steps=config["optimizer_params"]["decay_steps"], end_learning_rate=config["optimizer_params"]["end_learning_rate"], ) learning_rate_fn = WarmUp( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_schedule_fn=learning_rate_fn, warmup_steps=int( config["train_max_steps"] * config["optimizer_params"]["warmup_proportion"] ), ) optimizer = AdamWeightDecay( learning_rate=learning_rate_fn, weight_decay_rate=config["optimizer_params"]["weight_decay"], beta_1=0.9, beta_2=0.98, epsilon=1e-6, exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"], ) _ = optimizer.iterations # compile trainer trainer.compile(model=tacotron2, optimizer=optimizer) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
18,412
33.807183
96
py
TensorFlowTTS
TensorFlowTTS-master/examples/multiband_melgan_hf/train_multiband_melgan_hf.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train Multi-Band MelGAN + MPD.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import soundfile as sf import yaml from tensorflow.keras.mixed_precision import experimental as mixed_precision import tensorflow_tts from examples.melgan.audio_mel_dataset import AudioMelDataset from examples.hifigan.train_hifigan import TFHifiGANDiscriminator from examples.melgan.train_melgan import MelganTrainer, collater from tensorflow_tts.configs import ( MultiBandMelGANDiscriminatorConfig, MultiBandMelGANGeneratorConfig, HifiGANDiscriminatorConfig, ) from tensorflow_tts.losses import TFMultiResolutionSTFT from tensorflow_tts.models import ( TFPQMF, TFMelGANGenerator, TFMelGANMultiScaleDiscriminator, TFHifiGANMultiPeriodDiscriminator, ) from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class MultiBandMelganTrainer(MelganTrainer): """Multi-Band MelGAN Trainer class based on MelganTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_generator_mixed_precision=False, is_discriminator_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_generator_mixed_precision (bool): Use mixed precision for generator or not. is_discriminator_mixed_precision (bool): Use mixed precision for discriminator or not. """ super(MultiBandMelganTrainer, self).__init__( config=config, steps=steps, epochs=epochs, strategy=strategy, is_generator_mixed_precision=is_generator_mixed_precision, is_discriminator_mixed_precision=is_discriminator_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = [ "adversarial_loss", "subband_spectral_convergence_loss", "subband_log_magnitude_loss", "fullband_spectral_convergence_loss", "fullband_log_magnitude_loss", "gen_loss", "real_loss", "fake_loss", "dis_loss", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer, pqmf): super().compile(gen_model, dis_model, gen_optimizer, dis_optimizer) # define loss self.sub_band_stft_loss = TFMultiResolutionSTFT( **self.config["subband_stft_loss_params"] ) self.full_band_stft_loss = TFMultiResolutionSTFT( **self.config["stft_loss_params"] ) # define pqmf module self.pqmf = pqmf def compute_per_example_generator_losses(self, batch, outputs): """Compute per example generator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ dict_metrics_losses = {} per_example_losses = 0.0 audios = batch["audios"] y_mb_hat = outputs y_hat = self.pqmf.synthesis(y_mb_hat) y_mb = self.pqmf.analysis(tf.expand_dims(audios, -1)) y_mb = tf.transpose(y_mb, (0, 2, 1)) # [B, subbands, T//subbands] y_mb = tf.reshape(y_mb, (-1, tf.shape(y_mb)[-1])) # [B * subbands, T'] y_mb_hat = tf.transpose(y_mb_hat, (0, 2, 1)) # [B, subbands, T//subbands] y_mb_hat = tf.reshape( y_mb_hat, (-1, tf.shape(y_mb_hat)[-1]) ) # [B * subbands, T'] # calculate sub/full band spectral_convergence and log mag loss. sub_sc_loss, sub_mag_loss = calculate_2d_loss( y_mb, y_mb_hat, self.sub_band_stft_loss ) sub_sc_loss = tf.reduce_mean( tf.reshape(sub_sc_loss, [-1, self.pqmf.subbands]), -1 ) sub_mag_loss = tf.reduce_mean( tf.reshape(sub_mag_loss, [-1, self.pqmf.subbands]), -1 ) full_sc_loss, full_mag_loss = calculate_2d_loss( audios, tf.squeeze(y_hat, -1), self.full_band_stft_loss ) # define generator loss gen_loss = 0.5 * (sub_sc_loss + sub_mag_loss) + 0.5 * ( full_sc_loss + full_mag_loss ) if self.steps >= self.config["discriminator_train_start_steps"]: p_hat = self._discriminator(y_hat) p = self._discriminator(tf.expand_dims(audios, 2)) adv_loss = 0.0 for i in range(len(p_hat)): adv_loss += calculate_3d_loss( tf.ones_like(p_hat[i][-1]), p_hat[i][-1], loss_fn=self.mse_loss ) adv_loss /= i + 1 gen_loss += self.config["lambda_adv"] * adv_loss dict_metrics_losses.update( {"adversarial_loss": adv_loss}, ) dict_metrics_losses.update({"gen_loss": gen_loss}) dict_metrics_losses.update({"subband_spectral_convergence_loss": sub_sc_loss}) dict_metrics_losses.update({"subband_log_magnitude_loss": sub_mag_loss}) dict_metrics_losses.update({"fullband_spectral_convergence_loss": full_sc_loss}) dict_metrics_losses.update({"fullband_log_magnitude_loss": full_mag_loss}) per_example_losses = gen_loss return per_example_losses, dict_metrics_losses def compute_per_example_discriminator_losses(self, batch, gen_outputs): """Compute per example discriminator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ y_mb_hat = gen_outputs y_hat = self.pqmf.synthesis(y_mb_hat) ( per_example_losses, dict_metrics_losses, ) = super().compute_per_example_discriminator_losses(batch, y_hat) return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt y_mb_batch_ = self.one_step_predict(batch) # [B, T // subbands, subbands] y_batch = batch["audios"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: y_mb_batch_ = y_mb_batch_.values[0].numpy() y_batch = y_batch.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: y_mb_batch_ = y_mb_batch_.numpy() y_batch = y_batch.numpy() utt_ids = utt_ids.numpy() y_batch_ = self.pqmf.synthesis(y_mb_batch_).numpy() # [B, T, 1] # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (y, y_) in enumerate(zip(y_batch, y_batch_), 0): # convert to ndarray y, y_ = tf.reshape(y, [-1]).numpy(), tf.reshape(y_, [-1]).numpy() # plit figure and save it utt_id = utt_ids[idx] figname = os.path.join(dirname, f"{utt_id}.png") plt.subplot(2, 1, 1) plt.plot(y) plt.title("groundtruth speech") plt.subplot(2, 1, 2) plt.plot(y_) plt.title(f"generated speech @ {self.steps} steps") plt.tight_layout() plt.savefig(figname) plt.close() # save as wavefile y = np.clip(y, -1, 1) y_ = np.clip(y_, -1, 1) sf.write( figname.replace(".png", "_ref.wav"), y, self.config["sampling_rate"], "PCM_16", ) sf.write( figname.replace(".png", "_gen.wav"), y_, self.config["sampling_rate"], "PCM_16", ) def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train MultiBand MelGAN (See detail in examples/multiband_melgan/train_multiband_melgan.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="use norm mels for training or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--generator_mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--discriminator_mixed_precision", default=0, type=int, help="using mixed precision for discriminator or not.", ) parser.add_argument( "--postnets", default=0, type=int, help="using postnets instead of gt mels or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="path of .h5 mb-melgan generator and discriminator to load weights from. must be comma delineated, like ptgen.h5,ptdisc.h5", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.generator_mixed_precision == 1 or args.discriminator_mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.generator_mixed_precision = bool(args.generator_mixed_precision) args.discriminator_mixed_precision = bool(args.discriminator_mixed_precision) args.use_norm = bool(args.use_norm) args.postnets = bool(args.postnets) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify either --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["batch_max_steps"] // config[ "hop_size" ] + 2 * config["multiband_melgan_generator_params"].get("aux_context_window", 0) else: mel_length_threshold = None if config["format"] == "npy": audio_query = "*-wave.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" audio_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") if args.postnets is True: mel_query = "*-postnet.npy" logging.info("Using postnets") else: logging.info("Using GT Mels") # define train/valid dataset train_dataset = AudioMelDataset( root_dir=args.train_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant(config["batch_max_steps"], dtype=tf.int32), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = AudioMelDataset( root_dir=args.dev_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant( config["batch_max_steps_valid"], dtype=tf.int32 ), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = MultiBandMelganTrainer( steps=0, epochs=0, config=config, strategy=STRATEGY, is_generator_mixed_precision=args.generator_mixed_precision, is_discriminator_mixed_precision=args.discriminator_mixed_precision, ) with STRATEGY.scope(): # define generator and discriminator generator = TFMelGANGenerator( MultiBandMelGANGeneratorConfig( **config["multiband_melgan_generator_params"] ), name="multi_band_melgan_generator", ) multiscale_discriminator = TFMelGANMultiScaleDiscriminator( MultiBandMelGANDiscriminatorConfig( **config["multiband_melgan_discriminator_params"] ), name="multi_band_melgan_discriminator", ) multiperiod_discriminator = TFHifiGANMultiPeriodDiscriminator( HifiGANDiscriminatorConfig(**config["hifigan_discriminator_params"]), name="hifigan_multiperiod_discriminator", ) pqmf = TFPQMF( MultiBandMelGANGeneratorConfig( **config["multiband_melgan_generator_params"] ), dtype=tf.float32, name="pqmf", ) discriminator = TFHifiGANDiscriminator( multiperiod_discriminator, multiscale_discriminator, name="hifigan_discriminator", ) # dummy input to build model. fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) y_mb_hat = generator(fake_mels) y_hat = pqmf.synthesis(y_mb_hat) discriminator(y_hat) if len(args.pretrained) > 1: pt_splits = args.pretrained.split(",") generator.load_weights(pt_splits[0]) discriminator.load_weights(pt_splits[1]) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) generator.summary() discriminator.summary() # define optimizer generator_lr_fn = getattr( tf.keras.optimizers.schedules, config["generator_optimizer_params"]["lr_fn"] )(**config["generator_optimizer_params"]["lr_params"]) discriminator_lr_fn = getattr( tf.keras.optimizers.schedules, config["discriminator_optimizer_params"]["lr_fn"], )(**config["discriminator_optimizer_params"]["lr_params"]) gen_optimizer = tf.keras.optimizers.Adam( learning_rate=generator_lr_fn, amsgrad=config["generator_optimizer_params"]["amsgrad"], ) dis_optimizer = tf.keras.optimizers.Adam( learning_rate=discriminator_lr_fn, amsgrad=config["discriminator_optimizer_params"]["amsgrad"], ) _ = gen_optimizer.iterations _ = dis_optimizer.iterations trainer.compile( gen_model=generator, dis_model=discriminator, gen_optimizer=gen_optimizer, dis_optimizer=dis_optimizer, pqmf=pqmf, ) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
19,162
33.40395
137
py
TensorFlowTTS
TensorFlowTTS-master/examples/fastspeech/train_fastspeech.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train FastSpeech.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import argparse import logging import os import sys sys.path.append(".") import numpy as np import yaml import tensorflow_tts import tensorflow_tts.configs.fastspeech as FASTSPEECH_CONFIG from examples.fastspeech.fastspeech_dataset import CharactorDurationMelDataset from tensorflow_tts.models import TFFastSpeech from tensorflow_tts.optimizers import AdamWeightDecay, WarmUp from tensorflow_tts.trainers import Seq2SeqBasedTrainer from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class FastSpeechTrainer(Seq2SeqBasedTrainer): """FastSpeech Trainer class based on Seq2SeqBasedTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_mixed_precision (bool): Use mixed precision or not. """ super(FastSpeechTrainer, self).__init__( steps=steps, epochs=epochs, config=config, strategy=strategy, is_mixed_precision=is_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = ["duration_loss", "mel_loss_before", "mel_loss_after"] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() self.config = config def compile(self, model, optimizer): super().compile(model, optimizer) self.mse = tf.keras.losses.MeanSquaredError( reduction=tf.keras.losses.Reduction.NONE ) self.mae = tf.keras.losses.MeanAbsoluteError( reduction=tf.keras.losses.Reduction.NONE ) def compute_per_example_losses(self, batch, outputs): """Compute per example losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ mel_before, mel_after, duration_outputs = outputs log_duration = tf.math.log( tf.cast(tf.math.add(batch["duration_gts"], 1), tf.float32) ) duration_loss = self.mse(log_duration, duration_outputs) mel_loss_before = calculate_3d_loss(batch["mel_gts"], mel_before, self.mae) mel_loss_after = calculate_3d_loss(batch["mel_gts"], mel_after, self.mae) per_example_losses = duration_loss + mel_loss_before + mel_loss_after dict_metrics_losses = { "duration_loss": duration_loss, "mel_loss_before": mel_loss_before, "mel_loss_after": mel_loss_after, } return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt # predict with tf.function. outputs = self.one_step_predict(batch) mels_before, mels_after, *_ = outputs mel_gts = batch["mel_gts"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: mels_before = mels_before.values[0].numpy() mels_after = mels_after.values[0].numpy() mel_gts = mel_gts.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: mels_before = mels_before.numpy() mels_after = mels_after.numpy() mel_gts = mel_gts.numpy() utt_ids = utt_ids.numpy() # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (mel_gt, mel_before, mel_after) in enumerate( zip(mel_gts, mels_before, mels_after), 0 ): mel_gt = tf.reshape(mel_gt, (-1, 80)).numpy() # [length, 80] mel_before = tf.reshape(mel_before, (-1, 80)).numpy() # [length, 80] mel_after = tf.reshape(mel_after, (-1, 80)).numpy() # [length, 80] # plit figure and save it utt_id = utt_ids[idx].decode("utf-8") figname = os.path.join(dirname, f"{utt_id}.png") fig = plt.figure(figsize=(10, 8)) ax1 = fig.add_subplot(311) ax2 = fig.add_subplot(312) ax3 = fig.add_subplot(313) im = ax1.imshow(np.rot90(mel_gt), aspect="auto", interpolation="none") ax1.set_title("Target Mel-Spectrogram") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax1) ax2.set_title("Predicted Mel-before-Spectrogram") im = ax2.imshow(np.rot90(mel_before), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax2) ax3.set_title("Predicted Mel-after-Spectrogram") im = ax3.imshow(np.rot90(mel_after), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax3) plt.tight_layout() plt.savefig(figname) plt.close() def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train FastSpeech (See detail in tensorflow_tts/bin/train-fastspeech.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="usr norm-mels for train or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="pretrained checkpoint file to load weights from. Auto-skips non-matching layers", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.mixed_precision = bool(args.mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["mel_length_threshold"] else: mel_length_threshold = None if config["format"] == "npy": charactor_query = "*-ids.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" duration_query = "*-durations.npy" charactor_load_fn = np.load mel_load_fn = np.load duration_load_fn = np.load else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = CharactorDurationMelDataset( root_dir=args.train_dir, charactor_query=charactor_query, mel_query=mel_query, duration_query=duration_query, charactor_load_fn=charactor_load_fn, mel_load_fn=mel_load_fn, duration_load_fn=duration_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = CharactorDurationMelDataset( root_dir=args.dev_dir, charactor_query=charactor_query, mel_query=mel_query, duration_query=duration_query, charactor_load_fn=charactor_load_fn, mel_load_fn=mel_load_fn, duration_load_fn=duration_load_fn, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = FastSpeechTrainer( config=config, strategy=STRATEGY, steps=0, epochs=0, is_mixed_precision=args.mixed_precision, ) with STRATEGY.scope(): # define model fastspeech = TFFastSpeech( config=FASTSPEECH_CONFIG.FastSpeechConfig(**config["fastspeech_params"]) ) fastspeech._build() fastspeech.summary() if len(args.pretrained) > 1: fastspeech.load_weights(args.pretrained, by_name=True, skip_mismatch=True) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) # AdamW for fastspeech learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_steps=config["optimizer_params"]["decay_steps"], end_learning_rate=config["optimizer_params"]["end_learning_rate"], ) learning_rate_fn = WarmUp( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_schedule_fn=learning_rate_fn, warmup_steps=int( config["train_max_steps"] * config["optimizer_params"]["warmup_proportion"] ), ) optimizer = AdamWeightDecay( learning_rate=learning_rate_fn, weight_decay_rate=config["optimizer_params"]["weight_decay"], beta_1=0.9, beta_2=0.98, epsilon=1e-6, exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"], ) _ = optimizer.iterations # compile trainer trainer.compile(model=fastspeech, optimizer=optimizer) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
13,591
33.762148
95
py
TensorFlowTTS
TensorFlowTTS-master/examples/fastspeech2_libritts/train_fastspeech2.py
# -*- coding: utf-8 -*- # Copyright 2020 TensorFlowTTS Team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train FastSpeech2.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import yaml import json import tensorflow_tts from examples.fastspeech2_libritts.fastspeech2_dataset import ( CharactorDurationF0EnergyMelDataset, ) from tensorflow_tts.configs import FastSpeech2Config from tensorflow_tts.models import TFFastSpeech2 from tensorflow_tts.optimizers import AdamWeightDecay, WarmUp from tensorflow_tts.trainers import Seq2SeqBasedTrainer from tensorflow_tts.utils import ( calculate_2d_loss, calculate_3d_loss, return_strategy, TFGriffinLim, ) class FastSpeech2Trainer(Seq2SeqBasedTrainer): """FastSpeech2 Trainer class based on FastSpeechTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_mixed_precision=False, stats_path: str = "", dataset_config: str = "", ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_mixed_precision (bool): Use mixed precision or not. """ super(FastSpeech2Trainer, self).__init__( steps=steps, epochs=epochs, config=config, strategy=strategy, is_mixed_precision=is_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = [ "duration_loss", "f0_loss", "energy_loss", "mel_loss_before", "mel_loss_after", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() self.use_griffin = config.get("use_griffin", False) self.griffin_lim_tf = None if self.use_griffin: logging.info( f"Load griff stats from {stats_path} and config from {dataset_config}" ) self.griff_conf = yaml.load(open(dataset_config), Loader=yaml.Loader) self.prepare_grim(stats_path, self.griff_conf) def prepare_grim(self, stats_path, config): if not stats_path: raise KeyError("stats path need to exist") self.griffin_lim_tf = TFGriffinLim(stats_path, config) def compile(self, model, optimizer): super().compile(model, optimizer) self.mse = tf.keras.losses.MeanSquaredError( reduction=tf.keras.losses.Reduction.NONE ) self.mae = tf.keras.losses.MeanAbsoluteError( reduction=tf.keras.losses.Reduction.NONE ) def compute_per_example_losses(self, batch, outputs): """Compute per example losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ mel_before, mel_after, duration_outputs, f0_outputs, energy_outputs = outputs log_duration = tf.math.log( tf.cast(tf.math.add(batch["duration_gts"], 1), tf.float32) ) duration_loss = calculate_2d_loss(log_duration, duration_outputs, self.mse) f0_loss = calculate_2d_loss(batch["f0_gts"], f0_outputs, self.mse) energy_loss = calculate_2d_loss(batch["energy_gts"], energy_outputs, self.mse) mel_loss_before = calculate_3d_loss(batch["mel_gts"], mel_before, self.mae) mel_loss_after = calculate_3d_loss(batch["mel_gts"], mel_after, self.mae) per_example_losses = ( duration_loss + f0_loss + energy_loss + mel_loss_before + mel_loss_after ) dict_metrics_losses = { "duration_loss": duration_loss, "f0_loss": f0_loss, "energy_loss": energy_loss, "mel_loss_before": mel_loss_before, "mel_loss_after": mel_loss_after, } return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt # predict with tf.function. outputs = self.one_step_predict(batch) mels_before, mels_after, *_ = outputs mel_gts = batch["mel_gts"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: mels_before = mels_before.values[0].numpy() mels_after = mels_after.values[0].numpy() mel_gts = mel_gts.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: mels_before = mels_before.numpy() mels_after = mels_after.numpy() mel_gts = mel_gts.numpy() utt_ids = utt_ids.numpy() # check directory if self.use_griffin: griff_dir_name = os.path.join( self.config["outdir"], f"predictions/{self.steps}_wav" ) if not os.path.exists(griff_dir_name): os.makedirs(griff_dir_name) dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (mel_gt, mel_before, mel_after) in enumerate( zip(mel_gts, mels_before, mels_after), 0 ): if self.use_griffin: utt_id = utt_ids[idx] grif_before = self.griffin_lim_tf( tf.reshape(mel_before, [-1, 80])[tf.newaxis, :], n_iter=32 ) grif_after = self.griffin_lim_tf( tf.reshape(mel_after, [-1, 80])[tf.newaxis, :], n_iter=32 ) grif_gt = self.griffin_lim_tf( tf.reshape(mel_gt, [-1, 80])[tf.newaxis, :], n_iter=32 ) self.griffin_lim_tf.save_wav( grif_before, griff_dir_name, f"{utt_id}_before" ) self.griffin_lim_tf.save_wav( grif_after, griff_dir_name, f"{utt_id}_after" ) self.griffin_lim_tf.save_wav(grif_gt, griff_dir_name, f"{utt_id}_gt") utt_id = utt_ids[idx] mel_gt = tf.reshape(mel_gt, (-1, 80)).numpy() # [length, 80] mel_before = tf.reshape(mel_before, (-1, 80)).numpy() # [length, 80] mel_after = tf.reshape(mel_after, (-1, 80)).numpy() # [length, 80] # plit figure and save it figname = os.path.join(dirname, f"{utt_id}.png") fig = plt.figure(figsize=(10, 8)) ax1 = fig.add_subplot(311) ax2 = fig.add_subplot(312) ax3 = fig.add_subplot(313) im = ax1.imshow(np.rot90(mel_gt), aspect="auto", interpolation="none") ax1.set_title("Target Mel-Spectrogram") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax1) ax2.set_title("Predicted Mel-before-Spectrogram") im = ax2.imshow(np.rot90(mel_before), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax2) ax3.set_title("Predicted Mel-after-Spectrogram") im = ax3.imshow(np.rot90(mel_after), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax3) plt.tight_layout() plt.savefig(figname) plt.close() def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train FastSpeech (See detail in tensorflow_tts/bin/train-fastspeech.py)" ) parser.add_argument( "--train-dir", default="dump/train", type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default="dump/valid", type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="usr norm-mels for train or raw." ) parser.add_argument( "--f0-stat", default="./dump/stats_f0.npy", type=str, help="f0-stat path.", ) parser.add_argument( "--energy-stat", default="./dump/stats_energy.npy", type=str, help="energy-stat path.", ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--mixed_precision", default=1, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--dataset_config", default="preprocess/libritts_preprocess.yaml", type=str, ) parser.add_argument( "--dataset_stats", default="dump/stats.npy", type=str, ) parser.add_argument( "--dataset_mapping", default="dump/libritts_mapper.npy", type=str, ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="pretrained weights .h5 file to load weights from. Auto-skips non-matching layers", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.mixed_precision = bool(args.mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["mel_length_threshold"] else: mel_length_threshold = None if config["format"] == "npy": charactor_query = "*-ids.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" duration_query = "*-durations.npy" f0_query = "*-raw-f0.npy" energy_query = "*-raw-energy.npy" else: raise ValueError("Only npy are supported.") # load speakers map from dataset map with open(args.dataset_mapping) as f: dataset_mapping = json.load(f) speakers_map = dataset_mapping["speakers_map"] # Check n_speakers matches number of speakers in speakers_map n_speakers = config["fastspeech2_params"]["n_speakers"] assert n_speakers == len( speakers_map ), f"Number of speakers in dataset does not match n_speakers in config" # define train/valid dataset train_dataset = CharactorDurationF0EnergyMelDataset( root_dir=args.train_dir, charactor_query=charactor_query, mel_query=mel_query, duration_query=duration_query, f0_query=f0_query, energy_query=energy_query, f0_stat=args.f0_stat, energy_stat=args.energy_stat, mel_length_threshold=mel_length_threshold, speakers_map=speakers_map, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = CharactorDurationF0EnergyMelDataset( root_dir=args.dev_dir, charactor_query=charactor_query, mel_query=mel_query, duration_query=duration_query, f0_query=f0_query, energy_query=energy_query, f0_stat=args.f0_stat, energy_stat=args.energy_stat, mel_length_threshold=mel_length_threshold, speakers_map=speakers_map, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = FastSpeech2Trainer( config=config, strategy=STRATEGY, steps=0, epochs=0, is_mixed_precision=args.mixed_precision, stats_path=args.dataset_stats, dataset_config=args.dataset_config, ) with STRATEGY.scope(): # define model fastspeech = TFFastSpeech2( config=FastSpeech2Config(**config["fastspeech2_params"]) ) fastspeech._build() fastspeech.summary() if len(args.pretrained) > 1: fastspeech.load_weights(args.pretrained, by_name=True, skip_mismatch=True) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) # AdamW for fastspeech learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_steps=config["optimizer_params"]["decay_steps"], end_learning_rate=config["optimizer_params"]["end_learning_rate"], ) learning_rate_fn = WarmUp( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_schedule_fn=learning_rate_fn, warmup_steps=int( config["train_max_steps"] * config["optimizer_params"]["warmup_proportion"] ), ) optimizer = AdamWeightDecay( learning_rate=learning_rate_fn, weight_decay_rate=config["optimizer_params"]["weight_decay"], beta_1=0.9, beta_2=0.98, epsilon=1e-6, exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"], ) _ = optimizer.iterations # compile trainer trainer.compile(model=fastspeech, optimizer=optimizer) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
17,059
33.816327
96
py
TensorFlowTTS
TensorFlowTTS-master/examples/melgan/train_melgan.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train MelGAN.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import soundfile as sf import yaml from tqdm import tqdm import tensorflow_tts import tensorflow_tts.configs.melgan as MELGAN_CONFIG from examples.melgan.audio_mel_dataset import AudioMelDataset from tensorflow_tts.losses import TFMelSpectrogram from tensorflow_tts.models import TFMelGANGenerator, TFMelGANMultiScaleDiscriminator from tensorflow_tts.trainers import GanBasedTrainer from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class MelganTrainer(GanBasedTrainer): """Melgan Trainer class based on GanBasedTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_generator_mixed_precision=False, is_discriminator_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_generator_mixed_precision (bool): Use mixed precision for generator or not. is_discriminator_mixed_precision (bool): Use mixed precision for discriminator or not. """ super(MelganTrainer, self).__init__( steps, epochs, config, strategy, is_generator_mixed_precision, is_discriminator_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = [ "adversarial_loss", "fm_loss", "gen_loss", "real_loss", "fake_loss", "dis_loss", "mels_spectrogram_loss", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() self.config = config def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer): super().compile(gen_model, dis_model, gen_optimizer, dis_optimizer) # define loss self.mse_loss = tf.keras.losses.MeanSquaredError( reduction=tf.keras.losses.Reduction.NONE ) self.mae_loss = tf.keras.losses.MeanAbsoluteError( reduction=tf.keras.losses.Reduction.NONE ) self.mels_loss = TFMelSpectrogram() def compute_per_example_generator_losses(self, batch, outputs): """Compute per example generator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ audios = batch["audios"] y_hat = outputs p_hat = self._discriminator(y_hat) p = self._discriminator(tf.expand_dims(audios, 2)) adv_loss = 0.0 for i in range(len(p_hat)): adv_loss += calculate_3d_loss( tf.ones_like(p_hat[i][-1]), p_hat[i][-1], loss_fn=self.mse_loss ) adv_loss /= i + 1 # define feature-matching loss fm_loss = 0.0 for i in range(len(p_hat)): for j in range(len(p_hat[i]) - 1): fm_loss += calculate_3d_loss( p[i][j], p_hat[i][j], loss_fn=self.mae_loss ) fm_loss /= (i + 1) * (j + 1) adv_loss += self.config["lambda_feat_match"] * fm_loss per_example_losses = adv_loss dict_metrics_losses = { "adversarial_loss": adv_loss, "fm_loss": fm_loss, "gen_loss": adv_loss, "mels_spectrogram_loss": calculate_2d_loss( audios, tf.squeeze(y_hat, -1), loss_fn=self.mels_loss ), } return per_example_losses, dict_metrics_losses def compute_per_example_discriminator_losses(self, batch, gen_outputs): audios = batch["audios"] y_hat = gen_outputs y = tf.expand_dims(audios, 2) p = self._discriminator(y) p_hat = self._discriminator(y_hat) real_loss = 0.0 fake_loss = 0.0 for i in range(len(p)): real_loss += calculate_3d_loss( tf.ones_like(p[i][-1]), p[i][-1], loss_fn=self.mse_loss ) fake_loss += calculate_3d_loss( tf.zeros_like(p_hat[i][-1]), p_hat[i][-1], loss_fn=self.mse_loss ) real_loss /= i + 1 fake_loss /= i + 1 dis_loss = real_loss + fake_loss # calculate per_example_losses and dict_metrics_losses per_example_losses = dis_loss dict_metrics_losses = { "real_loss": real_loss, "fake_loss": fake_loss, "dis_loss": dis_loss, } return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt # generate y_batch_ = self.one_step_predict(batch) y_batch = batch["audios"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: y_batch_ = y_batch_.values[0].numpy() y_batch = y_batch.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: y_batch_ = y_batch_.numpy() y_batch = y_batch.numpy() utt_ids = utt_ids.numpy() # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (y, y_) in enumerate(zip(y_batch, y_batch_), 0): # convert to ndarray y, y_ = tf.reshape(y, [-1]).numpy(), tf.reshape(y_, [-1]).numpy() # plit figure and save it utt_id = utt_ids[idx] figname = os.path.join(dirname, f"{utt_id}.png") plt.subplot(2, 1, 1) plt.plot(y) plt.title("groundtruth speech") plt.subplot(2, 1, 2) plt.plot(y_) plt.title(f"generated speech @ {self.steps} steps") plt.tight_layout() plt.savefig(figname) plt.close() # save as wavefile y = np.clip(y, -1, 1) y_ = np.clip(y_, -1, 1) sf.write( figname.replace(".png", "_ref.wav"), y, self.config["sampling_rate"], "PCM_16", ) sf.write( figname.replace(".png", "_gen.wav"), y_, self.config["sampling_rate"], "PCM_16", ) def collater( items, batch_max_steps=tf.constant(8192, dtype=tf.int32), hop_size=tf.constant(256, dtype=tf.int32), ): """Initialize collater (mapping function) for Tensorflow Audio-Mel Dataset. Args: batch_max_steps (int): The maximum length of input signal in batch. hop_size (int): Hop size of auxiliary features. """ audio, mel = items["audios"], items["mels"] if batch_max_steps is None: batch_max_steps = (tf.shape(audio)[0] // hop_size) * hop_size batch_max_frames = batch_max_steps // hop_size if len(audio) < len(mel) * hop_size: audio = tf.pad(audio, [[0, len(mel) * hop_size - len(audio)]]) if len(mel) > batch_max_frames: # randomly pickup with the batch_max_steps length of the part interval_start = 0 interval_end = len(mel) - batch_max_frames start_frame = tf.random.uniform( shape=[], minval=interval_start, maxval=interval_end, dtype=tf.int32 ) start_step = start_frame * hop_size audio = audio[start_step : start_step + batch_max_steps] mel = mel[start_frame : start_frame + batch_max_frames, :] else: audio = tf.pad(audio, [[0, batch_max_steps - len(audio)]]) mel = tf.pad(mel, [[0, batch_max_frames - len(mel)], [0, 0]]) items = { "utt_ids": items["utt_ids"], "audios": audio, "mels": mel, "mel_lengths": len(mel), "audio_lengths": len(audio), } return items def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train MelGAN (See detail in tensorflow_tts/bin/train-melgan.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="use norm mels for training or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--generator_mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--discriminator_mixed_precision", default=0, type=int, help="using mixed precision for discriminator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="path of .h5 melgan generator to load weights from", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.generator_mixed_precision == 1 or args.discriminator_mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.generator_mixed_precision = bool(args.generator_mixed_precision) args.discriminator_mixed_precision = bool(args.discriminator_mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify either --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["batch_max_steps"] // config[ "hop_size" ] + 2 * config["melgan_generator_params"].get("aux_context_window", 0) else: mel_length_threshold = None if config["format"] == "npy": audio_query = "*-wave.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" audio_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = AudioMelDataset( root_dir=args.train_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant(config["batch_max_steps"], dtype=tf.int32), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = AudioMelDataset( root_dir=args.dev_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant( config["batch_max_steps_valid"], dtype=tf.int32 ), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = MelganTrainer( steps=0, epochs=0, config=config, strategy=STRATEGY, is_generator_mixed_precision=args.generator_mixed_precision, is_discriminator_mixed_precision=args.discriminator_mixed_precision, ) # define generator and discriminator with STRATEGY.scope(): generator = TFMelGANGenerator( MELGAN_CONFIG.MelGANGeneratorConfig(**config["melgan_generator_params"]), name="melgan_generator", ) discriminator = TFMelGANMultiScaleDiscriminator( MELGAN_CONFIG.MelGANDiscriminatorConfig( **config["melgan_discriminator_params"] ), name="melgan_discriminator", ) # dummy input to build model. fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) y_hat = generator(fake_mels) discriminator(y_hat) if len(args.pretrained) > 1: generator.load_weights(args.pretrained) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) generator.summary() discriminator.summary() gen_optimizer = tf.keras.optimizers.Adam(**config["generator_optimizer_params"]) dis_optimizer = tf.keras.optimizers.Adam( **config["discriminator_optimizer_params"] ) trainer.compile( gen_model=generator, dis_model=discriminator, gen_optimizer=gen_optimizer, dis_optimizer=dis_optimizer, ) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
16,989
31.48566
98
py
TensorFlowTTS
TensorFlowTTS-master/examples/melgan_stft/train_melgan_stft.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train MelGAN Multi Resolution STFT Loss.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import yaml import tensorflow_tts import tensorflow_tts.configs.melgan as MELGAN_CONFIG from examples.melgan.audio_mel_dataset import AudioMelDataset from examples.melgan.train_melgan import MelganTrainer, collater from tensorflow_tts.losses import TFMultiResolutionSTFT from tensorflow_tts.models import TFMelGANGenerator, TFMelGANMultiScaleDiscriminator from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class MultiSTFTMelganTrainer(MelganTrainer): """Multi STFT Melgan Trainer class based on MelganTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_generator_mixed_precision=False, is_discriminator_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_generator_mixed_precision (bool): Use mixed precision for generator or not. is_discriminator_mixed_precision (bool): Use mixed precision for discriminator or not. """ super(MultiSTFTMelganTrainer, self).__init__( config=config, steps=steps, epochs=epochs, strategy=strategy, is_generator_mixed_precision=is_generator_mixed_precision, is_discriminator_mixed_precision=is_discriminator_mixed_precision, ) self.list_metrics_name = [ "adversarial_loss", "fm_loss", "gen_loss", "real_loss", "fake_loss", "dis_loss", "spectral_convergence_loss", "log_magnitude_loss", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer): super().compile(gen_model, dis_model, gen_optimizer, dis_optimizer) # define loss self.stft_loss = TFMultiResolutionSTFT(**self.config["stft_loss_params"]) def compute_per_example_generator_losses(self, batch, outputs): """Compute per example generator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ dict_metrics_losses = {} per_example_losses = 0.0 audios = batch["audios"] y_hat = outputs # calculate multi-resolution stft loss sc_loss, mag_loss = calculate_2d_loss( audios, tf.squeeze(y_hat, -1), self.stft_loss ) # trick to prevent loss expoded here sc_loss = tf.where(sc_loss >= 15.0, 0.0, sc_loss) mag_loss = tf.where(mag_loss >= 15.0, 0.0, mag_loss) # compute generator loss gen_loss = 0.5 * (sc_loss + mag_loss) if self.steps >= self.config["discriminator_train_start_steps"]: p_hat = self._discriminator(y_hat) p = self._discriminator(tf.expand_dims(audios, 2)) adv_loss = 0.0 for i in range(len(p_hat)): adv_loss += calculate_3d_loss( tf.ones_like(p_hat[i][-1]), p_hat[i][-1], loss_fn=self.mse_loss ) adv_loss /= i + 1 # define feature-matching loss fm_loss = 0.0 for i in range(len(p_hat)): for j in range(len(p_hat[i]) - 1): fm_loss += calculate_3d_loss( p[i][j], p_hat[i][j], loss_fn=self.mae_loss ) fm_loss /= (i + 1) * (j + 1) adv_loss += self.config["lambda_feat_match"] * fm_loss gen_loss += self.config["lambda_adv"] * adv_loss dict_metrics_losses.update({"adversarial_loss": adv_loss}) dict_metrics_losses.update({"fm_loss": fm_loss}) dict_metrics_losses.update({"gen_loss": gen_loss}) dict_metrics_losses.update({"spectral_convergence_loss": sc_loss}) dict_metrics_losses.update({"log_magnitude_loss": mag_loss}) per_example_losses = gen_loss return per_example_losses, dict_metrics_losses def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train MelGAN (See detail in tensorflow_tts/bin/train-melgan.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="use norm mels for training or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--generator_mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--discriminator_mixed_precision", default=0, type=int, help="using mixed precision for discriminator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="path of .h5 melgan generator to load weights from", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.generator_mixed_precision == 1 or args.discriminator_mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.generator_mixed_precision = bool(args.generator_mixed_precision) args.discriminator_mixed_precision = bool(args.discriminator_mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify either --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["batch_max_steps"] // config[ "hop_size" ] + 2 * config["melgan_generator_params"].get("aux_context_window", 0) else: mel_length_threshold = None if config["format"] == "npy": audio_query = "*-wave.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" audio_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = AudioMelDataset( root_dir=args.train_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant(config["batch_max_steps"], dtype=tf.int32), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = AudioMelDataset( root_dir=args.dev_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant( config["batch_max_steps_valid"], dtype=tf.int32 ), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = MultiSTFTMelganTrainer( steps=0, epochs=0, config=config, strategy=STRATEGY, is_generator_mixed_precision=args.generator_mixed_precision, is_discriminator_mixed_precision=args.discriminator_mixed_precision, ) with STRATEGY.scope(): # define generator and discriminator generator = TFMelGANGenerator( MELGAN_CONFIG.MelGANGeneratorConfig(**config["melgan_generator_params"]), name="melgan_generator", ) discriminator = TFMelGANMultiScaleDiscriminator( MELGAN_CONFIG.MelGANDiscriminatorConfig( **config["melgan_discriminator_params"] ), name="melgan_discriminator", ) # dummy input to build model. fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) y_hat = generator(fake_mels) discriminator(y_hat) if len(args.pretrained) > 1: generator.load_weights(args.pretrained) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) generator.summary() discriminator.summary() # define optimizer generator_lr_fn = getattr( tf.keras.optimizers.schedules, config["generator_optimizer_params"]["lr_fn"] )(**config["generator_optimizer_params"]["lr_params"]) discriminator_lr_fn = getattr( tf.keras.optimizers.schedules, config["discriminator_optimizer_params"]["lr_fn"], )(**config["discriminator_optimizer_params"]["lr_params"]) gen_optimizer = tf.keras.optimizers.Adam( learning_rate=generator_lr_fn, amsgrad=False ) dis_optimizer = tf.keras.optimizers.Adam( learning_rate=discriminator_lr_fn, amsgrad=False ) trainer.compile( gen_model=generator, dis_model=discriminator, gen_optimizer=gen_optimizer, dis_optimizer=dis_optimizer, ) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
13,562
32.655087
98
py
TensorFlowTTS
TensorFlowTTS-master/examples/multiband_melgan/train_multiband_melgan.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train Multi-Band MelGAN.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import soundfile as sf import yaml from tensorflow.keras.mixed_precision import experimental as mixed_precision import tensorflow_tts from examples.melgan.audio_mel_dataset import AudioMelDataset from examples.melgan.train_melgan import MelganTrainer, collater from tensorflow_tts.configs import ( MultiBandMelGANDiscriminatorConfig, MultiBandMelGANGeneratorConfig, ) from tensorflow_tts.losses import TFMultiResolutionSTFT from tensorflow_tts.models import ( TFPQMF, TFMelGANGenerator, TFMelGANMultiScaleDiscriminator, ) from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class MultiBandMelganTrainer(MelganTrainer): """Multi-Band MelGAN Trainer class based on MelganTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_generator_mixed_precision=False, is_discriminator_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_generator_mixed_precision (bool): Use mixed precision for generator or not. is_discriminator_mixed_precision (bool): Use mixed precision for discriminator or not. """ super(MultiBandMelganTrainer, self).__init__( config=config, steps=steps, epochs=epochs, strategy=strategy, is_generator_mixed_precision=is_generator_mixed_precision, is_discriminator_mixed_precision=is_discriminator_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = [ "adversarial_loss", "subband_spectral_convergence_loss", "subband_log_magnitude_loss", "fullband_spectral_convergence_loss", "fullband_log_magnitude_loss", "gen_loss", "real_loss", "fake_loss", "dis_loss", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer, pqmf): super().compile(gen_model, dis_model, gen_optimizer, dis_optimizer) # define loss self.sub_band_stft_loss = TFMultiResolutionSTFT( **self.config["subband_stft_loss_params"] ) self.full_band_stft_loss = TFMultiResolutionSTFT( **self.config["stft_loss_params"] ) # define pqmf module self.pqmf = pqmf def compute_per_example_generator_losses(self, batch, outputs): """Compute per example generator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ dict_metrics_losses = {} per_example_losses = 0.0 audios = batch["audios"] y_mb_hat = outputs y_hat = self.pqmf.synthesis(y_mb_hat) y_mb = self.pqmf.analysis(tf.expand_dims(audios, -1)) y_mb = tf.transpose(y_mb, (0, 2, 1)) # [B, subbands, T//subbands] y_mb = tf.reshape(y_mb, (-1, tf.shape(y_mb)[-1])) # [B * subbands, T'] y_mb_hat = tf.transpose(y_mb_hat, (0, 2, 1)) # [B, subbands, T//subbands] y_mb_hat = tf.reshape( y_mb_hat, (-1, tf.shape(y_mb_hat)[-1]) ) # [B * subbands, T'] # calculate sub/full band spectral_convergence and log mag loss. sub_sc_loss, sub_mag_loss = calculate_2d_loss( y_mb, y_mb_hat, self.sub_band_stft_loss ) sub_sc_loss = tf.reduce_mean( tf.reshape(sub_sc_loss, [-1, self.pqmf.subbands]), -1 ) sub_mag_loss = tf.reduce_mean( tf.reshape(sub_mag_loss, [-1, self.pqmf.subbands]), -1 ) full_sc_loss, full_mag_loss = calculate_2d_loss( audios, tf.squeeze(y_hat, -1), self.full_band_stft_loss ) # define generator loss gen_loss = 0.5 * (sub_sc_loss + sub_mag_loss) + 0.5 * ( full_sc_loss + full_mag_loss ) if self.steps >= self.config["discriminator_train_start_steps"]: p_hat = self._discriminator(y_hat) p = self._discriminator(tf.expand_dims(audios, 2)) adv_loss = 0.0 for i in range(len(p_hat)): adv_loss += calculate_3d_loss( tf.ones_like(p_hat[i][-1]), p_hat[i][-1], loss_fn=self.mse_loss ) adv_loss /= i + 1 gen_loss += self.config["lambda_adv"] * adv_loss dict_metrics_losses.update({"adversarial_loss": adv_loss},) dict_metrics_losses.update({"gen_loss": gen_loss}) dict_metrics_losses.update({"subband_spectral_convergence_loss": sub_sc_loss}) dict_metrics_losses.update({"subband_log_magnitude_loss": sub_mag_loss}) dict_metrics_losses.update({"fullband_spectral_convergence_loss": full_sc_loss}) dict_metrics_losses.update({"fullband_log_magnitude_loss": full_mag_loss}) per_example_losses = gen_loss return per_example_losses, dict_metrics_losses def compute_per_example_discriminator_losses(self, batch, gen_outputs): """Compute per example discriminator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ y_mb_hat = gen_outputs y_hat = self.pqmf.synthesis(y_mb_hat) ( per_example_losses, dict_metrics_losses, ) = super().compute_per_example_discriminator_losses(batch, y_hat) return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt y_mb_batch_ = self.one_step_predict(batch) # [B, T // subbands, subbands] y_batch = batch["audios"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: y_mb_batch_ = y_mb_batch_.values[0].numpy() y_batch = y_batch.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: y_mb_batch_ = y_mb_batch_.numpy() y_batch = y_batch.numpy() utt_ids = utt_ids.numpy() y_batch_ = self.pqmf.synthesis(y_mb_batch_).numpy() # [B, T, 1] # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (y, y_) in enumerate(zip(y_batch, y_batch_), 0): # convert to ndarray y, y_ = tf.reshape(y, [-1]).numpy(), tf.reshape(y_, [-1]).numpy() # plit figure and save it utt_id = utt_ids[idx] figname = os.path.join(dirname, f"{utt_id}.png") plt.subplot(2, 1, 1) plt.plot(y) plt.title("groundtruth speech") plt.subplot(2, 1, 2) plt.plot(y_) plt.title(f"generated speech @ {self.steps} steps") plt.tight_layout() plt.savefig(figname) plt.close() # save as wavefile y = np.clip(y, -1, 1) y_ = np.clip(y_, -1, 1) sf.write( figname.replace(".png", "_ref.wav"), y, self.config["sampling_rate"], "PCM_16", ) sf.write( figname.replace(".png", "_gen.wav"), y_, self.config["sampling_rate"], "PCM_16", ) def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train MultiBand MelGAN (See detail in examples/multiband_melgan/train_multiband_melgan.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="use norm mels for training or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--generator_mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--discriminator_mixed_precision", default=0, type=int, help="using mixed precision for discriminator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="path of .h5 mb-melgan generator to load weights from", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.generator_mixed_precision == 1 or args.discriminator_mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.generator_mixed_precision = bool(args.generator_mixed_precision) args.discriminator_mixed_precision = bool(args.discriminator_mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify either --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["batch_max_steps"] // config[ "hop_size" ] + 2 * config["multiband_melgan_generator_params"].get("aux_context_window", 0) else: mel_length_threshold = None if config["format"] == "npy": audio_query = "*-wave.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" audio_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = AudioMelDataset( root_dir=args.train_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant(config["batch_max_steps"], dtype=tf.int32), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = AudioMelDataset( root_dir=args.dev_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant( config["batch_max_steps_valid"], dtype=tf.int32 ), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = MultiBandMelganTrainer( steps=0, epochs=0, config=config, strategy=STRATEGY, is_generator_mixed_precision=args.generator_mixed_precision, is_discriminator_mixed_precision=args.discriminator_mixed_precision, ) with STRATEGY.scope(): # define generator and discriminator generator = TFMelGANGenerator( MultiBandMelGANGeneratorConfig( **config["multiband_melgan_generator_params"] ), name="multi_band_melgan_generator", ) discriminator = TFMelGANMultiScaleDiscriminator( MultiBandMelGANDiscriminatorConfig( **config["multiband_melgan_discriminator_params"] ), name="multi_band_melgan_discriminator", ) pqmf = TFPQMF( MultiBandMelGANGeneratorConfig( **config["multiband_melgan_generator_params"] ), dtype=tf.float32, name="pqmf", ) # dummy input to build model. fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) y_mb_hat = generator(fake_mels) y_hat = pqmf.synthesis(y_mb_hat) discriminator(y_hat) if len(args.pretrained) > 1: generator.load_weights(args.pretrained) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) generator.summary() discriminator.summary() # define optimizer generator_lr_fn = getattr( tf.keras.optimizers.schedules, config["generator_optimizer_params"]["lr_fn"] )(**config["generator_optimizer_params"]["lr_params"]) discriminator_lr_fn = getattr( tf.keras.optimizers.schedules, config["discriminator_optimizer_params"]["lr_fn"], )(**config["discriminator_optimizer_params"]["lr_params"]) gen_optimizer = tf.keras.optimizers.Adam( learning_rate=generator_lr_fn, amsgrad=config["generator_optimizer_params"]["amsgrad"], ) dis_optimizer = tf.keras.optimizers.Adam( learning_rate=discriminator_lr_fn, amsgrad=config["discriminator_optimizer_params"]["amsgrad"], ) trainer.compile( gen_model=generator, dis_model=discriminator, gen_optimizer=gen_optimizer, dis_optimizer=dis_optimizer, pqmf=pqmf, ) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
18,014
33.379771
112
py
TensorFlowTTS
TensorFlowTTS-master/examples/fastspeech2/train_fastspeech2.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train FastSpeech2.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import yaml from tqdm import tqdm import tensorflow_tts from examples.fastspeech2.fastspeech2_dataset import CharactorDurationF0EnergyMelDataset from examples.fastspeech.train_fastspeech import FastSpeechTrainer from tensorflow_tts.configs import FastSpeech2Config from tensorflow_tts.models import TFFastSpeech2 from tensorflow_tts.optimizers import AdamWeightDecay, WarmUp from tensorflow_tts.trainers import Seq2SeqBasedTrainer from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy class FastSpeech2Trainer(Seq2SeqBasedTrainer): """FastSpeech2 Trainer class based on FastSpeechTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_mixed_precision (bool): Use mixed precision or not. """ super(FastSpeech2Trainer, self).__init__( steps=steps, epochs=epochs, config=config, strategy=strategy, is_mixed_precision=is_mixed_precision, ) # define metrics to aggregates data and use tf.summary logs them self.list_metrics_name = [ "duration_loss", "f0_loss", "energy_loss", "mel_loss_before", "mel_loss_after", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() def compile(self, model, optimizer): super().compile(model, optimizer) self.mse = tf.keras.losses.MeanSquaredError( reduction=tf.keras.losses.Reduction.NONE ) self.mae = tf.keras.losses.MeanAbsoluteError( reduction=tf.keras.losses.Reduction.NONE ) def compute_per_example_losses(self, batch, outputs): """Compute per example losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ mel_before, mel_after, duration_outputs, f0_outputs, energy_outputs = outputs log_duration = tf.math.log( tf.cast(tf.math.add(batch["duration_gts"], 1), tf.float32) ) duration_loss = calculate_2d_loss(log_duration, duration_outputs, self.mse) f0_loss = calculate_2d_loss(batch["f0_gts"], f0_outputs, self.mse) energy_loss = calculate_2d_loss(batch["energy_gts"], energy_outputs, self.mse) mel_loss_before = calculate_3d_loss(batch["mel_gts"], mel_before, self.mae) mel_loss_after = calculate_3d_loss(batch["mel_gts"], mel_after, self.mae) per_example_losses = ( duration_loss + f0_loss + energy_loss + mel_loss_before + mel_loss_after ) dict_metrics_losses = { "duration_loss": duration_loss, "f0_loss": f0_loss, "energy_loss": energy_loss, "mel_loss_before": mel_loss_before, "mel_loss_after": mel_loss_after, } return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt # predict with tf.function. outputs = self.one_step_predict(batch) mels_before, mels_after, *_ = outputs mel_gts = batch["mel_gts"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: mels_before = mels_before.values[0].numpy() mels_after = mels_after.values[0].numpy() mel_gts = mel_gts.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: mels_before = mels_before.numpy() mels_after = mels_after.numpy() mel_gts = mel_gts.numpy() utt_ids = utt_ids.numpy() # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (mel_gt, mel_before, mel_after) in enumerate( zip(mel_gts, mels_before, mels_after), 0 ): mel_gt = tf.reshape(mel_gt, (-1, 80)).numpy() # [length, 80] mel_before = tf.reshape(mel_before, (-1, 80)).numpy() # [length, 80] mel_after = tf.reshape(mel_after, (-1, 80)).numpy() # [length, 80] # plit figure and save it utt_id = utt_ids[idx] figname = os.path.join(dirname, f"{utt_id}.png") fig = plt.figure(figsize=(10, 8)) ax1 = fig.add_subplot(311) ax2 = fig.add_subplot(312) ax3 = fig.add_subplot(313) im = ax1.imshow(np.rot90(mel_gt), aspect="auto", interpolation="none") ax1.set_title("Target Mel-Spectrogram") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax1) ax2.set_title("Predicted Mel-before-Spectrogram") im = ax2.imshow(np.rot90(mel_before), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax2) ax3.set_title("Predicted Mel-after-Spectrogram") im = ax3.imshow(np.rot90(mel_after), aspect="auto", interpolation="none") fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax3) plt.tight_layout() plt.savefig(figname) plt.close() def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train FastSpeech (See detail in tensorflow_tts/bin/train-fastspeech.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="usr norm-mels for train or raw." ) parser.add_argument( "--f0-stat", default="./dump/stats_f0.npy", type=str, required=True, help="f0-stat path.", ) parser.add_argument( "--energy-stat", default="./dump/stats_energy.npy", type=str, required=True, help="energy-stat path.", ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="pretrained weights .h5 file to load weights from. Auto-skips non-matching layers", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.mixed_precision = bool(args.mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["mel_length_threshold"] else: mel_length_threshold = None if config["format"] == "npy": charactor_query = "*-ids.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" duration_query = "*-durations.npy" f0_query = "*-raw-f0.npy" energy_query = "*-raw-energy.npy" else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = CharactorDurationF0EnergyMelDataset( root_dir=args.train_dir, charactor_query=charactor_query, mel_query=mel_query, duration_query=duration_query, f0_query=f0_query, energy_query=energy_query, f0_stat=args.f0_stat, energy_stat=args.energy_stat, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = CharactorDurationF0EnergyMelDataset( root_dir=args.dev_dir, charactor_query=charactor_query, mel_query=mel_query, duration_query=duration_query, f0_query=f0_query, energy_query=energy_query, f0_stat=args.f0_stat, energy_stat=args.energy_stat, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = FastSpeech2Trainer( config=config, strategy=STRATEGY, steps=0, epochs=0, is_mixed_precision=args.mixed_precision, ) with STRATEGY.scope(): # define model fastspeech = TFFastSpeech2( config=FastSpeech2Config(**config["fastspeech2_params"]) ) fastspeech._build() fastspeech.summary() if len(args.pretrained) > 1: fastspeech.load_weights(args.pretrained, by_name=True, skip_mismatch=True) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) # AdamW for fastspeech learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_steps=config["optimizer_params"]["decay_steps"], end_learning_rate=config["optimizer_params"]["end_learning_rate"], ) learning_rate_fn = WarmUp( initial_learning_rate=config["optimizer_params"]["initial_learning_rate"], decay_schedule_fn=learning_rate_fn, warmup_steps=int( config["train_max_steps"] * config["optimizer_params"]["warmup_proportion"] ), ) optimizer = AdamWeightDecay( learning_rate=learning_rate_fn, weight_decay_rate=config["optimizer_params"]["weight_decay"], beta_1=0.9, beta_2=0.98, epsilon=1e-6, exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"], ) _ = optimizer.iterations # compile trainer trainer.compile(model=fastspeech, optimizer=optimizer) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
14,446
33.562201
96
py
TensorFlowTTS
TensorFlowTTS-master/examples/hifigan/train_hifigan.py
# -*- coding: utf-8 -*- # Copyright 2020 TensorFlowTTS Team # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train Hifigan.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import numpy as np import soundfile as sf import yaml from tqdm import tqdm import tensorflow_tts from examples.melgan.audio_mel_dataset import AudioMelDataset from examples.melgan.train_melgan import collater from examples.melgan_stft.train_melgan_stft import MultiSTFTMelganTrainer from tensorflow_tts.configs import ( HifiGANDiscriminatorConfig, HifiGANGeneratorConfig, MelGANDiscriminatorConfig, ) from tensorflow_tts.models import ( TFHifiGANGenerator, TFHifiGANMultiPeriodDiscriminator, TFMelGANMultiScaleDiscriminator, ) from tensorflow_tts.utils import return_strategy class TFHifiGANDiscriminator(tf.keras.Model): def __init__(self, multiperiod_dis, multiscale_dis, **kwargs): super().__init__(**kwargs) self.multiperiod_dis = multiperiod_dis self.multiscale_dis = multiscale_dis def call(self, x): outs = [] period_outs = self.multiperiod_dis(x) scale_outs = self.multiscale_dis(x) outs.extend(period_outs) outs.extend(scale_outs) return outs def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train Hifigan (See detail in examples/hifigan/train_hifigan.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="use norm mels for training or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--generator_mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--discriminator_mixed_precision", default=0, type=int, help="using mixed precision for discriminator or not.", ) parser.add_argument( "--pretrained", default="", type=str, nargs="?", help="path of .h5 melgan generator to load weights from", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.generator_mixed_precision == 1 or args.discriminator_mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.generator_mixed_precision = bool(args.generator_mixed_precision) args.discriminator_mixed_precision = bool(args.discriminator_mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify either --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["batch_max_steps"] // config[ "hop_size" ] + 2 * config["hifigan_generator_params"].get("aux_context_window", 0) else: mel_length_threshold = None if config["format"] == "npy": audio_query = "*-wave.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" audio_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = AudioMelDataset( root_dir=args.train_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant(config["batch_max_steps"], dtype=tf.int32), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = AudioMelDataset( root_dir=args.dev_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant( config["batch_max_steps_valid"], dtype=tf.int32 ), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = MultiSTFTMelganTrainer( steps=0, epochs=0, config=config, strategy=STRATEGY, is_generator_mixed_precision=args.generator_mixed_precision, is_discriminator_mixed_precision=args.discriminator_mixed_precision, ) with STRATEGY.scope(): # define generator and discriminator generator = TFHifiGANGenerator( HifiGANGeneratorConfig(**config["hifigan_generator_params"]), name="hifigan_generator", ) multiperiod_discriminator = TFHifiGANMultiPeriodDiscriminator( HifiGANDiscriminatorConfig(**config["hifigan_discriminator_params"]), name="hifigan_multiperiod_discriminator", ) multiscale_discriminator = TFMelGANMultiScaleDiscriminator( MelGANDiscriminatorConfig( **config["melgan_discriminator_params"], name="melgan_multiscale_discriminator", ) ) discriminator = TFHifiGANDiscriminator( multiperiod_discriminator, multiscale_discriminator, name="hifigan_discriminator", ) # dummy input to build model. fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) y_hat = generator(fake_mels) discriminator(y_hat) if len(args.pretrained) > 1: generator.load_weights(args.pretrained) logging.info( f"Successfully loaded pretrained weight from {args.pretrained}." ) generator.summary() discriminator.summary() # define optimizer generator_lr_fn = getattr( tf.keras.optimizers.schedules, config["generator_optimizer_params"]["lr_fn"] )(**config["generator_optimizer_params"]["lr_params"]) discriminator_lr_fn = getattr( tf.keras.optimizers.schedules, config["discriminator_optimizer_params"]["lr_fn"], )(**config["discriminator_optimizer_params"]["lr_params"]) gen_optimizer = tf.keras.optimizers.Adam( learning_rate=generator_lr_fn, amsgrad=config["generator_optimizer_params"]["amsgrad"], ) dis_optimizer = tf.keras.optimizers.Adam( learning_rate=discriminator_lr_fn, amsgrad=config["discriminator_optimizer_params"]["amsgrad"], ) trainer.compile( gen_model=generator, dis_model=discriminator, gen_optimizer=gen_optimizer, dis_optimizer=dis_optimizer, ) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
10,464
31.101227
88
py
TensorFlowTTS
TensorFlowTTS-master/examples/parallel_wavegan/train_parallel_wavegan.py
# -*- coding: utf-8 -*- # Copyright 2020 TensorFlowTTS Team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Train ParallelWavegan.""" import tensorflow as tf physical_devices = tf.config.list_physical_devices("GPU") for i in range(len(physical_devices)): tf.config.experimental.set_memory_growth(physical_devices[i], True) import sys sys.path.append(".") import argparse import logging import os import soundfile as sf import numpy as np import yaml import tensorflow_tts from examples.melgan.audio_mel_dataset import AudioMelDataset from examples.melgan.train_melgan import collater from tensorflow_tts.configs import ( ParallelWaveGANGeneratorConfig, ParallelWaveGANDiscriminatorConfig, ) from tensorflow_tts.models import ( TFParallelWaveGANGenerator, TFParallelWaveGANDiscriminator, ) from tensorflow_tts.trainers import GanBasedTrainer from tensorflow_tts.losses import TFMultiResolutionSTFT from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy from tensorflow_addons.optimizers import RectifiedAdam class ParallelWaveganTrainer(GanBasedTrainer): """ParallelWaveGAN Trainer class based on GanBasedTrainer.""" def __init__( self, config, strategy, steps=0, epochs=0, is_generator_mixed_precision=False, is_discriminator_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. is_generator_mixed_precision (bool): Use mixed precision for generator or not. is_discriminator_mixed_precision (bool): Use mixed precision for discriminator or not. """ super(ParallelWaveganTrainer, self).__init__( config=config, steps=steps, epochs=epochs, strategy=strategy, is_generator_mixed_precision=is_generator_mixed_precision, is_discriminator_mixed_precision=is_discriminator_mixed_precision, ) self.list_metrics_name = [ "adversarial_loss", "gen_loss", "real_loss", "fake_loss", "dis_loss", "spectral_convergence_loss", "log_magnitude_loss", ] self.init_train_eval_metrics(self.list_metrics_name) self.reset_states_train() self.reset_states_eval() def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer): super().compile(gen_model, dis_model, gen_optimizer, dis_optimizer) # define loss self.stft_loss = TFMultiResolutionSTFT(**self.config["stft_loss_params"]) self.mse_loss = tf.keras.losses.MeanSquaredError( reduction=tf.keras.losses.Reduction.NONE ) self.mae_loss = tf.keras.losses.MeanAbsoluteError( reduction=tf.keras.losses.Reduction.NONE ) def compute_per_example_generator_losses(self, batch, outputs): """Compute per example generator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ dict_metrics_losses = {} per_example_losses = 0.0 audios = batch["audios"] y_hat = outputs # calculate multi-resolution stft loss sc_loss, mag_loss = calculate_2d_loss( audios, tf.squeeze(y_hat, -1), self.stft_loss ) gen_loss = 0.5 * (sc_loss + mag_loss) if self.steps >= self.config["discriminator_train_start_steps"]: p_hat = self._discriminator(y_hat) p = self._discriminator(tf.expand_dims(audios, 2)) adv_loss = 0.0 adv_loss += calculate_3d_loss( tf.ones_like(p_hat), p_hat, loss_fn=self.mse_loss ) gen_loss += self.config["lambda_adv"] * adv_loss # update dict_metrics_losses dict_metrics_losses.update({"adversarial_loss": adv_loss}) dict_metrics_losses.update({"gen_loss": gen_loss}) dict_metrics_losses.update({"spectral_convergence_loss": sc_loss}) dict_metrics_losses.update({"log_magnitude_loss": mag_loss}) per_example_losses = gen_loss return per_example_losses, dict_metrics_losses def compute_per_example_discriminator_losses(self, batch, gen_outputs): audios = batch["audios"] y_hat = gen_outputs y = tf.expand_dims(audios, 2) p = self._discriminator(y) p_hat = self._discriminator(y_hat) real_loss = 0.0 fake_loss = 0.0 real_loss += calculate_3d_loss(tf.ones_like(p), p, loss_fn=self.mse_loss) fake_loss += calculate_3d_loss( tf.zeros_like(p_hat), p_hat, loss_fn=self.mse_loss ) dis_loss = real_loss + fake_loss # calculate per_example_losses and dict_metrics_losses per_example_losses = dis_loss dict_metrics_losses = { "real_loss": real_loss, "fake_loss": fake_loss, "dis_loss": dis_loss, } return per_example_losses, dict_metrics_losses def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" import matplotlib.pyplot as plt # generate y_batch_ = self.one_step_predict(batch) y_batch = batch["audios"] utt_ids = batch["utt_ids"] # convert to tensor. # here we just take a sample at first replica. try: y_batch_ = y_batch_.values[0].numpy() y_batch = y_batch.values[0].numpy() utt_ids = utt_ids.values[0].numpy() except Exception: y_batch_ = y_batch_.numpy() y_batch = y_batch.numpy() utt_ids = utt_ids.numpy() # check directory dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps") if not os.path.exists(dirname): os.makedirs(dirname) for idx, (y, y_) in enumerate(zip(y_batch, y_batch_), 0): # convert to ndarray y, y_ = tf.reshape(y, [-1]).numpy(), tf.reshape(y_, [-1]).numpy() # plit figure and save it utt_id = utt_ids[idx] figname = os.path.join(dirname, f"{utt_id}.png") plt.subplot(2, 1, 1) plt.plot(y) plt.title("groundtruth speech") plt.subplot(2, 1, 2) plt.plot(y_) plt.title(f"generated speech @ {self.steps} steps") plt.tight_layout() plt.savefig(figname) plt.close() # save as wavefile y = np.clip(y, -1, 1) y_ = np.clip(y_, -1, 1) sf.write( figname.replace(".png", "_ref.wav"), y, self.config["sampling_rate"], "PCM_16", ) sf.write( figname.replace(".png", "_gen.wav"), y_, self.config["sampling_rate"], "PCM_16", ) def main(): """Run training process.""" parser = argparse.ArgumentParser( description="Train ParallelWaveGan (See detail in tensorflow_tts/examples/parallel_wavegan/train_parallel_wavegan.py)" ) parser.add_argument( "--train-dir", default=None, type=str, help="directory including training data. ", ) parser.add_argument( "--dev-dir", default=None, type=str, help="directory including development data. ", ) parser.add_argument( "--use-norm", default=1, type=int, help="use norm mels for training or raw." ) parser.add_argument( "--outdir", type=str, required=True, help="directory to save checkpoints." ) parser.add_argument( "--config", type=str, required=True, help="yaml format configuration file." ) parser.add_argument( "--resume", default="", type=str, nargs="?", help='checkpoint file path to resume training. (default="")', ) parser.add_argument( "--verbose", type=int, default=1, help="logging level. higher is more logging. (default=1)", ) parser.add_argument( "--generator_mixed_precision", default=0, type=int, help="using mixed precision for generator or not.", ) parser.add_argument( "--discriminator_mixed_precision", default=0, type=int, help="using mixed precision for discriminator or not.", ) args = parser.parse_args() # return strategy STRATEGY = return_strategy() # set mixed precision config if args.generator_mixed_precision == 1 or args.discriminator_mixed_precision == 1: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) args.generator_mixed_precision = bool(args.generator_mixed_precision) args.discriminator_mixed_precision = bool(args.discriminator_mixed_precision) args.use_norm = bool(args.use_norm) # set logger if args.verbose > 1: logging.basicConfig( level=logging.DEBUG, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) elif args.verbose > 0: logging.basicConfig( level=logging.INFO, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) else: logging.basicConfig( level=logging.WARN, stream=sys.stdout, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) logging.warning("Skip DEBUG/INFO messages") # check directory existence if not os.path.exists(args.outdir): os.makedirs(args.outdir) # check arguments if args.train_dir is None: raise ValueError("Please specify --train-dir") if args.dev_dir is None: raise ValueError("Please specify either --valid-dir") # load and save config with open(args.config) as f: config = yaml.load(f, Loader=yaml.Loader) config.update(vars(args)) config["version"] = tensorflow_tts.__version__ with open(os.path.join(args.outdir, "config.yml"), "w") as f: yaml.dump(config, f, Dumper=yaml.Dumper) for key, value in config.items(): logging.info(f"{key} = {value}") # get dataset if config["remove_short_samples"]: mel_length_threshold = config["batch_max_steps"] // config[ "hop_size" ] + 2 * config["parallel_wavegan_generator_params"].get("aux_context_window", 0) else: mel_length_threshold = None if config["format"] == "npy": audio_query = "*-wave.npy" mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy" audio_load_fn = np.load mel_load_fn = np.load else: raise ValueError("Only npy are supported.") # define train/valid dataset train_dataset = AudioMelDataset( root_dir=args.train_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant(config["batch_max_steps"], dtype=tf.int32), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync * config["gradient_accumulation_steps"], ) valid_dataset = AudioMelDataset( root_dir=args.dev_dir, audio_query=audio_query, mel_query=mel_query, audio_load_fn=audio_load_fn, mel_load_fn=mel_load_fn, mel_length_threshold=mel_length_threshold, ).create( is_shuffle=config["is_shuffle"], map_fn=lambda items: collater( items, batch_max_steps=tf.constant( config["batch_max_steps_valid"], dtype=tf.int32 ), hop_size=tf.constant(config["hop_size"], dtype=tf.int32), ), allow_cache=config["allow_cache"], batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync, ) # define trainer trainer = ParallelWaveganTrainer( steps=0, epochs=0, config=config, strategy=STRATEGY, is_generator_mixed_precision=args.generator_mixed_precision, is_discriminator_mixed_precision=args.discriminator_mixed_precision, ) with STRATEGY.scope(): # define generator and discriminator generator = TFParallelWaveGANGenerator( ParallelWaveGANGeneratorConfig( **config["parallel_wavegan_generator_params"] ), name="parallel_wavegan_generator", ) discriminator = TFParallelWaveGANDiscriminator( ParallelWaveGANDiscriminatorConfig( **config["parallel_wavegan_discriminator_params"] ), name="parallel_wavegan_discriminator", ) # dummy input to build model. fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) y_hat = generator(fake_mels) discriminator(y_hat) generator.summary() discriminator.summary() # define optimizer generator_lr_fn = getattr( tf.keras.optimizers.schedules, config["generator_optimizer_params"]["lr_fn"] )(**config["generator_optimizer_params"]["lr_params"]) discriminator_lr_fn = getattr( tf.keras.optimizers.schedules, config["discriminator_optimizer_params"]["lr_fn"], )(**config["discriminator_optimizer_params"]["lr_params"]) gen_optimizer = RectifiedAdam(learning_rate=generator_lr_fn, amsgrad=False) dis_optimizer = RectifiedAdam(learning_rate=discriminator_lr_fn, amsgrad=False) trainer.compile( gen_model=generator, dis_model=discriminator, gen_optimizer=gen_optimizer, dis_optimizer=dis_optimizer, ) # start training try: trainer.fit( train_dataset, valid_dataset, saved_path=os.path.join(config["outdir"], "checkpoints/"), resume=args.resume, ) except KeyboardInterrupt: trainer.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.") if __name__ == "__main__": main()
15,699
32.052632
126
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/base_model.py
# -*- coding: utf-8 -*- # Copyright 2020 TensorFlowTTS Team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Base Model for all model.""" import tensorflow as tf import yaml import os import numpy as np from tensorflow_tts.utils.utils import MODEL_FILE_NAME, CONFIG_FILE_NAME class BaseModel(tf.keras.Model): def set_config(self, config): self.config = config def save_pretrained(self, saved_path): """Save config and weights to file""" os.makedirs(saved_path, exist_ok=True) self.config.save_pretrained(saved_path) self.save_weights(os.path.join(saved_path, MODEL_FILE_NAME))
1,131
32.294118
74
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/parallel_wavegan.py
# -*- coding: utf-8 -*- # Copyright 2020 The TensorFlowTTS Team and Tomoki Hayashi (@kan-bayashi) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Parallel-wavegan Modules. Based on pytorch implementation (https://github.com/kan-bayashi/ParallelWaveGAN)""" import tensorflow as tf from tensorflow_tts.models import BaseModel def get_initializer(initializer_seed=42): """Creates a `tf.initializers.he_normal` with the given seed. Args: initializer_seed: int, initializer seed. Returns: HeNormal initializer with seed = `initializer_seed`. """ return tf.keras.initializers.he_normal(seed=initializer_seed) class TFConv1d1x1(tf.keras.layers.Conv1D): """1x1 Conv1d with customized initialization.""" def __init__(self, filters, use_bias, padding, initializer_seed, **kwargs): """Initialize 1x1 Conv1d module.""" super().__init__( filters=filters, kernel_size=1, strides=1, padding=padding, dilation_rate=1, use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), **kwargs, ) class TFConv1d(tf.keras.layers.Conv1D): """Conv1d with customized initialization.""" def __init__(self, *args, **kwargs): """Initialize Conv1d module.""" initializer_seed = kwargs.pop("initializer_seed", 42) super().__init__( *args, **kwargs, kernel_initializer=get_initializer(initializer_seed) ) class TFResidualBlock(tf.keras.layers.Layer): """Residual block module in WaveNet.""" def __init__( self, kernel_size=3, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, dropout_rate=0.0, dilation_rate=1, use_bias=True, use_causal_conv=False, initializer_seed=42, **kwargs, ): """Initialize ResidualBlock module. Args: kernel_size (int): Kernel size of dilation convolution layer. residual_channels (int): Number of channels for residual connection. skip_channels (int): Number of channels for skip connection. aux_channels (int): Local conditioning channels i.e. auxiliary input dimension. dropout_rate (float): Dropout probability. dilation_rate (int): Dilation factor. use_bias (bool): Whether to add bias parameter in convolution layers. use_causal_conv (bool): Whether to use use_causal_conv or non-use_causal_conv convolution. initializer_seed (int32): initializer seed. """ super().__init__(**kwargs) self.dropout_rate = dropout_rate # no future time stamps available self.use_causal_conv = use_causal_conv # dilation conv self.conv = TFConv1d( filters=gate_channels, kernel_size=kernel_size, padding="same" if self.use_causal_conv is False else "causal", strides=1, dilation_rate=dilation_rate, use_bias=use_bias, initializer_seed=initializer_seed, ) # local conditionong if aux_channels > 0: self.conv1x1_aux = TFConv1d1x1( gate_channels, use_bias=False, padding="same", initializer_seed=initializer_seed, name="conv1x1_aux", ) else: self.conv1x1_aux = None # conv output is split into two groups gate_out_channels = gate_channels // 2 self.conv1x1_out = TFConv1d1x1( residual_channels, use_bias=use_bias, padding="same", initializer_seed=initializer_seed, name="conv1x1_out", ) self.conv1x1_skip = TFConv1d1x1( skip_channels, use_bias=use_bias, padding="same", initializer_seed=initializer_seed, name="conv1x1_skip", ) self.dropout = tf.keras.layers.Dropout(rate=self.dropout_rate) def call(self, x, c, training=False): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, residual_channels, T). c (Tensor): Local conditioning auxiliary tensor (B, aux_channels, T). Returns: Tensor: Output tensor for residual connection (B, T, residual_channels). Tensor: Output tensor for skip connection (B, T, skip_channels). """ residual = x x = self.dropout(x, training=training) x = self.conv(x) # split into two part for gated activation xa, xb = tf.split(x, 2, axis=-1) # local conditioning if c is not None: assert self.conv1x1_aux is not None c = self.conv1x1_aux(c) ca, cb = tf.split(c, 2, axis=-1) xa, xb = xa + ca, xb + cb x = tf.nn.tanh(xa) * tf.nn.sigmoid(xb) # for skip connection s = self.conv1x1_skip(x) # for residual connection x = self.conv1x1_out(x) x = (x + residual) * tf.math.sqrt(0.5) return x, s class TFStretch1d(tf.keras.layers.Layer): """Stretch2d module.""" def __init__(self, x_scale, y_scale, method="nearest", **kwargs): """Initialize Stretch2d module. Args: x_scale (int): X scaling factor (Time axis in spectrogram). y_scale (int): Y scaling factor (Frequency axis in spectrogram). method (str): Interpolation method. """ super().__init__(**kwargs) self.x_scale = x_scale self.y_scale = y_scale self.method = method def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, T, C, 1). Returns: Tensor: Interpolated tensor (B, T * x_scale, C * y_scale, 1) """ x_shape = tf.shape(x) new_size = (x_shape[1] * self.x_scale, x_shape[2] * self.y_scale) x = tf.image.resize(x, method=self.method, size=new_size) return x class TFUpsampleNetWork(tf.keras.layers.Layer): """Upsampling network module.""" def __init__( self, output_channels, upsample_scales, nonlinear_activation=None, nonlinear_activation_params={}, interpolate_mode="nearest", freq_axis_kernel_size=1, use_causal_conv=False, **kwargs, ): """Initialize upsampling network module. Args: output_channels (int): output feature channels. upsample_scales (list): List of upsampling scales. nonlinear_activation (str): Activation function name. nonlinear_activation_params (dict): Arguments for specified activation function. interpolate_mode (str): Interpolation mode. freq_axis_kernel_size (int): Kernel size in the direction of frequency axis. """ super().__init__(**kwargs) self.use_causal_conv = use_causal_conv self.up_layers = [] for scale in upsample_scales: # interpolation layer stretch = TFStretch1d( scale, 1, interpolate_mode, name="stretch_._{}".format(scale) ) # ->> outputs: [B, T * scale, C * 1, 1] self.up_layers += [stretch] # conv layer assert ( freq_axis_kernel_size - 1 ) % 2 == 0, "Not support even number freq axis kernel size." kernel_size = scale * 2 + 1 conv = tf.keras.layers.Conv2D( filters=1, kernel_size=(kernel_size, freq_axis_kernel_size), padding="causal" if self.use_causal_conv is True else "same", use_bias=False, ) # ->> outputs: [B, T * scale, C * 1, 1] self.up_layers += [conv] # nonlinear if nonlinear_activation is not None: nonlinear = getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ) self.up_layers += [nonlinear] def call(self, c): """Calculate forward propagation. Args: c : Input tensor (B, T, C). Returns: Tensor: Upsampled tensor (B, T', C), where T' = T * prod(upsample_scales). """ c = tf.expand_dims(c, -1) # [B, T, C, 1] for f in self.up_layers: c = f(c) return tf.squeeze(c, -1) # [B, T, C] class TFConvInUpsampleNetWork(tf.keras.layers.Layer): """Convolution + upsampling network module.""" def __init__( self, upsample_scales, nonlinear_activation=None, nonlinear_activation_params={}, interpolate_mode="nearest", freq_axis_kernel_size=1, aux_channels=80, aux_context_window=0, use_causal_conv=False, initializer_seed=42, **kwargs, ): """Initialize convolution + upsampling network module. Args: upsample_scales (list): List of upsampling scales. nonlinear_activation (str): Activation function name. nonlinear_activation_params (dict): Arguments for specified activation function. mode (str): Interpolation mode. freq_axis_kernel_size (int): Kernel size in the direction of frequency axis. aux_channels (int): Number of channels of pre-convolutional layer. aux_context_window (int): Context window size of the pre-convolutional layer. use_causal_conv (bool): Whether to use causal structure. """ super().__init__(**kwargs) self.aux_context_window = aux_context_window self.use_causal_conv = use_causal_conv and aux_context_window > 0 # To capture wide-context information in conditional features kernel_size = ( aux_context_window + 1 if use_causal_conv else 2 * aux_context_window + 1 ) self.conv_in = TFConv1d( filters=aux_channels, kernel_size=kernel_size, padding="same", use_bias=False, initializer_seed=initializer_seed, name="conv_in", ) self.upsample = TFUpsampleNetWork( output_channels=aux_channels, upsample_scales=upsample_scales, nonlinear_activation=nonlinear_activation, nonlinear_activation_params=nonlinear_activation_params, interpolate_mode=interpolate_mode, freq_axis_kernel_size=freq_axis_kernel_size, use_causal_conv=use_causal_conv, name="upsample_network", ) def call(self, c): """Calculate forward propagation. Args: c : Input tensor (B, T', C). Returns: Tensor: Upsampled tensor (B, T, C), where T = (T' - aux_context_window * 2) * prod(upsample_scales). Note: The length of inputs considers the context window size. """ c_ = self.conv_in(c) return self.upsample(c_) class TFParallelWaveGANGenerator(BaseModel): """Parallel WaveGAN Generator module.""" def __init__(self, config, **kwargs): super().__init__(**kwargs) self.out_channels = config.out_channels self.aux_channels = config.aux_channels self.n_layers = config.n_layers self.stacks = config.stacks self.kernel_size = config.kernel_size self.upsample_params = config.upsample_params # check the number of layers and stacks assert self.n_layers % self.stacks == 0 n_layers_per_stack = self.n_layers // self.stacks # define first convolution self.first_conv = TFConv1d1x1( filters=config.residual_channels, use_bias=True, padding="same", initializer_seed=config.initializer_seed, name="first_convolution", ) # define conv + upsampling network if config.upsample_conditional_features: self.upsample_params.update({"use_causal_conv": config.use_causal_conv}) self.upsample_params.update( { "aux_channels": config.aux_channels, "aux_context_window": config.aux_context_window, } ) self.upsample_net = TFConvInUpsampleNetWork(**self.upsample_params) else: self.upsample_net = None # define residual blocks self.conv_layers = [] for layer in range(self.n_layers): dilation_rate = 2 ** (layer % n_layers_per_stack) conv = TFResidualBlock( kernel_size=config.kernel_size, residual_channels=config.residual_channels, gate_channels=config.gate_channels, skip_channels=config.skip_channels, aux_channels=config.aux_channels, dilation_rate=dilation_rate, dropout_rate=config.dropout_rate, use_bias=config.use_bias, use_causal_conv=config.use_causal_conv, initializer_seed=config.initializer_seed, name="residual_block_._{}".format(layer), ) self.conv_layers += [conv] # define output layers self.last_conv_layers = [ tf.keras.layers.ReLU(), TFConv1d1x1( filters=config.skip_channels, use_bias=config.use_bias, padding="same", initializer_seed=config.initializer_seed, ), tf.keras.layers.ReLU(), TFConv1d1x1( filters=config.out_channels, use_bias=True, padding="same", initializer_seed=config.initializer_seed, ), tf.keras.layers.Activation("tanh"), ] def _build(self): mels = tf.random.uniform(shape=[2, 20, 80], dtype=tf.float32) self(mels, training=tf.cast(True, tf.bool)) def call(self, mels, training=False, **kwargs): """Calculate forward propagation. Args: mels (Tensor): Local conditioning auxiliary features (B, T', C). Returns: Tensor: Output tensor (B, T, 1) """ # perform upsampling if mels is not None and self.upsample_net is not None: c = self.upsample_net(mels) # random noise x # enccode to hidden representation x = tf.expand_dims(tf.random.normal(shape=tf.shape(c)[0:2]), axis=2) x = self.first_conv(x) skips = 0 for f in self.conv_layers: x, h = f(x, c, training=training) skips += h skips *= tf.math.sqrt(1.0 / len(self.conv_layers)) # apply final layers x = skips for f in self.last_conv_layers: x = f(x) return x @tf.function( experimental_relax_shapes=True, input_signature=[ tf.TensorSpec(shape=[None, None, 80], dtype=tf.float32, name="mels"), ], ) def inference(self, mels): """Calculate forward propagation. Args: c (Tensor): Local conditioning auxiliary features (B, T', C). Returns: Tensor: Output tensor (B, T, 1) """ # perform upsampling if mels is not None and self.upsample_net is not None: c = self.upsample_net(mels) # enccode to hidden representation x = tf.expand_dims(tf.random.normal(shape=tf.shape(c)[0:2]), axis=2) x = self.first_conv(x) skips = 0 for f in self.conv_layers: x, h = f(x, c, training=False) skips += h skips *= tf.math.sqrt(1.0 / len(self.conv_layers)) # apply final layers x = skips for f in self.last_conv_layers: x = f(x) return x class TFParallelWaveGANDiscriminator(BaseModel): """Parallel WaveGAN Discriminator module.""" def __init__(self, config, **kwargs): super().__init__(**kwargs) assert (config.kernel_size - 1) % 2 == 0, "Not support even number kernel size." assert config.dilation_factor > 0, "Dilation factor must be > 0." self.conv_layers = [] for i in range(config.n_layers - 1): if i == 0: dilation_rate = 1 else: dilation_rate = ( i if config.dilation_factor == 1 else config.dilation_factor ** i ) self.conv_layers += [ TFConv1d( filters=config.conv_channels, kernel_size=config.kernel_size, padding="same", dilation_rate=dilation_rate, use_bias=config.use_bias, initializer_seed=config.initializer_seed, ) ] self.conv_layers += [ getattr(tf.keras.layers, config.nonlinear_activation)( **config.nonlinear_activation_params ) ] self.conv_layers += [ TFConv1d( filters=config.out_channels, kernel_size=config.kernel_size, padding="same", use_bias=config.use_bias, initializer_seed=config.initializer_seed, ) ] if config.apply_sigmoid_at_last: self.conv_layers += [ tf.keras.layers.Activation("sigmoid"), ] def _build(self): x = tf.random.uniform(shape=[2, 16000, 1]) self(x) def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, T, 1). Returns: Tensor: Output tensor (B, T, 1) """ for f in self.conv_layers: x = f(x) return x
18,663
32.508079
112
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/melgan.py
# -*- coding: utf-8 -*- # Copyright 2020 The MelGAN Authors and Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """MelGAN Modules.""" import numpy as np import tensorflow as tf from tensorflow_tts.models import BaseModel from tensorflow_tts.utils import GroupConv1D, WeightNormalization def get_initializer(initializer_seed=42): """Creates a `tf.initializers.glorot_normal` with the given seed. Args: initializer_seed: int, initializer seed. Returns: GlorotNormal initializer with seed = `initializer_seed`. """ return tf.keras.initializers.GlorotNormal(seed=initializer_seed) class TFReflectionPad1d(tf.keras.layers.Layer): """Tensorflow ReflectionPad1d module.""" def __init__(self, padding_size, padding_type="REFLECT", **kwargs): """Initialize TFReflectionPad1d module. Args: padding_size (int) padding_type (str) ("CONSTANT", "REFLECT", or "SYMMETRIC". Default is "REFLECT") """ super().__init__(**kwargs) self.padding_size = padding_size self.padding_type = padding_type def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, T, C). Returns: Tensor: Padded tensor (B, T + 2 * padding_size, C). """ return tf.pad( x, [[0, 0], [self.padding_size, self.padding_size], [0, 0]], self.padding_type, ) class TFConvTranspose1d(tf.keras.layers.Layer): """Tensorflow ConvTranspose1d module.""" def __init__( self, filters, kernel_size, strides, padding, is_weight_norm, initializer_seed, **kwargs ): """Initialize TFConvTranspose1d( module. Args: filters (int): Number of filters. kernel_size (int): kernel size. strides (int): Stride width. padding (str): Padding type ("same" or "valid"). """ super().__init__(**kwargs) self.conv1d_transpose = tf.keras.layers.Conv2DTranspose( filters=filters, kernel_size=(kernel_size, 1), strides=(strides, 1), padding="same", kernel_initializer=get_initializer(initializer_seed), ) if is_weight_norm: self.conv1d_transpose = WeightNormalization(self.conv1d_transpose) def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, T, C). Returns: Tensor: Output tensor (B, T', C'). """ x = tf.expand_dims(x, 2) x = self.conv1d_transpose(x) x = tf.squeeze(x, 2) return x class TFResidualStack(tf.keras.layers.Layer): """Tensorflow ResidualStack module.""" def __init__( self, kernel_size, filters, dilation_rate, use_bias, nonlinear_activation, nonlinear_activation_params, is_weight_norm, initializer_seed, **kwargs ): """Initialize TFResidualStack module. Args: kernel_size (int): Kernel size. filters (int): Number of filters. dilation_rate (int): Dilation rate. use_bias (bool): Whether to add bias parameter in convolution layers. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. """ super().__init__(**kwargs) self.blocks = [ getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ), TFReflectionPad1d((kernel_size - 1) // 2 * dilation_rate), tf.keras.layers.Conv1D( filters=filters, kernel_size=kernel_size, dilation_rate=dilation_rate, use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), ), getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ), tf.keras.layers.Conv1D( filters=filters, kernel_size=1, use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), ), ] self.shortcut = tf.keras.layers.Conv1D( filters=filters, kernel_size=1, use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), name="shortcut", ) # apply weightnorm if is_weight_norm: self._apply_weightnorm(self.blocks) self.shortcut = WeightNormalization(self.shortcut) def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, T, C). Returns: Tensor: Output tensor (B, T, C). """ _x = tf.identity(x) for layer in self.blocks: _x = layer(_x) shortcut = self.shortcut(x) return shortcut + _x def _apply_weightnorm(self, list_layers): """Try apply weightnorm for all layer in list_layers.""" for i in range(len(list_layers)): try: layer_name = list_layers[i].name.lower() if "conv1d" in layer_name or "dense" in layer_name: list_layers[i] = WeightNormalization(list_layers[i]) except Exception: pass class TFMelGANGenerator(BaseModel): """Tensorflow MelGAN generator module.""" def __init__(self, config, **kwargs): """Initialize TFMelGANGenerator module. Args: config: config object of Melgan generator. """ super().__init__(**kwargs) # check hyper parameter is valid or not assert config.filters >= np.prod(config.upsample_scales) assert config.filters % (2 ** len(config.upsample_scales)) == 0 # add initial layer layers = [] layers += [ TFReflectionPad1d( (config.kernel_size - 1) // 2, padding_type=config.padding_type, name="first_reflect_padding", ), tf.keras.layers.Conv1D( filters=config.filters, kernel_size=config.kernel_size, use_bias=config.use_bias, kernel_initializer=get_initializer(config.initializer_seed), ), ] for i, upsample_scale in enumerate(config.upsample_scales): # add upsampling layer layers += [ getattr(tf.keras.layers, config.nonlinear_activation)( **config.nonlinear_activation_params ), TFConvTranspose1d( filters=config.filters // (2 ** (i + 1)), kernel_size=upsample_scale * 2, strides=upsample_scale, padding="same", is_weight_norm=config.is_weight_norm, initializer_seed=config.initializer_seed, name="conv_transpose_._{}".format(i), ), ] # ad residual stack layer for j in range(config.stacks): layers += [ TFResidualStack( kernel_size=config.stack_kernel_size, filters=config.filters // (2 ** (i + 1)), dilation_rate=config.stack_kernel_size ** j, use_bias=config.use_bias, nonlinear_activation=config.nonlinear_activation, nonlinear_activation_params=config.nonlinear_activation_params, is_weight_norm=config.is_weight_norm, initializer_seed=config.initializer_seed, name="residual_stack_._{}._._{}".format(i, j), ) ] # add final layer layers += [ getattr(tf.keras.layers, config.nonlinear_activation)( **config.nonlinear_activation_params ), TFReflectionPad1d( (config.kernel_size - 1) // 2, padding_type=config.padding_type, name="last_reflect_padding", ), tf.keras.layers.Conv1D( filters=config.out_channels, kernel_size=config.kernel_size, use_bias=config.use_bias, kernel_initializer=get_initializer(config.initializer_seed), dtype=tf.float32, ), ] if config.use_final_nolinear_activation: layers += [tf.keras.layers.Activation("tanh", dtype=tf.float32)] if config.is_weight_norm is True: self._apply_weightnorm(layers) self.melgan = tf.keras.models.Sequential(layers) def call(self, mels, **kwargs): """Calculate forward propagation. Args: c (Tensor): Input tensor (B, T, channels) Returns: Tensor: Output tensor (B, T ** prod(upsample_scales), out_channels) """ return self.inference(mels) @tf.function( input_signature=[ tf.TensorSpec(shape=[None, None, 80], dtype=tf.float32, name="mels") ] ) def inference(self, mels): return self.melgan(mels) @tf.function( input_signature=[ tf.TensorSpec(shape=[1, None, 80], dtype=tf.float32, name="mels") ] ) def inference_tflite(self, mels): return self.melgan(mels) def _apply_weightnorm(self, list_layers): """Try apply weightnorm for all layer in list_layers.""" for i in range(len(list_layers)): try: layer_name = list_layers[i].name.lower() if "conv1d" in layer_name or "dense" in layer_name: list_layers[i] = WeightNormalization(list_layers[i]) except Exception: pass def _build(self): """Build model by passing fake input.""" fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) self(fake_mels) class TFMelGANDiscriminator(tf.keras.layers.Layer): """Tensorflow MelGAN generator module.""" def __init__( self, out_channels=1, kernel_sizes=[5, 3], filters=16, max_downsample_filters=1024, use_bias=True, downsample_scales=[4, 4, 4, 4], nonlinear_activation="LeakyReLU", nonlinear_activation_params={"alpha": 0.2}, padding_type="REFLECT", is_weight_norm=True, initializer_seed=0.02, **kwargs ): """Initilize MelGAN discriminator module. Args: out_channels (int): Number of output channels. kernel_sizes (list): List of two kernel sizes. The prod will be used for the first conv layer, and the first and the second kernel sizes will be used for the last two layers. For example if kernel_sizes = [5, 3], the first layer kernel size will be 5 * 3 = 15. the last two layers' kernel size will be 5 and 3, respectively. filters (int): Initial number of filters for conv layer. max_downsample_filters (int): Maximum number of filters for downsampling layers. use_bias (bool): Whether to add bias parameter in convolution layers. downsample_scales (list): List of downsampling scales. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. padding_type (str): Padding type (support only "REFLECT", "CONSTANT", "SYMMETRIC") """ super().__init__(**kwargs) discriminator = [] # check kernel_size is valid assert len(kernel_sizes) == 2 assert kernel_sizes[0] % 2 == 1 assert kernel_sizes[1] % 2 == 1 # add first layer discriminator = [ TFReflectionPad1d( (np.prod(kernel_sizes) - 1) // 2, padding_type=padding_type ), tf.keras.layers.Conv1D( filters=filters, kernel_size=int(np.prod(kernel_sizes)), use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), ), getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ), ] # add downsample layers in_chs = filters with tf.keras.utils.CustomObjectScope({"GroupConv1D": GroupConv1D}): for downsample_scale in downsample_scales: out_chs = min(in_chs * downsample_scale, max_downsample_filters) discriminator += [ GroupConv1D( filters=out_chs, kernel_size=downsample_scale * 10 + 1, strides=downsample_scale, padding="same", use_bias=use_bias, groups=in_chs // 4, kernel_initializer=get_initializer(initializer_seed), ) ] discriminator += [ getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ) ] in_chs = out_chs # add final layers out_chs = min(in_chs * 2, max_downsample_filters) discriminator += [ tf.keras.layers.Conv1D( filters=out_chs, kernel_size=kernel_sizes[0], padding="same", use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), ) ] discriminator += [ getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ) ] discriminator += [ tf.keras.layers.Conv1D( filters=out_channels, kernel_size=kernel_sizes[1], padding="same", use_bias=use_bias, kernel_initializer=get_initializer(initializer_seed), ) ] if is_weight_norm is True: self._apply_weightnorm(discriminator) self.disciminator = discriminator def call(self, x, **kwargs): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, T, 1). Returns: List: List of output tensors of each layer. """ outs = [] for f in self.disciminator: x = f(x) outs += [x] return outs def _apply_weightnorm(self, list_layers): """Try apply weightnorm for all layer in list_layers.""" for i in range(len(list_layers)): try: layer_name = list_layers[i].name.lower() if "conv1d" in layer_name or "dense" in layer_name: list_layers[i] = WeightNormalization(list_layers[i]) except Exception: pass class TFMelGANMultiScaleDiscriminator(BaseModel): """MelGAN multi-scale discriminator module.""" def __init__(self, config, **kwargs): """Initilize MelGAN multi-scale discriminator module. Args: config: config object for melgan discriminator """ super().__init__(**kwargs) self.discriminator = [] # add discriminator for i in range(config.scales): self.discriminator += [ TFMelGANDiscriminator( out_channels=config.out_channels, kernel_sizes=config.kernel_sizes, filters=config.filters, max_downsample_filters=config.max_downsample_filters, use_bias=config.use_bias, downsample_scales=config.downsample_scales, nonlinear_activation=config.nonlinear_activation, nonlinear_activation_params=config.nonlinear_activation_params, padding_type=config.padding_type, is_weight_norm=config.is_weight_norm, initializer_seed=config.initializer_seed, name="melgan_discriminator_scale_._{}".format(i), ) ] self.pooling = getattr(tf.keras.layers, config.downsample_pooling)( **config.downsample_pooling_params ) def call(self, x, **kwargs): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, T, 1). Returns: List: List of list of each discriminator outputs, which consists of each layer output tensors. """ outs = [] for f in self.discriminator: outs += [f(x)] x = self.pooling(x) return outs
17,807
34.687375
106
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/tacotron2.py
# -*- coding: utf-8 -*- # Copyright 2020 The Tacotron-2 Authors, Minh Nguyen (@dathudeptrai), Eren Gölge (@erogol) and Jae Yoo (@jaeyoo) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tacotron-2 Modules.""" import collections import numpy as np import tensorflow as tf # TODO: once https://github.com/tensorflow/addons/pull/1964 is fixed, # uncomment this line. # from tensorflow_addons.seq2seq import dynamic_decode from tensorflow_addons.seq2seq import BahdanauAttention, Decoder, Sampler from tensorflow_tts.utils import dynamic_decode from tensorflow_tts.models import BaseModel def get_initializer(initializer_range=0.02): """Creates a `tf.initializers.truncated_normal` with the given range. Args: initializer_range: float, initializer range for stddev. Returns: TruncatedNormal initializer with stddev = `initializer_range`. """ return tf.keras.initializers.TruncatedNormal(stddev=initializer_range) def gelu(x): """Gaussian Error Linear unit.""" cdf = 0.5 * (1.0 + tf.math.erf(x / tf.math.sqrt(2.0))) return x * cdf def gelu_new(x): """Smoother gaussian Error Linear Unit.""" cdf = 0.5 * (1.0 + tf.tanh((np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3))))) return x * cdf def swish(x): """Swish activation function.""" return tf.nn.swish(x) def mish(x): return x * tf.math.tanh(tf.math.softplus(x)) ACT2FN = { "identity": tf.keras.layers.Activation("linear"), "tanh": tf.keras.layers.Activation("tanh"), "gelu": tf.keras.layers.Activation(gelu), "relu": tf.keras.activations.relu, "swish": tf.keras.layers.Activation(swish), "gelu_new": tf.keras.layers.Activation(gelu_new), "mish": tf.keras.layers.Activation(mish), } class TFEmbedding(tf.keras.layers.Embedding): """Faster version of embedding.""" def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) def call(self, inputs): inputs = tf.cast(tf.expand_dims(inputs, -1), tf.int32) outputs = tf.gather_nd(self.embeddings, inputs) return outputs class TFTacotronConvBatchNorm(tf.keras.layers.Layer): """Tacotron-2 Convolutional Batchnorm module.""" def __init__( self, filters, kernel_size, dropout_rate, activation=None, name_idx=None ): super().__init__() self.conv1d = tf.keras.layers.Conv1D( filters, kernel_size, kernel_initializer=get_initializer(0.02), padding="same", name="conv_._{}".format(name_idx), ) self.norm = tf.keras.layers.experimental.SyncBatchNormalization( axis=-1, name="batch_norm_._{}".format(name_idx) ) self.dropout = tf.keras.layers.Dropout( rate=dropout_rate, name="dropout_._{}".format(name_idx) ) self.act = ACT2FN[activation] def call(self, inputs, training=False): outputs = self.conv1d(inputs) outputs = self.norm(outputs, training=training) outputs = self.act(outputs) outputs = self.dropout(outputs, training=training) return outputs class TFTacotronEmbeddings(tf.keras.layers.Layer): """Construct character/phoneme/positional/speaker embeddings.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.vocab_size = config.vocab_size self.embedding_hidden_size = config.embedding_hidden_size self.initializer_range = config.initializer_range self.config = config if config.n_speakers > 1: self.speaker_embeddings = TFEmbedding( config.n_speakers, config.embedding_hidden_size, embeddings_initializer=get_initializer(self.initializer_range), name="speaker_embeddings", ) self.speaker_fc = tf.keras.layers.Dense( units=config.embedding_hidden_size, name="speaker_fc" ) self.LayerNorm = tf.keras.layers.LayerNormalization( epsilon=config.layer_norm_eps, name="LayerNorm" ) self.dropout = tf.keras.layers.Dropout(config.embedding_dropout_prob) def build(self, input_shape): """Build shared character/phoneme embedding layers.""" with tf.name_scope("character_embeddings"): self.character_embeddings = self.add_weight( "weight", shape=[self.vocab_size, self.embedding_hidden_size], initializer=get_initializer(self.initializer_range), ) super().build(input_shape) def call(self, inputs, training=False): """Get character embeddings of inputs. Args: 1. character, Tensor (int32) shape [batch_size, length]. 2. speaker_id, Tensor (int32) shape [batch_size] Returns: Tensor (float32) shape [batch_size, length, embedding_size]. """ return self._embedding(inputs, training=training) def _embedding(self, inputs, training=False): """Applies embedding based on inputs tensor.""" input_ids, speaker_ids = inputs # create embeddings inputs_embeds = tf.gather(self.character_embeddings, input_ids) embeddings = inputs_embeds if self.config.n_speakers > 1: speaker_embeddings = self.speaker_embeddings(speaker_ids) speaker_features = tf.math.softplus(self.speaker_fc(speaker_embeddings)) # extended speaker embeddings extended_speaker_features = speaker_features[:, tf.newaxis, :] # sum all embedding embeddings += extended_speaker_features # apply layer-norm and dropout for embeddings. embeddings = self.LayerNorm(embeddings) embeddings = self.dropout(embeddings, training=training) return embeddings class TFTacotronEncoderConvs(tf.keras.layers.Layer): """Tacotron-2 Encoder Convolutional Batchnorm module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.conv_batch_norm = [] for i in range(config.n_conv_encoder): conv = TFTacotronConvBatchNorm( filters=config.encoder_conv_filters, kernel_size=config.encoder_conv_kernel_sizes, activation=config.encoder_conv_activation, dropout_rate=config.encoder_conv_dropout_rate, name_idx=i, ) self.conv_batch_norm.append(conv) def call(self, inputs, training=False): """Call logic.""" outputs = inputs for conv in self.conv_batch_norm: outputs = conv(outputs, training=training) return outputs class TFTacotronEncoder(tf.keras.layers.Layer): """Tacotron-2 Encoder.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.embeddings = TFTacotronEmbeddings(config, name="embeddings") self.convbn = TFTacotronEncoderConvs(config, name="conv_batch_norm") self.bilstm = tf.keras.layers.Bidirectional( tf.keras.layers.LSTM( units=config.encoder_lstm_units, return_sequences=True ), name="bilstm", ) if config.n_speakers > 1: self.encoder_speaker_embeddings = TFEmbedding( config.n_speakers, config.embedding_hidden_size, embeddings_initializer=get_initializer(config.initializer_range), name="encoder_speaker_embeddings", ) self.encoder_speaker_fc = tf.keras.layers.Dense( units=config.encoder_lstm_units * 2, name="encoder_speaker_fc" ) self.config = config def call(self, inputs, training=False): """Call logic.""" input_ids, speaker_ids, input_mask = inputs # create embedding and mask them since we sum # speaker embedding to all character embedding. input_embeddings = self.embeddings([input_ids, speaker_ids], training=training) # pass embeddings to convolution batch norm conv_outputs = self.convbn(input_embeddings, training=training) # bi-lstm. outputs = self.bilstm(conv_outputs, mask=input_mask) if self.config.n_speakers > 1: encoder_speaker_embeddings = self.encoder_speaker_embeddings(speaker_ids) encoder_speaker_features = tf.math.softplus( self.encoder_speaker_fc(encoder_speaker_embeddings) ) # extended encoderspeaker embeddings extended_encoder_speaker_features = encoder_speaker_features[ :, tf.newaxis, : ] # sum to encoder outputs outputs += extended_encoder_speaker_features return outputs class Tacotron2Sampler(Sampler): """Tacotron2 sampler for Seq2Seq training.""" def __init__( self, config, ): super().__init__() self.config = config # create schedule factor. # the input of a next decoder cell is calculated by formular: # next_inputs = ratio * prev_groundtruth_outputs + (1.0 - ratio) * prev_predicted_outputs. self._ratio = tf.constant(1.0, dtype=tf.float32) self._reduction_factor = self.config.reduction_factor def setup_target(self, targets, mel_lengths): """Setup ground-truth mel outputs for decoder.""" self.mel_lengths = mel_lengths self.set_batch_size(tf.shape(targets)[0]) self.targets = targets[ :, self._reduction_factor - 1 :: self._reduction_factor, : ] self.max_lengths = tf.tile([tf.shape(self.targets)[1]], [self._batch_size]) @property def batch_size(self): return self._batch_size @property def sample_ids_shape(self): return tf.TensorShape([]) @property def sample_ids_dtype(self): return tf.int32 @property def reduction_factor(self): return self._reduction_factor def initialize(self): """Return (Finished, next_inputs).""" return ( tf.tile([False], [self._batch_size]), tf.tile([[0.0]], [self._batch_size, self.config.n_mels]), ) def sample(self, time, outputs, state): return tf.tile([0], [self._batch_size]) def next_inputs( self, time, outputs, state, sample_ids, stop_token_prediction, training=False, **kwargs, ): if training: finished = time + 1 >= self.max_lengths next_inputs = ( self._ratio * self.targets[:, time, :] + (1.0 - self._ratio) * outputs[:, -self.config.n_mels :] ) next_state = state return (finished, next_inputs, next_state) else: stop_token_prediction = tf.nn.sigmoid(stop_token_prediction) finished = tf.cast(tf.round(stop_token_prediction), tf.bool) finished = tf.reduce_all(finished) next_inputs = outputs[:, -self.config.n_mels :] next_state = state return (finished, next_inputs, next_state) def set_batch_size(self, batch_size): self._batch_size = batch_size class TFTacotronLocationSensitiveAttention(BahdanauAttention): """Tacotron-2 Location Sensitive Attention module.""" def __init__( self, config, memory, mask_encoder=True, memory_sequence_length=None, is_cumulate=True, ): """Init variables.""" memory_length = memory_sequence_length if (mask_encoder is True) else None super().__init__( units=config.attention_dim, memory=memory, memory_sequence_length=memory_length, probability_fn="softmax", name="LocationSensitiveAttention", ) self.location_convolution = tf.keras.layers.Conv1D( filters=config.attention_filters, kernel_size=config.attention_kernel, padding="same", use_bias=False, name="location_conv", ) self.location_layer = tf.keras.layers.Dense( units=config.attention_dim, use_bias=False, name="location_layer" ) self.v = tf.keras.layers.Dense(1, use_bias=True, name="scores_attention") self.config = config self.is_cumulate = is_cumulate self.use_window = False def setup_window(self, win_front=2, win_back=4): self.win_front = tf.constant(win_front, tf.int32) self.win_back = tf.constant(win_back, tf.int32) self._indices = tf.expand_dims(tf.range(tf.shape(self.keys)[1]), 0) self._indices = tf.tile( self._indices, [tf.shape(self.keys)[0], 1] ) # [batch_size, max_time] self.use_window = True def _compute_window_mask(self, max_alignments): """Compute window mask for inference. Args: max_alignments (int): [batch_size] """ expanded_max_alignments = tf.expand_dims(max_alignments, 1) # [batch_size, 1] low = expanded_max_alignments - self.win_front high = expanded_max_alignments + self.win_back mlow = tf.cast((self._indices < low), tf.float32) mhigh = tf.cast((self._indices > high), tf.float32) mask = mlow + mhigh return mask # [batch_size, max_length] def __call__(self, inputs, training=False): query, state, prev_max_alignments = inputs processed_query = self.query_layer(query) if self.query_layer else query processed_query = tf.expand_dims(processed_query, 1) expanded_alignments = tf.expand_dims(state, axis=2) f = self.location_convolution(expanded_alignments) processed_location_features = self.location_layer(f) energy = self._location_sensitive_score( processed_query, processed_location_features, self.keys ) # mask energy on inference steps. if self.use_window is True: window_mask = self._compute_window_mask(prev_max_alignments) energy = energy + window_mask * -1e20 alignments = self.probability_fn(energy, state) if self.is_cumulate: state = alignments + state else: state = alignments expanded_alignments = tf.expand_dims(alignments, 2) context = tf.reduce_sum(expanded_alignments * self.values, 1) return context, alignments, state def _location_sensitive_score(self, W_query, W_fil, W_keys): """Calculate location sensitive energy.""" return tf.squeeze(self.v(tf.nn.tanh(W_keys + W_query + W_fil)), -1) def get_initial_state(self, batch_size, size): """Get initial alignments.""" return tf.zeros(shape=[batch_size, size], dtype=tf.float32) def get_initial_context(self, batch_size): """Get initial attention.""" return tf.zeros( shape=[batch_size, self.config.encoder_lstm_units * 2], dtype=tf.float32 ) class TFTacotronPrenet(tf.keras.layers.Layer): """Tacotron-2 prenet.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.prenet_dense = [ tf.keras.layers.Dense( units=config.prenet_units, activation=ACT2FN[config.prenet_activation], name="dense_._{}".format(i), ) for i in range(config.n_prenet_layers) ] self.dropout = tf.keras.layers.Dropout( rate=config.prenet_dropout_rate, name="dropout" ) def call(self, inputs, training=False): """Call logic.""" outputs = inputs for layer in self.prenet_dense: outputs = layer(outputs) outputs = self.dropout(outputs, training=True) return outputs class TFTacotronPostnet(tf.keras.layers.Layer): """Tacotron-2 postnet.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.conv_batch_norm = [] for i in range(config.n_conv_postnet): conv = TFTacotronConvBatchNorm( filters=config.postnet_conv_filters, kernel_size=config.postnet_conv_kernel_sizes, dropout_rate=config.postnet_dropout_rate, activation="identity" if i + 1 == config.n_conv_postnet else "tanh", name_idx=i, ) self.conv_batch_norm.append(conv) def call(self, inputs, training=False): """Call logic.""" outputs = inputs for _, conv in enumerate(self.conv_batch_norm): outputs = conv(outputs, training=training) return outputs TFTacotronDecoderCellState = collections.namedtuple( "TFTacotronDecoderCellState", [ "attention_lstm_state", "decoder_lstms_state", "context", "time", "state", "alignment_history", "max_alignments", ], ) TFDecoderOutput = collections.namedtuple( "TFDecoderOutput", ("mel_output", "token_output", "sample_id") ) class TFTacotronDecoderCell(tf.keras.layers.AbstractRNNCell): """Tacotron-2 custom decoder cell.""" def __init__(self, config, enable_tflite_convertible=False, **kwargs): """Init variables.""" super().__init__(**kwargs) self.enable_tflite_convertible = enable_tflite_convertible self.prenet = TFTacotronPrenet(config, name="prenet") # define lstm cell on decoder. # TODO(@dathudeptrai) switch to zone-out lstm. self.attention_lstm = tf.keras.layers.LSTMCell( units=config.decoder_lstm_units, name="attention_lstm_cell" ) lstm_cells = [] for i in range(config.n_lstm_decoder): lstm_cell = tf.keras.layers.LSTMCell( units=config.decoder_lstm_units, name="lstm_cell_._{}".format(i) ) lstm_cells.append(lstm_cell) self.decoder_lstms = tf.keras.layers.StackedRNNCells( lstm_cells, name="decoder_lstms" ) # define attention layer. if config.attention_type == "lsa": # create location-sensitive attention. self.attention_layer = TFTacotronLocationSensitiveAttention( config, memory=None, mask_encoder=True, memory_sequence_length=None, is_cumulate=True, ) else: raise ValueError("Only lsa (location-sensitive attention) is supported") # frame, stop projection layer. self.frame_projection = tf.keras.layers.Dense( units=config.n_mels * config.reduction_factor, name="frame_projection" ) self.stop_projection = tf.keras.layers.Dense( units=config.reduction_factor, name="stop_projection" ) self.config = config def set_alignment_size(self, alignment_size): self.alignment_size = alignment_size @property def output_size(self): """Return output (mel) size.""" return self.frame_projection.units @property def state_size(self): """Return hidden state size.""" return TFTacotronDecoderCellState( attention_lstm_state=self.attention_lstm.state_size, decoder_lstms_state=self.decoder_lstms.state_size, time=tf.TensorShape([]), attention=self.config.attention_dim, state=self.alignment_size, alignment_history=(), max_alignments=tf.TensorShape([1]), ) def get_initial_state(self, batch_size): """Get initial states.""" initial_attention_lstm_cell_states = self.attention_lstm.get_initial_state( None, batch_size, dtype=tf.float32 ) initial_decoder_lstms_cell_states = self.decoder_lstms.get_initial_state( None, batch_size, dtype=tf.float32 ) initial_context = tf.zeros( shape=[batch_size, self.config.encoder_lstm_units * 2], dtype=tf.float32 ) initial_state = self.attention_layer.get_initial_state( batch_size, size=self.alignment_size ) if self.enable_tflite_convertible: initial_alignment_history = () else: initial_alignment_history = tf.TensorArray( dtype=tf.float32, size=0, dynamic_size=True ) return TFTacotronDecoderCellState( attention_lstm_state=initial_attention_lstm_cell_states, decoder_lstms_state=initial_decoder_lstms_cell_states, time=tf.zeros([], dtype=tf.int32), context=initial_context, state=initial_state, alignment_history=initial_alignment_history, max_alignments=tf.zeros([batch_size], dtype=tf.int32), ) def call(self, inputs, states, training=False): """Call logic.""" decoder_input = inputs # 1. apply prenet for decoder_input. prenet_out = self.prenet(decoder_input, training=training) # [batch_size, dim] # 2. concat prenet_out and prev context vector # then use it as input of attention lstm layer. attention_lstm_input = tf.concat([prenet_out, states.context], axis=-1) attention_lstm_output, next_attention_lstm_state = self.attention_lstm( attention_lstm_input, states.attention_lstm_state ) # 3. compute context, alignment and cumulative alignment. prev_state = states.state if not self.enable_tflite_convertible: prev_alignment_history = states.alignment_history prev_max_alignments = states.max_alignments context, alignments, state = self.attention_layer( [attention_lstm_output, prev_state, prev_max_alignments], training=training, ) # 4. run decoder lstm(s) decoder_lstms_input = tf.concat([attention_lstm_output, context], axis=-1) decoder_lstms_output, next_decoder_lstms_state = self.decoder_lstms( decoder_lstms_input, states.decoder_lstms_state ) # 5. compute frame feature and stop token. projection_inputs = tf.concat([decoder_lstms_output, context], axis=-1) decoder_outputs = self.frame_projection(projection_inputs) stop_inputs = tf.concat([decoder_lstms_output, decoder_outputs], axis=-1) stop_tokens = self.stop_projection(stop_inputs) # 6. save alignment history to visualize. if self.enable_tflite_convertible: alignment_history = () else: alignment_history = prev_alignment_history.write(states.time, alignments) # 7. return new states. new_states = TFTacotronDecoderCellState( attention_lstm_state=next_attention_lstm_state, decoder_lstms_state=next_decoder_lstms_state, time=states.time + 1, context=context, state=state, alignment_history=alignment_history, max_alignments=tf.argmax(alignments, -1, output_type=tf.int32), ) return (decoder_outputs, stop_tokens), new_states class TFTacotronDecoder(Decoder): """Tacotron-2 Decoder.""" def __init__( self, decoder_cell, decoder_sampler, output_layer=None, enable_tflite_convertible=False, ): """Initial variables.""" self.cell = decoder_cell self.sampler = decoder_sampler self.output_layer = output_layer self.enable_tflite_convertible = enable_tflite_convertible def setup_decoder_init_state(self, decoder_init_state): self.initial_state = decoder_init_state def initialize(self, **kwargs): return self.sampler.initialize() + (self.initial_state,) @property def output_size(self): return TFDecoderOutput( mel_output=tf.nest.map_structure( lambda shape: tf.TensorShape(shape), self.cell.output_size ), token_output=tf.TensorShape(self.sampler.reduction_factor), sample_id=tf.TensorShape([1]) if self.enable_tflite_convertible else self.sampler.sample_ids_shape, # tf.TensorShape([]) ) @property def output_dtype(self): return TFDecoderOutput(tf.float32, tf.float32, self.sampler.sample_ids_dtype) @property def batch_size(self): return self.sampler._batch_size def step(self, time, inputs, state, training=False): (mel_outputs, stop_tokens), cell_state = self.cell( inputs, state, training=training ) if self.output_layer is not None: mel_outputs = self.output_layer(mel_outputs) sample_ids = self.sampler.sample( time=time, outputs=mel_outputs, state=cell_state ) (finished, next_inputs, next_state) = self.sampler.next_inputs( time=time, outputs=mel_outputs, state=cell_state, sample_ids=sample_ids, stop_token_prediction=stop_tokens, training=training, ) outputs = TFDecoderOutput(mel_outputs, stop_tokens, sample_ids) return (outputs, next_state, next_inputs, finished) class TFTacotron2(BaseModel): """Tensorflow tacotron-2 model.""" def __init__(self, config, **kwargs): """Initalize tacotron-2 layers.""" enable_tflite_convertible = kwargs.pop("enable_tflite_convertible", False) super().__init__(self, **kwargs) self.encoder = TFTacotronEncoder(config, name="encoder") self.decoder_cell = TFTacotronDecoderCell( config, name="decoder_cell", enable_tflite_convertible=enable_tflite_convertible, ) self.decoder = TFTacotronDecoder( self.decoder_cell, Tacotron2Sampler(config), enable_tflite_convertible=enable_tflite_convertible, ) self.postnet = TFTacotronPostnet(config, name="post_net") self.post_projection = tf.keras.layers.Dense( units=config.n_mels, name="residual_projection" ) self.use_window_mask = False self.maximum_iterations = 4000 self.enable_tflite_convertible = enable_tflite_convertible self.config = config def setup_window(self, win_front, win_back): """Call only for inference.""" self.use_window_mask = True self.win_front = win_front self.win_back = win_back def setup_maximum_iterations(self, maximum_iterations): """Call only for inference.""" self.maximum_iterations = maximum_iterations def _build(self): input_ids = np.array([[1, 2, 3, 4, 5, 6, 7, 8, 9]]) input_lengths = np.array([9]) speaker_ids = np.array([0]) mel_outputs = np.random.normal(size=(1, 50, 80)).astype(np.float32) mel_lengths = np.array([50]) self( input_ids, input_lengths, speaker_ids, mel_outputs, mel_lengths, 10, training=True, ) def call( self, input_ids, input_lengths, speaker_ids, mel_gts, mel_lengths, maximum_iterations=None, use_window_mask=False, win_front=2, win_back=3, training=False, **kwargs, ): """Call logic.""" # create input-mask based on input_lengths input_mask = tf.sequence_mask( input_lengths, maxlen=tf.reduce_max(input_lengths), name="input_sequence_masks", ) # Encoder Step. encoder_hidden_states = self.encoder( [input_ids, speaker_ids, input_mask], training=training ) batch_size = tf.shape(encoder_hidden_states)[0] alignment_size = tf.shape(encoder_hidden_states)[1] # Setup some initial placeholders for decoder step. Include: # 1. mel_gts, mel_lengths for teacher forcing mode. # 2. alignment_size for attention size. # 3. initial state for decoder cell. # 4. memory (encoder hidden state) for attention mechanism. self.decoder.sampler.setup_target(targets=mel_gts, mel_lengths=mel_lengths) self.decoder.cell.set_alignment_size(alignment_size) self.decoder.setup_decoder_init_state( self.decoder.cell.get_initial_state(batch_size) ) self.decoder.cell.attention_layer.setup_memory( memory=encoder_hidden_states, memory_sequence_length=input_lengths, # use for mask attention. ) if use_window_mask: self.decoder.cell.attention_layer.setup_window( win_front=win_front, win_back=win_back ) # run decode step. ( (frames_prediction, stop_token_prediction, _), final_decoder_state, _, ) = dynamic_decode( self.decoder, maximum_iterations=maximum_iterations, enable_tflite_convertible=self.enable_tflite_convertible, training=training, ) decoder_outputs = tf.reshape( frames_prediction, [batch_size, -1, self.config.n_mels] ) stop_token_prediction = tf.reshape(stop_token_prediction, [batch_size, -1]) residual = self.postnet(decoder_outputs, training=training) residual_projection = self.post_projection(residual) mel_outputs = decoder_outputs + residual_projection if self.enable_tflite_convertible: mask = tf.math.not_equal( tf.cast( tf.reduce_sum(tf.abs(decoder_outputs), axis=-1), dtype=tf.int32 ), 0, ) decoder_outputs = tf.expand_dims( tf.boolean_mask(decoder_outputs, mask), axis=0 ) mel_outputs = tf.expand_dims(tf.boolean_mask(mel_outputs, mask), axis=0) alignment_history = () else: alignment_history = tf.transpose( final_decoder_state.alignment_history.stack(), [1, 2, 0] ) return decoder_outputs, mel_outputs, stop_token_prediction, alignment_history @tf.function( experimental_relax_shapes=True, input_signature=[ tf.TensorSpec([None, None], dtype=tf.int32, name="input_ids"), tf.TensorSpec([None,], dtype=tf.int32, name="input_lengths"), tf.TensorSpec([None,], dtype=tf.int32, name="speaker_ids"), ], ) def inference(self, input_ids, input_lengths, speaker_ids, **kwargs): """Call logic.""" # create input-mask based on input_lengths input_mask = tf.sequence_mask( input_lengths, maxlen=tf.reduce_max(input_lengths), name="input_sequence_masks", ) # Encoder Step. encoder_hidden_states = self.encoder( [input_ids, speaker_ids, input_mask], training=False ) batch_size = tf.shape(encoder_hidden_states)[0] alignment_size = tf.shape(encoder_hidden_states)[1] # Setup some initial placeholders for decoder step. Include: # 1. batch_size for inference. # 2. alignment_size for attention size. # 3. initial state for decoder cell. # 4. memory (encoder hidden state) for attention mechanism. # 5. window front/back to solve long sentence synthesize problems. (call after setup memory.) self.decoder.sampler.set_batch_size(batch_size) self.decoder.cell.set_alignment_size(alignment_size) self.decoder.setup_decoder_init_state( self.decoder.cell.get_initial_state(batch_size) ) self.decoder.cell.attention_layer.setup_memory( memory=encoder_hidden_states, memory_sequence_length=input_lengths, # use for mask attention. ) if self.use_window_mask: self.decoder.cell.attention_layer.setup_window( win_front=self.win_front, win_back=self.win_back ) # run decode step. ( (frames_prediction, stop_token_prediction, _), final_decoder_state, _, ) = dynamic_decode( self.decoder, maximum_iterations=self.maximum_iterations, training=False ) decoder_outputs = tf.reshape( frames_prediction, [batch_size, -1, self.config.n_mels] ) stop_token_predictions = tf.reshape(stop_token_prediction, [batch_size, -1]) residual = self.postnet(decoder_outputs, training=False) residual_projection = self.post_projection(residual) mel_outputs = decoder_outputs + residual_projection alignment_historys = tf.transpose( final_decoder_state.alignment_history.stack(), [1, 2, 0] ) return decoder_outputs, mel_outputs, stop_token_predictions, alignment_historys @tf.function( experimental_relax_shapes=True, input_signature=[ tf.TensorSpec([1, None], dtype=tf.int32, name="input_ids"), tf.TensorSpec([1,], dtype=tf.int32, name="input_lengths"), tf.TensorSpec([1,], dtype=tf.int32, name="speaker_ids"), ], ) def inference_tflite(self, input_ids, input_lengths, speaker_ids, **kwargs): """Call logic.""" # create input-mask based on input_lengths input_mask = tf.sequence_mask( input_lengths, maxlen=tf.reduce_max(input_lengths), name="input_sequence_masks", ) # Encoder Step. encoder_hidden_states = self.encoder( [input_ids, speaker_ids, input_mask], training=False ) batch_size = tf.shape(encoder_hidden_states)[0] alignment_size = tf.shape(encoder_hidden_states)[1] # Setup some initial placeholders for decoder step. Include: # 1. batch_size for inference. # 2. alignment_size for attention size. # 3. initial state for decoder cell. # 4. memory (encoder hidden state) for attention mechanism. # 5. window front/back to solve long sentence synthesize problems. (call after setup memory.) self.decoder.sampler.set_batch_size(batch_size) self.decoder.cell.set_alignment_size(alignment_size) self.decoder.setup_decoder_init_state( self.decoder.cell.get_initial_state(batch_size) ) self.decoder.cell.attention_layer.setup_memory( memory=encoder_hidden_states, memory_sequence_length=input_lengths, # use for mask attention. ) if self.use_window_mask: self.decoder.cell.attention_layer.setup_window( win_front=self.win_front, win_back=self.win_back ) # run decode step. ( (frames_prediction, stop_token_prediction, _), final_decoder_state, _, ) = dynamic_decode( self.decoder, maximum_iterations=self.maximum_iterations, enable_tflite_convertible=self.enable_tflite_convertible, training=False, ) decoder_outputs = tf.reshape( frames_prediction, [batch_size, -1, self.config.n_mels] ) stop_token_predictions = tf.reshape(stop_token_prediction, [batch_size, -1]) residual = self.postnet(decoder_outputs, training=False) residual_projection = self.post_projection(residual) mel_outputs = decoder_outputs + residual_projection if self.enable_tflite_convertible: mask = tf.math.not_equal( tf.cast( tf.reduce_sum(tf.abs(decoder_outputs), axis=-1), dtype=tf.int32 ), 0, ) decoder_outputs = tf.expand_dims( tf.boolean_mask(decoder_outputs, mask), axis=0 ) mel_outputs = tf.expand_dims(tf.boolean_mask(mel_outputs, mask), axis=0) alignment_historys = () else: alignment_historys = tf.transpose( final_decoder_state.alignment_history.stack(), [1, 2, 0] ) return decoder_outputs, mel_outputs, stop_token_predictions, alignment_historys
37,180
34.716619
112
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/mb_melgan.py
# -*- coding: utf-8 -*- # Copyright 2020 The Multi-band MelGAN Authors , Minh Nguyen (@dathudeptrai) and Tomoki Hayashi (@kan-bayashi) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================ # # Compatible with https://github.com/kan-bayashi/ParallelWaveGAN/blob/master/parallel_wavegan/layers/pqmf.py. """Multi-band MelGAN Modules.""" import numpy as np import tensorflow as tf from scipy.signal import kaiser from tensorflow_tts.models import BaseModel from tensorflow_tts.models import TFMelGANGenerator def design_prototype_filter(taps=62, cutoff_ratio=0.15, beta=9.0): """Design prototype filter for PQMF. This method is based on `A Kaiser window approach for the design of prototype filters of cosine modulated filterbanks`_. Args: taps (int): The number of filter taps. cutoff_ratio (float): Cut-off frequency ratio. beta (float): Beta coefficient for kaiser window. Returns: ndarray: Impluse response of prototype filter (taps + 1,). .. _`A Kaiser window approach for the design of prototype filters of cosine modulated filterbanks`: https://ieeexplore.ieee.org/abstract/document/681427 """ # check the arguments are valid assert taps % 2 == 0, "The number of taps mush be even number." assert 0.0 < cutoff_ratio < 1.0, "Cutoff ratio must be > 0.0 and < 1.0." # make initial filter omega_c = np.pi * cutoff_ratio with np.errstate(invalid="ignore"): h_i = np.sin(omega_c * (np.arange(taps + 1) - 0.5 * taps)) / ( np.pi * (np.arange(taps + 1) - 0.5 * taps) ) # fix nan due to indeterminate form h_i[taps // 2] = np.cos(0) * cutoff_ratio # apply kaiser window w = kaiser(taps + 1, beta) h = h_i * w return h class TFPQMF(tf.keras.layers.Layer): """PQMF module.""" def __init__(self, config, **kwargs): """Initilize PQMF module. Args: config (class): MultiBandMelGANGeneratorConfig """ super().__init__(**kwargs) subbands = config.subbands taps = config.taps cutoff_ratio = config.cutoff_ratio beta = config.beta # define filter coefficient h_proto = design_prototype_filter(taps, cutoff_ratio, beta) h_analysis = np.zeros((subbands, len(h_proto))) h_synthesis = np.zeros((subbands, len(h_proto))) for k in range(subbands): h_analysis[k] = ( 2 * h_proto * np.cos( (2 * k + 1) * (np.pi / (2 * subbands)) * (np.arange(taps + 1) - (taps / 2)) + (-1) ** k * np.pi / 4 ) ) h_synthesis[k] = ( 2 * h_proto * np.cos( (2 * k + 1) * (np.pi / (2 * subbands)) * (np.arange(taps + 1) - (taps / 2)) - (-1) ** k * np.pi / 4 ) ) # [subbands, 1, taps + 1] == [filter_width, in_channels, out_channels] analysis_filter = np.expand_dims(h_analysis, 1) analysis_filter = np.transpose(analysis_filter, (2, 1, 0)) synthesis_filter = np.expand_dims(h_synthesis, 0) synthesis_filter = np.transpose(synthesis_filter, (2, 1, 0)) # filter for downsampling & upsampling updown_filter = np.zeros((subbands, subbands, subbands), dtype=np.float32) for k in range(subbands): updown_filter[0, k, k] = 1.0 self.subbands = subbands self.taps = taps self.analysis_filter = analysis_filter.astype(np.float32) self.synthesis_filter = synthesis_filter.astype(np.float32) self.updown_filter = updown_filter.astype(np.float32) @tf.function( experimental_relax_shapes=True, input_signature=[tf.TensorSpec(shape=[None, None, 1], dtype=tf.float32)], ) def analysis(self, x): """Analysis with PQMF. Args: x (Tensor): Input tensor (B, T, 1). Returns: Tensor: Output tensor (B, T // subbands, subbands). """ x = tf.pad(x, [[0, 0], [self.taps // 2, self.taps // 2], [0, 0]]) x = tf.nn.conv1d(x, self.analysis_filter, stride=1, padding="VALID") x = tf.nn.conv1d(x, self.updown_filter, stride=self.subbands, padding="VALID") return x @tf.function( experimental_relax_shapes=True, input_signature=[tf.TensorSpec(shape=[None, None, None], dtype=tf.float32)], ) def synthesis(self, x): """Synthesis with PQMF. Args: x (Tensor): Input tensor (B, T // subbands, subbands). Returns: Tensor: Output tensor (B, T, 1). """ x = tf.nn.conv1d_transpose( x, self.updown_filter * self.subbands, strides=self.subbands, output_shape=( tf.shape(x)[0], tf.shape(x)[1] * self.subbands, self.subbands, ), ) x = tf.pad(x, [[0, 0], [self.taps // 2, self.taps // 2], [0, 0]]) return tf.nn.conv1d(x, self.synthesis_filter, stride=1, padding="VALID") class TFMBMelGANGenerator(TFMelGANGenerator): """Tensorflow MBMelGAN generator module.""" def __init__(self, config, **kwargs): super().__init__(config, **kwargs) self.pqmf = TFPQMF(config=config, dtype=tf.float32, name="pqmf") def call(self, mels, **kwargs): """Calculate forward propagation. Args: c (Tensor): Input tensor (B, T, channels) Returns: Tensor: Output tensor (B, T ** prod(upsample_scales), out_channels) """ return self.inference(mels) @tf.function( input_signature=[ tf.TensorSpec(shape=[None, None, 80], dtype=tf.float32, name="mels") ] ) def inference(self, mels): mb_audios = self.melgan(mels) return self.pqmf.synthesis(mb_audios) @tf.function( input_signature=[ tf.TensorSpec(shape=[1, None, 80], dtype=tf.float32, name="mels") ] ) def inference_tflite(self, mels): mb_audios = self.melgan(mels) return self.pqmf.synthesis(mb_audios)
6,890
34.704663
110
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/hifigan.py
# -*- coding: utf-8 -*- # Copyright 2020 The Hifigan Authors and TensorflowTTS Team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Hifi Modules.""" import numpy as np import tensorflow as tf from tensorflow_tts.models.melgan import TFReflectionPad1d from tensorflow_tts.models.melgan import TFConvTranspose1d from tensorflow_tts.utils import GroupConv1D from tensorflow_tts.utils import WeightNormalization from tensorflow_tts.models import BaseModel from tensorflow_tts.models import TFMelGANGenerator class TFHifiResBlock(tf.keras.layers.Layer): """Tensorflow Hifigan resblock 1 module.""" def __init__( self, kernel_size, filters, dilation_rate, use_bias, nonlinear_activation, nonlinear_activation_params, is_weight_norm, initializer_seed, **kwargs ): """Initialize TFHifiResBlock module. Args: kernel_size (int): Kernel size. filters (int): Number of filters. dilation_rate (list): List dilation rate. use_bias (bool): Whether to add bias parameter in convolution layers. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. is_weight_norm (bool): Whether to use weight norm or not. """ super().__init__(**kwargs) self.blocks_1 = [] self.blocks_2 = [] for i in range(len(dilation_rate)): self.blocks_1.append( [ TFReflectionPad1d((kernel_size - 1) // 2 * dilation_rate[i]), tf.keras.layers.Conv1D( filters=filters, kernel_size=kernel_size, dilation_rate=dilation_rate[i], use_bias=use_bias, ), ] ) self.blocks_2.append( [ TFReflectionPad1d((kernel_size - 1) // 2 * 1), tf.keras.layers.Conv1D( filters=filters, kernel_size=kernel_size, dilation_rate=1, use_bias=use_bias, ), ] ) self.activation = getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ) # apply weightnorm if is_weight_norm: self._apply_weightnorm(self.blocks_1) self._apply_weightnorm(self.blocks_2) def call(self, x, training=False): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, T, C). Returns: Tensor: Output tensor (B, T, C). """ for c1, c2 in zip(self.blocks_1, self.blocks_2): xt = self.activation(x) for c in c1: xt = c(xt) xt = self.activation(xt) for c in c2: xt = c(xt) x = xt + x return x def _apply_weightnorm(self, list_layers): """Try apply weightnorm for all layer in list_layers.""" for i in range(len(list_layers)): try: layer_name = list_layers[i].name.lower() if "conv1d" in layer_name or "dense" in layer_name: list_layers[i] = WeightNormalization(list_layers[i]) except Exception: pass class TFMultiHifiResBlock(tf.keras.layers.Layer): """Tensorflow Multi Hifigan resblock 1 module.""" def __init__(self, list_resblock, **kwargs): super().__init__(**kwargs) self.list_resblock = list_resblock def call(self, x, training=False): xs = None for resblock in self.list_resblock: if xs is None: xs = resblock(x, training=training) else: xs += resblock(x, training=training) return xs / len(self.list_resblock) class TFHifiGANGenerator(BaseModel): def __init__(self, config, **kwargs): super().__init__(**kwargs) # check hyper parameter is valid or not assert ( config.stacks == len(config.stack_kernel_size) == len(config.stack_dilation_rate) ) # add initial layer layers = [] layers += [ TFReflectionPad1d( (config.kernel_size - 1) // 2, padding_type=config.padding_type, name="first_reflect_padding", ), tf.keras.layers.Conv1D( filters=config.filters, kernel_size=config.kernel_size, use_bias=config.use_bias, ), ] for i, upsample_scale in enumerate(config.upsample_scales): # add upsampling layer layers += [ getattr(tf.keras.layers, config.nonlinear_activation)( **config.nonlinear_activation_params ), TFConvTranspose1d( filters=config.filters // (2 ** (i + 1)), kernel_size=upsample_scale * 2, strides=upsample_scale, padding="same", is_weight_norm=config.is_weight_norm, initializer_seed=config.initializer_seed, name="conv_transpose_._{}".format(i), ), ] # add residual stack layer layers += [ TFMultiHifiResBlock( list_resblock=[ TFHifiResBlock( kernel_size=config.stack_kernel_size[j], filters=config.filters // (2 ** (i + 1)), dilation_rate=config.stack_dilation_rate[j], use_bias=config.use_bias, nonlinear_activation=config.nonlinear_activation, nonlinear_activation_params=config.nonlinear_activation_params, is_weight_norm=config.is_weight_norm, initializer_seed=config.initializer_seed, name="hifigan_resblock_._{}".format(j), ) for j in range(config.stacks) ], name="multi_hifigan_resblock_._{}".format(i), ) ] # add final layer layers += [ getattr(tf.keras.layers, config.nonlinear_activation)( **config.nonlinear_activation_params ), TFReflectionPad1d( (config.kernel_size - 1) // 2, padding_type=config.padding_type, name="last_reflect_padding", ), tf.keras.layers.Conv1D( filters=config.out_channels, kernel_size=config.kernel_size, use_bias=config.use_bias, dtype=tf.float32, ), ] if config.use_final_nolinear_activation: layers += [tf.keras.layers.Activation("tanh", dtype=tf.float32)] if config.is_weight_norm is True: self._apply_weightnorm(layers) self.hifigan = tf.keras.models.Sequential(layers) def call(self, mels, **kwargs): """Calculate forward propagation. Args: c (Tensor): Input tensor (B, T, channels) Returns: Tensor: Output tensor (B, T ** prod(upsample_scales), out_channels) """ return self.inference(mels) @tf.function( input_signature=[ tf.TensorSpec(shape=[None, None, 80], dtype=tf.float32, name="mels") ] ) def inference(self, mels): return self.hifigan(mels) @tf.function( input_signature=[ tf.TensorSpec(shape=[1, None, 80], dtype=tf.float32, name="mels") ] ) def inference_tflite(self, mels): return self.hifigan(mels) def _apply_weightnorm(self, list_layers): """Try apply weightnorm for all layer in list_layers.""" for i in range(len(list_layers)): try: layer_name = list_layers[i].name.lower() if "conv1d" in layer_name or "dense" in layer_name: list_layers[i] = WeightNormalization(list_layers[i]) except Exception: pass def _build(self): """Build model by passing fake input.""" fake_mels = tf.random.uniform(shape=[1, 100, 80], dtype=tf.float32) self(fake_mels) class TFHifiGANPeriodDiscriminator(tf.keras.layers.Layer): """Tensorflow Hifigan period discriminator module.""" def __init__( self, period, out_channels=1, n_layers=5, kernel_size=5, strides=3, filters=8, filter_scales=4, max_filters=1024, nonlinear_activation="LeakyReLU", nonlinear_activation_params={"alpha": 0.2}, initializer_seed=42, is_weight_norm=False, **kwargs ): super().__init__(**kwargs) self.period = period self.out_filters = out_channels self.convs = [] for i in range(n_layers): self.convs.append( tf.keras.layers.Conv2D( filters=min(filters * (filter_scales ** (i + 1)), max_filters), kernel_size=(kernel_size, 1), strides=(strides, 1), padding="same", ) ) self.conv_post = tf.keras.layers.Conv2D( filters=out_channels, kernel_size=(3, 1), padding="same", ) self.activation = getattr(tf.keras.layers, nonlinear_activation)( **nonlinear_activation_params ) if is_weight_norm: self._apply_weightnorm(self.convs) self.conv_post = WeightNormalization(self.conv_post) def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, T, 1). Returns: List: List of output tensors. """ shape = tf.shape(x) n_pad = tf.convert_to_tensor(0, dtype=tf.int32) if shape[1] % self.period != 0: n_pad = self.period - (shape[1] % self.period) x = tf.pad(x, [[0, 0], [0, n_pad], [0, 0]], "REFLECT") x = tf.reshape( x, [shape[0], (shape[1] + n_pad) // self.period, self.period, x.shape[2]] ) for layer in self.convs: x = layer(x) x = self.activation(x) x = self.conv_post(x) x = tf.reshape(x, [shape[0], -1, self.out_filters]) return [x] def _apply_weightnorm(self, list_layers): """Try apply weightnorm for all layer in list_layers.""" for i in range(len(list_layers)): try: layer_name = list_layers[i].name.lower() if "conv1d" in layer_name or "dense" in layer_name: list_layers[i] = WeightNormalization(list_layers[i]) except Exception: pass class TFHifiGANMultiPeriodDiscriminator(BaseModel): """Tensorflow Hifigan Multi Period discriminator module.""" def __init__(self, config, **kwargs): super().__init__(**kwargs) self.discriminator = [] # add discriminator for i in range(len(config.period_scales)): self.discriminator += [ TFHifiGANPeriodDiscriminator( config.period_scales[i], out_channels=config.out_channels, n_layers=config.n_layers, kernel_size=config.kernel_size, strides=config.strides, filters=config.filters, filter_scales=config.filter_scales, max_filters=config.max_filters, nonlinear_activation=config.nonlinear_activation, nonlinear_activation_params=config.nonlinear_activation_params, initializer_seed=config.initializer_seed, is_weight_norm=config.is_weight_norm, name="hifigan_period_discriminator_._{}".format(i), ) ] def call(self, x): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, T, 1). Returns: List: list of each discriminator outputs """ outs = [] for f in self.discriminator: outs += [f(x)] return outs
13,272
33.928947
91
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/fastspeech2.py
# -*- coding: utf-8 -*- # Copyright 2020 The FastSpeech2 Authors and Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tensorflow Model modules for FastSpeech2.""" import tensorflow as tf from tensorflow_tts.models.fastspeech import TFFastSpeech, get_initializer class TFFastSpeechVariantPredictor(tf.keras.layers.Layer): """FastSpeech duration predictor module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.conv_layers = [] for i in range(config.variant_prediction_num_conv_layers): self.conv_layers.append( tf.keras.layers.Conv1D( config.variant_predictor_filter, config.variant_predictor_kernel_size, padding="same", name="conv_._{}".format(i), ) ) self.conv_layers.append(tf.keras.layers.Activation(tf.nn.relu)) self.conv_layers.append( tf.keras.layers.LayerNormalization( epsilon=config.layer_norm_eps, name="LayerNorm_._{}".format(i) ) ) self.conv_layers.append( tf.keras.layers.Dropout(config.variant_predictor_dropout_rate) ) self.conv_layers_sequence = tf.keras.Sequential(self.conv_layers) self.output_layer = tf.keras.layers.Dense(1) if config.n_speakers > 1: self.decoder_speaker_embeddings = tf.keras.layers.Embedding( config.n_speakers, config.encoder_self_attention_params.hidden_size, embeddings_initializer=get_initializer(config.initializer_range), name="speaker_embeddings", ) self.speaker_fc = tf.keras.layers.Dense( units=config.encoder_self_attention_params.hidden_size, name="speaker_fc", ) self.config = config def call(self, inputs, training=False): """Call logic.""" encoder_hidden_states, speaker_ids, attention_mask = inputs attention_mask = tf.cast( tf.expand_dims(attention_mask, 2), encoder_hidden_states.dtype ) if self.config.n_speakers > 1: speaker_embeddings = self.decoder_speaker_embeddings(speaker_ids) speaker_features = tf.math.softplus(self.speaker_fc(speaker_embeddings)) # extended speaker embeddings extended_speaker_features = speaker_features[:, tf.newaxis, :] encoder_hidden_states += extended_speaker_features # mask encoder hidden states masked_encoder_hidden_states = encoder_hidden_states * attention_mask # pass though first layer outputs = self.conv_layers_sequence(masked_encoder_hidden_states) outputs = self.output_layer(outputs) masked_outputs = outputs * attention_mask outputs = tf.squeeze(masked_outputs, -1) return outputs class TFFastSpeech2(TFFastSpeech): """TF Fastspeech module.""" def __init__(self, config, **kwargs): """Init layers for fastspeech.""" super().__init__(config, **kwargs) self.f0_predictor = TFFastSpeechVariantPredictor( config, dtype=tf.float32, name="f0_predictor" ) self.energy_predictor = TFFastSpeechVariantPredictor( config, dtype=tf.float32, name="energy_predictor", ) self.duration_predictor = TFFastSpeechVariantPredictor( config, dtype=tf.float32, name="duration_predictor" ) # define f0_embeddings and energy_embeddings self.f0_embeddings = tf.keras.layers.Conv1D( filters=config.encoder_self_attention_params.hidden_size, kernel_size=9, padding="same", name="f0_embeddings", ) self.f0_dropout = tf.keras.layers.Dropout(0.5) self.energy_embeddings = tf.keras.layers.Conv1D( filters=config.encoder_self_attention_params.hidden_size, kernel_size=9, padding="same", name="energy_embeddings", ) self.energy_dropout = tf.keras.layers.Dropout(0.5) def _build(self): """Dummy input for building model.""" # fake inputs input_ids = tf.convert_to_tensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]], tf.int32) speaker_ids = tf.convert_to_tensor([0], tf.int32) duration_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32) f0_gts = tf.convert_to_tensor( [[10, 10, 10, 10, 10, 10, 10, 10, 10, 10]], tf.float32 ) energy_gts = tf.convert_to_tensor( [[10, 10, 10, 10, 10, 10, 10, 10, 10, 10]], tf.float32 ) self( input_ids=input_ids, speaker_ids=speaker_ids, duration_gts=duration_gts, f0_gts=f0_gts, energy_gts=energy_gts, ) def call( self, input_ids, speaker_ids, duration_gts, f0_gts, energy_gts, training=False, **kwargs, ): """Call logic.""" attention_mask = tf.math.not_equal(input_ids, 0) embedding_output = self.embeddings([input_ids, speaker_ids], training=training) encoder_output = self.encoder( [embedding_output, attention_mask], training=training ) last_encoder_hidden_states = encoder_output[0] # energy predictor, here use last_encoder_hidden_states, u can use more hidden_states layers # rather than just use last_hidden_states of encoder for energy_predictor. duration_outputs = self.duration_predictor( [last_encoder_hidden_states, speaker_ids, attention_mask] ) # [batch_size, length] f0_outputs = self.f0_predictor( [last_encoder_hidden_states, speaker_ids, attention_mask], training=training ) energy_outputs = self.energy_predictor( [last_encoder_hidden_states, speaker_ids, attention_mask], training=training ) f0_embedding = self.f0_embeddings( tf.expand_dims(f0_gts, 2) ) # [barch_size, mel_length, feature] energy_embedding = self.energy_embeddings( tf.expand_dims(energy_gts, 2) ) # [barch_size, mel_length, feature] # apply dropout both training/inference f0_embedding = self.f0_dropout(f0_embedding, training=True) energy_embedding = self.energy_dropout(energy_embedding, training=True) # sum features last_encoder_hidden_states += f0_embedding + energy_embedding length_regulator_outputs, encoder_masks = self.length_regulator( [last_encoder_hidden_states, duration_gts], training=training ) # create decoder positional embedding decoder_pos = tf.range( 1, tf.shape(length_regulator_outputs)[1] + 1, dtype=tf.int32 ) masked_decoder_pos = tf.expand_dims(decoder_pos, 0) * encoder_masks decoder_output = self.decoder( [length_regulator_outputs, speaker_ids, encoder_masks, masked_decoder_pos], training=training, ) last_decoder_hidden_states = decoder_output[0] # here u can use sum or concat more than 1 hidden states layers from decoder. mels_before = self.mel_dense(last_decoder_hidden_states) mels_after = ( self.postnet([mels_before, encoder_masks], training=training) + mels_before ) outputs = ( mels_before, mels_after, duration_outputs, f0_outputs, energy_outputs, ) return outputs def _inference( self, input_ids, speaker_ids, speed_ratios, f0_ratios, energy_ratios, **kwargs, ): """Call logic.""" attention_mask = tf.math.not_equal(input_ids, 0) embedding_output = self.embeddings([input_ids, speaker_ids], training=False) encoder_output = self.encoder( [embedding_output, attention_mask], training=False ) last_encoder_hidden_states = encoder_output[0] # expand ratios speed_ratios = tf.expand_dims(speed_ratios, 1) # [B, 1] f0_ratios = tf.expand_dims(f0_ratios, 1) # [B, 1] energy_ratios = tf.expand_dims(energy_ratios, 1) # [B, 1] # energy predictor, here use last_encoder_hidden_states, u can use more hidden_states layers # rather than just use last_hidden_states of encoder for energy_predictor. duration_outputs = self.duration_predictor( [last_encoder_hidden_states, speaker_ids, attention_mask] ) # [batch_size, length] duration_outputs = tf.nn.relu(tf.math.exp(duration_outputs) - 1.0) duration_outputs = tf.cast( tf.math.round(duration_outputs * speed_ratios), tf.int32 ) f0_outputs = self.f0_predictor( [last_encoder_hidden_states, speaker_ids, attention_mask], training=False ) f0_outputs *= f0_ratios energy_outputs = self.energy_predictor( [last_encoder_hidden_states, speaker_ids, attention_mask], training=False ) energy_outputs *= energy_ratios f0_embedding = self.f0_dropout( self.f0_embeddings(tf.expand_dims(f0_outputs, 2)), training=True ) energy_embedding = self.energy_dropout( self.energy_embeddings(tf.expand_dims(energy_outputs, 2)), training=True ) # sum features last_encoder_hidden_states += f0_embedding + energy_embedding length_regulator_outputs, encoder_masks = self.length_regulator( [last_encoder_hidden_states, duration_outputs], training=False ) # create decoder positional embedding decoder_pos = tf.range( 1, tf.shape(length_regulator_outputs)[1] + 1, dtype=tf.int32 ) masked_decoder_pos = tf.expand_dims(decoder_pos, 0) * encoder_masks decoder_output = self.decoder( [length_regulator_outputs, speaker_ids, encoder_masks, masked_decoder_pos], training=False, ) last_decoder_hidden_states = decoder_output[0] # here u can use sum or concat more than 1 hidden states layers from decoder. mel_before = self.mel_dense(last_decoder_hidden_states) mel_after = ( self.postnet([mel_before, encoder_masks], training=False) + mel_before ) outputs = (mel_before, mel_after, duration_outputs, f0_outputs, energy_outputs) return outputs def setup_inference_fn(self): self.inference = tf.function( self._inference, experimental_relax_shapes=True, input_signature=[ tf.TensorSpec(shape=[None, None], dtype=tf.int32, name="input_ids"), tf.TensorSpec(shape=[None,], dtype=tf.int32, name="speaker_ids"), tf.TensorSpec(shape=[None,], dtype=tf.float32, name="speed_ratios"), tf.TensorSpec(shape=[None,], dtype=tf.float32, name="f0_ratios"), tf.TensorSpec(shape=[None,], dtype=tf.float32, name="energy_ratios"), ], ) self.inference_tflite = tf.function( self._inference, experimental_relax_shapes=True, input_signature=[ tf.TensorSpec(shape=[1, None], dtype=tf.int32, name="input_ids"), tf.TensorSpec(shape=[1,], dtype=tf.int32, name="speaker_ids"), tf.TensorSpec(shape=[1,], dtype=tf.float32, name="speed_ratios"), tf.TensorSpec(shape=[1,], dtype=tf.float32, name="f0_ratios"), tf.TensorSpec(shape=[1,], dtype=tf.float32, name="energy_ratios"), ], )
12,399
38.616613
100
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/models/fastspeech.py
# -*- coding: utf-8 -*- # Copyright 2020 The FastSpeech Authors, The HuggingFace Inc. team and Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tensorflow Model modules for FastSpeech.""" import numpy as np import tensorflow as tf from tensorflow_tts.models import BaseModel def get_initializer(initializer_range=0.02): """Creates a `tf.initializers.truncated_normal` with the given range. Args: initializer_range: float, initializer range for stddev. Returns: TruncatedNormal initializer with stddev = `initializer_range`. """ return tf.keras.initializers.TruncatedNormal(stddev=initializer_range) def gelu(x): """Gaussian Error Linear unit.""" cdf = 0.5 * (1.0 + tf.math.erf(x / tf.math.sqrt(2.0))) return x * cdf def gelu_new(x): """Smoother gaussian Error Linear Unit.""" cdf = 0.5 * (1.0 + tf.tanh((np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3))))) return x * cdf def swish(x): """Swish activation function.""" return tf.nn.swish(x) def mish(x): return x * tf.math.tanh(tf.math.softplus(x)) ACT2FN = { "identity": tf.keras.layers.Activation("linear"), "tanh": tf.keras.layers.Activation("tanh"), "gelu": tf.keras.layers.Activation(gelu), "relu": tf.keras.activations.relu, "swish": tf.keras.layers.Activation(swish), "gelu_new": tf.keras.layers.Activation(gelu_new), "mish": tf.keras.layers.Activation(mish), } class TFEmbedding(tf.keras.layers.Embedding): """Faster version of embedding.""" def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) def call(self, inputs): inputs = tf.cast(inputs, tf.int32) outputs = tf.gather(self.embeddings, inputs) return outputs class TFFastSpeechEmbeddings(tf.keras.layers.Layer): """Construct charactor/phoneme/positional/speaker embeddings.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.vocab_size = config.vocab_size self.hidden_size = config.encoder_self_attention_params.hidden_size self.initializer_range = config.initializer_range self.config = config self.position_embeddings = TFEmbedding( config.max_position_embeddings + 1, self.hidden_size, weights=[ self._sincos_embedding( self.hidden_size, self.config.max_position_embeddings ) ], name="position_embeddings", trainable=False, ) if config.n_speakers > 1: self.encoder_speaker_embeddings = TFEmbedding( config.n_speakers, self.hidden_size, embeddings_initializer=get_initializer(self.initializer_range), name="speaker_embeddings", ) self.speaker_fc = tf.keras.layers.Dense( units=self.hidden_size, name="speaker_fc" ) def build(self, input_shape): """Build shared charactor/phoneme embedding layers.""" with tf.name_scope("charactor_embeddings"): self.charactor_embeddings = self.add_weight( "weight", shape=[self.vocab_size, self.hidden_size], initializer=get_initializer(self.initializer_range), ) super().build(input_shape) def call(self, inputs, training=False): """Get charactor embeddings of inputs. Args: 1. charactor, Tensor (int32) shape [batch_size, length]. 2. speaker_id, Tensor (int32) shape [batch_size] Returns: Tensor (float32) shape [batch_size, length, embedding_size]. """ return self._embedding(inputs, training=training) def _embedding(self, inputs, training=False): """Applies embedding based on inputs tensor.""" input_ids, speaker_ids = inputs input_shape = tf.shape(input_ids) seq_length = input_shape[1] position_ids = tf.range(1, seq_length + 1, dtype=tf.int32)[tf.newaxis, :] # create embeddings inputs_embeds = tf.gather(self.charactor_embeddings, input_ids) position_embeddings = self.position_embeddings(position_ids) # sum embedding embeddings = inputs_embeds + tf.cast(position_embeddings, inputs_embeds.dtype) if self.config.n_speakers > 1: speaker_embeddings = self.encoder_speaker_embeddings(speaker_ids) speaker_features = tf.math.softplus(self.speaker_fc(speaker_embeddings)) # extended speaker embeddings extended_speaker_features = speaker_features[:, tf.newaxis, :] embeddings += extended_speaker_features return embeddings def _sincos_embedding( self, hidden_size, max_positional_embedding, ): position_enc = np.array( [ [ pos / np.power(10000, 2.0 * (i // 2) / hidden_size) for i in range(hidden_size) ] for pos in range(max_positional_embedding + 1) ] ) position_enc[:, 0::2] = np.sin(position_enc[:, 0::2]) position_enc[:, 1::2] = np.cos(position_enc[:, 1::2]) # pad embedding. position_enc[0] = 0.0 return position_enc def resize_positional_embeddings(self, new_size): self.position_embeddings = TFEmbedding( new_size + 1, self.hidden_size, weights=[self._sincos_embedding(self.hidden_size, new_size)], name="position_embeddings", trainable=False, ) class TFFastSpeechSelfAttention(tf.keras.layers.Layer): """Self attention module for fastspeech.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) if config.hidden_size % config.num_attention_heads != 0: raise ValueError( "The hidden size (%d) is not a multiple of the number of attention " "heads (%d)" % (config.hidden_size, config.num_attention_heads) ) self.output_attentions = config.output_attentions self.num_attention_heads = config.num_attention_heads self.all_head_size = self.num_attention_heads * config.attention_head_size self.query = tf.keras.layers.Dense( self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query", ) self.key = tf.keras.layers.Dense( self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key", ) self.value = tf.keras.layers.Dense( self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value", ) self.dropout = tf.keras.layers.Dropout(config.attention_probs_dropout_prob) self.config = config def transpose_for_scores(self, x, batch_size): """Transpose to calculate attention scores.""" x = tf.reshape( x, (batch_size, -1, self.num_attention_heads, self.config.attention_head_size), ) return tf.transpose(x, perm=[0, 2, 1, 3]) def call(self, inputs, training=False): """Call logic.""" hidden_states, attention_mask = inputs batch_size = tf.shape(hidden_states)[0] mixed_query_layer = self.query(hidden_states) mixed_key_layer = self.key(hidden_states) mixed_value_layer = self.value(hidden_states) query_layer = self.transpose_for_scores(mixed_query_layer, batch_size) key_layer = self.transpose_for_scores(mixed_key_layer, batch_size) value_layer = self.transpose_for_scores(mixed_value_layer, batch_size) attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True) dk = tf.cast( tf.shape(key_layer)[-1], attention_scores.dtype ) # scale attention_scores attention_scores = attention_scores / tf.math.sqrt(dk) if attention_mask is not None: # extended_attention_masks for self attention encoder. extended_attention_mask = attention_mask[:, tf.newaxis, tf.newaxis, :] extended_attention_mask = tf.cast( extended_attention_mask, attention_scores.dtype ) extended_attention_mask = (1.0 - extended_attention_mask) * -1e9 attention_scores = attention_scores + extended_attention_mask # Normalize the attention scores to probabilities. attention_probs = tf.nn.softmax(attention_scores, axis=-1) attention_probs = self.dropout(attention_probs, training=training) context_layer = tf.matmul(attention_probs, value_layer) context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3]) context_layer = tf.reshape(context_layer, (batch_size, -1, self.all_head_size)) outputs = ( (context_layer, attention_probs) if self.output_attentions else (context_layer,) ) return outputs class TFFastSpeechSelfOutput(tf.keras.layers.Layer): """Fastspeech output of self attention module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.dense = tf.keras.layers.Dense( config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense", ) self.LayerNorm = tf.keras.layers.LayerNormalization( epsilon=config.layer_norm_eps, name="LayerNorm" ) self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob) def call(self, inputs, training=False): """Call logic.""" hidden_states, input_tensor = inputs hidden_states = self.dense(hidden_states) hidden_states = self.dropout(hidden_states, training=training) hidden_states = self.LayerNorm(hidden_states + input_tensor) return hidden_states class TFFastSpeechAttention(tf.keras.layers.Layer): """Fastspeech attention module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.self_attention = TFFastSpeechSelfAttention(config, name="self") self.dense_output = TFFastSpeechSelfOutput(config, name="output") def call(self, inputs, training=False): input_tensor, attention_mask = inputs self_outputs = self.self_attention( [input_tensor, attention_mask], training=training ) attention_output = self.dense_output( [self_outputs[0], input_tensor], training=training ) masked_attention_output = attention_output * tf.cast( tf.expand_dims(attention_mask, 2), dtype=attention_output.dtype ) outputs = (masked_attention_output,) + self_outputs[ 1: ] # add attentions if we output them return outputs class TFFastSpeechIntermediate(tf.keras.layers.Layer): """Intermediate representation module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.conv1d_1 = tf.keras.layers.Conv1D( config.intermediate_size, kernel_size=config.intermediate_kernel_size, kernel_initializer=get_initializer(config.initializer_range), padding="same", name="conv1d_1", ) self.conv1d_2 = tf.keras.layers.Conv1D( config.hidden_size, kernel_size=config.intermediate_kernel_size, kernel_initializer=get_initializer(config.initializer_range), padding="same", name="conv1d_2", ) if isinstance(config.hidden_act, str): self.intermediate_act_fn = ACT2FN[config.hidden_act] else: self.intermediate_act_fn = config.hidden_act def call(self, inputs): """Call logic.""" hidden_states, attention_mask = inputs hidden_states = self.conv1d_1(hidden_states) hidden_states = self.intermediate_act_fn(hidden_states) hidden_states = self.conv1d_2(hidden_states) masked_hidden_states = hidden_states * tf.cast( tf.expand_dims(attention_mask, 2), dtype=hidden_states.dtype ) return masked_hidden_states class TFFastSpeechOutput(tf.keras.layers.Layer): """Output module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.LayerNorm = tf.keras.layers.LayerNormalization( epsilon=config.layer_norm_eps, name="LayerNorm" ) self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob) def call(self, inputs, training=False): """Call logic.""" hidden_states, input_tensor = inputs hidden_states = self.dropout(hidden_states, training=training) hidden_states = self.LayerNorm(hidden_states + input_tensor) return hidden_states class TFFastSpeechLayer(tf.keras.layers.Layer): """Fastspeech module (FFT module on the paper).""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.attention = TFFastSpeechAttention(config, name="attention") self.intermediate = TFFastSpeechIntermediate(config, name="intermediate") self.bert_output = TFFastSpeechOutput(config, name="output") def call(self, inputs, training=False): """Call logic.""" hidden_states, attention_mask = inputs attention_outputs = self.attention( [hidden_states, attention_mask], training=training ) attention_output = attention_outputs[0] intermediate_output = self.intermediate( [attention_output, attention_mask], training=training ) layer_output = self.bert_output( [intermediate_output, attention_output], training=training ) masked_layer_output = layer_output * tf.cast( tf.expand_dims(attention_mask, 2), dtype=layer_output.dtype ) outputs = (masked_layer_output,) + attention_outputs[ 1: ] # add attentions if we output them return outputs class TFFastSpeechEncoder(tf.keras.layers.Layer): """Fast Speech encoder module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.output_attentions = config.output_attentions self.output_hidden_states = config.output_hidden_states self.layer = [ TFFastSpeechLayer(config, name="layer_._{}".format(i)) for i in range(config.num_hidden_layers) ] def call(self, inputs, training=False): """Call logic.""" hidden_states, attention_mask = inputs all_hidden_states = () all_attentions = () for _, layer_module in enumerate(self.layer): if self.output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) layer_outputs = layer_module( [hidden_states, attention_mask], training=training ) hidden_states = layer_outputs[0] if self.output_attentions: all_attentions = all_attentions + (layer_outputs[1],) # Add last layer if self.output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) outputs = (hidden_states,) if self.output_hidden_states: outputs = outputs + (all_hidden_states,) if self.output_attentions: outputs = outputs + (all_attentions,) return outputs # outputs, (hidden states), (attentions) class TFFastSpeechDecoder(TFFastSpeechEncoder): """Fast Speech decoder module.""" def __init__(self, config, **kwargs): self.is_compatible_encoder = kwargs.pop("is_compatible_encoder", True) super().__init__(config, **kwargs) self.config = config # create decoder positional embedding self.decoder_positional_embeddings = TFEmbedding( config.max_position_embeddings + 1, config.hidden_size, weights=[self._sincos_embedding()], name="position_embeddings", trainable=False, ) if self.is_compatible_encoder is False: self.project_compatible_decoder = tf.keras.layers.Dense( units=config.hidden_size, name="project_compatible_decoder" ) if config.n_speakers > 1: self.decoder_speaker_embeddings = TFEmbedding( config.n_speakers, config.hidden_size, embeddings_initializer=get_initializer(config.initializer_range), name="speaker_embeddings", ) self.speaker_fc = tf.keras.layers.Dense( units=config.hidden_size, name="speaker_fc" ) def call(self, inputs, training=False): hidden_states, speaker_ids, encoder_mask, decoder_pos = inputs if self.is_compatible_encoder is False: hidden_states = self.project_compatible_decoder(hidden_states) # calculate new hidden states. hidden_states += tf.cast( self.decoder_positional_embeddings(decoder_pos), hidden_states.dtype ) if self.config.n_speakers > 1: speaker_embeddings = self.decoder_speaker_embeddings(speaker_ids) speaker_features = tf.math.softplus(self.speaker_fc(speaker_embeddings)) # extended speaker embeddings extended_speaker_features = speaker_features[:, tf.newaxis, :] hidden_states += extended_speaker_features return super().call([hidden_states, encoder_mask], training=training) def _sincos_embedding(self): position_enc = np.array( [ [ pos / np.power(10000, 2.0 * (i // 2) / self.config.hidden_size) for i in range(self.config.hidden_size) ] for pos in range(self.config.max_position_embeddings + 1) ] ) position_enc[:, 0::2] = np.sin(position_enc[:, 0::2]) position_enc[:, 1::2] = np.cos(position_enc[:, 1::2]) # pad embedding. position_enc[0] = 0.0 return position_enc class TFTacotronPostnet(tf.keras.layers.Layer): """Tacotron-2 postnet.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.conv_batch_norm = [] for i in range(config.n_conv_postnet): conv = tf.keras.layers.Conv1D( filters=config.postnet_conv_filters if i < config.n_conv_postnet - 1 else config.num_mels, kernel_size=config.postnet_conv_kernel_sizes, padding="same", name="conv_._{}".format(i), ) batch_norm = tf.keras.layers.BatchNormalization( axis=-1, name="batch_norm_._{}".format(i) ) self.conv_batch_norm.append((conv, batch_norm)) self.dropout = tf.keras.layers.Dropout( rate=config.postnet_dropout_rate, name="dropout" ) self.activation = [tf.nn.tanh] * (config.n_conv_postnet - 1) + [tf.identity] def call(self, inputs, training=False): """Call logic.""" outputs, mask = inputs extended_mask = tf.cast(tf.expand_dims(mask, axis=2), outputs.dtype) for i, (conv, bn) in enumerate(self.conv_batch_norm): outputs = conv(outputs) outputs = bn(outputs) outputs = self.activation[i](outputs) outputs = self.dropout(outputs, training=training) return outputs * extended_mask class TFFastSpeechDurationPredictor(tf.keras.layers.Layer): """FastSpeech duration predictor module.""" def __init__(self, config, **kwargs): """Init variables.""" super().__init__(**kwargs) self.conv_layers = [] for i in range(config.num_duration_conv_layers): self.conv_layers.append( tf.keras.layers.Conv1D( config.duration_predictor_filters, config.duration_predictor_kernel_sizes, padding="same", name="conv_._{}".format(i), ) ) self.conv_layers.append( tf.keras.layers.LayerNormalization( epsilon=config.layer_norm_eps, name="LayerNorm_._{}".format(i) ) ) self.conv_layers.append(tf.keras.layers.Activation(tf.nn.relu6)) self.conv_layers.append( tf.keras.layers.Dropout(config.duration_predictor_dropout_probs) ) self.conv_layers_sequence = tf.keras.Sequential(self.conv_layers) self.output_layer = tf.keras.layers.Dense(1) def call(self, inputs, training=False): """Call logic.""" encoder_hidden_states, attention_mask = inputs attention_mask = tf.cast( tf.expand_dims(attention_mask, 2), encoder_hidden_states.dtype ) # mask encoder hidden states masked_encoder_hidden_states = encoder_hidden_states * attention_mask # pass though first layer outputs = self.conv_layers_sequence(masked_encoder_hidden_states) outputs = self.output_layer(outputs) masked_outputs = outputs * attention_mask return tf.squeeze(tf.nn.relu6(masked_outputs), -1) # make sure positive value. class TFFastSpeechLengthRegulator(tf.keras.layers.Layer): """FastSpeech lengthregulator module.""" def __init__(self, config, **kwargs): """Init variables.""" self.enable_tflite_convertible = kwargs.pop("enable_tflite_convertible", False) super().__init__(**kwargs) self.config = config def call(self, inputs, training=False): """Call logic. Args: 1. encoder_hidden_states, Tensor (float32) shape [batch_size, length, hidden_size] 2. durations_gt, Tensor (float32/int32) shape [batch_size, length] """ encoder_hidden_states, durations_gt = inputs outputs, encoder_masks = self._length_regulator( encoder_hidden_states, durations_gt ) return outputs, encoder_masks def _length_regulator(self, encoder_hidden_states, durations_gt): """Length regulator logic.""" sum_durations = tf.reduce_sum(durations_gt, axis=-1) # [batch_size] max_durations = tf.reduce_max(sum_durations) input_shape = tf.shape(encoder_hidden_states) batch_size = input_shape[0] hidden_size = input_shape[-1] # initialize output hidden states and encoder masking. if self.enable_tflite_convertible: # There is only 1 batch in inference, so we don't have to use # `tf.While` op with 3-D output tensor. repeats = durations_gt[0] real_length = tf.reduce_sum(repeats) pad_size = max_durations - real_length # masks : [max_durations] masks = tf.sequence_mask([real_length], max_durations, dtype=tf.int32) repeat_encoder_hidden_states = tf.repeat( encoder_hidden_states[0], repeats=repeats, axis=0 ) repeat_encoder_hidden_states = tf.expand_dims( tf.pad(repeat_encoder_hidden_states, [[0, pad_size], [0, 0]]), 0 ) # [1, max_durations, hidden_size] outputs = repeat_encoder_hidden_states encoder_masks = masks else: outputs = tf.zeros( shape=[0, max_durations, hidden_size], dtype=encoder_hidden_states.dtype ) encoder_masks = tf.zeros(shape=[0, max_durations], dtype=tf.int32) def condition( i, batch_size, outputs, encoder_masks, encoder_hidden_states, durations_gt, max_durations, ): return tf.less(i, batch_size) def body( i, batch_size, outputs, encoder_masks, encoder_hidden_states, durations_gt, max_durations, ): repeats = durations_gt[i] real_length = tf.reduce_sum(repeats) pad_size = max_durations - real_length masks = tf.sequence_mask([real_length], max_durations, dtype=tf.int32) repeat_encoder_hidden_states = tf.repeat( encoder_hidden_states[i], repeats=repeats, axis=0 ) repeat_encoder_hidden_states = tf.expand_dims( tf.pad(repeat_encoder_hidden_states, [[0, pad_size], [0, 0]]), 0 ) # [1, max_durations, hidden_size] outputs = tf.concat([outputs, repeat_encoder_hidden_states], axis=0) encoder_masks = tf.concat([encoder_masks, masks], axis=0) return [ i + 1, batch_size, outputs, encoder_masks, encoder_hidden_states, durations_gt, max_durations, ] # initialize iteration i. i = tf.constant(0, dtype=tf.int32) _, _, outputs, encoder_masks, _, _, _, = tf.while_loop( condition, body, [ i, batch_size, outputs, encoder_masks, encoder_hidden_states, durations_gt, max_durations, ], shape_invariants=[ i.get_shape(), batch_size.get_shape(), tf.TensorShape( [ None, None, self.config.encoder_self_attention_params.hidden_size, ] ), tf.TensorShape([None, None]), encoder_hidden_states.get_shape(), durations_gt.get_shape(), max_durations.get_shape(), ], ) return outputs, encoder_masks class TFFastSpeech(BaseModel): """TF Fastspeech module.""" def __init__(self, config, **kwargs): """Init layers for fastspeech.""" self.enable_tflite_convertible = kwargs.pop("enable_tflite_convertible", False) super().__init__(**kwargs) self.embeddings = TFFastSpeechEmbeddings(config, name="embeddings") self.encoder = TFFastSpeechEncoder( config.encoder_self_attention_params, name="encoder" ) self.duration_predictor = TFFastSpeechDurationPredictor( config, dtype=tf.float32, name="duration_predictor" ) self.length_regulator = TFFastSpeechLengthRegulator( config, enable_tflite_convertible=self.enable_tflite_convertible, name="length_regulator", ) self.decoder = TFFastSpeechDecoder( config.decoder_self_attention_params, is_compatible_encoder=config.encoder_self_attention_params.hidden_size == config.decoder_self_attention_params.hidden_size, name="decoder", ) self.mel_dense = tf.keras.layers.Dense( units=config.num_mels, dtype=tf.float32, name="mel_before" ) self.postnet = TFTacotronPostnet( config=config, dtype=tf.float32, name="postnet" ) self.setup_inference_fn() def _build(self): """Dummy input for building model.""" # fake inputs input_ids = tf.convert_to_tensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]], tf.int32) speaker_ids = tf.convert_to_tensor([0], tf.int32) duration_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32) self(input_ids, speaker_ids, duration_gts) def resize_positional_embeddings(self, new_size): self.embeddings.resize_positional_embeddings(new_size) self._build() def call( self, input_ids, speaker_ids, duration_gts, training=False, **kwargs, ): """Call logic.""" attention_mask = tf.math.not_equal(input_ids, 0) embedding_output = self.embeddings([input_ids, speaker_ids], training=training) encoder_output = self.encoder( [embedding_output, attention_mask], training=training ) last_encoder_hidden_states = encoder_output[0] # duration predictor, here use last_encoder_hidden_states, u can use more hidden_states layers # rather than just use last_hidden_states of encoder for duration_predictor. duration_outputs = self.duration_predictor( [last_encoder_hidden_states, attention_mask] ) # [batch_size, length] length_regulator_outputs, encoder_masks = self.length_regulator( [last_encoder_hidden_states, duration_gts], training=training ) # create decoder positional embedding decoder_pos = tf.range( 1, tf.shape(length_regulator_outputs)[1] + 1, dtype=tf.int32 ) masked_decoder_pos = tf.expand_dims(decoder_pos, 0) * encoder_masks decoder_output = self.decoder( [length_regulator_outputs, speaker_ids, encoder_masks, masked_decoder_pos], training=training, ) last_decoder_hidden_states = decoder_output[0] # here u can use sum or concat more than 1 hidden states layers from decoder. mel_before = self.mel_dense(last_decoder_hidden_states) mel_after = ( self.postnet([mel_before, encoder_masks], training=training) + mel_before ) outputs = (mel_before, mel_after, duration_outputs) return outputs def _inference(self, input_ids, speaker_ids, speed_ratios, **kwargs): """Call logic.""" attention_mask = tf.math.not_equal(input_ids, 0) embedding_output = self.embeddings([input_ids, speaker_ids], training=False) encoder_output = self.encoder( [embedding_output, attention_mask], training=False ) last_encoder_hidden_states = encoder_output[0] # duration predictor, here use last_encoder_hidden_states, u can use more hidden_states layers # rather than just use last_hidden_states of encoder for duration_predictor. duration_outputs = self.duration_predictor( [last_encoder_hidden_states, attention_mask] ) # [batch_size, length] duration_outputs = tf.math.exp(duration_outputs) - 1.0 if speed_ratios is None: speed_ratios = tf.convert_to_tensor(np.array([1.0]), dtype=tf.float32) speed_ratios = tf.expand_dims(speed_ratios, 1) duration_outputs = tf.cast( tf.math.round(duration_outputs * speed_ratios), tf.int32 ) length_regulator_outputs, encoder_masks = self.length_regulator( [last_encoder_hidden_states, duration_outputs], training=False ) # create decoder positional embedding decoder_pos = tf.range( 1, tf.shape(length_regulator_outputs)[1] + 1, dtype=tf.int32 ) masked_decoder_pos = tf.expand_dims(decoder_pos, 0) * encoder_masks decoder_output = self.decoder( [length_regulator_outputs, speaker_ids, encoder_masks, masked_decoder_pos], training=False, ) last_decoder_hidden_states = decoder_output[0] # here u can use sum or concat more than 1 hidden states layers from decoder. mel_before = self.mel_dense(last_decoder_hidden_states) mel_after = ( self.postnet([mel_before, encoder_masks], training=False) + mel_before ) outputs = (mel_before, mel_after, duration_outputs) return outputs def setup_inference_fn(self): self.inference = tf.function( self._inference, experimental_relax_shapes=True, input_signature=[ tf.TensorSpec(shape=[None, None], dtype=tf.int32, name="input_ids"), tf.TensorSpec(shape=[None,], dtype=tf.int32, name="speaker_ids"), tf.TensorSpec(shape=[None,], dtype=tf.float32, name="speed_ratios"), ], ) self.inference_tflite = tf.function( self._inference, experimental_relax_shapes=True, input_signature=[ tf.TensorSpec(shape=[1, None], dtype=tf.int32, name="input_ids"), tf.TensorSpec(shape=[1,], dtype=tf.int32, name="speaker_ids"), tf.TensorSpec(shape=[1,], dtype=tf.float32, name="speed_ratios"), ], )
33,971
36.372937
102
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/optimizers/adamweightdecay.py
# -*- coding: utf-8 -*- # Copyright 2019 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """AdamW for training self-attention.""" import re import tensorflow as tf class WarmUp(tf.keras.optimizers.schedules.LearningRateSchedule): """Applys a warmup schedule on a given learning rate decay schedule.""" def __init__( self, initial_learning_rate, decay_schedule_fn, warmup_steps, power=1.0, name=None, ): super(WarmUp, self).__init__() self.initial_learning_rate = initial_learning_rate self.warmup_steps = warmup_steps self.power = power self.decay_schedule_fn = decay_schedule_fn self.name = name def __call__(self, step): with tf.name_scope(self.name or "WarmUp") as name: # Implements polynomial warmup. i.e., if global_step < warmup_steps, the # learning rate will be `global_step/num_warmup_steps * init_lr`. global_step_float = tf.cast(step, tf.float32) warmup_steps_float = tf.cast(self.warmup_steps, tf.float32) warmup_percent_done = global_step_float / warmup_steps_float warmup_learning_rate = self.initial_learning_rate * tf.math.pow( warmup_percent_done, self.power ) return tf.cond( global_step_float < warmup_steps_float, lambda: warmup_learning_rate, lambda: self.decay_schedule_fn(step), name=name, ) def get_config(self): return { "initial_learning_rate": self.initial_learning_rate, "decay_schedule_fn": self.decay_schedule_fn, "warmup_steps": self.warmup_steps, "power": self.power, "name": self.name, } class AdamWeightDecay(tf.keras.optimizers.Adam): """Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact with the m and v parameters in strange ways. Instead we want ot decay the weights in a manner that doesn't interact with the m/v parameters. This is equivalent to adding the square of the weights to the loss with plain (non-momentum) SGD. """ def __init__( self, learning_rate=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-7, amsgrad=False, weight_decay_rate=0.0, include_in_weight_decay=None, exclude_from_weight_decay=None, name="AdamWeightDecay", **kwargs ): super(AdamWeightDecay, self).__init__( learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs ) self.weight_decay_rate = weight_decay_rate self._include_in_weight_decay = include_in_weight_decay self._exclude_from_weight_decay = exclude_from_weight_decay @classmethod def from_config(cls, config): """Creates an optimizer from its config with WarmUp custom object.""" custom_objects = {"WarmUp": WarmUp} return super(AdamWeightDecay, cls).from_config( config, custom_objects=custom_objects ) def _prepare_local(self, var_device, var_dtype, apply_state): super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype, apply_state) apply_state["weight_decay_rate"] = tf.constant( self.weight_decay_rate, name="adam_weight_decay_rate" ) def _decay_weights_op(self, var, learning_rate, apply_state): do_decay = self._do_use_weight_decay(var.name) if do_decay: return var.assign_sub( learning_rate * var * apply_state["weight_decay_rate"], use_locking=self._use_locking, ) return tf.no_op() def apply_gradients(self, grads_and_vars, clip_norm=0.5, **kwargs): grads, tvars = list(zip(*grads_and_vars)) (grads, _) = tf.clip_by_global_norm(grads, clip_norm=clip_norm) return super(AdamWeightDecay, self).apply_gradients(zip(grads, tvars), **kwargs) def _get_lr(self, var_device, var_dtype, apply_state): """Retrieves the learning rate with the given state.""" if apply_state is None: return self._decayed_lr_t[var_dtype], {} apply_state = apply_state or {} coefficients = apply_state.get((var_device, var_dtype)) if coefficients is None: coefficients = self._fallback_apply_state(var_device, var_dtype) apply_state[(var_device, var_dtype)] = coefficients return coefficients["lr_t"], dict(apply_state=apply_state) def _resource_apply_dense(self, grad, var, apply_state=None): lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state) decay = self._decay_weights_op(var, lr_t, apply_state) with tf.control_dependencies([decay]): return super(AdamWeightDecay, self)._resource_apply_dense( grad, var, **kwargs ) def _resource_apply_sparse(self, grad, var, indices, apply_state=None): lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state) decay = self._decay_weights_op(var, lr_t, apply_state) with tf.control_dependencies([decay]): return super(AdamWeightDecay, self)._resource_apply_sparse( grad, var, indices, **kwargs ) def get_config(self): config = super(AdamWeightDecay, self).get_config() config.update( {"weight_decay_rate": self.weight_decay_rate,} ) return config def _do_use_weight_decay(self, param_name): """Whether to use L2 weight decay for `param_name`.""" if self.weight_decay_rate == 0: return False if self._include_in_weight_decay: for r in self._include_in_weight_decay: if re.search(r, param_name) is not None: return True if self._exclude_from_weight_decay: for r in self._exclude_from_weight_decay: if re.search(r, param_name) is not None: return False return True
6,854
37.511236
88
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/utils/utils.py
# -*- coding: utf-8 -*- # Copyright 2019 Tomoki Hayashi # MIT License (https://opensource.org/licenses/MIT) """Utility functions.""" import fnmatch import os import re import tempfile from pathlib import Path import tensorflow as tf MODEL_FILE_NAME = "model.h5" CONFIG_FILE_NAME = "config.yml" PROCESSOR_FILE_NAME = "processor.json" LIBRARY_NAME = "tensorflow_tts" CACHE_DIRECTORY = os.path.join(Path.home(), ".cache", LIBRARY_NAME) def find_files(root_dir, query="*.wav", include_root_dir=True): """Find files recursively. Args: root_dir (str): Root root_dir to find. query (str): Query to find. include_root_dir (bool): If False, root_dir name is not included. Returns: list: List of found filenames. """ files = [] for root, _, filenames in os.walk(root_dir, followlinks=True): for filename in fnmatch.filter(filenames, query): files.append(os.path.join(root, filename)) if not include_root_dir: files = [file_.replace(root_dir + "/", "") for file_ in files] return files def _path_requires_gfile(filepath): """Checks if the given path requires use of GFile API. Args: filepath (str): Path to check. Returns: bool: True if the given path needs GFile API to access, such as "s3://some/path" and "gs://some/path". """ # If the filepath contains a protocol (e.g. "gs://"), it should be handled # using TensorFlow GFile API. return bool(re.match(r"^[a-z]+://", filepath)) def save_weights(model, filepath): """Save model weights. Same as model.save_weights(filepath), but supports saving to S3 or GCS buckets using TensorFlow GFile API. Args: model (tf.keras.Model): Model to save. filepath (str): Path to save the model weights to. """ if not _path_requires_gfile(filepath): model.save_weights(filepath) return # Save to a local temp file and copy to the desired path using GFile API. _, ext = os.path.splitext(filepath) with tempfile.NamedTemporaryFile(suffix=ext) as temp_file: model.save_weights(temp_file.name) # To preserve the original semantics, we need to overwrite the target # file. tf.io.gfile.copy(temp_file.name, filepath, overwrite=True) def load_weights(model, filepath): """Load model weights. Same as model.load_weights(filepath), but supports loading from S3 or GCS buckets using TensorFlow GFile API. Args: model (tf.keras.Model): Model to load weights to. filepath (str): Path to the weights file. """ if not _path_requires_gfile(filepath): model.load_weights(filepath) return # Make a local copy and load it. _, ext = os.path.splitext(filepath) with tempfile.NamedTemporaryFile(suffix=ext) as temp_file: # The target temp_file should be created above, so we need to overwrite. tf.io.gfile.copy(filepath, temp_file.name, overwrite=True) model.load_weights(temp_file.name)
3,053
30.163265
80
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/utils/group_conv.py
# -*- coding: utf-8 -*- # This code is copy from https://github.com/tensorflow/tensorflow/pull/36773. """Group Convolution Modules.""" from tensorflow.python.framework import tensor_shape from tensorflow.python.keras import activations, constraints, initializers, regularizers from tensorflow.python.keras.engine.base_layer import Layer from tensorflow.python.keras.engine.input_spec import InputSpec from tensorflow.python.keras.layers import Conv1D, SeparableConv1D from tensorflow.python.keras.utils import conv_utils from tensorflow.python.ops import array_ops, nn, nn_ops class Convolution(object): """Helper class for convolution. Note that this class assumes that shapes of input and filter passed to __call__ are compatible with input_shape and filter_shape passed to the constructor. Arguments input_shape: static shape of input. i.e. input.get_shape(). filter_shape: static shape of the filter. i.e. filter.get_shape(). padding: see convolution. strides: see convolution. dilation_rate: see convolution. name: see convolution. data_format: see convolution. """ def __init__( self, input_shape, filter_shape, padding, strides=None, dilation_rate=None, name=None, data_format=None, ): """Helper function for convolution.""" num_total_dims = filter_shape.ndims if num_total_dims is None: num_total_dims = input_shape.ndims if num_total_dims is None: raise ValueError("rank of input or filter must be known") num_spatial_dims = num_total_dims - 2 try: input_shape.with_rank(num_spatial_dims + 2) except ValueError: raise ValueError("input tensor must have rank %d" % (num_spatial_dims + 2)) try: filter_shape.with_rank(num_spatial_dims + 2) except ValueError: raise ValueError("filter tensor must have rank %d" % (num_spatial_dims + 2)) if data_format is None or not data_format.startswith("NC"): input_channels_dim = tensor_shape.dimension_at_index( input_shape, num_spatial_dims + 1 ) spatial_dims = range(1, num_spatial_dims + 1) else: input_channels_dim = tensor_shape.dimension_at_index(input_shape, 1) spatial_dims = range(2, num_spatial_dims + 2) filter_dim = tensor_shape.dimension_at_index(filter_shape, num_spatial_dims) if not (input_channels_dim % filter_dim).is_compatible_with(0): raise ValueError( "number of input channels is not divisible by corresponding " "dimension of filter, {} % {} != 0".format( input_channels_dim, filter_dim ) ) strides, dilation_rate = nn_ops._get_strides_and_dilation_rate( num_spatial_dims, strides, dilation_rate ) self.input_shape = input_shape self.filter_shape = filter_shape self.data_format = data_format self.strides = strides self.padding = padding self.name = name self.dilation_rate = dilation_rate self.conv_op = nn_ops._WithSpaceToBatch( input_shape, dilation_rate=dilation_rate, padding=padding, build_op=self._build_op, filter_shape=filter_shape, spatial_dims=spatial_dims, data_format=data_format, ) def _build_op(self, _, padding): return nn_ops._NonAtrousConvolution( self.input_shape, filter_shape=self.filter_shape, padding=padding, data_format=self.data_format, strides=self.strides, name=self.name, ) def __call__(self, inp, filter): return self.conv_op(inp, filter) class Conv(Layer): """Abstract N-D convolution layer (private, used as implementation base). This layer creates a convolution kernel that is convolved (actually cross-correlated) with the layer input to produce a tensor of outputs. If `use_bias` is True (and a `bias_initializer` is provided), a bias vector is created and added to the outputs. Finally, if `activation` is not `None`, it is applied to the outputs as well. Note: layer attributes cannot be modified after the layer has been called once (except the `trainable` attribute). Arguments: rank: An integer, the rank of the convolution, e.g. "2" for 2D convolution. filters: Integer, the dimensionality of the output space (i.e. the number of filters in the convolution). kernel_size: An integer or tuple/list of n integers, specifying the length of the convolution window. strides: An integer or tuple/list of n integers, specifying the stride length of the convolution. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"same"`, or `"causal"` (case-insensitive). data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch_size, ..., channels)` while `channels_first` corresponds to inputs with shape `(batch_size, channels, ...)`. dilation_rate: An integer or tuple/list of n integers, specifying the dilation rate to use for dilated convolution. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any `strides` value != 1. groups: Integer, the number of channel groups controlling the connections between inputs and outputs. Input channels and `filters` must both be divisible by `groups`. For example, - At `groups=1`, all inputs are convolved to all outputs. - At `groups=2`, the operation becomes equivalent to having two convolutional layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. - At `groups=input_channels`, each input channel is convolved with its own set of filters, of size `input_channels / filters` activation: Activation function to use. If you don't specify anything, no activation is applied. use_bias: Boolean, whether the layer uses a bias. kernel_initializer: An initializer for the convolution kernel. bias_initializer: An initializer for the bias vector. If None, the default initializer will be used. kernel_regularizer: Optional regularizer for the convolution kernel. bias_regularizer: Optional regularizer for the bias vector. activity_regularizer: Optional regularizer function for the output. kernel_constraint: Optional projection function to be applied to the kernel after being updated by an `Optimizer` (e.g. used to implement norm constraints or value constraints for layer weights). The function must take as input the unprojected variable and must return the projected variable (which must have the same shape). Constraints are not safe to use when doing asynchronous distributed training. bias_constraint: Optional projection function to be applied to the bias after being updated by an `Optimizer`. trainable: Boolean, if `True` the weights of this layer will be marked as trainable (and listed in `layer.trainable_weights`). name: A string, the name of the layer. """ def __init__( self, rank, filters, kernel_size, strides=1, padding="valid", data_format=None, dilation_rate=1, groups=1, activation=None, use_bias=True, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, trainable=True, name=None, **kwargs ): super(Conv, self).__init__( trainable=trainable, name=name, activity_regularizer=regularizers.get(activity_regularizer), **kwargs ) self.rank = rank if filters is not None and not isinstance(filters, int): filters = int(filters) self.filters = filters self.groups = groups or 1 if filters is not None and filters % self.groups != 0: raise ValueError( "The number of filters must be evenly divisible by the number of " "groups. Received: groups={}, filters={}".format(groups, filters) ) self.kernel_size = conv_utils.normalize_tuple(kernel_size, rank, "kernel_size") if not all(self.kernel_size): raise ValueError( "The argument `kernel_size` cannot contain 0(s). " "Received: %s" % (kernel_size,) ) self.strides = conv_utils.normalize_tuple(strides, rank, "strides") self.padding = conv_utils.normalize_padding(padding) if self.padding == "causal" and not isinstance(self, (Conv1D, SeparableConv1D)): raise ValueError( "Causal padding is only supported for `Conv1D`" "and ``SeparableConv1D`." ) self.data_format = conv_utils.normalize_data_format(data_format) self.dilation_rate = conv_utils.normalize_tuple( dilation_rate, rank, "dilation_rate" ) self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.kernel_constraint = constraints.get(kernel_constraint) self.bias_constraint = constraints.get(bias_constraint) self.input_spec = InputSpec(ndim=self.rank + 2) def build(self, input_shape): input_shape = tensor_shape.TensorShape(input_shape) input_channel = self._get_input_channel(input_shape) if input_channel % self.groups != 0: raise ValueError( "The number of input channels must be evenly divisible by the number " "of groups. Received groups={}, but the input has {} channels " "(full input shape is {}).".format( self.groups, input_channel, input_shape ) ) kernel_shape = self.kernel_size + (input_channel // self.groups, self.filters) self.kernel = self.add_weight( name="kernel", shape=kernel_shape, initializer=self.kernel_initializer, regularizer=self.kernel_regularizer, constraint=self.kernel_constraint, trainable=True, dtype=self.dtype, ) if self.use_bias: self.bias = self.add_weight( name="bias", shape=(self.filters,), initializer=self.bias_initializer, regularizer=self.bias_regularizer, constraint=self.bias_constraint, trainable=True, dtype=self.dtype, ) else: self.bias = None channel_axis = self._get_channel_axis() self.input_spec = InputSpec( ndim=self.rank + 2, axes={channel_axis: input_channel} ) self._build_conv_op_input_shape = input_shape self._build_input_channel = input_channel self._padding_op = self._get_padding_op() self._conv_op_data_format = conv_utils.convert_data_format( self.data_format, self.rank + 2 ) self._convolution_op = Convolution( input_shape, filter_shape=self.kernel.shape, dilation_rate=self.dilation_rate, strides=self.strides, padding=self._padding_op, data_format=self._conv_op_data_format, ) self.built = True def call(self, inputs): if self._recreate_conv_op(inputs): self._convolution_op = Convolution( inputs.get_shape(), filter_shape=self.kernel.shape, dilation_rate=self.dilation_rate, strides=self.strides, padding=self._padding_op, data_format=self._conv_op_data_format, ) self._build_conv_op_input_shape = inputs.get_shape() # Apply causal padding to inputs for Conv1D. if self.padding == "causal" and self.__class__.__name__ == "Conv1D": inputs = array_ops.pad(inputs, self._compute_causal_padding()) outputs = self._convolution_op(inputs, self.kernel) if self.use_bias: if self.data_format == "channels_first": if self.rank == 1: # nn.bias_add does not accept a 1D input tensor. bias = array_ops.reshape(self.bias, (1, self.filters, 1)) outputs += bias else: outputs = nn.bias_add(outputs, self.bias, data_format="NCHW") else: outputs = nn.bias_add(outputs, self.bias, data_format="NHWC") if self.activation is not None: return self.activation(outputs) return outputs def compute_output_shape(self, input_shape): input_shape = tensor_shape.TensorShape(input_shape).as_list() if self.data_format == "channels_last": space = input_shape[1:-1] new_space = [] for i in range(len(space)): new_dim = conv_utils.conv_output_length( space[i], self.kernel_size[i], padding=self.padding, stride=self.strides[i], dilation=self.dilation_rate[i], ) new_space.append(new_dim) return tensor_shape.TensorShape( [input_shape[0]] + new_space + [self.filters] ) else: space = input_shape[2:] new_space = [] for i in range(len(space)): new_dim = conv_utils.conv_output_length( space[i], self.kernel_size[i], padding=self.padding, stride=self.strides[i], dilation=self.dilation_rate[i], ) new_space.append(new_dim) return tensor_shape.TensorShape([input_shape[0], self.filters] + new_space) def get_config(self): config = { "filters": self.filters, "kernel_size": self.kernel_size, "strides": self.strides, "padding": self.padding, "data_format": self.data_format, "dilation_rate": self.dilation_rate, "groups": self.groups, "activation": activations.serialize(self.activation), "use_bias": self.use_bias, "kernel_initializer": initializers.serialize(self.kernel_initializer), "bias_initializer": initializers.serialize(self.bias_initializer), "kernel_regularizer": regularizers.serialize(self.kernel_regularizer), "bias_regularizer": regularizers.serialize(self.bias_regularizer), "activity_regularizer": regularizers.serialize(self.activity_regularizer), "kernel_constraint": constraints.serialize(self.kernel_constraint), "bias_constraint": constraints.serialize(self.bias_constraint), } base_config = super(Conv, self).get_config() return dict(list(base_config.items()) + list(config.items())) def _compute_causal_padding(self): """Calculates padding for 'causal' option for 1-d conv layers.""" left_pad = self.dilation_rate[0] * (self.kernel_size[0] - 1) if self.data_format == "channels_last": causal_padding = [[0, 0], [left_pad, 0], [0, 0]] else: causal_padding = [[0, 0], [0, 0], [left_pad, 0]] return causal_padding def _get_channel_axis(self): if self.data_format == "channels_first": return 1 else: return -1 def _get_input_channel(self, input_shape): channel_axis = self._get_channel_axis() if input_shape.dims[channel_axis].value is None: raise ValueError( "The channel dimension of the inputs " "should be defined. Found `None`." ) return int(input_shape[channel_axis]) def _get_padding_op(self): if self.padding == "causal": op_padding = "valid" else: op_padding = self.padding if not isinstance(op_padding, (list, tuple)): op_padding = op_padding.upper() return op_padding def _recreate_conv_op(self, inputs): """Recreate conv_op if necessary. Check if the input_shape in call() is different from that in build(). For the values that are not None, if they are different, recreate the _convolution_op to avoid the stateful behavior. Args: inputs: The input data to call() method. Returns: `True` or `False` to indicate whether to recreate the conv_op. """ call_input_shape = inputs.get_shape() for axis in range(1, len(call_input_shape)): if ( call_input_shape[axis] is not None and self._build_conv_op_input_shape[axis] is not None and call_input_shape[axis] != self._build_conv_op_input_shape[axis] ): return True return False class GroupConv1D(Conv): """1D convolution layer (e.g. temporal convolution). This layer creates a convolution kernel that is convolved with the layer input over a single spatial (or temporal) dimension to produce a tensor of outputs. If `use_bias` is True, a bias vector is created and added to the outputs. Finally, if `activation` is not `None`, it is applied to the outputs as well. When using this layer as the first layer in a model, provide an `input_shape` argument (tuple of integers or `None`, e.g. `(10, 128)` for sequences of 10 vectors of 128-dimensional vectors, or `(None, 128)` for variable-length sequences of 128-dimensional vectors. Examples: >>> # The inputs are 128-length vectors with 10 timesteps, and the batch size >>> # is 4. >>> input_shape = (4, 10, 128) >>> x = tf.random.normal(input_shape) >>> y = tf.keras.layers.Conv1D( ... 32, 3, activation='relu',input_shape=input_shape)(x) >>> print(y.shape) (4, 8, 32) Arguments: filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of a single integer, specifying the length of the 1D convolution window. strides: An integer or tuple/list of a single integer, specifying the stride length of the convolution. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"causal"` or `"same"` (case-insensitive). `"causal"` results in causal (dilated) convolutions, e.g. `output[t]` does not depend on `input[t+1:]`. Useful when modeling temporal data where the model should not violate the temporal order. See [WaveNet: A Generative Model for Raw Audio, section 2.1](https://arxiv.org/abs/1609.03499). data_format: A string, one of `channels_last` (default) or `channels_first`. groups: Integer, the number of channel groups controlling the connections between inputs and outputs. Input channels and `filters` must both be divisible by `groups`. For example, - At `groups=1`, all inputs are convolved to all outputs. - At `groups=2`, the operation becomes equivalent to having two convolutional layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. - At `groups=input_channels`, each input channel is convolved with its own set of filters, of size `input_channels / filters` dilation_rate: an integer or tuple/list of a single integer, specifying the dilation rate to use for dilated convolution. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any `strides` value != 1. activation: Activation function to use. If you don't specify anything, no activation is applied ( see `keras.activations`). use_bias: Boolean, whether the layer uses a bias vector. kernel_initializer: Initializer for the `kernel` weights matrix ( see `keras.initializers`). bias_initializer: Initializer for the bias vector ( see `keras.initializers`). kernel_regularizer: Regularizer function applied to the `kernel` weights matrix (see `keras.regularizers`). bias_regularizer: Regularizer function applied to the bias vector ( see `keras.regularizers`). activity_regularizer: Regularizer function applied to the output of the layer (its "activation") ( see `keras.regularizers`). kernel_constraint: Constraint function applied to the kernel matrix ( see `keras.constraints`). bias_constraint: Constraint function applied to the bias vector ( see `keras.constraints`). Input shape: 3D tensor with shape: `(batch_size, steps, input_dim)` Output shape: 3D tensor with shape: `(batch_size, new_steps, filters)` `steps` value might have changed due to padding or strides. Returns: A tensor of rank 3 representing `activation(conv1d(inputs, kernel) + bias)`. Raises: ValueError: when both `strides` > 1 and `dilation_rate` > 1. """ def __init__( self, filters, kernel_size, strides=1, padding="valid", data_format="channels_last", dilation_rate=1, groups=1, activation=None, use_bias=True, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs ): super().__init__( rank=1, filters=filters, kernel_size=kernel_size, strides=strides, padding=padding, data_format=data_format, dilation_rate=dilation_rate, groups=groups, activation=activations.get(activation), use_bias=use_bias, kernel_initializer=initializers.get(kernel_initializer), bias_initializer=initializers.get(bias_initializer), kernel_regularizer=regularizers.get(kernel_regularizer), bias_regularizer=regularizers.get(bias_regularizer), activity_regularizer=regularizers.get(activity_regularizer), kernel_constraint=constraints.get(kernel_constraint), bias_constraint=constraints.get(bias_constraint), **kwargs )
23,944
41.989228
88
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/utils/griffin_lim.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Griffin-Lim phase reconstruction algorithm from mel spectrogram.""" import os import librosa import numpy as np import soundfile as sf import tensorflow as tf from sklearn.preprocessing import StandardScaler def griffin_lim_lb( mel_spec, stats_path, dataset_config, n_iter=32, output_dir=None, wav_name="lb" ): """Generate wave from mel spectrogram with Griffin-Lim algorithm using Librosa. Args: mel_spec (ndarray): array representing the mel spectrogram. stats_path (str): path to the `stats.npy` file containing norm statistics. dataset_config (Dict): dataset configuration parameters. n_iter (int): number of iterations for GL. output_dir (str): output directory where audio file will be saved. wav_name (str): name of the output file. Returns: gl_lb (ndarray): generated wave. """ scaler = StandardScaler() scaler.mean_, scaler.scale_ = np.load(stats_path) mel_spec = np.power(10.0, scaler.inverse_transform(mel_spec)).T mel_basis = librosa.filters.mel( dataset_config["sampling_rate"], n_fft=dataset_config["fft_size"], n_mels=dataset_config["num_mels"], fmin=dataset_config["fmin"], fmax=dataset_config["fmax"], ) mel_to_linear = np.maximum(1e-10, np.dot(np.linalg.pinv(mel_basis), mel_spec)) gl_lb = librosa.griffinlim( mel_to_linear, n_iter=n_iter, hop_length=dataset_config["hop_size"], win_length=dataset_config["win_length"] or dataset_config["fft_size"], ) if output_dir: output_path = os.path.join(output_dir, f"{wav_name}.wav") sf.write(output_path, gl_lb, dataset_config["sampling_rate"], "PCM_16") return gl_lb class TFGriffinLim(tf.keras.layers.Layer): """Griffin-Lim algorithm for phase reconstruction from mel spectrogram magnitude.""" def __init__(self, stats_path, dataset_config, normalized: bool = True): """Init GL params. Args: stats_path (str): path to the `stats.npy` file containing norm statistics. dataset_config (Dict): dataset configuration parameters. """ super().__init__() self.normalized = normalized if normalized: scaler = StandardScaler() scaler.mean_, scaler.scale_ = np.load(stats_path) self.scaler = scaler self.ds_config = dataset_config self.mel_basis = librosa.filters.mel( self.ds_config["sampling_rate"], n_fft=self.ds_config["fft_size"], n_mels=self.ds_config["num_mels"], fmin=self.ds_config["fmin"], fmax=self.ds_config["fmax"], ) # [num_mels, fft_size // 2 + 1] def save_wav(self, gl_tf, output_dir, wav_name): """Generate WAV file and save it. Args: gl_tf (tf.Tensor): reconstructed signal from GL algorithm. output_dir (str): output directory where audio file will be saved. wav_name (str): name of the output file. """ encode_fn = lambda x: tf.audio.encode_wav(x, self.ds_config["sampling_rate"]) gl_tf = tf.expand_dims(gl_tf, -1) if not isinstance(wav_name, list): wav_name = [wav_name] if len(gl_tf.shape) > 2: bs, *_ = gl_tf.shape assert bs == len(wav_name), "Batch and 'wav_name' have different size." tf_wav = tf.map_fn(encode_fn, gl_tf, dtype=tf.string) for idx in tf.range(bs): output_path = os.path.join(output_dir, f"{wav_name[idx]}.wav") tf.io.write_file(output_path, tf_wav[idx]) else: tf_wav = encode_fn(gl_tf) tf.io.write_file(os.path.join(output_dir, f"{wav_name[0]}.wav"), tf_wav) @tf.function( input_signature=[ tf.TensorSpec(shape=[None, None, None], dtype=tf.float32), tf.TensorSpec(shape=[], dtype=tf.int32), ] ) def call(self, mel_spec, n_iter=32): """Apply GL algorithm to batched mel spectrograms. Args: mel_spec (tf.Tensor): normalized mel spectrogram. n_iter (int): number of iterations to run GL algorithm. Returns: (tf.Tensor): reconstructed signal from GL algorithm. """ # de-normalize mel spectogram if self.normalized: mel_spec = tf.math.pow( 10.0, mel_spec * self.scaler.scale_ + self.scaler.mean_ ) else: mel_spec = tf.math.pow( 10.0, mel_spec ) # TODO @dathudeptrai check if its ok without it wavs were too quiet inverse_mel = tf.linalg.pinv(self.mel_basis) # [:, num_mels] @ [fft_size // 2 + 1, num_mels].T mel_to_linear = tf.linalg.matmul(mel_spec, inverse_mel, transpose_b=True) mel_to_linear = tf.cast(tf.math.maximum(1e-10, mel_to_linear), tf.complex64) init_phase = tf.cast( tf.random.uniform(tf.shape(mel_to_linear), maxval=1), tf.complex64 ) phase = tf.math.exp(2j * np.pi * init_phase) for _ in tf.range(n_iter): inverse = tf.signal.inverse_stft( mel_to_linear * phase, frame_length=self.ds_config["win_length"] or self.ds_config["fft_size"], frame_step=self.ds_config["hop_size"], fft_length=self.ds_config["fft_size"], window_fn=tf.signal.inverse_stft_window_fn(self.ds_config["hop_size"]), ) phase = tf.signal.stft( inverse, self.ds_config["win_length"] or self.ds_config["fft_size"], self.ds_config["hop_size"], self.ds_config["fft_size"], ) phase /= tf.cast(tf.maximum(1e-10, tf.abs(phase)), tf.complex64) return tf.signal.inverse_stft( mel_to_linear * phase, frame_length=self.ds_config["win_length"] or self.ds_config["fft_size"], frame_step=self.ds_config["hop_size"], fft_length=self.ds_config["fft_size"], window_fn=tf.signal.inverse_stft_window_fn(self.ds_config["hop_size"]), )
6,824
39.868263
88
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/utils/weight_norm.py
# -*- coding: utf-8 -*- # Copyright 2019 The TensorFlow Probability Authors and Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Weight Norm Modules.""" import warnings import tensorflow as tf class WeightNormalization(tf.keras.layers.Wrapper): """Layer wrapper to decouple magnitude and direction of the layer's weights. This wrapper reparameterizes a layer by decoupling the weight's magnitude and direction. This speeds up convergence by improving the conditioning of the optimization problem. It has an optional data-dependent initialization scheme, in which initial values of weights are set as functions of the first minibatch of data. Both the weight normalization and data- dependent initialization are described in [Salimans and Kingma (2016)][1]. #### Example ```python net = WeightNorm(tf.keras.layers.Conv2D(2, 2, activation='relu'), input_shape=(32, 32, 3), data_init=True)(x) net = WeightNorm(tf.keras.layers.Conv2DTranspose(16, 5, activation='relu'), data_init=True) net = WeightNorm(tf.keras.layers.Dense(120, activation='relu'), data_init=True)(net) net = WeightNorm(tf.keras.layers.Dense(num_classes), data_init=True)(net) ``` #### References [1]: Tim Salimans and Diederik P. Kingma. Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks. In _30th Conference on Neural Information Processing Systems_, 2016. https://arxiv.org/abs/1602.07868 """ def __init__(self, layer, data_init=True, **kwargs): """Initialize WeightNorm wrapper. Args: layer: A `tf.keras.layers.Layer` instance. Supported layer types are `Dense`, `Conv2D`, and `Conv2DTranspose`. Layers with multiple inputs are not supported. data_init: `bool`, if `True` use data dependent variable initialization. **kwargs: Additional keyword args passed to `tf.keras.layers.Wrapper`. Raises: ValueError: If `layer` is not a `tf.keras.layers.Layer` instance. """ if not isinstance(layer, tf.keras.layers.Layer): raise ValueError( "Please initialize `WeightNorm` layer with a `tf.keras.layers.Layer` " "instance. You passed: {input}".format(input=layer) ) layer_type = type(layer).__name__ if layer_type not in [ "Dense", "Conv2D", "Conv2DTranspose", "Conv1D", "GroupConv1D", ]: warnings.warn( "`WeightNorm` is tested only for `Dense`, `Conv2D`, `Conv1D`, `GroupConv1D`, " "`GroupConv2D`, and `Conv2DTranspose` layers. You passed a layer of type `{}`".format( layer_type ) ) super().__init__(layer, **kwargs) self.data_init = data_init self._track_trackable(layer, name="layer") self.filter_axis = -2 if layer_type == "Conv2DTranspose" else -1 def _compute_weights(self): """Generate weights with normalization.""" # Determine the axis along which to expand `g` so that `g` broadcasts to # the shape of `v`. new_axis = -self.filter_axis - 3 self.layer.kernel = tf.nn.l2_normalize( self.v, axis=self.kernel_norm_axes ) * tf.expand_dims(self.g, new_axis) def _init_norm(self): """Set the norm of the weight vector.""" kernel_norm = tf.sqrt( tf.reduce_sum(tf.square(self.v), axis=self.kernel_norm_axes) ) self.g.assign(kernel_norm) def _data_dep_init(self, inputs): """Data dependent initialization.""" # Normalize kernel first so that calling the layer calculates # `tf.dot(v, x)/tf.norm(v)` as in (5) in ([Salimans and Kingma, 2016][1]). self._compute_weights() activation = self.layer.activation self.layer.activation = None use_bias = self.layer.bias is not None if use_bias: bias = self.layer.bias self.layer.bias = tf.zeros_like(bias) # Since the bias is initialized as zero, setting the activation to zero and # calling the initialized layer (with normalized kernel) yields the correct # computation ((5) in Salimans and Kingma (2016)) x_init = self.layer(inputs) norm_axes_out = list(range(x_init.shape.rank - 1)) m_init, v_init = tf.nn.moments(x_init, norm_axes_out) scale_init = 1.0 / tf.sqrt(v_init + 1e-10) self.g.assign(self.g * scale_init) if use_bias: self.layer.bias = bias self.layer.bias.assign(-m_init * scale_init) self.layer.activation = activation def build(self, input_shape=None): """Build `Layer`. Args: input_shape: The shape of the input to `self.layer`. Raises: ValueError: If `Layer` does not contain a `kernel` of weights """ if not self.layer.built: self.layer.build(input_shape) if not hasattr(self.layer, "kernel"): raise ValueError( "`WeightNorm` must wrap a layer that" " contains a `kernel` for weights" ) self.kernel_norm_axes = list(range(self.layer.kernel.shape.ndims)) self.kernel_norm_axes.pop(self.filter_axis) self.v = self.layer.kernel # to avoid a duplicate `kernel` variable after `build` is called self.layer.kernel = None self.g = self.add_weight( name="g", shape=(int(self.v.shape[self.filter_axis]),), initializer="ones", dtype=self.v.dtype, trainable=True, ) self.initialized = self.add_weight( name="initialized", dtype=tf.bool, trainable=False ) self.initialized.assign(False) super().build() def call(self, inputs): """Call `Layer`.""" if not self.initialized: if self.data_init: self._data_dep_init(inputs) else: # initialize `g` as the norm of the initialized kernel self._init_norm() self.initialized.assign(True) self._compute_weights() output = self.layer(inputs) return output def compute_output_shape(self, input_shape): return tf.TensorShape(self.layer.compute_output_shape(input_shape).as_list())
7,216
38.010811
102
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/losses/stft.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """STFT-based loss modules.""" import tensorflow as tf class TFSpectralConvergence(tf.keras.layers.Layer): """Spectral convergence loss.""" def __init__(self): """Initialize.""" super().__init__() def call(self, y_mag, x_mag): """Calculate forward propagation. Args: y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins). x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins). Returns: Tensor: Spectral convergence loss value. """ return tf.norm(y_mag - x_mag, ord="fro", axis=(-2, -1)) / tf.norm( y_mag, ord="fro", axis=(-2, -1) ) class TFLogSTFTMagnitude(tf.keras.layers.Layer): """Log STFT magnitude loss module.""" def __init__(self): """Initialize.""" super().__init__() def call(self, y_mag, x_mag): """Calculate forward propagation. Args: y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins). x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins). Returns: Tensor: Spectral convergence loss value. """ return tf.abs(tf.math.log(y_mag) - tf.math.log(x_mag)) class TFSTFT(tf.keras.layers.Layer): """STFT loss module.""" def __init__(self, frame_length=600, frame_step=120, fft_length=1024): """Initialize.""" super().__init__() self.frame_length = frame_length self.frame_step = frame_step self.fft_length = fft_length self.spectral_convergenge_loss = TFSpectralConvergence() self.log_stft_magnitude_loss = TFLogSTFTMagnitude() def call(self, y, x): """Calculate forward propagation. Args: y (Tensor): Groundtruth signal (B, T). x (Tensor): Predicted signal (B, T). Returns: Tensor: Spectral convergence loss value (pre-reduce). Tensor: Log STFT magnitude loss value (pre-reduce). """ x_mag = tf.abs( tf.signal.stft( signals=x, frame_length=self.frame_length, frame_step=self.frame_step, fft_length=self.fft_length, ) ) y_mag = tf.abs( tf.signal.stft( signals=y, frame_length=self.frame_length, frame_step=self.frame_step, fft_length=self.fft_length, ) ) # add small number to prevent nan value. # compatible with pytorch version. x_mag = tf.clip_by_value(tf.math.sqrt(x_mag ** 2 + 1e-7), 1e-7, 1e3) y_mag = tf.clip_by_value(tf.math.sqrt(y_mag ** 2 + 1e-7), 1e-7, 1e3) sc_loss = self.spectral_convergenge_loss(y_mag, x_mag) mag_loss = self.log_stft_magnitude_loss(y_mag, x_mag) return sc_loss, mag_loss class TFMultiResolutionSTFT(tf.keras.layers.Layer): """Multi resolution STFT loss module.""" def __init__( self, fft_lengths=[1024, 2048, 512], frame_lengths=[600, 1200, 240], frame_steps=[120, 240, 50], ): """Initialize Multi resolution STFT loss module. Args: frame_lengths (list): List of FFT sizes. frame_steps (list): List of hop sizes. fft_lengths (list): List of window lengths. """ super().__init__() assert len(frame_lengths) == len(frame_steps) == len(fft_lengths) self.stft_losses = [] for frame_length, frame_step, fft_length in zip( frame_lengths, frame_steps, fft_lengths ): self.stft_losses.append(TFSTFT(frame_length, frame_step, fft_length)) def call(self, y, x): """Calculate forward propagation. Args: y (Tensor): Groundtruth signal (B, T). x (Tensor): Predicted signal (B, T). Returns: Tensor: Multi resolution spectral convergence loss value. Tensor: Multi resolution log STFT magnitude loss value. """ sc_loss = 0.0 mag_loss = 0.0 for f in self.stft_losses: sc_l, mag_l = f(y, x) sc_loss += tf.reduce_mean(sc_l, axis=list(range(1, len(sc_l.shape)))) mag_loss += tf.reduce_mean(mag_l, axis=list(range(1, len(mag_l.shape)))) sc_loss /= len(self.stft_losses) mag_loss /= len(self.stft_losses) return sc_loss, mag_loss
5,179
33.533333
97
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/losses/spectrogram.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Spectrogram-based loss modules.""" import tensorflow as tf class TFMelSpectrogram(tf.keras.layers.Layer): """Mel Spectrogram loss.""" def __init__( self, n_mels=80, f_min=80.0, f_max=7600, frame_length=1024, frame_step=256, fft_length=1024, sample_rate=16000, **kwargs ): """Initialize.""" super().__init__(**kwargs) self.frame_length = frame_length self.frame_step = frame_step self.fft_length = fft_length self.linear_to_mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix( n_mels, fft_length // 2 + 1, sample_rate, f_min, f_max ) def _calculate_log_mels_spectrogram(self, signals): """Calculate forward propagation. Args: signals (Tensor): signal (B, T). Returns: Tensor: Mel spectrogram (B, T', 80) """ stfts = tf.signal.stft( signals, frame_length=self.frame_length, frame_step=self.frame_step, fft_length=self.fft_length, ) linear_spectrograms = tf.abs(stfts) mel_spectrograms = tf.tensordot( linear_spectrograms, self.linear_to_mel_weight_matrix, 1 ) mel_spectrograms.set_shape( linear_spectrograms.shape[:-1].concatenate( self.linear_to_mel_weight_matrix.shape[-1:] ) ) log_mel_spectrograms = tf.math.log(mel_spectrograms + 1e-6) # prevent nan. return log_mel_spectrograms def call(self, y, x): """Calculate forward propagation. Args: y (Tensor): Groundtruth signal (B, T). x (Tensor): Predicted signal (B, T). Returns: Tensor: Mean absolute Error Spectrogram Loss. """ y_mels = self._calculate_log_mels_spectrogram(y) x_mels = self._calculate_log_mels_spectrogram(x) return tf.reduce_mean( tf.abs(y_mels - x_mels), axis=list(range(1, len(x_mels.shape))) )
2,697
31.902439
83
py
TensorFlowTTS
TensorFlowTTS-master/tensorflow_tts/trainers/base_trainer.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Based Trainer.""" import abc import logging import os import tensorflow as tf from tqdm import tqdm from tensorflow_tts.optimizers import GradientAccumulator from tensorflow_tts.utils import utils class BasedTrainer(metaclass=abc.ABCMeta): """Customized trainer module for all models.""" def __init__(self, steps, epochs, config): self.steps = steps self.epochs = epochs self.config = config self.finish_train = False self.writer = tf.summary.create_file_writer(config["outdir"]) self.train_data_loader = None self.eval_data_loader = None self.train_metrics = None self.eval_metrics = None self.list_metrics_name = None def init_train_eval_metrics(self, list_metrics_name): """Init train and eval metrics to save it to tensorboard.""" self.train_metrics = {} self.eval_metrics = {} for name in list_metrics_name: self.train_metrics.update( {name: tf.keras.metrics.Mean(name="train_" + name, dtype=tf.float32)} ) self.eval_metrics.update( {name: tf.keras.metrics.Mean(name="eval_" + name, dtype=tf.float32)} ) def reset_states_train(self): """Reset train metrics after save it to tensorboard.""" for metric in self.train_metrics.keys(): self.train_metrics[metric].reset_states() def reset_states_eval(self): """Reset eval metrics after save it to tensorboard.""" for metric in self.eval_metrics.keys(): self.eval_metrics[metric].reset_states() def update_train_metrics(self, dict_metrics_losses): for name, value in dict_metrics_losses.items(): self.train_metrics[name].update_state(value) def update_eval_metrics(self, dict_metrics_losses): for name, value in dict_metrics_losses.items(): self.eval_metrics[name].update_state(value) def set_train_data_loader(self, train_dataset): """Set train data loader (MUST).""" self.train_data_loader = train_dataset def get_train_data_loader(self): """Get train data loader.""" return self.train_data_loader def set_eval_data_loader(self, eval_dataset): """Set eval data loader (MUST).""" self.eval_data_loader = eval_dataset def get_eval_data_loader(self): """Get eval data loader.""" return self.eval_data_loader @abc.abstractmethod def compile(self): pass @abc.abstractmethod def create_checkpoint_manager(self, saved_path=None, max_to_keep=10): """Create checkpoint management.""" pass def run(self): """Run training.""" self.tqdm = tqdm( initial=self.steps, total=self.config["train_max_steps"], desc="[train]" ) while True: self._train_epoch() if self.finish_train: break self.tqdm.close() logging.info("Finish training.") @abc.abstractmethod def save_checkpoint(self): """Save checkpoint.""" pass @abc.abstractmethod def load_checkpoint(self, pretrained_path): """Load checkpoint.""" pass def _train_epoch(self): """Train model one epoch.""" for train_steps_per_epoch, batch in enumerate(self.train_data_loader, 1): # one step training self._train_step(batch) # check interval self._check_log_interval() self._check_eval_interval() self._check_save_interval() # check wheter training is finished if self.finish_train: return # update self.epochs += 1 self.train_steps_per_epoch = train_steps_per_epoch logging.info( f"(Steps: {self.steps}) Finished {self.epochs} epoch training " f"({self.train_steps_per_epoch} steps per epoch)." ) @abc.abstractmethod def _eval_epoch(self): """One epoch evaluation.""" pass @abc.abstractmethod def _train_step(self, batch): """One step training.""" pass @abc.abstractmethod def _check_log_interval(self): """Save log interval.""" pass @abc.abstractmethod def fit(self): pass def _check_eval_interval(self): """Evaluation interval step.""" if self.steps % self.config["eval_interval_steps"] == 0: self._eval_epoch() def _check_save_interval(self): """Save interval checkpoint.""" if self.steps % self.config["save_interval_steps"] == 0: self.save_checkpoint() logging.info(f"Successfully saved checkpoint @ {self.steps} steps.") def generate_and_save_intermediate_result(self, batch): """Generate and save intermediate result.""" pass def _write_to_tensorboard(self, list_metrics, stage="train"): """Write variables to tensorboard.""" with self.writer.as_default(): for key, value in list_metrics.items(): tf.summary.scalar(stage + "/" + key, value.result(), step=self.steps) self.writer.flush() class GanBasedTrainer(BasedTrainer): """Customized trainer module for GAN TTS training (MelGAN, GAN-TTS, ParallelWaveGAN).""" def __init__( self, steps, epochs, config, strategy, is_generator_mixed_precision=False, is_discriminator_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. """ super().__init__(steps, epochs, config) self._is_generator_mixed_precision = is_generator_mixed_precision self._is_discriminator_mixed_precision = is_discriminator_mixed_precision self._strategy = strategy self._already_apply_input_signature = False self._generator_gradient_accumulator = GradientAccumulator() self._discriminator_gradient_accumulator = GradientAccumulator() self._generator_gradient_accumulator.reset() self._discriminator_gradient_accumulator.reset() def init_train_eval_metrics(self, list_metrics_name): with self._strategy.scope(): super().init_train_eval_metrics(list_metrics_name) def get_n_gpus(self): return self._strategy.num_replicas_in_sync def _get_train_element_signature(self): return self.train_data_loader.element_spec def _get_eval_element_signature(self): return self.eval_data_loader.element_spec def set_gen_model(self, generator_model): """Set generator class model (MUST).""" self._generator = generator_model def get_gen_model(self): """Get generator model.""" return self._generator def set_dis_model(self, discriminator_model): """Set discriminator class model (MUST).""" self._discriminator = discriminator_model def get_dis_model(self): """Get discriminator model.""" return self._discriminator def set_gen_optimizer(self, generator_optimizer): """Set generator optimizer (MUST).""" self._gen_optimizer = generator_optimizer if self._is_generator_mixed_precision: self._gen_optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer( self._gen_optimizer, "dynamic" ) def get_gen_optimizer(self): """Get generator optimizer.""" return self._gen_optimizer def set_dis_optimizer(self, discriminator_optimizer): """Set discriminator optimizer (MUST).""" self._dis_optimizer = discriminator_optimizer if self._is_discriminator_mixed_precision: self._dis_optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer( self._dis_optimizer, "dynamic" ) def get_dis_optimizer(self): """Get discriminator optimizer.""" return self._dis_optimizer def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer): self.set_gen_model(gen_model) self.set_dis_model(dis_model) self.set_gen_optimizer(gen_optimizer) self.set_dis_optimizer(dis_optimizer) def _train_step(self, batch): if self._already_apply_input_signature is False: train_element_signature = self._get_train_element_signature() eval_element_signature = self._get_eval_element_signature() self.one_step_forward = tf.function( self._one_step_forward, input_signature=[train_element_signature] ) self.one_step_evaluate = tf.function( self._one_step_evaluate, input_signature=[eval_element_signature] ) self.one_step_predict = tf.function( self._one_step_predict, input_signature=[eval_element_signature] ) self._already_apply_input_signature = True # run one_step_forward self.one_step_forward(batch) # update counts self.steps += 1 self.tqdm.update(1) self._check_train_finish() def _one_step_forward(self, batch): per_replica_losses = self._strategy.run( self._one_step_forward_per_replica, args=(batch,) ) return self._strategy.reduce( tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None ) @abc.abstractmethod def compute_per_example_generator_losses(self, batch, outputs): """Compute per example generator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ per_example_losses = 0.0 dict_metrics_losses = {} return per_example_losses, dict_metrics_losses @abc.abstractmethod def compute_per_example_discriminator_losses(self, batch, gen_outputs): """Compute per example discriminator losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ per_example_losses = 0.0 dict_metrics_losses = {} return per_example_losses, dict_metrics_losses def _calculate_generator_gradient_per_batch(self, batch): outputs = self._generator(**batch, training=True) ( per_example_losses, dict_metrics_losses, ) = self.compute_per_example_generator_losses(batch, outputs) per_replica_gen_losses = tf.nn.compute_average_loss( per_example_losses, global_batch_size=self.config["batch_size"] * self.get_n_gpus() * self.config["gradient_accumulation_steps"], ) if self._is_generator_mixed_precision: scaled_per_replica_gen_losses = self._gen_optimizer.get_scaled_loss( per_replica_gen_losses ) if self._is_generator_mixed_precision: scaled_gradients = tf.gradients( scaled_per_replica_gen_losses, self._generator.trainable_variables ) gradients = self._gen_optimizer.get_unscaled_gradients(scaled_gradients) else: gradients = tf.gradients( per_replica_gen_losses, self._generator.trainable_variables ) # gradient accumulate for generator here if self.config["gradient_accumulation_steps"] > 1: self._generator_gradient_accumulator(gradients) # accumulate loss into metrics self.update_train_metrics(dict_metrics_losses) if self.config["gradient_accumulation_steps"] == 1: return gradients, per_replica_gen_losses else: return per_replica_gen_losses def _calculate_discriminator_gradient_per_batch(self, batch): ( per_example_losses, dict_metrics_losses, ) = self.compute_per_example_discriminator_losses( batch, self._generator(**batch, training=True) ) per_replica_dis_losses = tf.nn.compute_average_loss( per_example_losses, global_batch_size=self.config["batch_size"] * self.get_n_gpus() * self.config["gradient_accumulation_steps"], ) if self._is_discriminator_mixed_precision: scaled_per_replica_dis_losses = self._dis_optimizer.get_scaled_loss( per_replica_dis_losses ) if self._is_discriminator_mixed_precision: scaled_gradients = tf.gradients( scaled_per_replica_dis_losses, self._discriminator.trainable_variables, ) gradients = self._dis_optimizer.get_unscaled_gradients(scaled_gradients) else: gradients = tf.gradients( per_replica_dis_losses, self._discriminator.trainable_variables ) # accumulate loss into metrics self.update_train_metrics(dict_metrics_losses) # gradient accumulate for discriminator here if self.config["gradient_accumulation_steps"] > 1: self._discriminator_gradient_accumulator(gradients) if self.config["gradient_accumulation_steps"] == 1: return gradients, per_replica_dis_losses else: return per_replica_dis_losses def _one_step_forward_per_replica(self, batch): per_replica_gen_losses = 0.0 per_replica_dis_losses = 0.0 if self.config["gradient_accumulation_steps"] == 1: ( gradients, per_replica_gen_losses, ) = self._calculate_generator_gradient_per_batch(batch) self._gen_optimizer.apply_gradients( zip(gradients, self._generator.trainable_variables) ) else: # gradient acummulation here. for i in tf.range(self.config["gradient_accumulation_steps"]): reduced_batch = { k: v[ i * self.config["batch_size"] : (i + 1) * self.config["batch_size"] ] for k, v in batch.items() } # run 1 step accumulate reduced_batch_losses = self._calculate_generator_gradient_per_batch( reduced_batch ) # sum per_replica_losses per_replica_gen_losses += reduced_batch_losses gradients = self._generator_gradient_accumulator.gradients self._gen_optimizer.apply_gradients( zip(gradients, self._generator.trainable_variables) ) self._generator_gradient_accumulator.reset() # one step discriminator # recompute y_hat after 1 step generator for discriminator training. if self.steps >= self.config["discriminator_train_start_steps"]: if self.config["gradient_accumulation_steps"] == 1: ( gradients, per_replica_dis_losses, ) = self._calculate_discriminator_gradient_per_batch(batch) self._dis_optimizer.apply_gradients( zip(gradients, self._discriminator.trainable_variables) ) else: # gradient acummulation here. for i in tf.range(self.config["gradient_accumulation_steps"]): reduced_batch = { k: v[ i * self.config["batch_size"] : (i + 1) * self.config["batch_size"] ] for k, v in batch.items() } # run 1 step accumulate reduced_batch_losses = ( self._calculate_discriminator_gradient_per_batch(reduced_batch) ) # sum per_replica_losses per_replica_dis_losses += reduced_batch_losses gradients = self._discriminator_gradient_accumulator.gradients self._dis_optimizer.apply_gradients( zip(gradients, self._discriminator.trainable_variables) ) self._discriminator_gradient_accumulator.reset() return per_replica_gen_losses + per_replica_dis_losses def _eval_epoch(self): """Evaluate model one epoch.""" logging.info(f"(Steps: {self.steps}) Start evaluation.") # calculate loss for each batch for eval_steps_per_epoch, batch in enumerate( tqdm(self.eval_data_loader, desc="[eval]"), 1 ): # eval one step self.one_step_evaluate(batch) if eval_steps_per_epoch <= self.config["num_save_intermediate_results"]: # save intermedia self.generate_and_save_intermediate_result(batch) logging.info( f"(Steps: {self.steps}) Finished evaluation " f"({eval_steps_per_epoch} steps per epoch)." ) # average loss for key in self.eval_metrics.keys(): logging.info( f"(Steps: {self.steps}) eval_{key} = {self.eval_metrics[key].result():.4f}." ) # record self._write_to_tensorboard(self.eval_metrics, stage="eval") # reset self.reset_states_eval() def _one_step_evaluate_per_replica(self, batch): ################################################ # one step generator. outputs = self._generator(**batch, training=False) _, dict_metrics_losses = self.compute_per_example_generator_losses( batch, outputs ) # accumulate loss into metrics self.update_eval_metrics(dict_metrics_losses) ################################################ # one step discriminator if self.steps >= self.config["discriminator_train_start_steps"]: _, dict_metrics_losses = self.compute_per_example_discriminator_losses( batch, outputs ) # accumulate loss into metrics self.update_eval_metrics(dict_metrics_losses) ################################################ def _one_step_evaluate(self, batch): self._strategy.run(self._one_step_evaluate_per_replica, args=(batch,)) def _one_step_predict_per_replica(self, batch): outputs = self._generator(**batch, training=False) return outputs def _one_step_predict(self, batch): outputs = self._strategy.run(self._one_step_predict_per_replica, args=(batch,)) return outputs @abc.abstractmethod def generate_and_save_intermediate_result(self, batch): return def create_checkpoint_manager(self, saved_path=None, max_to_keep=10): """Create checkpoint management.""" if saved_path is None: saved_path = self.config["outdir"] + "/checkpoints/" os.makedirs(saved_path, exist_ok=True) self.saved_path = saved_path self.ckpt = tf.train.Checkpoint( steps=tf.Variable(1), epochs=tf.Variable(1), gen_optimizer=self.get_gen_optimizer(), dis_optimizer=self.get_dis_optimizer(), ) self.ckp_manager = tf.train.CheckpointManager( self.ckpt, saved_path, max_to_keep=max_to_keep ) def save_checkpoint(self): """Save checkpoint.""" self.ckpt.steps.assign(self.steps) self.ckpt.epochs.assign(self.epochs) self.ckp_manager.save(checkpoint_number=self.steps) utils.save_weights( self._generator, self.saved_path + "generator-{}.h5".format(self.steps) ) utils.save_weights( self._discriminator, self.saved_path + "discriminator-{}.h5".format(self.steps) ) def load_checkpoint(self, pretrained_path): """Load checkpoint.""" self.ckpt.restore(pretrained_path) self.steps = self.ckpt.steps.numpy() self.epochs = self.ckpt.epochs.numpy() self._gen_optimizer = self.ckpt.gen_optimizer # re-assign iterations (global steps) for gen_optimizer. self._gen_optimizer.iterations.assign(tf.cast(self.steps, tf.int64)) # re-assign iterations (global steps) for dis_optimizer. try: discriminator_train_start_steps = self.config[ "discriminator_train_start_steps" ] discriminator_train_start_steps = tf.math.maximum( 0, self.steps - discriminator_train_start_steps ) except Exception: discriminator_train_start_steps = self.steps self._dis_optimizer = self.ckpt.dis_optimizer self._dis_optimizer.iterations.assign( tf.cast(discriminator_train_start_steps, tf.int64) ) # load weights. utils.load_weights( self._generator, self.saved_path + "generator-{}.h5".format(self.steps) ) utils.load_weights( self._discriminator, self.saved_path + "discriminator-{}.h5".format(self.steps) ) def _check_train_finish(self): """Check training finished.""" if self.steps >= self.config["train_max_steps"]: self.finish_train = True if ( self.steps != 0 and self.steps == self.config["discriminator_train_start_steps"] ): self.finish_train = True logging.info( f"Finished training only generator at {self.steps}steps, pls resume and continue training." ) def _check_log_interval(self): """Log to tensorboard.""" if self.steps % self.config["log_interval_steps"] == 0: for metric_name in self.list_metrics_name: logging.info( f"(Step: {self.steps}) train_{metric_name} = {self.train_metrics[metric_name].result():.4f}." ) self._write_to_tensorboard(self.train_metrics, stage="train") # reset self.reset_states_train() def fit(self, train_data_loader, valid_data_loader, saved_path, resume=None): self.set_train_data_loader(train_data_loader) self.set_eval_data_loader(valid_data_loader) self.train_data_loader = self._strategy.experimental_distribute_dataset( self.train_data_loader ) self.eval_data_loader = self._strategy.experimental_distribute_dataset( self.eval_data_loader ) with self._strategy.scope(): self.create_checkpoint_manager(saved_path=saved_path, max_to_keep=10000) if len(resume) > 1: self.load_checkpoint(resume) logging.info(f"Successfully resumed from {resume}.") self.run() class Seq2SeqBasedTrainer(BasedTrainer, metaclass=abc.ABCMeta): """Customized trainer module for Seq2Seq TTS training (Tacotron, FastSpeech).""" def __init__( self, steps, epochs, config, strategy, is_mixed_precision=False, ): """Initialize trainer. Args: steps (int): Initial global steps. epochs (int): Initial global epochs. config (dict): Config dict loaded from yaml format configuration file. strategy (tf.distribute): Strategy for distributed training. is_mixed_precision (bool): Use mixed_precision training or not. """ super().__init__(steps, epochs, config) self._is_mixed_precision = is_mixed_precision self._strategy = strategy self._model = None self._optimizer = None self._trainable_variables = None # check if we already apply input_signature for train_step. self._already_apply_input_signature = False # create gradient accumulator self._gradient_accumulator = GradientAccumulator() self._gradient_accumulator.reset() def init_train_eval_metrics(self, list_metrics_name): with self._strategy.scope(): super().init_train_eval_metrics(list_metrics_name) def set_model(self, model): """Set generator class model (MUST).""" self._model = model def get_model(self): """Get generator model.""" return self._model def set_optimizer(self, optimizer): """Set optimizer (MUST).""" self._optimizer = optimizer if self._is_mixed_precision: self._optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer( self._optimizer, "dynamic" ) def get_optimizer(self): """Get optimizer.""" return self._optimizer def get_n_gpus(self): return self._strategy.num_replicas_in_sync def compile(self, model, optimizer): self.set_model(model) self.set_optimizer(optimizer) self._trainable_variables = self._train_vars() def _train_vars(self): if self.config["var_train_expr"]: list_train_var = self.config["var_train_expr"].split("|") return [ v for v in self._model.trainable_variables if self._check_string_exist(list_train_var, v.name) ] return self._model.trainable_variables def _check_string_exist(self, list_string, inp_string): for string in list_string: if string in inp_string: return True return False def _get_train_element_signature(self): return self.train_data_loader.element_spec def _get_eval_element_signature(self): return self.eval_data_loader.element_spec def _train_step(self, batch): if self._already_apply_input_signature is False: train_element_signature = self._get_train_element_signature() eval_element_signature = self._get_eval_element_signature() self.one_step_forward = tf.function( self._one_step_forward, input_signature=[train_element_signature] ) self.one_step_evaluate = tf.function( self._one_step_evaluate, input_signature=[eval_element_signature] ) self.one_step_predict = tf.function( self._one_step_predict, input_signature=[eval_element_signature] ) self._already_apply_input_signature = True # run one_step_forward self.one_step_forward(batch) # update counts self.steps += 1 self.tqdm.update(1) self._check_train_finish() def _one_step_forward(self, batch): per_replica_losses = self._strategy.run( self._one_step_forward_per_replica, args=(batch,) ) return self._strategy.reduce( tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None ) def _calculate_gradient_per_batch(self, batch): outputs = self._model(**batch, training=True) per_example_losses, dict_metrics_losses = self.compute_per_example_losses( batch, outputs ) per_replica_losses = tf.nn.compute_average_loss( per_example_losses, global_batch_size=self.config["batch_size"] * self.get_n_gpus() * self.config["gradient_accumulation_steps"], ) if self._is_mixed_precision: scaled_per_replica_losses = self._optimizer.get_scaled_loss( per_replica_losses ) if self._is_mixed_precision: scaled_gradients = tf.gradients( scaled_per_replica_losses, self._trainable_variables ) gradients = self._optimizer.get_unscaled_gradients(scaled_gradients) else: gradients = tf.gradients(per_replica_losses, self._trainable_variables) # gradient accumulate here if self.config["gradient_accumulation_steps"] > 1: self._gradient_accumulator(gradients) # accumulate loss into metrics self.update_train_metrics(dict_metrics_losses) if self.config["gradient_accumulation_steps"] == 1: return gradients, per_replica_losses else: return per_replica_losses def _one_step_forward_per_replica(self, batch): if self.config["gradient_accumulation_steps"] == 1: gradients, per_replica_losses = self._calculate_gradient_per_batch(batch) self._optimizer.apply_gradients( zip(gradients, self._trainable_variables), 1.0 ) else: # gradient acummulation here. per_replica_losses = 0.0 for i in tf.range(self.config["gradient_accumulation_steps"]): reduced_batch = { k: v[ i * self.config["batch_size"] : (i + 1) * self.config["batch_size"] ] for k, v in batch.items() } # run 1 step accumulate reduced_batch_losses = self._calculate_gradient_per_batch(reduced_batch) # sum per_replica_losses per_replica_losses += reduced_batch_losses gradients = self._gradient_accumulator.gradients self._optimizer.apply_gradients( zip(gradients, self._trainable_variables), 1.0 ) self._gradient_accumulator.reset() return per_replica_losses @abc.abstractmethod def compute_per_example_losses(self, batch, outputs): """Compute per example losses and return dict_metrics_losses Note that all element of the loss MUST has a shape [batch_size] and the keys of dict_metrics_losses MUST be in self.list_metrics_name. Args: batch: dictionary batch input return from dataloader outputs: outputs of the model Returns: per_example_losses: per example losses for each GPU, shape [B] dict_metrics_losses: dictionary loss. """ per_example_losses = 0.0 dict_metrics_losses = {} return per_example_losses, dict_metrics_losses def _eval_epoch(self): """Evaluate model one epoch.""" logging.info(f"(Steps: {self.steps}) Start evaluation.") # calculate loss for each batch for eval_steps_per_epoch, batch in enumerate( tqdm(self.eval_data_loader, desc="[eval]"), 1 ): # eval one step self.one_step_evaluate(batch) if eval_steps_per_epoch <= self.config["num_save_intermediate_results"]: # save intermedia self.generate_and_save_intermediate_result(batch) logging.info( f"(Steps: {self.steps}) Finished evaluation " f"({eval_steps_per_epoch} steps per epoch)." ) # average loss for key in self.eval_metrics.keys(): logging.info( f"(Steps: {self.steps}) eval_{key} = {self.eval_metrics[key].result():.4f}." ) # record self._write_to_tensorboard(self.eval_metrics, stage="eval") # reset self.reset_states_eval() def _one_step_evaluate_per_replica(self, batch): outputs = self._model(**batch, training=False) _, dict_metrics_losses = self.compute_per_example_losses(batch, outputs) self.update_eval_metrics(dict_metrics_losses) def _one_step_evaluate(self, batch): self._strategy.run(self._one_step_evaluate_per_replica, args=(batch,)) def _one_step_predict_per_replica(self, batch): outputs = self._model(**batch, training=False) return outputs def _one_step_predict(self, batch): outputs = self._strategy.run(self._one_step_predict_per_replica, args=(batch,)) return outputs @abc.abstractmethod def generate_and_save_intermediate_result(self, batch): return def create_checkpoint_manager(self, saved_path=None, max_to_keep=10): """Create checkpoint management.""" if saved_path is None: saved_path = self.config["outdir"] + "/checkpoints/" os.makedirs(saved_path, exist_ok=True) self.saved_path = saved_path self.ckpt = tf.train.Checkpoint( steps=tf.Variable(1), epochs=tf.Variable(1), optimizer=self.get_optimizer() ) self.ckp_manager = tf.train.CheckpointManager( self.ckpt, saved_path, max_to_keep=max_to_keep ) def save_checkpoint(self): """Save checkpoint.""" self.ckpt.steps.assign(self.steps) self.ckpt.epochs.assign(self.epochs) self.ckp_manager.save(checkpoint_number=self.steps) utils.save_weights( self._model, self.saved_path + "model-{}.h5".format(self.steps) ) def load_checkpoint(self, pretrained_path): """Load checkpoint.""" self.ckpt.restore(pretrained_path) self.steps = self.ckpt.steps.numpy() self.epochs = self.ckpt.epochs.numpy() self._optimizer = self.ckpt.optimizer # re-assign iterations (global steps) for optimizer. self._optimizer.iterations.assign(tf.cast(self.steps, tf.int64)) # load weights. utils.load_weights( self._model, self.saved_path + "model-{}.h5".format(self.steps) ) def _check_train_finish(self): """Check training finished.""" if self.steps >= self.config["train_max_steps"]: self.finish_train = True def _check_log_interval(self): """Log to tensorboard.""" if self.steps % self.config["log_interval_steps"] == 0: for metric_name in self.list_metrics_name: logging.info( f"(Step: {self.steps}) train_{metric_name} = {self.train_metrics[metric_name].result():.4f}." ) self._write_to_tensorboard(self.train_metrics, stage="train") # reset self.reset_states_train() def fit(self, train_data_loader, valid_data_loader, saved_path, resume=None): self.set_train_data_loader(train_data_loader) self.set_eval_data_loader(valid_data_loader) self.train_data_loader = self._strategy.experimental_distribute_dataset( self.train_data_loader ) self.eval_data_loader = self._strategy.experimental_distribute_dataset( self.eval_data_loader ) with self._strategy.scope(): self.create_checkpoint_manager(saved_path=saved_path, max_to_keep=10000) if len(resume) > 1: self.load_checkpoint(resume) logging.info(f"Successfully resumed from {resume}.") self.run()
36,562
35.165183
113
py
TensorFlowTTS
TensorFlowTTS-master/test/test_fastspeech.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import logging import os import pytest import tensorflow as tf from tensorflow_tts.configs import FastSpeechConfig from tensorflow_tts.models import TFFastSpeech os.environ["CUDA_VISIBLE_DEVICES"] = "" logging.basicConfig( level=logging.DEBUG, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) @pytest.mark.parametrize("new_size", [100, 200, 300]) def test_fastspeech_resize_positional_embeddings(new_size): config = FastSpeechConfig() fastspeech = TFFastSpeech(config, name="fastspeech") fastspeech._build() fastspeech.save_weights("./test.h5") fastspeech.resize_positional_embeddings(new_size) fastspeech.load_weights("./test.h5", by_name=True, skip_mismatch=True) @pytest.mark.parametrize("num_hidden_layers,n_speakers", [(2, 1), (3, 2), (4, 3)]) def test_fastspeech_trainable(num_hidden_layers, n_speakers): config = FastSpeechConfig( encoder_num_hidden_layers=num_hidden_layers, decoder_num_hidden_layers=num_hidden_layers + 1, n_speakers=n_speakers, ) fastspeech = TFFastSpeech(config, name="fastspeech") optimizer = tf.keras.optimizers.Adam(lr=0.001) # fake inputs input_ids = tf.convert_to_tensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]], tf.int32) attention_mask = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32) speaker_ids = tf.convert_to_tensor([0], tf.int32) duration_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32) mel_gts = tf.random.uniform(shape=[1, 10, 80], dtype=tf.float32) @tf.function def one_step_training(): with tf.GradientTape() as tape: mel_outputs_before, _, duration_outputs = fastspeech( input_ids, speaker_ids, duration_gts, training=True ) duration_loss = tf.keras.losses.MeanSquaredError()( duration_gts, duration_outputs ) mel_loss = tf.keras.losses.MeanSquaredError()(mel_gts, mel_outputs_before) loss = duration_loss + mel_loss gradients = tape.gradient(loss, fastspeech.trainable_variables) optimizer.apply_gradients(zip(gradients, fastspeech.trainable_variables)) tf.print(loss) import time for i in range(2): if i == 1: start = time.time() one_step_training() print(time.time() - start)
2,995
34.247059
86
py
TensorFlowTTS
TensorFlowTTS-master/test/test_tacotron2.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import logging import os import time import yaml import numpy as np import pytest import tensorflow as tf from tensorflow_tts.configs import Tacotron2Config from tensorflow_tts.models import TFTacotron2 from tensorflow_tts.utils import return_strategy from examples.tacotron2.train_tacotron2 import Tacotron2Trainer os.environ["CUDA_VISIBLE_DEVICES"] = "-1" logging.basicConfig( level=logging.WARNING, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) @pytest.mark.parametrize( "var_train_expr, config_path", [ ("embeddings|decoder_cell", "./examples/tacotron2/conf/tacotron2.v1.yaml"), (None, "./examples/tacotron2/conf/tacotron2.v1.yaml"), ( "embeddings|decoder_cell", "./examples/tacotron2/conf/tacotron2.baker.v1.yaml", ), ("embeddings|decoder_cell", "./examples/tacotron2/conf/tacotron2.kss.v1.yaml"), ], ) def test_tacotron2_train_some_layers(var_train_expr, config_path): config = Tacotron2Config(n_speakers=5, reduction_factor=1) model = TFTacotron2(config, name="tacotron2") model._build() optimizer = tf.keras.optimizers.Adam(lr=0.001) with open(config_path) as f: config = yaml.load(f, Loader=yaml.Loader) config.update({"outdir": "./"}) config.update({"var_train_expr": var_train_expr}) STRATEGY = return_strategy() trainer = Tacotron2Trainer( config=config, strategy=STRATEGY, steps=0, epochs=0, is_mixed_precision=False, ) trainer.compile(model, optimizer) len_trainable_vars = len(trainer._trainable_variables) all_trainable_vars = len(model.trainable_variables) if var_train_expr is None: tf.debugging.assert_equal(len_trainable_vars, all_trainable_vars) else: tf.debugging.assert_less(len_trainable_vars, all_trainable_vars) @pytest.mark.parametrize( "n_speakers, n_chars, max_input_length, max_mel_length, batch_size", [(2, 15, 25, 50, 2),], ) def test_tacotron2_trainable( n_speakers, n_chars, max_input_length, max_mel_length, batch_size ): config = Tacotron2Config(n_speakers=n_speakers, reduction_factor=1) model = TFTacotron2(config, name="tacotron2") model._build() # fake input input_ids = tf.random.uniform( [batch_size, max_input_length], maxval=n_chars, dtype=tf.int32 ) speaker_ids = tf.convert_to_tensor([0] * batch_size, tf.int32) mel_gts = tf.random.uniform(shape=[batch_size, max_mel_length, 80]) mel_lengths = np.random.randint( max_mel_length, high=max_mel_length + 1, size=[batch_size] ) mel_lengths[-1] = max_mel_length mel_lengths = tf.convert_to_tensor(mel_lengths, dtype=tf.int32) stop_tokens = np.zeros((batch_size, max_mel_length), np.float32) stop_tokens = tf.convert_to_tensor(stop_tokens) optimizer = tf.keras.optimizers.Adam(lr=0.001) binary_crossentropy = tf.keras.losses.BinaryCrossentropy(from_logits=True) @tf.function(experimental_relax_shapes=True) def one_step_training(input_ids, speaker_ids, mel_gts, mel_lengths): with tf.GradientTape() as tape: mel_preds, post_mel_preds, stop_preds, alignment_history = model( input_ids, tf.constant([max_input_length, max_input_length]), speaker_ids, mel_gts, mel_lengths, training=True, ) loss_before = tf.keras.losses.MeanSquaredError()(mel_gts, mel_preds) loss_after = tf.keras.losses.MeanSquaredError()(mel_gts, post_mel_preds) stop_gts = tf.expand_dims( tf.range(tf.reduce_max(mel_lengths), dtype=tf.int32), 0 ) # [1, max_len] stop_gts = tf.tile(stop_gts, [tf.shape(mel_lengths)[0], 1]) # [B, max_len] stop_gts = tf.cast( tf.math.greater_equal(stop_gts, tf.expand_dims(mel_lengths, 1) - 1), tf.float32, ) # calculate stop_token loss stop_token_loss = binary_crossentropy(stop_gts, stop_preds) loss = stop_token_loss + loss_before + loss_after gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) return loss, alignment_history for i in range(2): if i == 1: start = time.time() loss, alignment_history = one_step_training( input_ids, speaker_ids, mel_gts, mel_lengths ) print(f" > loss: {loss}") total_runtime = time.time() - start print(f" > Total run-time: {total_runtime}") print(f" > Avg run-time: {total_runtime/10}")
5,329
34.533333
87
py
TensorFlowTTS
TensorFlowTTS-master/test/test_melgan_layers.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import logging import os import numpy as np import pytest import tensorflow as tf from tensorflow_tts.models.melgan import ( TFConvTranspose1d, TFReflectionPad1d, TFResidualStack, ) os.environ["CUDA_VISIBLE_DEVICES"] = "" logging.basicConfig( level=logging.DEBUG, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) @pytest.mark.parametrize("padding_size", [(3), (5)]) def test_padding(padding_size): fake_input_1d = tf.random.normal(shape=[4, 8000, 256], dtype=tf.float32) out = TFReflectionPad1d(padding_size=padding_size)(fake_input_1d) assert np.array_equal( tf.keras.backend.int_shape(out), [4, 8000 + 2 * padding_size, 256] ) @pytest.mark.parametrize( "filters,kernel_size,strides,padding,is_weight_norm", [(512, 40, 8, "same", False), (768, 15, 8, "same", True)], ) def test_convtranpose1d(filters, kernel_size, strides, padding, is_weight_norm): fake_input_1d = tf.random.normal(shape=[4, 8000, 256], dtype=tf.float32) conv1d_transpose = TFConvTranspose1d( filters=filters, kernel_size=kernel_size, strides=strides, padding=padding, is_weight_norm=is_weight_norm, initializer_seed=42, ) out = conv1d_transpose(fake_input_1d) assert np.array_equal(tf.keras.backend.int_shape(out), [4, 8000 * strides, filters]) @pytest.mark.parametrize( "kernel_size,filters,dilation_rate,use_bias,nonlinear_activation,nonlinear_activation_params,is_weight_norm", [ (3, 256, 1, True, "LeakyReLU", {"alpha": 0.3}, True), (3, 256, 3, True, "ReLU", {}, False), ], ) def test_residualblock( kernel_size, filters, dilation_rate, use_bias, nonlinear_activation, nonlinear_activation_params, is_weight_norm, ): fake_input_1d = tf.random.normal(shape=[4, 8000, 256], dtype=tf.float32) residual_block = TFResidualStack( kernel_size=kernel_size, filters=filters, dilation_rate=dilation_rate, use_bias=use_bias, nonlinear_activation=nonlinear_activation, nonlinear_activation_params=nonlinear_activation_params, is_weight_norm=is_weight_norm, initializer_seed=42, ) out = residual_block(fake_input_1d) assert np.array_equal(tf.keras.backend.int_shape(out), [4, 8000, filters])
2,965
30.892473
113
py
TensorFlowTTS
TensorFlowTTS-master/test/test_fastspeech2.py
# -*- coding: utf-8 -*- # Copyright 2020 Minh Nguyen (@dathudeptrai) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import logging import os import yaml import pytest import tensorflow as tf from tensorflow_tts.configs import FastSpeech2Config from tensorflow_tts.models import TFFastSpeech2 from tensorflow_tts.utils import return_strategy from examples.fastspeech2.train_fastspeech2 import FastSpeech2Trainer os.environ["CUDA_VISIBLE_DEVICES"] = "" logging.basicConfig( level=logging.DEBUG, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s", ) @pytest.mark.parametrize("new_size", [100, 200, 300]) def test_fastspeech_resize_positional_embeddings(new_size): config = FastSpeech2Config() fastspeech2 = TFFastSpeech2(config, name="fastspeech") fastspeech2._build() fastspeech2.save_weights("./test.h5") fastspeech2.resize_positional_embeddings(new_size) fastspeech2.load_weights("./test.h5", by_name=True, skip_mismatch=True) @pytest.mark.parametrize( "var_train_expr, config_path", [ (None, "./examples/fastspeech2/conf/fastspeech2.v1.yaml"), ("embeddings|encoder", "./examples/fastspeech2/conf/fastspeech2.v1.yaml"), ("embeddings|encoder", "./examples/fastspeech2/conf/fastspeech2.v2.yaml"), ("embeddings|encoder", "./examples/fastspeech2/conf/fastspeech2.baker.v2.yaml"), ("embeddings|encoder", "./examples/fastspeech2/conf/fastspeech2.kss.v1.yaml"), ("embeddings|encoder", "./examples/fastspeech2/conf/fastspeech2.kss.v2.yaml"), ], ) def test_fastspeech2_train_some_layers(var_train_expr, config_path): config = FastSpeech2Config(n_speakers=5) model = TFFastSpeech2(config) model._build() optimizer = tf.keras.optimizers.Adam(lr=0.001) with open(config_path) as f: config = yaml.load(f, Loader=yaml.Loader) config.update({"outdir": "./"}) config.update({"var_train_expr": var_train_expr}) STRATEGY = return_strategy() trainer = FastSpeech2Trainer( config=config, strategy=STRATEGY, steps=0, epochs=0, is_mixed_precision=False, ) trainer.compile(model, optimizer) len_trainable_vars = len(trainer._trainable_variables) all_trainable_vars = len(model.trainable_variables) if var_train_expr is None: tf.debugging.assert_equal(len_trainable_vars, all_trainable_vars) else: tf.debugging.assert_less(len_trainable_vars, all_trainable_vars) @pytest.mark.parametrize("num_hidden_layers,n_speakers", [(2, 1), (3, 2), (4, 3)]) def test_fastspeech_trainable(num_hidden_layers, n_speakers): config = FastSpeech2Config( encoder_num_hidden_layers=num_hidden_layers, decoder_num_hidden_layers=num_hidden_layers + 1, n_speakers=n_speakers, ) fastspeech2 = TFFastSpeech2(config, name="fastspeech") optimizer = tf.keras.optimizers.Adam(lr=0.001) # fake inputs input_ids = tf.convert_to_tensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]], tf.int32) attention_mask = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32) speaker_ids = tf.convert_to_tensor([0], tf.int32) duration_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32) f0_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.float32) energy_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.float32) mel_gts = tf.random.uniform(shape=[1, 10, 80], dtype=tf.float32) @tf.function def one_step_training(): with tf.GradientTape() as tape: mel_outputs_before, _, duration_outputs, _, _ = fastspeech2( input_ids, speaker_ids, duration_gts, f0_gts, energy_gts, training=True, ) duration_loss = tf.keras.losses.MeanSquaredError()( duration_gts, duration_outputs ) mel_loss = tf.keras.losses.MeanSquaredError()(mel_gts, mel_outputs_before) loss = duration_loss + mel_loss gradients = tape.gradient(loss, fastspeech2.trainable_variables) optimizer.apply_gradients(zip(gradients, fastspeech2.trainable_variables)) tf.print(loss) import time for i in range(2): if i == 1: start = time.time() one_step_training() print(time.time() - start)
4,805
35.969231
88
py
sgmcmc_ssm_code
sgmcmc_ssm_code-master/sgmcmc_ssm/sgmcmc_sampler.py
import numpy as np import pandas as pd import time from datetime import timedelta import logging from .evaluator import BaseEvaluator logger = logging.getLogger(name=__name__) NOISE_NUGGET=1e-9 # SGMCMCSampler class SGMCMCSampler(object): """ Base Class for SGMCMC with Time Series """ def __init__(self, **kwargs): raise NotImplementedError() ## Init Functions def setup(self, **kwargs): # Depreciated raise NotImplementedError() def prior_init(self): self.parameters = self.prior.sample_prior() return self.parameters ## Loglikelihood Functions def exact_loglikelihood(self, tqdm=None): """ Return the exact loglikelihood given the current parameters """ loglikelihood = self.message_helper.marginal_loglikelihood( observations=self.observations, parameters=self.parameters, forward_message=self.forward_message, backward_message=self.backward_message, tqdm=tqdm, ) return loglikelihood def exact_logjoint(self, return_loglike=False, tqdm=None): """ Return the loglikelihood + logprior given the current parameters """ loglikelihood = self.exact_loglikelihood(tqdm=tqdm) logprior = self.prior.logprior(self.parameters) if return_loglike: return dict( logjoint=loglikelihood + logprior, loglikelihood=loglikelihood, ) else: return loglikelihood + logprior def predictive_loglikelihood(self, kind='marginal', num_steps_ahead=10, subsequence_length=-1, minibatch_size=1, buffer_length=10, num_samples=1000, parameters=None, observations=None, **kwargs): """ Return the predictive loglikelihood given the parameters """ if parameters is None: parameters = self.parameters observations = self._get_observations(observations) T = observations.shape[0] if kind == 'marginal': pred_loglikelihood = 0.0 for s in range(0, minibatch_size): out = self._random_subsequence_and_buffers(buffer_length, subsequence_length=subsequence_length, T=T) forward_message = self.message_helper.forward_message( observations[ out['left_buffer_start']:out['subsequence_start'] ], self.parameters, forward_message=self.forward_message, tqdm=kwargs.get('tqdm', None), ) # Noisy Loglikelihood should use only forward pass # E.g. log Pr(y) \approx \sum_s log Pr(y_s | y<min(s)) pred_loglikelihood_S = ( self.message_helper.predictive_loglikelihood( observations=observations, parameters=parameters, forward_message=forward_message, backward_message=self.backward_message, lag=num_steps_ahead, tqdm=kwargs.get('tqdm', None), )) pred_loglikelihood += ( pred_loglikelihood_S * (T-num_steps_ahead)/( out['subsequence_end'] - out['subsequence_start'] - \ num_steps_ahead )) pred_loglikelihood *= 1.0/minibatch_size return pred_loglikelihood elif kind == 'pf': if kwargs.get("N", None) is None: kwargs['N'] = num_samples pred_loglikelihood = np.zeros(num_steps_ahead+1) for s in range(0, minibatch_size): out = self._random_subsequence_and_buffers( buffer_length=buffer_length, subsequence_length=subsequence_length, T=T) relative_start = (out['subsequence_start'] - out['left_buffer_start']) relative_end = (out['subsequence_end'] - out['left_buffer_start']) buffer_ = observations[ out['left_buffer_start']: out['right_buffer_end'] ] pred_loglike_add = ( self.message_helper .pf_predictive_loglikelihood_estimate( observations=buffer_, parameters=self.parameters, num_steps_ahead=num_steps_ahead, subsequence_start=relative_start, subsequence_end=relative_end, **kwargs) ) for ll in range(num_steps_ahead+1): pred_loglikelihood[ll] += pred_loglike_add[ll] * (T-ll)/( out['subsequence_end'] - out['subsequence_start']-ll ) pred_loglikelihood *= 1.0/minibatch_size return pred_loglikelihood else: raise ValueError("Unrecognized kind = {0}".format(kind)) def noisy_loglikelihood(self, kind='marginal', subsequence_length=-1, minibatch_size=1, buffer_length=10, num_samples=None, observations=None, **kwargs): """ Subsequence Approximation to loglikelihood Args: kind (string): how to estimate the loglikelihood subsequence_length (int): length of subsequence used in evaluation minibatch_size (int): number of subsequences buffer_length (int): length of each subsequence buffer """ observations = self._get_observations(observations) T = observations.shape[0] noisy_loglikelihood = 0.0 if kind == 'marginal': for s in range(0, minibatch_size): out = self._random_subsequence_and_buffers(buffer_length, subsequence_length=subsequence_length, T=T) forward_message = self.message_helper.forward_message( observations[ out['left_buffer_start']:out['subsequence_start'] ], self.parameters, forward_message=self.forward_message, tqdm=kwargs.get('tqdm', None), ) # Noisy Loglikelihood should use only forward pass # E.g. log Pr(y) \approx \sum_s log Pr(y_s | y<min(s)) noisy_loglikelihood += ( self.message_helper.marginal_loglikelihood( observations=observations[ out['subsequence_start']:out['subsequence_end'] ], parameters=self.parameters, weights=out['weights'], forward_message=forward_message, backward_message=self.backward_message, tqdm=kwargs.get('tqdm', None), ) - forward_message['log_constant'] ) elif kind == 'complete': for s in range(0, minibatch_size): out = self._random_subsequence_and_buffers( buffer_length=buffer_length, subsequence_length=subsequence_length, T=T) buffer_ = observations[ out['left_buffer_start']:out['right_buffer_end'] ] # Draw Samples: latent_buffer = self.sample_x( parameters=self.parameters, observations=buffer_, num_samples=num_samples, ) relative_start = out['subsequence_start']-out['left_buffer_start'] relative_end = out['subsequence_end']-out['left_buffer_start'] forward_message = {} if relative_start > 0: forward_message = dict( x_prev = latent_buffer[relative_start-1] ) noisy_loglikelihood += \ self.message_helper.complete_data_loglikelihood( observations=observations[ out['subsequence_start']:out['subsequence_end'] ], latent_vars=latent_buffer[relative_start:relative_end], weights=out['weights'], parameters=self.parameters, forward_message=forward_message, ) elif kind == 'pf': if kwargs.get("N", None) is None: kwargs['N'] = num_samples noisy_loglikelihood = 0.0 for s in range(0, minibatch_size): out = self._random_subsequence_and_buffers( buffer_length=buffer_length, subsequence_length=subsequence_length, T=T) relative_start = (out['subsequence_start'] - out['left_buffer_start']) relative_end = (out['subsequence_end'] - out['left_buffer_start']) buffer_ = observations[ out['left_buffer_start']: out['right_buffer_end'] ] noisy_loglikelihood += ( self.message_helper .pf_loglikelihood_estimate( observations=buffer_, parameters=self.parameters, weights=out['weights'], subsequence_start=relative_start, subsequence_end=relative_end, **kwargs) ) else: raise ValueError("Unrecognized kind = {0}".format(kind)) noisy_loglikelihood *= 1.0/minibatch_size if np.isnan(noisy_loglikelihood): raise ValueError("NaNs in loglikelihood") return noisy_loglikelihood def noisy_logjoint(self, return_loglike=False, **kwargs): """ Return the loglikelihood + logprior given the current parameters """ loglikelihood = self.noisy_loglikelihood(**kwargs) logprior = self.prior.logprior(self.parameters) if return_loglike: return dict( logjoint=loglikelihood + logprior, loglikelihood=loglikelihood, ) else: return loglikelihood + logprior ## Gradient Functions def _random_subsequence_and_buffers(self, buffer_length, subsequence_length, T=None): """ Get a subsequence and the forward and backward message approx""" if T is None: T = self._get_T() if buffer_length == -1: buffer_length = T if (subsequence_length == -1) or (T-subsequence_length <= 0): subsequence_start = 0 subsequence_end = T weights = None else: subsequence_start, subsequence_end, weights = \ random_subsequence_and_weights( S=subsequence_length, T=T, partition_style=self.options.get('partition_style'), ) left_buffer_start = max(0, subsequence_start - buffer_length) right_buffer_end = min(T, subsequence_end + buffer_length) out = dict( subsequence_start = subsequence_start, subsequence_end = subsequence_end, left_buffer_start = left_buffer_start, right_buffer_end = right_buffer_end, weights = weights, ) return out def _single_noisy_grad_loglikelihood(self, buffer_dict, kind='marginal', num_samples=None, observations=None, parameters=None, **kwargs): # buffer_dict is the output of _random_subsequence_and_buffers observations = self._get_observations(observations, check_shape=False) if parameters is None: parameters = self.parameters T = observations.shape[0] if kind == 'marginal': forward_message = self.message_helper.forward_message( observations[ buffer_dict['left_buffer_start']: buffer_dict['subsequence_start'] ], parameters, forward_message=self.forward_message) backward_message = self.message_helper.backward_message( observations[ buffer_dict['subsequence_end']: buffer_dict['right_buffer_end'] ], parameters, backward_message=self.backward_message, ) noisy_grad = ( self.message_helper .gradient_marginal_loglikelihood( observations=observations[ buffer_dict['subsequence_start']: buffer_dict['subsequence_end'] ], parameters=parameters, weights=buffer_dict['weights'], forward_message=forward_message, backward_message=backward_message, **kwargs ) ) elif kind == 'complete': buffer_ = observations[ buffer_dict['left_buffer_start']: buffer_dict['right_buffer_end'] ] # Draw Samples: latent_buffer = self.sample_x( parameters=parameters, observations=buffer_, num_samples=num_samples, ) relative_start = (buffer_dict['subsequence_start'] - buffer_dict['left_buffer_start']) relative_end = (buffer_dict['subsequence_end'] - buffer_dict['left_buffer_start']) forward_message = {} if relative_start > 0: forward_message = dict( x_prev = latent_buffer[relative_start-1] ) noisy_grad = ( self.message_helper .gradient_complete_data_loglikelihood( observations=observations[ buffer_dict['subsequence_start']: buffer_dict['subsequence_end'] ], latent_vars=latent_buffer[relative_start:relative_end], parameters=parameters, weights=buffer_dict['weights'], forward_message=forward_message, **kwargs) ) elif kind == 'pf': if kwargs.get("N", None) is None: kwargs['N'] = num_samples relative_start = (buffer_dict['subsequence_start'] - buffer_dict['left_buffer_start']) relative_end = (buffer_dict['subsequence_end'] - buffer_dict['left_buffer_start']) buffer_ = observations[ buffer_dict['left_buffer_start']: buffer_dict['right_buffer_end'] ] noisy_grad = ( self.message_helper .pf_gradient_estimate( observations=buffer_, parameters=self.parameters, subsequence_start=relative_start, subsequence_end=relative_end, weights=buffer_dict['weights'], **kwargs) ) else: raise ValueError("Unrecognized kind = {0}".format(kind)) return noisy_grad def _noisy_grad_loglikelihood(self, subsequence_length=-1, minibatch_size=1, buffer_length=0, observations=None, buffer_dicts=None, **kwargs): observations = self._get_observations(observations, check_shape=False) T = observations.shape[0] if buffer_dicts is None: buffer_dicts = [ self._random_subsequence_and_buffers( buffer_length=buffer_length, subsequence_length=subsequence_length, T=T) for _ in range(minibatch_size) ] elif len(buffer_dicts) != minibatch_size: raise ValueError("len(buffer_dicts != minibatch_size") noisy_grad = {var: np.zeros_like(value) for var, value in self.parameters.as_dict().items()} for s in range(0, minibatch_size): noisy_grad_add = self._single_noisy_grad_loglikelihood( buffer_dict=buffer_dicts[s], observations=observations, **kwargs, ) for var in noisy_grad: noisy_grad[var] += noisy_grad_add[var] * 1.0/minibatch_size if np.any(np.isnan(noisy_grad[var])): raise ValueError("NaNs in gradient of {0}".format(var)) if np.linalg.norm(noisy_grad[var]) > 1e16: logger.warning("Norm of noisy_grad_loglike[{1} > 1e16: {0}".format( noisy_grad[var], var)) return noisy_grad def noisy_gradient(self, preconditioner=None, is_scaled=True, **kwargs): """ Noisy Gradient Estimate noisy_gradient = -grad tilde{U}(theta) = grad marginal loglike + grad logprior Monte Carlo Estimate of gradient (using buffering) Args: preconditioner (object): preconditioner for gradients is_scaled (boolean): scale gradient by 1/T **kwargs: arguments for `self._noisy_grad_loglikelihood()` For example: minibatch_size, buffer_length, use_analytic Returns: noisy_gradient (dict): dict of gradient vectors """ noisy_grad_loglike = \ self._noisy_grad_loglikelihood(**kwargs) noisy_grad_prior = self.prior.grad_logprior( parameters=kwargs.get('parameters',self.parameters), **kwargs ) noisy_gradient = {var: noisy_grad_prior[var] + noisy_grad_loglike[var] for var in noisy_grad_prior} if preconditioner is None: if is_scaled: for var in noisy_gradient: noisy_gradient[var] /= self._get_T(**kwargs) else: scale = 1.0/self._get_T(**kwargs) if is_scaled else 1.0 noisy_gradient = preconditioner.precondition(noisy_gradient, parameters=kwargs.get('parameters',self.parameters), scale=scale) return noisy_gradient ## Sampler/Optimizer Step Functions def step_sgd(self, epsilon, **kwargs): """ One step of Stochastic Gradient Descent (Learns the MAP, not a sample from the posterior) Args: epsilon (double): step size **kwargs (kwargs): to pass to self.noisy_gradient minibatch_size (int): number of subsequences to sample from buffer_length (int): length of buffer to use Returns: parameters (Parameters): sampled parameters after one step """ delta = self.noisy_gradient(**kwargs) for var in self.parameters.var_dict: self.parameters.var_dict[var] += epsilon * delta[var] return self.parameters def step_precondition_sgd(self, epsilon, preconditioner, **kwargs): """ One Step of Preconditioned Stochastic Gradient Descent Args: epsilon (double): step size preconditioner (object): preconditioner **kwargs (kwargs): to pass to self.noisy_gradient minibatch_size (int): number of subsequences to sample from buffer_length (int): length of buffer to use Returns: parameters (Parameters): sampled parameters after one step """ delta = self.noisy_gradient(preconditioner=preconditioner, **kwargs) for var in self.parameters.var_dict: self.parameters.var_dict[var] += epsilon * delta[var] return self.parameters def step_adagrad(self, epsilon, **kwargs): """ One step of adagrad (Learns the MAP, not a sample from the posterior) Args: epsilon (double): step size **kwargs (kwargs): to pass to self.noisy_gradient """ if not hasattr(self, "_adagrad_moments"): self._adagrad_moments = dict(t=0, G=0.0) g = self.parameters.from_dict_to_vector(self.noisy_gradient(**kwargs)) t = self._adagrad_moments['t'] + 1 G = self._adagrad_moments['G'] + g**2 delta_vec = g/np.sqrt(G + NOISE_NUGGET) delta = self.parameters.from_vector_to_dict(delta_vec, **self.parameters.dim) for var in self.parameters.var_dict: self.parameters.var_dict[var] += epsilon * delta[var] self._adagrad_moments['t'] = t self._adagrad_moments['G'] = G return self.parameters def _get_sgmcmc_noise(self, is_scaled=True, preconditioner=None, **kwargs): if is_scaled: scale = 1.0 / self._get_T(**kwargs) else: scale = 1.0 if preconditioner is not None: white_noise = preconditioner.precondition_noise( parameters=self.parameters, scale=scale, ) else: white_noise = {var: np.random.normal( loc=0, scale=np.sqrt(scale), size=value.shape ) for var, value in self.parameters.as_dict().items()} return white_noise def sample_sgld(self, epsilon, **kwargs): """ One Step of Stochastic Gradient Langevin Dynamics Args: epsilon (double): step size **kwargs (kwargs): to pass to self.noisy_gradient Returns: parameters (Parameters): sampled parameters after one step """ if "preconditioner" in kwargs: raise ValueError("Use SGRLD instead") delta = self.noisy_gradient(**kwargs) white_noise = self._get_sgmcmc_noise(**kwargs) for var in self.parameters.var_dict: self.parameters.var_dict[var] += \ epsilon * delta[var] + np.sqrt(2.0*epsilon) * white_noise[var] return self.parameters def sample_sgld_cv(self, epsilon, centering_parameters, centering_gradient, **kwargs): """ One Step of Stochastic Gradient Langevin Dynamics with Control Variates grad = full_gradient(centering_parameters) + \ sub_gradient(parameters) - sub_gradient(centering_gradient) Args: epsilon (double): step size centering_parameters (Parameters): centering parameters centering_gradient (dict): full data grad of centering_parameters **kwargs (kwargs): to pass to self.noisy_gradient Returns: parameters (Parameters): sampled parameters after one step """ if "preconditioner" in kwargs: raise ValueError("Use SGRLD instead") buffer_dicts = [ self._random_subsequence_and_buffers( buffer_length=kwargs.get('buffer_length', 0), subsequence_length=kwargs.get('subsequence_length', -1), T=self._get_T(**kwargs), ) for _ in range(kwargs.get('minibatch_size', 1)) ] cur_subseq_grad = self.noisy_gradient( buffer_dicts=buffer_dicts, **kwargs) centering_subseq_grad = self.noisy_gradient( parameters=centering_parameters, buffer_dicts=buffer_dicts, **kwargs) delta = {} for var in cur_subseq_grad.keys(): delta[var] = centering_gradient[var] + \ cur_subseq_grad[var] - centering_subseq_grad[var] white_noise = self._get_sgmcmc_noise(**kwargs) for var in self.parameters.var_dict: self.parameters.var_dict[var] += \ epsilon * delta[var] + np.sqrt(2.0*epsilon) * white_noise[var] return self.parameters def sample_sgrld(self, epsilon, preconditioner, **kwargs): """ One Step of Stochastic Gradient Riemannian Langevin Dynamics theta += epsilon * (D(theta) * grad_logjoint + correction_term) + \ N(0, 2 epsilon D(theta)) Args: epsilon (double): step size preconditioner (object): preconditioner Returns: parameters (Parameters): sampled parameters after one step """ if kwargs.get("is_scaled", True): scale = 1.0 / self._get_T(**kwargs) else: scale = 1.0 delta = self.noisy_gradient(preconditioner=preconditioner, **kwargs) white_noise = self._get_sgmcmc_noise( preconditioner=preconditioner, **kwargs) correction = preconditioner.correction_term( self.parameters, scale=scale) for var in self.parameters.var_dict: self.parameters.var_dict[var] += \ epsilon * (delta[var] + correction[var]) + \ np.sqrt(2.0*epsilon) * white_noise[var] return self.parameters def sample_gibbs(self): """ One Step of Blocked Gibbs Sampler Returns: parameters (Parameters): sampled parameters after one step """ raise NotImplementedError() def project_parameters(self, **kwargs): """ Project parameters to valid values + fix constants See **kwargs in __init__ for more details """ self.parameters.project_parameters(**self.options, **kwargs) return self.parameters ## Fit Functions def fit(self, iter_type, num_iters, output_all=False, observations=None, init_parameters=None, tqdm=None, catch_interrupt=False, **kwargs): """ Run multiple learning / inference steps Args: iter_type (string): 'SGD', 'ADAGRAD', 'SGLD', 'SGRLD', 'Gibbs', etc. num_iters (int): number of steps output_all (bool): whether to output each iteration's parameters observations (ndarray): observations to fit on, optional init_parameters (Parameters): initial parameters, optional tqdm (tqdm): progress bar wrapper catch_interrupt (bool): terminate early on Ctrl-C **kwargs: for each iter e.g. steps_per_iter, epsilon, minibatch_size, subsequence_length, buffer_length, preconditioner, pf_kwargs, etc. see documentation for get_iter_step() Returns: (depends on output_all arg) parameters (Parameters): parameters_list (list of Parameters): length num_iters+1 """ if observations is not None: self.observations = observations if init_parameters is not None: self.parameters = init_parameters.copy() iter_func_names, iter_func_kwargs = \ self.get_iter_step(iter_type, tqdm=tqdm, **kwargs) if output_all: parameters_list = [None]*(num_iters+1) parameters_list[0] = self.parameters.copy() # Fit Loop pbar = range(1, num_iters+1) if tqdm is not None: pbar = tqdm(pbar) pbar.set_description("fit using {0} iters".format(iter_type)) for it in pbar: # Run iter funcs try: for func_name, func_kwargs in zip(iter_func_names, iter_func_kwargs): getattr(self, func_name)(**func_kwargs) if output_all: parameters_list[it] = self.parameters.copy() except KeyboardInterrupt as e: if catch_interrupt: logger.warning("Interrupt in fit:\n{0}\n".format(e) + \ "Stopping early after {0} iters".format(it)) if output_all: return parameters_list[:it] else: return self.parameters.copy() else: raise e if output_all: return parameters_list else: return self.parameters.copy() def fit_timed(self, iter_type, max_time=60, min_save_time=1, observations=None, init_parameters=None, tqdm=None, tqdm_iter=False, catch_interrupt=False, **kwargs): """ Run multiple learning / inference steps Args: iter_type (string): 'SGD', 'ADAGRAD', 'SGLD', 'SGRLD', 'Gibbs', etc. max_time (float): maxium time in seconds to run fit min_save_time (float): min time between saved parameters observations (ndarray): observations to fit on, optional init_parameters (Parameters): initial parameters, optional tqdm (tqdm): progress bar wrapper catch_interrupt (bool): terminate early on Ctrl-C **kwargs: for each iter e.g. steps_per_iter, epsilon, minibatch_size, subsequence_length, buffer_length, preconditioner, pf_kwargs, etc. see documentation for get_iter_step() Returns: parameters_list (list of Parameters): times (list of float): fit time for each parameter """ parameters_list, times, _ = self.fit_evaluate( iter_type=iter_type, max_time=max_time, min_save_time=min_save_time, observations=observations, init_parameters=init_parameters, tqdm=tqdm, tqdm_iter=tqdm_iter, catch_interrupt=catch_interrupt, **kwargs) return parameters_list['parameters'].tolist(), times['time'].tolist() def fit_evaluate(self, iter_type, metric_functions=None, max_num_iters=None, max_time=60, min_save_time=1, observations=None, init_parameters=None, tqdm=None, tqdm_iter=False, catch_interrupt=False, total_max_time=None, **kwargs): """ Run multiple learning / inference steps with evaluator Args: iter_type (string): 'SGD', 'ADAGRAD', 'SGLD', 'SGRLD', 'Gibbs', etc. metric_functions (func or list of funcs): evaluation functions Each function takes a sampler and returns a dict or list of dict dict(metric=string, variable=string, value=double) for each See metric_functions.py for examples max_num_iters (int): maximum number of iterations to save max_time (float): maxium time in seconds to run sampler does *not* include time used by evaluator min_save_time (float): min time between saved parameters observations (ndarray): observations to fit on, optional init_parameters (Parameters): initial parameters, optional tqdm (tqdm): progress bar wrapper tqdm_iter (bool): progress bar for each iteration catch_interrupt (bool): terminate early on Ctrl-C total_max_time (float): maximum time in seconds to run fit_evaluate **kwargs: for each iter e.g. steps_per_iter, epsilon, minibatch_size, subsequence_length, buffer_length, preconditioner, pf_kwargs, etc. see documentation for get_iter_step() Returns: parameters_list (pd.DataFrame): parameters saved columns: iteration: number of iter_func_kwargs steps called parameters: Parameters times (pd.DataFrame): fit time for each saved parameter columns: iteration: number of iter_func_kwargs steps called time: time used by iter_func_kwargs metrics (pd.DataFrame): metric for each saved parameter columns: iteration: number of iter_func_kwargs steps called metric: name of metric variable: name of variable value: value of metric for variable """ if observations is not None: self.observations = observations if init_parameters is not None: self.parameters = init_parameters.copy() evaluator = BaseEvaluator( sampler=self, metric_functions=metric_functions, ) iter_func_names, iter_func_kwargs = \ self.get_iter_step(iter_type, tqdm=tqdm, **kwargs) if tqdm_iter: iter_func_kwargs[0]['tqdm'] = tqdm num_iters = max_time//min_save_time if max_num_iters is not None: num_iters = min(num_iters, max_num_iters) iteration = 0 total_time = 0 parameters_list = [None]*(num_iters+1) times = np.zeros(num_iters+1)*np.nan iterations = np.zeros(num_iters+1, dtype=int) fit_start_time = time.time() last_save_time = time.time() parameters_list[0] = self.parameters.copy() times[0] = total_time iterations[0] = iteration evaluator.eval_metric_functions(iteration=iteration) # Fit Loop pbar = range(1, num_iters+1) if tqdm is not None: pbar = tqdm(pbar) for it in pbar: if tqdm is not None: pbar.set_description("fit using {0}".format(iter_type) + \ " on iter {0}".format(iteration) ) try: for step in range(1000): # Run iter funcs for func_name, func_kwargs in zip(iter_func_names, iter_func_kwargs): getattr(self, func_name)(**func_kwargs) if time.time() - last_save_time > min_save_time: parameters_list[it] = self.parameters.copy() total_time += time.time() - last_save_time times[it] = total_time iteration += step + 1 iterations[it] = iteration evaluator.eval_metric_functions(iteration=iteration) last_save_time = time.time() break except KeyboardInterrupt as e: if catch_interrupt: logger.warning("Interrupt in fit_timed:\n{0}\n".format(e) + \ "Stopping early after {0} iters".format(it)) break else: raise e if total_time > max_time: # Break it total time on iter_funcs exceeds max time break if total_max_time is not None: if fit_start_time - time.time() > total_max_time: # Break it total time on fit_evalute exceeds total max time break valid = np.sum(~np.isnan(times)) parameters_list = pd.DataFrame(dict( iteration = iterations[0:valid], parameters = parameters_list[0:valid], )) times = pd.DataFrame(dict( iteration = iterations[0:valid], time = times[0:valid], )) metric = evaluator.get_metrics() return parameters_list, times, metric def get_iter_step(self, iter_type, steps_per_iteration=1, **kwargs): # Returns iter_func_names, iter_func_kwargs project_kwargs = kwargs.get("project_kwargs",{}) if iter_type == 'Gibbs': iter_func_names = ["sample_gibbs", "project_parameters"] iter_func_kwargs = [{}, project_kwargs] elif iter_type == 'custom': iter_func_names = kwargs.get("iter_func_names") iter_func_kwargs = kwargs.get("iter_func_kwargs") elif iter_type in ['SGD', 'ADAGRAD', 'SGLD', 'SGRD', 'SGRLD']: grad_kwargs = dict( epsilon = kwargs['epsilon'], subsequence_length = kwargs['subsequence_length'], buffer_length = kwargs['buffer_length'], minibatch_size = kwargs.get('minibatch_size', 1), kind = kwargs.get("kind", "marginal"), num_samples = kwargs.get("num_samples", None), **kwargs.get("pf_kwargs", {}) ) if 'num_sequences' in kwargs: grad_kwargs['num_sequences'] = kwargs['num_sequences'] if 'use_scir' in kwargs: grad_kwargs['use_scir'] = kwargs['use_scir'] if iter_type == 'SGD': iter_func_names = ['step_sgd', 'project_parameters'] iter_func_kwargs = [grad_kwargs, project_kwargs] elif iter_type == 'ADAGRAD': iter_func_names = ['step_adagrad', 'project_parameters'] iter_func_kwargs = [grad_kwargs, project_kwargs] elif iter_type == 'SGLD': iter_func_names = ['sample_sgld', 'project_parameters'] iter_func_kwargs = [grad_kwargs, project_kwargs] elif iter_type == 'SGRD': grad_kwargs['preconditioner'] = self._get_preconditioner( kwargs.get('preconditioner') ) iter_func_names = ['step_precondition_sgd', 'project_parameters'] iter_func_kwargs = [grad_kwargs, project_kwargs] elif iter_type == 'SGRLD': grad_kwargs['preconditioner'] = self._get_preconditioner( kwargs.get('preconditioner') ) iter_func_names = ['sample_sgrld', 'project_parameters'] iter_func_kwargs = [grad_kwargs, project_kwargs] else: raise ValueError("Unrecognized iter_type {0}".format(iter_type)) iter_func_names = iter_func_names * steps_per_iteration iter_func_kwargs = iter_func_kwargs * steps_per_iteration return iter_func_names, iter_func_kwargs def _get_preconditioner(self, preconditioner=None): if preconditioner is None: raise NotImplementedError("No Default Preconditioner for {}".format( self.name)) return preconditioner ## Predict Functions def predict(self, target='latent', distr=None, lag=None, return_distr=None, num_samples=None, kind='analytic', observations=None, parameters=None, **kwargs): """ Make predictions based on fit Args: target (string): variable to predict 'latent' - latent variables 'y' - observation variables distr (string): what distribution to sample/target 'marginal' - marginal (default for return_distr) 'joint' - joint (default for sampling) lag (int): distribution is p(U_t | Y_{1:t+lag}) default/None -> use all observations return_distr (bool): return distribution (default is True if num_samples is None otherwise True) num_samples (int): number of samples return kind (string): how to calculate distribution 'analytic' - use message passing 'pf' - use particle filter/smoother observations (ndarray): observations to use parameters (Parameters): parameters kwargs: key word arguments tqdm (tqdm): progress bar see message_helper.latent_var_distr, message_helper.y_distr, message_helper.latent_var_sample, message_helper.y_sample, message_helper.pf_latent_var_distr, message_helper.pf_y_distr, for more details Returns: Depends on target, return_distr, num_samples """ observations = self._get_observations(observations) if parameters is None: parameters = self.parameters if return_distr is None: if kind == 'pf': return_distr = True else: return_distr = (num_samples is None) if kind == 'analytic': if return_distr: if distr is None: distr = 'marginal' if target == 'latent': return self.message_helper.latent_var_distr( distr=distr, lag=lag, observations=observations, parameters=parameters, **kwargs, ) elif target == 'y': return self.message_helper.y_distr( distr=distr, lag=lag, observations=observations, parameters=parameters, **kwargs, ) else: raise ValueError("Unrecognized target '{0}'".format(target)) else: if distr is None: distr = 'joint' if target == 'latent': return self.message_helper.latent_var_sample( distr=distr, lag=lag, num_samples=num_samples, observations=observations, parameters=parameters, **kwargs, ) elif target == 'y': return self.message_helper.y_sample( distr=distr, lag=lag, num_samples=num_samples, observations=observations, parameters=parameters, **kwargs, ) else: raise ValueError("Unrecognized target '{0}'".format(target)) elif kind == 'pf': if return_distr: if target == 'latent': return self.message_helper.pf_latent_var_distr( lag=lag, observations=observations, parameters=parameters, **kwargs, ) elif target == 'y': return self.message_helper.pf_y_distr( distr=distr, lag=lag, observations=observations, parameters=parameters, **kwargs, ) else: raise ValueError("Unrecognized target '{0}'".format(target)) else: raise ValueError("return_distr must be True for kind = pf") else: raise ValueError("Unrecognized kind == '{0}'".format(kind)) def simulate(self, T, init_message=None, return_distr=False, num_samples=None, kind='analytic', observations=None, parameters=None, **kwargs): """ Simulate dynamics Args: T (int): length of simulated data init_message (dict): initial forward message return_distr (bool): return distribution (default is False) num_samples (int): number of samples return kind (string): how to calculate distribution 'analytic' - use message passing 'pf' - use particle filter/smoother observations (ndarray): observations parameters (Parameters): parameters Returns: dict with key values depending on return_distr and num_samples latent_vars (ndarray): simulated latent vars observations (ndarray): simulated observations latent_mean/latent_prob/latent_cov observation_mean/observation_prob/observations_cov """ observations = self._get_observations(observations) if parameters is None: parameters = self.parameters if kind == 'analytic': if init_message is None: init_message = self.message_helper.forward_message( observations=observations, parameters=parameters, ) if return_distr: return self.message_helper.simulate_distr( T=T, parameters=parameters, init_message=init_message, **kwargs ) else: return self.message_helper.simulate( T=T, parameters=parameters, init_message=init_message, num_samples=num_samples, **kwargs ) elif kind == 'pf': raise NotImplementedError() else: raise ValueError("Unrecognized kind == '{0}'".format(kind)) ## Attributes + Misc Helper Functions @property def observations(self): return self._observations @observations.setter def observations(self, observations): self._check_observation_shape(observations) self._observations = observations return def _check_observation_shape(self, observations): return def _get_observations(self, observations, check_shape=True): if observations is None: observations = self.observations if observations is None: raise ValueError("observations not specified") elif check_shape: self._check_observation_shape(observations) return observations def _get_T(self, **kwargs): T = kwargs.get('T') if T is None: observations = kwargs.get('observations') observations = self._get_observations(observations) T = observations.shape[0] return T # SeqSGMCMCSampler class SeqSGMCMCSampler(object): """ Mixin for handling a list of sequences """ def _get_T(self, **kwargs): T = kwargs.get('T') if T is None: observations = kwargs.get('observations') observations = self._get_observations(observations) T = np.sum(np.shape(observation)[0] for observation in observations) return T def _check_observation_shape(self, observations): if observations is not None: for ii, observation in enumerate(observations): try: super()._check_observation_shape(observations=observation) except ValueError as e: raise ValueError("Error in observations[{0}] :\n{1}".format( ii, e)) def exact_loglikelihood(self, observations=None, tqdm=None): """ Return exact loglikelihood over all observation sequences """ observations = self._get_observations(observations) loglikelihood = 0 pbar = observations if tqdm is not None: pbar = tqdm(pbar) pbar.set_description("Seq Loglikelihood") for observation in pbar: loglikelihood += self.message_helper.marginal_loglikelihood( observations=observation, parameters=self.parameters, forward_message=self.forward_message, backward_message=self.backward_message, tqdm=tqdm, ) return loglikelihood def noisy_loglikelihood(self, num_sequences=-1, observations=None, tqdm=None, **kwargs): """ Subsequence Approximation to loglikelihood Args: num_sequences (int): how many observation sequences to use (default = -1) is to use all observation sequences """ observations = self._get_observations(observations) loglikelihood = 0 S = 0.0 sequence_indices = np.arange(len(observations)) if num_sequences != -1: sequence_indices = np.random.choice( sequence_indices, num_sequences, replace=False, ) pbar = sequence_indices if tqdm is not None: pbar = tqdm(pbar) pbar.set_description("Seq Loglikelihood") for sequence_index in pbar: S += observations[sequence_index].shape[0] loglikelihood += super().noisy_loglikelihood( observations=observations[sequence_index], tqdm=tqdm, **kwargs) if num_sequences != -1: loglikelihood *= self._get_T(**kwargs) / S return loglikelihood def predictive_loglikelihood(self, num_sequences=-1, observations=None, tqdm=None, **kwargs): """ Return the predictive loglikelihood given the parameters """ observations = self._get_observations(observations) predictive_loglikelihood = 0 S = 0.0 sequence_indices = np.arange(len(observations)) if num_sequences != -1: sequence_indices = np.random.choice( sequence_indices, num_sequences, replace=False, ) pbar = sequence_indices if tqdm is not None: pbar = tqdm(pbar) pbar.set_description("Seq Pred Loglikelihood") for sequence_index in pbar: S += observations[sequence_index].shape[0] predictive_loglikelihood += super().predictive_loglikelihood( observations=observations[sequence_index], tqdm=tqdm, **kwargs) if num_sequences != -1: predictive_loglikelihood *= self._get_T(**kwargs) / S return predictive_loglikelihood def _noisy_grad_loglikelihood(self, num_sequences=-1, **kwargs): """ Subsequence approximation to gradient of loglikelihood Args: num_sequences (int): how many observation sequences to use (default = -1) is to use all observation sequences """ noisy_grad_loglike = None S = 0.0 sequence_indices = np.arange(len(self.observations)) if num_sequences != -1: sequence_indices = np.random.choice( sequence_indices, num_sequences, replace=False, ) for sequence_index in sequence_indices: noisy_grad_index = super()._noisy_grad_loglikelihood( observations=self.observations[sequence_index], **kwargs) S += self.observations[sequence_index].shape[0] if noisy_grad_loglike is None: noisy_grad_loglike = {var: noisy_grad_index[var] for var in noisy_grad_index.keys() } else: noisy_grad_loglike = { var: noisy_grad_loglike[var] + noisy_grad_index[var] for var in noisy_grad_index.keys() } if num_sequences != -1: noisy_grad_loglike = { var: noisy_grad_loglike[var] * self._get_T(**kwargs) / S for var in noisy_grad_index.keys() } return noisy_grad_loglike def predict(self, target='latent', distr=None, lag=None, return_distr=None, num_samples=None, kind='analytic', observations=None, parameters=None, tqdm=None, **kwargs): """ Make predictions based on fit Args: target (string): variable to predict 'latent' - latent variables 'y' - observation variables distr (string): what distribution to sample/target 'marginal' - marginal (default for return_distr) 'joint' - joint (default for sampling) lag (int): distribution is p(U_t | Y_{1:t+lag}) default/None -> use all observations return_distr (bool): return distribution (default is True if num_samples is None otherwise True) num_samples (int): number of samples return kind (string): how to calculate distribution 'analytic' - use message passing 'pf' - use particle filter/smoother observations (list of ndarray): observations to use parameters (Parameters): parameters kwargs: key word arguments tqdm (tqdm): progress bar see message_helper.latent_var_distr, message_helper.y_distr, message_helper.latent_var_sample, message_helper.y_sample, message_helper.pf_latent_var_distr, message_helper.pf_y_distr, for more details Returns: Depends on target, return_distr, num_samples """ observations = self._get_observations(observations) if parameters is None: parameters = self.parameters if return_distr is None: if kind == 'pf': return_distr = True else: return_distr = (num_samples is None) output = [] if tqdm is not None: kwargs['tqdm'] = tqdm observations = tqdm(observations, desc='sequence #') if kind == 'analytic': if return_distr: if distr is None: distr = 'marginal' if target == 'latent': for observation in observations: output.append( self.message_helper.latent_var_distr( distr=distr, lag=lag, observations=observation, parameters=parameters, **kwargs, ) ) elif target == 'y': for observation in observations: output.append( self.message_helper.y_distr( distr=distr, lag=lag, observations=observation, parameters=parameters, **kwargs, ) ) else: raise ValueError("Unrecognized target '{0}'".format(target)) else: if distr is None: distr = 'joint' if target == 'latent': for observation in observations: output.append( self.message_helper.latent_var_sample( distr=distr, lag=lag, num_samples=num_samples, observations=observation, parameters=parameters, **kwargs, ) ) elif target == 'y': for observation in observations: output.append( self.message_helper.y_sample( distr=distr, lag=lag, num_samples=num_samples, observations=observation, parameters=parameters, **kwargs, ) ) else: raise ValueError("Unrecognized target '{0}'".format(target)) elif kind == 'pf': if return_distr: if target == 'latent': for observation in observations: output.append( self.message_helper.pf_latent_var_distr( lag=lag, observations=observation, parameters=parameters, **kwargs, ) ) elif target == 'y': for observation in observations: output.append( self.message_helper.pf_y_distr( distr=distr, lag=lag, observations=observation, parameters=parameters, **kwargs, ) ) else: raise ValueError("Unrecognized target '{0}'".format(target)) else: raise ValueError("return_distr must be True for kind = pf") else: raise ValueError("Unrecognized kind == '{0}'".format(kind)) return output # SGMCMC Helper class SGMCMCHelper(object): """ Base Class for SGMCMC Helper """ def __init__(self, **kwargs): raise NotImplementedError() ## Message Passing Functions def forward_message(self, observations, parameters, forward_message=None, **kwargs): """ Calculate forward messages over the observations Pr(u_t | y_{<=t}) for y_t in observations Args: observations (ndarray): observations parameters (parameters): parameters forward_message (dict): latent state prior Pr(u_{-1} | y_{<=-1}) Returns: forward_message (dict): same format as forward_message """ if forward_message is None: forward_message = self.default_forward_message if np.shape(observations)[0] == 0: return forward_message forward_message = self._forward_messages( observations=observations, parameters=parameters, forward_message=forward_message, only_return_last=True, **kwargs ) return forward_message def backward_message(self, observations, parameters, backward_message=None, **kwargs): """ Calculate backward messages over the observations Pr(y_{>t} | u_t) for y_t in observations Args: observations (ndarray): observations parameters (parameters): parameters backward_message (dict): backward message Pr(y_{>T-1} | u_{T-1}) Returns: backward_message (dict): same format as forward_message """ if backward_message is None: backward_message = self.default_backward_message if np.shape(observations)[0] == 0: return backward_message backward_message = self._backward_messages( observations=observations, parameters=parameters, backward_message=backward_message, only_return_last=True, **kwargs ) return backward_message def forward_pass(self, observations, parameters, forward_message=None, include_init_message=False, **kwargs): """ Calculate forward messages over the observations Pr(u_t | y_{<=t}) for y_t in observations Args: observations (ndarray): observations parameters (parameters): parameters forward_message (dict): latent state prior Pr(u_{-1} | y_{<=-1}) include_init_message (boolean) whether to include t = -1 Returns: forward_messages (list of dict): same format as forward_message """ if forward_message is None: forward_message = self.default_forward_message if np.shape(observations)[0] == 0: if include_init_message: return [forward_message] else: return [] forward_messages = self._forward_messages( observations=observations, parameters=parameters, forward_message=forward_message, **kwargs ) if include_init_message: return forward_messages else: return forward_messages[1:] def backward_pass(self, observations, parameters, backward_message=None, include_init_message=False, **kwargs): """ Calculate backward message over the observations Pr(y_{>t} | u_t) for y_t in observations Args: observations (ndarray): observations parameters (parameters): parameters backward_message (dict): backward message Pr(y_{>T-1} | u_{T-1}) include_init_message (boolean) whether to include t = -1 Returns: backward_messages (list of dict): same format as backward_message """ if backward_message is None: backward_message = self.default_backward_message if np.shape(observations)[0] == 0: if include_init_message: return [backward_message] else: return [] backward_messages = self._backward_messages( observations=observations, parameters=parameters, backward_message=backward_message, **kwargs ) if include_init_message: return backward_messages else: return backward_messages[1:] def _forward_messages(self, observations, parameters, forward_message, weights=None, only_return_last=False, **kwargs): raise NotImplementedError() def _backward_messages(self, observations, parameters, backward_message, weights=None, only_return_last=False, **kwargs): raise NotImplementedError() def _forward_message(self, observations, parameters, forward_message, **kwargs): return self._forward_messages(observations, parameters, forward_message, only_return_last=True, **kwargs) def _backward_message(self, observations, parameters, backward_message, **kwargs): return self._backward_messages(observations, parameters, backward_message, only_return_last=True, **kwargs) ## Loglikelihood Functions def marginal_loglikelihood(self, observations, parameters, forward_message=None, backward_message=None, weights=None, tqdm=None): """ Calculate the marginal loglikelihood Pr(y | theta) Args: observations (ndarray): observations parameters (Parameters): parameters forward_message (dict): latent state forward message backward_message (dict): latent state backward message weights (ndarray): optional, weights for loglikelihood calculation Returns: marginal_loglikelihood (float): marginal loglikelihood """ raise NotImplementedError() def predictive_loglikelihood(self, observations, parameters, forward_message=None, backward_message=None, lag=1): """ Calculate the predictive loglikelihood pred_loglikelihood = sum_t Pr(y_{t+lag} | y_{<t} theta) Args: observations (ndarray): observations parameters (Parameters): parameters forward_message (dict): latent state forward message backward_message (dict): latent state backward message lag (int): how many steps ahead to predict Returns: pred_loglikelihood (float): predictive loglikelihood """ raise NotImplementedError() def complete_data_loglikelihood(self, observations, latent_vars, parameters, forward_message=None, weights=None, **kwargs): """ Calculate the complete data loglikelihood Pr(y, u | theta) Args: observations (ndarray): observations latent_vars (ndarray): latent vars parameters (Parameters): parameters forward_message (dict): latent state forward message weights (ndarray): optional, weights for loglikelihood calculation Returns: complete_data_loglikelihood (float): complete data loglikelihood """ raise NotImplementedError() ## Gradient Functions def gradient_marginal_loglikelihood(self, observations, parameters, forward_message=None, backward_message=None, weights=None, **kwargs): """ Gradient Calculation Gradient of log Pr(y_[0:T) | y_<0, y_>=T, parameters) Args: observations (ndarray): num_obs observations parameters (Parameters): parameters forward_message (dict): Pr(u_-1, y_<0 | parameters) backward_message (dict): Pr(y_>T | u_T, parameters) weights (ndarray): how to weight terms Returns grad (dict): grad of variables in parameters """ raise NotImplementedError() def gradient_complete_data_loglikelihood(self, observations, latent_vars, parameters, forward_message=None, weights=None, **kwargs): """ Gradient Calculation Gradient of log Pr(y_[0:T), u_[0:T) | y_<0, parameters) Args: observations (ndarray): num_obs observations latent_vars (ndarray): num_obs latent vars parameters (Parameters): parameters forward_message (dict): Pr(u_-1, y_<0 | parameters) weights (ndarray): how to weight terms Returns grad (dict): grad of variables in parameters """ raise NotImplementedError() ## Gibbs Functions def parameters_gibbs_sample(self, observations, latent_vars, prior, **kwargs): """ Gibbs sample parameters based on data Samples parameters from the posterior conditional distribution theta ~ Pr(theta | y, u) Args: observations (ndarray): num_obs observations latent_vars (ndarray): num_obs latent variables prior (prior): prior Returns sample_parameters (parameters): sampled parameters """ sufficient_stat = self.calc_gibbs_sufficient_statistic( observations, latent_vars, **kwargs, ) sample_parameters = prior.sample_posterior( sufficient_stat, **kwargs, ) return sample_parameters def calc_gibbs_sufficient_statistic(self, observations, latent_vars, **kwargs): """ Gibbs Sample Sufficient Statistics Args: observations (ndarray): num_obs observations latent_vars (ndarray): latent vars Returns: sufficient_stat (dict of dict) keys are parameter values are dict for parameter's sufficient statistics """ raise NotImplementedError() ## Predict Functions def latent_var_distr(self, observations, parameters, distr='marginal', lag=None, forward_message=None, backward_message=None, tqdm=None, **kwargs): """ Sample latent vars distribution conditional on observations Returns distribution for (u_t | y_{<= t+lag}, theta) Args: observations (ndarray): observations parameters (LGSSMParameters): parameters lag (int): what observations to condition on, None = all forward_message (dict): forward message backward_message (dict): backward message Returns: Depends on latent var type, Gaussian -> mean, cov; Discrete -> prob """ raise NotImplementedError() def latent_var_sample(self, observations, parameters, distr='joint', lag=None, num_samples=None, forward_message=None, backward_message=None, include_init=False, tqdm=None, **kwargs): """ Sample latent vars conditional on observations Samples u_t ~ u_t | y_{<= t+lag}, theta Args: observations (ndarray): observations parameters (LGSSMParameters): parameters lag (int): what observations to condition on, None = all num_samples (int, optional) number of samples forward_message (dict): forward message backward_message (dict): backward message include_init (bool, optional): whether to sample u_{-1} | y Returns: sampled_latent_vars : shape depends on num_samples parameters last dimension is num_samples """ raise NotImplementedError() def y_distr(self, observations, parameters, distr='marginal', lag=None, forward_message=None, backward_message=None, latent_var=None, tqdm=None, **kwargs): """ Sample observation distribution conditional on observations Returns distribution for (y_t* | y_{<= t+lag}, theta) Args: observations (ndarray): observations parameters (LGSSMParameters): parameters lag (int): what observations to condition on, None = all forward_message (dict): forward message backward_message (dict): backward message latent_var (ndarray): latent vars if provided, will return (y_t* | u_t, theta) instead Returns: Depends on observation type, Gaussian -> mean, cov; Discrete -> prob """ raise NotImplementedError() def y_sample(self, observations, parameters, distr='joint', lag=None, num_samples=None, forward_message=None, backward_message=None, latent_var=None, tqdm=None, **kwargs): """ Sample new observations conditional on observations Samples y_t* ~ y_t* | y_{<= t+lag}, theta Args: observations (ndarray): observations parameters (LGSSMParameters): parameters lag (int): what observations to condition on, None = all num_samples (int, optional) number of samples forward_message (dict): forward message backward_message (dict): backward message latent_var (ndarray): latent vars if provided, will sample from (y_t* | u_t, theta) instead must match num_samples parameters Returns: sampled_observations : shape depends on num_samples parameters last dimension is num_samples """ raise NotImplementedError() def simulate_distr(self, T, parameters, init_message=None, tqdm=None): raise NotImplementedError() def simulate(self, T, parameters, init_message=None, num_samples=None, tqdm=None): raise NotImplementedError() ## PF Functions def pf_loglikelihood_estimate(self, observations, parameters, subsequence_start=0, subsequence_end=None, weights=None, pf="poyiadjis_N", N=1000, kernel='prior', forward_message=None, **kwargs): """ Particle Filter Marginal Log-Likelihood Estimate Args: observations (ndarray): num_obs bufferd observations parameters (Parameters): parameters weights (ndarray): weights (to correct storchastic approx) subsequence_start (int): relative start of subsequence (0:subsequence_start are left buffer) subsequence_end (int): relative end of subsequence (subsequence_end: is right buffer) pf (string): particle filter name "nemeth" - use Nemeth et al. O(N) "poyiadjis_N" - use Poyiadjis et al. O(N) "poyiadjis_N2" - use Poyiadjis et al. O(N^2) "paris" - use PaRIS Olsson + Westborn O(N log N) N (int): number of particles used by particle filter kernel (string): kernel to use "prior" - bootstrap filter P(u_t | u_{t-1}) "optimal" - bootstrap filter P(u_t | u_{t-1}, Y_t) forward_message (dict): prior for buffered subsequence **kwargs - additional keyword args for individual filters Return: loglikelihood (double): marignal log likelihood estimate """ raise NotImplementedError() def pf_predictive_loglikelihood_estimate(self, observations, parameters, num_steps_ahead=1, subsequence_start=0, subsequence_end=None, weights=None, pf="filter", N=1000, kernel=None, forward_message=None, **kwargs): """ Particle Filter Predictive Log-Likelihoood Estimate Returns predictive log-likleihood for k = [0,1, ...,num_steps_ahead] Args: observations (ndarray): num_obs bufferd observations parameters (Parameters): parameters num_steps_ahead (int): number of steps subsequence_start (int): relative start of subsequence (0:subsequence_start are left buffer) subsequence_end (int): relative end of subsequence (subsequence_end: is right buffer) N (int): number of particles used by particle filter kernel (string): kernel to use forward_message (dict): prior for buffered subsequence **kwargs - additional keyword args for individual filters Return: predictive_loglikelihood (num_steps_ahead + 1 ndarray) """ raise NotImplementedError() def pf_gradient_estimate(self, observations, parameters, subsequence_start=0, subsequence_end=None, weights=None, pf="poyiadjis_N", N=1000, kernel=None, forward_message=None, **kwargs): """ Particle Smoother Gradient Estimate Args: observations (ndarray): num_obs bufferd observations parameters (Parameters): parameters subsequence_start (int): relative start of subsequence (0:subsequence_start are left buffer) subsequence_end (int): relative end of subsequence (subsequence_end: is right buffer) weights (ndarray): weights (to correct storchastic approx) pf (string): particle filter name "nemeth" - use Nemeth et al. O(N) "poyiadjis_N" - use Poyiadjis et al. O(N) "poyiadjis_N2" - use Poyiadjis et al. O(N^2) "paris" - use PaRIS Olsson + Westborn O(N log N) N (int): number of particles used by particle filter kernel (string): kernel to use "prior" - bootstrap filter P(u_t | u_{t-1}) "optimal" - bootstrap filter P(u_t | u_{t-1}, Y_t) forward_message (dict): prior for buffered subsequence **kwargs - additional keyword args for individual filters Return: grad (dict): grad of variables in parameters """ raise NotImplementedError() def pf_latent_var_distr(self, observations, parameters, lag=None, subsequence_start=0, subsequence_end=None, pf="poyiadjis_N", N=1000, kernel=None, forward_message=None, **kwargs): """ Sample latent vars distribution conditional on observations Returns distribution for (u_t | y_{<= t+lag}, theta) Estimated using particle filter/smoother Args: observations (ndarray): observations parameters (LGSSMParameters): parameters lag (int): what observations to condition on, None = all subsequence_start (int): relative start of subsequence (0:subsequence_start are left buffer) subsequence_end (int): relative end of subsequence (subsequence_end: is right buffer) pf (string): particle filter name "nemeth" - use Nemeth et al. O(N) "poyiadjis_N" - use Poyiadjis et al. O(N) "poyiadjis_N2" - use Poyiadjis et al. O(N^2) "paris" - use PaRIS Olsson + Westborn O(N log N) N (int): number of particles used by particle filter kernel (string): kernel to use "prior" - bootstrap filter P(u_t | u_{t-1}) "optimal" - bootstrap filter P(u_t | u_{t-1}, Y_t) forward_message (dict): prior for buffered subsequence **kwargs - additional keyword args for individual filters Returns: Depends on latent var type, Gaussian -> mean, cov; Discrete -> prob """ raise NotImplementedError() def pf_y_distr(self, observations, parameters, distr='marginal', lag=None, subsequence_start=0, subsequence_end=None, pf="poyiadjis_N", N=1000, kernel=None, forward_message=None, **kwargs): """ Sample observation distribution conditional on observations Returns distribution for (u_t | y_{<= t+lag}, theta) Estimated using particle filter/smoother Args: observations (ndarray): observations parameters (LGSSMParameters): parameters lag (int): what observations to condition on, None = all subsequence_start (int): relative start of subsequence (0:subsequence_start are left buffer) subsequence_end (int): relative end of subsequence (subsequence_end: is right buffer) pf (string): particle filter name "nemeth" - use Nemeth et al. O(N) "poyiadjis_N" - use Poyiadjis et al. O(N) "poyiadjis_N2" - use Poyiadjis et al. O(N^2) "paris" - use PaRIS Olsson + Westborn O(N log N) N (int): number of particles used by particle filter kernel (string): kernel to use "prior" - bootstrap filter P(u_t | u_{t-1}) "optimal" - bootstrap filter P(u_t | u_{t-1}, Y_t) forward_message (dict): prior for buffered subsequence **kwargs - additional keyword args for individual filters Returns: Depends on latent var type, Gaussian -> mean, cov; Discrete -> prob """ raise NotImplementedError() # Helper Function for Sampling Subsequences def random_subsequence_and_weights(S, T, partition_style=None): """ Get Subsequence + Weights Args: S (int): length of subsequence T (int): length of full sequence partition_style (string): what type of partition 'strict' - strict partition, with weights 'uniform' - uniformly, with weights 'naive' - uniformly, with incorrect weights (not recommended) Returns: subsequence_start (int): start of subsequence (inclusive) subsequence_end (int): end of subsequence (exclusive) weights (ndarray): weights for [start,end) """ if partition_style is None: partition_style = 'uniform' if partition_style == 'strict': if T % S != 0: raise ValueError("S {0} does not evenly divide T {1}".format(S, T) ) subsequence_start = np.random.choice(np.arange(0, T//S)) * S subsequence_end = subsequence_start + S weights = np.ones(S, dtype=float)*T/S elif partition_style == 'uniform': subsequence_start = np.random.randint(0, T-S+1) subsequence_end = subsequence_start + S t = np.arange(subsequence_start, subsequence_end) if subsequence_end <= 2*S: num_sequences = np.min(np.array([ t+1, np.ones_like(t)*min(S, T-S+1) ]), axis=0) elif subsequence_start >= T-2*S-1: num_sequences = np.min(np.array([ T-t, np.ones_like(t)*min(S, T-S+1) ]), axis=0) else: num_sequences = np.ones(S)*S weights = np.ones(S, dtype=float)*(T-S+1)/num_sequences elif partition_style == 'naive': # Not recommended because the weights are incorrect subsequence_start = np.random.randint(0, T-S+1) subsequence_end = subsequence_start + S weights = np.ones(S, dtype=float)*T/S else: raise ValueError("Unrecognized partition_style = '{0}'".format( partition_style)) return int(subsequence_start), int(subsequence_end), weights
82,434
39.789213
86
py
CCasGNN
CCasGNN-main/layers.py
#encoding: utf-8 import torch from torch_geometric.nn import GCNConv, GATConv from math import sqrt class Positional_GAT(torch.nn.Module): def __init__(self, in_channels, out_channels, n_heads, location_embedding_dim, filters_1, filters_2, dropout): super(Positional_GAT, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.n_heads = n_heads self.filters_1 = filters_1 self.filters_2 = filters_2 self.dropout = dropout self.location_embedding_dim = location_embedding_dim self.setup_layers() def setup_layers(self): self.GAT_1 = GATConv(in_channels=self.in_channels,out_channels=self.filters_1, heads=self.n_heads, dropout=0.1) self.GAT_2 = GATConv(in_channels=self.filters_1 * self.n_heads + self.location_embedding_dim, out_channels=self.out_channels, heads=self.n_heads, dropout=0.1, concat=False) def forward(self, edge_indices, features, location_embedding): features = torch.cat((features, location_embedding), dim=-1) features = self.GAT_1(features, edge_indices) features = torch.nn.functional.relu(features) features = torch.nn.functional.dropout(features, p=self.dropout, training=self.training) features = torch.cat((features, location_embedding), dim=-1) features = self.GAT_2(features, edge_indices) return features class Positional_GCN(torch.nn.Module): def __init__(self, in_channels, out_channels, location_embedding_dim, filters_1, filters_2, dropout): """ GCN function :param args: Arguments object. :param in_channel: Nodes' input feature dimensions :param out_channel: Nodes embedding dimension :param bais: """ super(Positional_GCN, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.filters_1 = filters_1 self.filters_2 = filters_2 self.dropout = dropout self.location_embedding_dim = location_embedding_dim self.setup_layers() def setup_layers (self): self.convolution_1 = GCNConv(self.in_channels, self.filters_1) self.convolution_2 = GCNConv(self.filters_1 + self.location_embedding_dim, self.out_channels) def forward (self, edge_indices, features, location_embedding): """ making convolution :param edge_indices: 2 * edge_number :param features: N * feature_size :return: """ features = torch.cat((features, location_embedding), dim=-1) features = self.convolution_1(features, edge_indices) features = torch.nn.functional.relu(features) features = torch.nn.functional.dropout(features, p=self.dropout, training=self.training) features = torch.cat((features, location_embedding), dim=-1) features = self.convolution_2(features, edge_indices) return features class MultiHeadGraphAttention(torch.nn.Module): def __init__(self, num_heads, dim_in, dim_k, dim_v): super(MultiHeadGraphAttention, self).__init__() #"dim_k and dim_v must be multiple of num_heads" assert dim_k % num_heads == 0 and dim_v % num_heads == 0 self.num_heads = num_heads self.dim_in = dim_in self.dim_k = dim_k self.dim_v = dim_v self.linear_q = torch.nn.Linear(dim_in, dim_k, bias=False) self.linear_k = torch.nn.Linear(dim_in, dim_k, bias=False) self.linear_v = torch.nn.Linear(dim_in, dim_v, bias=False) self.leaky_relu = torch.nn.LeakyReLU(negative_slope=0.2) self._nor_fact = 1 / sqrt(dim_k // num_heads) def forward(self, x): # x: tensor of shape (batch, n, dim_in) batch, n, dim_in = x.shape assert dim_in == self.dim_in nh = self.num_heads dk = self.dim_k // nh #dim_k of each head dv = self.dim_v // nh q = self.linear_q(x).reshape(batch, n, nh, dk).transpose(1,2) # (batch, nh, n, dk) k = self.linear_k(x).reshape(batch, n, nh, dk).transpose(1,2) v = self.linear_v(x).reshape(batch, n, nh, dv).transpose(1,2) dist = torch.matmul(q, k.transpose(2,3)) * self._nor_fact # batch, nh, n, n # label = torch.where(dist == 0, torch.tensor(1), torch.tensor(0)) # dist.data.masked_fill_(label, -float("inf")) dist = self.leaky_relu(dist) # batch, nh, n, n # dist = torch.where(torch.isnan(dist), torch.full_like(dist,0), dist) att = torch.matmul(dist, v) # batch, nh, n, dv att = att.transpose(1,2).reshape(batch, n, self.dim_v) return att class dens_Net(torch.nn.Module): def __init__(self,dens_hiddensize, dens_dropout, dens_inputsize, dens_outputsize): super(dens_Net, self).__init__() self.inputsize = dens_inputsize self.dens_hiddensize = dens_hiddensize self.dens_dropout = dens_dropout self.outputsize = dens_outputsize self.setup_layers() def setup_layers(self): self.dens_net = torch.nn.Sequential( torch.nn.Linear(self.inputsize, self.dens_hiddensize), torch.nn.Dropout(p=self.dens_dropout), torch.nn.Linear(self.dens_hiddensize, self.dens_hiddensize), torch.nn.Dropout(p=self.dens_dropout), torch.nn.Linear(self.dens_hiddensize, self.outputsize) ) def forward(self, x1, x2): return torch.nn.functional.relu(self.dens_net(x1)), torch.nn.functional.relu(self.dens_net(x2)) # return torch.nn.functional.relu(self.dens_net(x1)) class fuse_gate(torch.nn.Module): def __init__(self, batch_size, in_dim): super(fuse_gate, self).__init__() self.indim = in_dim self.batch_size = batch_size self.setup_layers() def setup_layers(self): self.omega = torch.nn.Parameter(torch.tensor([[0.5],[0.5]])) def forward(self, x): omega = self.omega.transpose(1,0) prediction = torch.matmul(omega, x) return prediction, self.omega[0], self.omega[1]
6,309
41.635135
180
py
CCasGNN
CCasGNN-main/CCasGNN.py
#encoding: utf-8 import torch import json import numpy as np import copy import time import sys import math from layers import Positional_GCN, MultiHeadGraphAttention, dens_Net, Positional_GAT, fuse_gate import scipy.stats as sci class CCasGNN(torch.nn.Module): def __init__(self, args): super(CCasGNN, self).__init__() self.args = args self.number_of_features = self.args.number_of_features self.number_of_nodes = self.args.number_of_nodes self._setup_layers() def _setup_GCN_layers(self): self.GCN_layers = Positional_GCN(in_channels=self.args.user_embedding_dim + self.args.location_embedding_dim, out_channels=self.args.gcn_out_channel, location_embedding_dim=self.args.location_embedding_dim, filters_1=self.args.gcn_filters_1, filters_2=self.args.gcn_filters_2, dropout=self.args.gcn_dropout) #self.args.user_embedding_dim def _setup_GAT_layers(self): self.GAT_layers = Positional_GAT(in_channels=self.args.user_embedding_dim + self.args.location_embedding_dim, out_channels=self.args.gcn_out_channel, n_heads=self.args.gat_n_heads, location_embedding_dim=self.args.location_embedding_dim, filters_1=self.args.gcn_filters_1, filters_2=self.args.gcn_filters_2, dropout=self.args.gcn_dropout) #self.number_of_features + self.args.location_embedding_dim def _setup_MultiHead_att_layers(self): self.MultiHead_att_layers = MultiHeadGraphAttention(num_heads=self.args.att_num_heads, dim_in=self.args.gcn_out_channel, dim_k=self.args.att_dim_k, dim_v=self.args.att_dim_v) def _setup_dens_layers (self): self.dens_layers = dens_Net(dens_inputsize=self.args.gcn_out_channel, dens_hiddensize=self.args.dens_hiddensize, dens_dropout=self.args.dens_dropout, dens_outputsize=self.args.dens_outsize ) #self.args.attn_out_dim def _setup_fuse_layers(self): self.fuse_layers = fuse_gate(batch_size=1, in_dim=2) def _setup_layers(self): self._setup_GCN_layers() self._setup_MultiHead_att_layers() self._setup_dens_layers() self._setup_GAT_layers() self._setup_fuse_layers() def forward(self,data): true_nodes_num = data["true_nodes_num"] features = data['features'][:true_nodes_num] edges = data['edges'] undirected_edges = data['undirected_edges'] location_embedding = data['location_embedding'][:true_nodes_num] GAT_representation = torch.nn.functional.relu(self.GAT_layers(edges, features, location_embedding)) GAT_representation = GAT_representation[:true_nodes_num] #nodes_num * feature_num GAT_representation = torch.mean(GAT_representation, dim=0, keepdim=False) GCN_representation = torch.nn.functional.relu(self.GCN_layers(undirected_edges, features, location_embedding)) GCN_representation = GCN_representation[:true_nodes_num] #nodes_num * feature_num GCN_representation = GCN_representation.unsqueeze(dim=0) #batch_size * nodes_num * feature_num # GCN_att_representation = self.MultiHead_att_layers(GCN_representation) #batch_size * nodes_num * feature_num GCN_att_representation = torch.mean(GCN_att_representation, dim=1, keepdim=False) #batch_size * feature_num GCN_squeeze_att_representation = GCN_att_representation.squeeze(dim=0) GAT_pred, GCN_pred = self.dens_layers(GAT_representation, GCN_squeeze_att_representation) model_predict = torch.cat((GAT_pred, GCN_pred), dim=0) prediction, omega1, omega2 = self.fuse_layers(model_predict) return prediction, omega1, omega2, GAT_pred, GCN_pred class CCasGNN_Trainer(torch.nn.Module): def __init__(self, args): super(CCasGNN_Trainer, self).__init__() self.args = args self.setup_model() def setup_model(self): self.load_graph_data() self.model = CCasGNN(self.args) def load_graph_data(self): self.number_of_nodes = self.args.number_of_nodes self.number_of_features = self.args.number_of_features self.graph_data = json.load(open(self.args.graph_file_path, 'r')) N = len(self.graph_data) #the number of graphs train_start, valid_start, test_start = \ 0, int(N * self.args.train_ratio), int(N * (self.args.train_ratio + self.args.valid_ratio)) train_graph_data = self.graph_data[0:valid_start] #list type [dict,dict,...] valid_graph_data = self.graph_data[valid_start:test_start] test_graph_data = self.graph_data[test_start:N] self.train_batches, self.valid_batches, self.test_batches = [], [], [] for i in range(0, len(train_graph_data), self.args.batch_size): self.train_batches.append(train_graph_data[i:i+self.args.batch_size]) for j in range(0, len(valid_graph_data), self.args.batch_size): self.valid_batches.append(valid_graph_data[j:j+self.args.batch_size]) for k in range(0, len(test_graph_data), self.args.batch_size): self.test_batches.append(test_graph_data[k:k+self.args.batch_size]) def create_edges(self,data): """ create an Edge matrix :param data: :return: Edge matrix """ self.nodes_map = [str(nodes_id) for nodes_id in data['nodes']] self.true_nodes_num = len(data['nodes']) edges = [[self.nodes_map.index(str(edge[0])), self.nodes_map.index(str(edge[1]))] for edge in data['edges']] undirected_edges = edges + [[self.nodes_map.index(str(edge[1])), self.nodes_map.index(str(edge[0]))] for edge in data['edges']] return torch.t(torch.LongTensor(edges)), torch.t(torch.LongTensor(undirected_edges)) def create_location_embedding(self, omega=0.001): location_dim = self.args.location_embedding_dim location_emb = torch.zeros(self.number_of_nodes, location_dim) for i in range(self.number_of_nodes): for j in range(location_dim): if j % 2 == 0: location_emb[i][j] = math.sin(i * math.pow(omega, j / location_dim)) else: location_emb[i][j] = math.cos(i * math.pow(omega, (j - 1) / location_dim)) return location_emb def create_target(self,data): return torch.tensor([data['activated_size']]) def create_features(self,data): features = np.zeros((self.number_of_nodes, self.args.user_embedding_dim)) # features = np.zeros((self.number_of_nodes, self.number_of_features)) for nodes_id in data['nodes']: features[self.nodes_map.index(str(nodes_id))][:self.args.user_embedding_dim] = data['nodes_embedding'][str(nodes_id)] # features[self.nodes_map.index(str(nodes_id))][self.args.user_embedding_dim:] = data['location_embedding'][str(nodes_id)] features = torch.FloatTensor(features) return features def create_input_data(self, data): """ :param data: one data in the train/valid/test graph data :return: to_pass_forward: Data dictionary """ to_pass_forward = dict() activated_size = self.create_target(data) edges, undirected_edges = self.create_edges(data) features = self.create_features(data) user_embedding = self.create_user_embedding(data) location_embedding = self.create_location_embedding(omega=0.001) to_pass_forward["edges"] = edges to_pass_forward["undirected_edges"] = undirected_edges to_pass_forward["features"] = features to_pass_forward["true_nodes_num"] = self.true_nodes_num to_pass_forward['location_embedding'] = location_embedding return to_pass_forward, activated_size def create_forward_data(self, data_batches): data_x, data_y = [], [] for data_batch in data_batches: data_x_tmp, data_y_tmp = [], [] for each_data in data_batch: input_data, target = self.create_input_data(each_data) data_x_tmp.append(input_data) data_y_tmp.append(target) data_x.append(copy.deepcopy(data_x_tmp)) data_y.append(copy.deepcopy(data_y_tmp)) return data_x, data_y def fit(self): print('\nLoading data.\n') self.model.train() train_data_x, train_data_y = self.create_forward_data(self.train_batches) valid_data_x, valid_data_y = self.create_forward_data(self.valid_batches) test_data_x, test_data_y = self.create_forward_data(self.test_batches) optimizer = torch.optim.Adam(self.model.parameters(), lr=self.args.learning_rate, weight_decay=self.args.weight_decay) time_start = time.time() print('\nTraining started.\n') for epoch in range(self.args.epochs): losses = 0. average_loss = 0. for step, (train_x_batch, train_y_batch) in enumerate(zip(train_data_x, train_data_y)): optimizer.zero_grad() GAT_prediction_tensor = torch.tensor([]) GCN_prediction_tensor = torch.tensor([]) target_tensor = torch.tensor([]) for k, (train_x, train_y) in enumerate(zip(train_x_batch, train_y_batch)): prediction = self.model(train_x) GAT_prediction_tensor = torch.cat((GAT_prediction_tensor, prediction[3].float()), 0) GCN_prediction_tensor = torch.cat((GCN_prediction_tensor, prediction[4].float()), 0) target_tensor = torch.cat((target_tensor, torch.log2(train_y.float() + 1)), 0) omega1 = prediction[1].data.float() omega2 = prediction[2].data.float() GAT_loss = torch.nn.functional.mse_loss(target_tensor,GAT_prediction_tensor) GCN_loss = torch.nn.functional.mse_loss(target_tensor,GCN_prediction_tensor) loss = omega1 * GAT_loss + omega2 * GCN_loss loss.backward() optimizer.step() losses = losses + loss.item() average_loss = losses / (step + 1) print('CCasGNN train MSLE loss in ', epoch + 1, ' epoch = ', average_loss) time_now = time.time() print('the rest of running time about:', (((time_now-time_start)/ (epoch+1)) * (self.args.epochs - epoch)) / 60, ' minutes') print('\n') if (epoch + 1) % self.args.check_point == 0: print('epoch ',epoch + 1, ' evaluating.') self.evaluation(valid_data_x, valid_data_y) self.test(test_data_x, test_data_y) def evaluation(self, valid_x_batches, valid_y_batches): self.model.eval() losses = 0. average_loss = 0. for step, (valid_x_batch, valid_y_batch) in enumerate(zip(valid_x_batches, valid_y_batches)): loss = 0. prediction_tensor = torch.tensor([]) target_tensor = torch.tensor([]) for (valid_x, valid_y) in zip(valid_x_batch, valid_y_batch): prediction = self.model(valid_x) prediction_tensor = torch.cat((prediction_tensor, prediction[0].float()), 0) target_tensor = torch.cat((target_tensor, torch.log2(valid_y.float() + 1)), 0) loss = torch.nn.functional.mse_loss(target_tensor, prediction_tensor) losses = losses + loss.item() average_loss = losses / (step + 1) print('#####CCasGNN valid MSLE loss in this epoch = ', average_loss) print('\n') def test(self, test_x_batches, test_y_batches): print("\n\nScoring.\n") self.model.eval() losses = 0. average_loss = 0. all_test_tensor = torch.tensor([]) all_true_tensor = torch.tensor([]) for step, (test_x_batch, test_y_batch) in enumerate(zip(test_x_batches, test_y_batches)): loss = 0. prediction_tensor = torch.tensor([]) target_tensor = torch.tensor([]) for (test_x, test_y) in zip(test_x_batch, test_y_batch): prediction = self.model(test_x) prediction_tensor = torch.cat((prediction_tensor, prediction[0].float()), 0) all_test_tensor = torch.cat((all_test_tensor, prediction[0].float()), dim=0) target_tensor = torch.cat((target_tensor, torch.log2(test_y.float() + 1)), 0) all_true_tensor = torch.cat((all_true_tensor, torch.log2(test_y.float() + 1)), dim=0) loss = torch.nn.functional.mse_loss(target_tensor, prediction_tensor) losses = losses + loss.item() average_loss = losses / (step + 1) all_test_np = all_test_tensor.detach().numpy() all_true_np = all_true_tensor.detach().numpy() sub_np = all_test_np - all_true_np print('correlation: ', sci.pearsonr(sub_np, all_true_np)) print('#####CCasGNN test MSLE loss = ', average_loss) print('\n')
13,767
50.373134
136
py
pyEPR
pyEPR-master/docs/source/conf.py
# Configuration file for the Sphinx documentation builder. # # This file only contains a selection of the most common options. For a full # list see the documentation: # http://www.sphinx-doc.org/en/master/config # -- Path setup -------------------------------------------------------------- # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. # # import os # import sys # sys.path.insert(0, os.path.abspath('.')) # -- Project information ----------------------------------------------------- project = 'pyEPR' copyright = '2020, Zlatko Minev, Zaki Leghtas, and the pyEPR Team' author = 'Zlatko Minev, Zaki Leghtas, and the pyEPR Team' import sys import os sys.path.insert(0, os.path.abspath("../../pyEPR")) print(sys.path) # The full version, including alpha/beta/rc tags import pyEPR version = pyEPR.__version__ release = version import sphinx_rtd_theme # -- General configuration --------------------------------------------------- # Add any Sphinx extension module names here, as strings. They can be # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = [ "sphinx.ext.intersphinx", "sphinx.ext.autodoc", "sphinx.ext.autosummary", 'sphinx.ext.coverage', 'sphinx.ext.napoleon', # parse both NumPy and Google style docstrings 'sphinx.ext.viewcode', 'sphinx.ext.githubpages', "sphinx.ext.mathjax", "sphinx_rtd_theme", #'sphinx_automodapi.automodapi', "IPython.sphinxext.ipython_directive", "IPython.sphinxext.ipython_console_highlighting", "matplotlib.sphinxext.plot_directive", #'numpydoc' ] # https://github.com/readthedocs/readthedocs.org/issues/2569 master_doc = 'index' # Add any paths that contain templates here, relative to this directory. templates_path = ['_templates'] # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This pattern also affects html_static_path and html_extra_path. exclude_patterns = ["**.ipynb_checkpoints"] # If true, the current module name will be prepended to all description # unit titles (such as .. function::). add_module_names = False # If true, sectionauthor and moduleauthor directives will be shown in the # output. They are ignored by default. show_authors = True # The name of the Pygments (syntax highlighting) style to use. pygments_style = 'sphinx' numpydoc_show_class_members = True napoleon_numpy_docstring = True napoleon_use_admonition_for_notes = True # -- Options for HTML output ------------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = 'sphinx_rtd_theme' #'default' # 'sphinx_rtd_theme' #'alabaster' "sphinxdoc" 'classic' if 0: import os on_rtd = os.environ.get('READTHEDOCS') == 'True' if on_rtd: html_theme = 'default' else: html_theme = 'nature' # -- Options for HTML output ---------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = 'sphinx_rtd_theme' full_logo= True # Theme options are theme-specific and customize the look and feel of a theme # further. For a list of options available for each theme, see the # documentation. html_theme_options = { 'canonical_url': '', #'logo_only': False, 'display_version': True, 'prev_next_buttons_location': 'bottom', 'style_external_links': False, #'style_nav_header_background': 'white', # Toc options 'collapse_navigation': False, 'sticky_navigation': True, 'navigation_depth': 4, 'includehidden': True, 'titles_only': False } # Add any paths that contain custom themes here, relative to this directory. # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". html_static_path = ['_static'] # If false, no module index is generated. html_use_modindex = True html_show_sourcelink = True # Sort members by type #autodoc_member_order = 'groupwise' # The name of the Pygments (syntax highlighting) style to use. pygments_style = "sphinx" # ----------------------------------------------------------------------------- # Autosummary # ----------------------------------------------------------------------------- autosummary_generate = True # ----------------------------------------------------------------------------- # Autodoc # ----------------------------------------------------------------------------- #The supported options are # 'members', 'member-order', 'undoc-members', 'private-members', # 'special-members', 'inherited-members', 'show-inheritance', 'ignore-module-all', # 'imported-members' and 'exclude-members'. autodoc_default_options = { 'inherited-members': None, #'member-order': 'bysource', 'member-order': 'alphabetical', #This value selects if automatically documented members are sorted alphabetical (value 'alphabetical'), by member type (value 'groupwise') or by source order (value 'bysource'). The default is alphabetical. 'undoc-members': True, # Members without docstrings will be left out, unless you give the undoc-members flag option: 'exclude-members': '__weakref__', 'show-inheritance' : True # , a list of base classes will be inserted just below the class signature (when used with automodule, this will be inserted for every class that is documented in the module). } # If true, figures, tables and code-blocks are automatically numbered if they # have a caption. numfig = True # A dictionary mapping 'figure', 'table', 'code-block' and 'section' to # strings that are used for format of figure numbers. As a special character, # %s will be replaced to figure number. numfig_format = { 'table': 'Table %s' } # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. # # This is also used if you do content translation via gettext catalogs. # Usually you set "language" from the command line for these cases. language = None # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This pattern also affects html_static_path and html_extra_path. exclude_patterns = [] # The name of the Pygments (syntax highlighting) style to use. pygments_style = 'colorful' # A boolean that decides whether module names are prepended to all object names # (for object types where a “module” of some kind is defined), e.g. for # py:function directives. add_module_names = True # A list of prefixes that are ignored for sorting the Python module index # (e.g., if this is set to ['foo.'], then foo.bar is shown under B, not F). # This can be handy if you document a project that consists of a single # package. Works only for the HTML builder currently. #modindex_common_prefix = ['pyEPR.']
7,220
33.716346
242
py
FUNIT
FUNIT-master/test_k_shot.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import os import numpy as np from PIL import Image import torch import torch.backends.cudnn as cudnn from torchvision import transforms from utils import get_config from trainer import Trainer import argparse os.environ['CUDA_VISIBLE_DEVICES']='1' parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, default='configs/funit_animals.yaml') parser.add_argument('--ckpt', type=str, default='pretrained/animal119_gen_00200000.pt') parser.add_argument('--class_image_folder', type=str, default='images/n02138411') parser.add_argument('--input', type=str, default='images/input_content.jpg') parser.add_argument('--output', type=str, default='images/output.jpg') opts = parser.parse_args() cudnn.benchmark = True opts.vis = True config = get_config(opts.config) config['batch_size'] = 1 config['gpus'] = 1 trainer = Trainer(config) trainer.cuda() trainer.load_ckpt(opts.ckpt) trainer.eval() transform_list = [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))] transform_list = [transforms.Resize((128, 128))] + transform_list transform = transforms.Compose(transform_list) print('Compute average class codes for images in %s' % opts.class_image_folder) images = os.listdir(opts.class_image_folder) for i, f in enumerate(images): fn = os.path.join(opts.class_image_folder, f) img = Image.open(fn).convert('RGB') img_tensor = transform(img).unsqueeze(0).cuda() with torch.no_grad(): class_code = trainer.model.compute_k_style(img_tensor, 1) if i == 0: new_class_code = class_code else: new_class_code += class_code final_class_code = new_class_code / len(images) image = Image.open(opts.input) image = image.convert('RGB') content_img = transform(image).unsqueeze(0) print('Compute translation for %s' % opts.input) with torch.no_grad(): output_image = trainer.model.translate_simple(content_img, final_class_code) image = output_image.detach().cpu().squeeze().numpy() image = np.transpose(image, (1, 2, 0)) image = ((image + 1) * 0.5 * 255.0) output_img = Image.fromarray(np.uint8(image)) output_img.save(opts.output, 'JPEG', quality=99) print('Save output to %s' % opts.output)
2,618
31.7375
80
py
FUNIT
FUNIT-master/utils.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import os import yaml import time import torch from torch.utils.data import DataLoader from torchvision import transforms import torchvision.utils as vutils from data import ImageLabelFilelist def update_average(model_tgt, model_src, beta=0.999): with torch.no_grad(): param_dict_src = dict(model_src.named_parameters()) for p_name, p_tgt in model_tgt.named_parameters(): p_src = param_dict_src[p_name] assert(p_src is not p_tgt) p_tgt.copy_(beta*p_tgt + (1. - beta)*p_src) def loader_from_list( root, file_list, batch_size, new_size=None, height=128, width=128, crop=True, num_workers=4, shuffle=True, center_crop=False, return_paths=False, drop_last=True): transform_list = [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))] if center_crop: transform_list = [transforms.CenterCrop((height, width))] + \ transform_list if crop else transform_list else: transform_list = [transforms.RandomCrop((height, width))] + \ transform_list if crop else transform_list transform_list = [transforms.Resize(new_size)] + transform_list \ if new_size is not None else transform_list if not center_crop: transform_list = [transforms.RandomHorizontalFlip()] + transform_list transform = transforms.Compose(transform_list) dataset = ImageLabelFilelist(root, file_list, transform, return_paths=return_paths) loader = DataLoader(dataset, batch_size, shuffle=shuffle, drop_last=drop_last, num_workers=num_workers) return loader def get_evaluation_loaders(conf, shuffle_content=False): batch_size = conf['batch_size'] num_workers = conf['num_workers'] new_size = conf['new_size'] width = conf['crop_image_width'] height = conf['crop_image_height'] content_loader = loader_from_list( root=conf['data_folder_train'], file_list=conf['data_list_train'], batch_size=batch_size, new_size=new_size, height=height, width=width, crop=True, num_workers=num_workers, shuffle=shuffle_content, center_crop=True, return_paths=True, drop_last=False) class_loader = loader_from_list( root=conf['data_folder_test'], file_list=conf['data_list_test'], batch_size=batch_size * conf['k_shot'], new_size=new_size, height=height, width=width, crop=True, num_workers=1, shuffle=False, center_crop=True, return_paths=True, drop_last=False) return content_loader, class_loader def get_train_loaders(conf): batch_size = conf['batch_size'] num_workers = conf['num_workers'] new_size = conf['new_size'] width = conf['crop_image_width'] height = conf['crop_image_height'] train_content_loader = loader_from_list( root=conf['data_folder_train'], file_list=conf['data_list_train'], batch_size=batch_size, new_size=new_size, height=height, width=width, crop=True, num_workers=num_workers) train_class_loader = loader_from_list( root=conf['data_folder_train'], file_list=conf['data_list_train'], batch_size=batch_size, new_size=new_size, height=height, width=width, crop=True, num_workers=num_workers) test_content_loader = loader_from_list( root=conf['data_folder_test'], file_list=conf['data_list_test'], batch_size=batch_size, new_size=new_size, height=height, width=width, crop=True, num_workers=1) test_class_loader = loader_from_list( root=conf['data_folder_test'], file_list=conf['data_list_test'], batch_size=batch_size, new_size=new_size, height=height, width=width, crop=True, num_workers=1) return (train_content_loader, train_class_loader, test_content_loader, test_class_loader) def get_config(config): with open(config, 'r') as stream: return yaml.load(stream, Loader=yaml.FullLoader) def make_result_folders(output_directory): image_directory = os.path.join(output_directory, 'images') if not os.path.exists(image_directory): print("Creating directory: {}".format(image_directory)) os.makedirs(image_directory) checkpoint_directory = os.path.join(output_directory, 'checkpoints') if not os.path.exists(checkpoint_directory): print("Creating directory: {}".format(checkpoint_directory)) os.makedirs(checkpoint_directory) return checkpoint_directory, image_directory def __write_images(im_outs, dis_img_n, file_name): im_outs = [images.expand(-1, 3, -1, -1) for images in im_outs] image_tensor = torch.cat([images[:dis_img_n] for images in im_outs], 0) image_grid = vutils.make_grid(image_tensor.data, nrow=dis_img_n, padding=0, normalize=True) vutils.save_image(image_grid, file_name, nrow=1) def write_1images(image_outputs, image_directory, postfix): display_image_num = image_outputs[0].size(0) __write_images(image_outputs, display_image_num, '%s/gen_%s.jpg' % (image_directory, postfix)) def _write_row(html_file, it, fn, all_size): html_file.write("<h3>iteration [%d] (%s)</h3>" % (it, fn.split('/')[-1])) html_file.write(""" <p><a href="%s"> <img src="%s" style="width:%dpx"> </a><br> <p> """ % (fn, fn, all_size)) return def write_html(filename, it, img_save_it, img_dir, all_size=1536): html_file = open(filename, "w") html_file.write(''' <!DOCTYPE html> <html> <head> <title>Experiment name = %s</title> <meta http-equiv="refresh" content="30"> </head> <body> ''' % os.path.basename(filename)) html_file.write("<h3>current</h3>") _write_row(html_file, it, '%s/gen_train_current.jpg' % img_dir, all_size) for j in range(it, img_save_it - 1, -1): _write_row(html_file, j, '%s/gen_train_%08d.jpg' % (img_dir, j), all_size) html_file.write("</body></html>") html_file.close() def write_loss(iterations, trainer, train_writer): members = [attr for attr in dir(trainer) if ((not callable(getattr(trainer, attr)) and not attr.startswith("__")) and ('loss' in attr or 'grad' in attr or 'nwd' in attr or 'accuracy' in attr))] for m in members: train_writer.add_scalar(m, getattr(trainer, m), iterations + 1) class Timer: def __init__(self, msg): self.msg = msg self.start_time = None def __enter__(self): self.start_time = time.time() def __exit__(self, exc_type, exc_value, exc_tb): print(self.msg % (time.time() - self.start_time))
7,743
32.37931
77
py
FUNIT
FUNIT-master/data.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import os.path from PIL import Image import torch.utils.data as data def default_loader(path): return Image.open(path).convert('RGB') def default_filelist_reader(filelist): im_list = [] with open(filelist, 'r') as rf: for line in rf.readlines(): im_path = line.strip() im_list.append(im_path) return im_list class ImageLabelFilelist(data.Dataset): def __init__(self, root, filelist, transform=None, filelist_reader=default_filelist_reader, loader=default_loader, return_paths=False): self.root = root self.im_list = filelist_reader(os.path.join(filelist)) self.transform = transform self.loader = loader self.classes = sorted( list(set([path.split('/')[0] for path in self.im_list]))) self.class_to_idx = {self.classes[i]: i for i in range(len(self.classes))} self.imgs = [(im_path, self.class_to_idx[im_path.split('/')[0]]) for im_path in self.im_list] self.return_paths = return_paths print('Data loader') print("\tRoot: %s" % root) print("\tList: %s" % filelist) print("\tNumber of classes: %d" % (len(self.classes))) def __getitem__(self, index): im_path, label = self.imgs[index] path = os.path.join(self.root, im_path) img = self.loader(path) if self.transform is not None: img = self.transform(img) if self.return_paths: return img, label, path else: return img, label def __len__(self): return len(self.imgs)
1,913
29.870968
76
py
FUNIT
FUNIT-master/networks.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import numpy as np import torch from torch import nn from torch import autograd from blocks import LinearBlock, Conv2dBlock, ResBlocks, ActFirstResBlock def assign_adain_params(adain_params, model): # assign the adain_params to the AdaIN layers in model for m in model.modules(): if m.__class__.__name__ == "AdaptiveInstanceNorm2d": mean = adain_params[:, :m.num_features] std = adain_params[:, m.num_features:2*m.num_features] m.bias = mean.contiguous().view(-1) m.weight = std.contiguous().view(-1) if adain_params.size(1) > 2*m.num_features: adain_params = adain_params[:, 2*m.num_features:] def get_num_adain_params(model): # return the number of AdaIN parameters needed by the model num_adain_params = 0 for m in model.modules(): if m.__class__.__name__ == "AdaptiveInstanceNorm2d": num_adain_params += 2*m.num_features return num_adain_params class GPPatchMcResDis(nn.Module): def __init__(self, hp): super(GPPatchMcResDis, self).__init__() assert hp['n_res_blks'] % 2 == 0, 'n_res_blk must be multiples of 2' self.n_layers = hp['n_res_blks'] // 2 nf = hp['nf'] cnn_f = [Conv2dBlock(3, nf, 7, 1, 3, pad_type='reflect', norm='none', activation='none')] for i in range(self.n_layers - 1): nf_out = np.min([nf * 2, 1024]) cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')] cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')] cnn_f += [nn.ReflectionPad2d(1)] cnn_f += [nn.AvgPool2d(kernel_size=3, stride=2)] nf = np.min([nf * 2, 1024]) nf_out = np.min([nf * 2, 1024]) cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')] cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')] cnn_c = [Conv2dBlock(nf_out, hp['num_classes'], 1, 1, norm='none', activation='lrelu', activation_first=True)] self.cnn_f = nn.Sequential(*cnn_f) self.cnn_c = nn.Sequential(*cnn_c) def forward(self, x, y): assert(x.size(0) == y.size(0)) feat = self.cnn_f(x) out = self.cnn_c(feat) index = torch.LongTensor(range(out.size(0))).cuda() out = out[index, y, :, :] return out, feat def calc_dis_fake_loss(self, input_fake, input_label): resp_fake, gan_feat = self.forward(input_fake, input_label) total_count = torch.tensor(np.prod(resp_fake.size()), dtype=torch.float).cuda() fake_loss = torch.nn.ReLU()(1.0 + resp_fake).mean() correct_count = (resp_fake < 0).sum() fake_accuracy = correct_count.type_as(fake_loss) / total_count return fake_loss, fake_accuracy, resp_fake def calc_dis_real_loss(self, input_real, input_label): resp_real, gan_feat = self.forward(input_real, input_label) total_count = torch.tensor(np.prod(resp_real.size()), dtype=torch.float).cuda() real_loss = torch.nn.ReLU()(1.0 - resp_real).mean() correct_count = (resp_real >= 0).sum() real_accuracy = correct_count.type_as(real_loss) / total_count return real_loss, real_accuracy, resp_real def calc_gen_loss(self, input_fake, input_fake_label): resp_fake, gan_feat = self.forward(input_fake, input_fake_label) total_count = torch.tensor(np.prod(resp_fake.size()), dtype=torch.float).cuda() loss = -resp_fake.mean() correct_count = (resp_fake >= 0).sum() accuracy = correct_count.type_as(loss) / total_count return loss, accuracy, gan_feat def calc_grad2(self, d_out, x_in): batch_size = x_in.size(0) grad_dout = autograd.grad(outputs=d_out.mean(), inputs=x_in, create_graph=True, retain_graph=True, only_inputs=True)[0] grad_dout2 = grad_dout.pow(2) assert (grad_dout2.size() == x_in.size()) reg = grad_dout2.sum()/batch_size return reg class FewShotGen(nn.Module): def __init__(self, hp): super(FewShotGen, self).__init__() nf = hp['nf'] nf_mlp = hp['nf_mlp'] down_class = hp['n_downs_class'] down_content = hp['n_downs_content'] n_mlp_blks = hp['n_mlp_blks'] n_res_blks = hp['n_res_blks'] latent_dim = hp['latent_dim'] self.enc_class_model = ClassModelEncoder(down_class, 3, nf, latent_dim, norm='none', activ='relu', pad_type='reflect') self.enc_content = ContentEncoder(down_content, n_res_blks, 3, nf, 'in', activ='relu', pad_type='reflect') self.dec = Decoder(down_content, n_res_blks, self.enc_content.output_dim, 3, res_norm='adain', activ='relu', pad_type='reflect') self.mlp = MLP(latent_dim, get_num_adain_params(self.dec), nf_mlp, n_mlp_blks, norm='none', activ='relu') def forward(self, one_image, model_set): # reconstruct an image content, model_codes = self.encode(one_image, model_set) model_code = torch.mean(model_codes, dim=0).unsqueeze(0) images_trans = self.decode(content, model_code) return images_trans def encode(self, one_image, model_set): # extract content code from the input image content = self.enc_content(one_image) # extract model code from the images in the model set class_codes = self.enc_class_model(model_set) class_code = torch.mean(class_codes, dim=0).unsqueeze(0) return content, class_code def decode(self, content, model_code): # decode content and style codes to an image adain_params = self.mlp(model_code) assign_adain_params(adain_params, self.dec) images = self.dec(content) return images class ClassModelEncoder(nn.Module): def __init__(self, downs, ind_im, dim, latent_dim, norm, activ, pad_type): super(ClassModelEncoder, self).__init__() self.model = [] self.model += [Conv2dBlock(ind_im, dim, 7, 1, 3, norm=norm, activation=activ, pad_type=pad_type)] for i in range(2): self.model += [Conv2dBlock(dim, 2 * dim, 4, 2, 1, norm=norm, activation=activ, pad_type=pad_type)] dim *= 2 for i in range(downs - 2): self.model += [Conv2dBlock(dim, dim, 4, 2, 1, norm=norm, activation=activ, pad_type=pad_type)] self.model += [nn.AdaptiveAvgPool2d(1)] self.model += [nn.Conv2d(dim, latent_dim, 1, 1, 0)] self.model = nn.Sequential(*self.model) self.output_dim = dim def forward(self, x): return self.model(x) class ContentEncoder(nn.Module): def __init__(self, downs, n_res, input_dim, dim, norm, activ, pad_type): super(ContentEncoder, self).__init__() self.model = [] self.model += [Conv2dBlock(input_dim, dim, 7, 1, 3, norm=norm, activation=activ, pad_type=pad_type)] for i in range(downs): self.model += [Conv2dBlock(dim, 2 * dim, 4, 2, 1, norm=norm, activation=activ, pad_type=pad_type)] dim *= 2 self.model += [ResBlocks(n_res, dim, norm=norm, activation=activ, pad_type=pad_type)] self.model = nn.Sequential(*self.model) self.output_dim = dim def forward(self, x): return self.model(x) class Decoder(nn.Module): def __init__(self, ups, n_res, dim, out_dim, res_norm, activ, pad_type): super(Decoder, self).__init__() self.model = [] self.model += [ResBlocks(n_res, dim, res_norm, activ, pad_type=pad_type)] for i in range(ups): self.model += [nn.Upsample(scale_factor=2), Conv2dBlock(dim, dim // 2, 5, 1, 2, norm='in', activation=activ, pad_type=pad_type)] dim //= 2 self.model += [Conv2dBlock(dim, out_dim, 7, 1, 3, norm='none', activation='tanh', pad_type=pad_type)] self.model = nn.Sequential(*self.model) def forward(self, x): return self.model(x) class MLP(nn.Module): def __init__(self, in_dim, out_dim, dim, n_blk, norm, activ): super(MLP, self).__init__() self.model = [] self.model += [LinearBlock(in_dim, dim, norm=norm, activation=activ)] for i in range(n_blk - 2): self.model += [LinearBlock(dim, dim, norm=norm, activation=activ)] self.model += [LinearBlock(dim, out_dim, norm='none', activation='none')] self.model = nn.Sequential(*self.model) def forward(self, x): return self.model(x.view(x.size(0), -1))
10,860
39.830827
78
py
FUNIT
FUNIT-master/funit_model.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import copy import torch import torch.nn as nn from networks import FewShotGen, GPPatchMcResDis def recon_criterion(predict, target): return torch.mean(torch.abs(predict - target)) class FUNITModel(nn.Module): def __init__(self, hp): super(FUNITModel, self).__init__() self.gen = FewShotGen(hp['gen']) self.dis = GPPatchMcResDis(hp['dis']) self.gen_test = copy.deepcopy(self.gen) def forward(self, co_data, cl_data, hp, mode): xa = co_data[0].cuda() la = co_data[1].cuda() xb = cl_data[0].cuda() lb = cl_data[1].cuda() if mode == 'gen_update': # INSERT here the object detection # then, for each objects, keep the bounding box and encode it, # then merge (in some way, maybe replace the features), in the encoding # of the whole image, the encoding of each objects using its bounding box # information c_xa = self.gen.enc_content(xa) s_xa = self.gen.enc_class_model(xa) s_xb = self.gen.enc_class_model(xb) xt = self.gen.decode(c_xa, s_xb) # translation xr = self.gen.decode(c_xa, s_xa) # reconstruction l_adv_t, gacc_t, xt_gan_feat = self.dis.calc_gen_loss(xt, lb) l_adv_r, gacc_r, xr_gan_feat = self.dis.calc_gen_loss(xr, la) _, xb_gan_feat = self.dis(xb, lb) _, xa_gan_feat = self.dis(xa, la) l_c_rec = recon_criterion(xr_gan_feat.mean(3).mean(2), xa_gan_feat.mean(3).mean(2)) l_m_rec = recon_criterion(xt_gan_feat.mean(3).mean(2), xb_gan_feat.mean(3).mean(2)) l_x_rec = recon_criterion(xr, xa) l_adv = 0.5 * (l_adv_t + l_adv_r) acc = 0.5 * (gacc_t + gacc_r) l_total = (hp['gan_w'] * l_adv + hp['r_w'] * l_x_rec + hp[ 'fm_w'] * (l_c_rec + l_m_rec)) l_total.backward() return l_total, l_adv, l_x_rec, l_c_rec, l_m_rec, acc elif mode == 'dis_update': xb.requires_grad_() l_real_pre, acc_r, resp_r = self.dis.calc_dis_real_loss(xb, lb) l_real = hp['gan_w'] * l_real_pre l_real.backward(retain_graph=True) l_reg_pre = self.dis.calc_grad2(resp_r, xb) l_reg = 10 * l_reg_pre l_reg.backward() with torch.no_grad(): c_xa = self.gen.enc_content(xa) s_xb = self.gen.enc_class_model(xb) xt = self.gen.decode(c_xa, s_xb) l_fake_p, acc_f, resp_f = self.dis.calc_dis_fake_loss(xt.detach(), lb) l_fake = hp['gan_w'] * l_fake_p l_fake.backward() l_total = l_fake + l_real + l_reg acc = 0.5 * (acc_f + acc_r) return l_total, l_fake_p, l_real_pre, l_reg_pre, acc else: assert 0, 'Not support operation' def test(self, co_data, cl_data): self.eval() self.gen.eval() self.gen_test.eval() xa = co_data[0].cuda() xb = cl_data[0].cuda() c_xa_current = self.gen.enc_content(xa) s_xa_current = self.gen.enc_class_model(xa) s_xb_current = self.gen.enc_class_model(xb) xt_current = self.gen.decode(c_xa_current, s_xb_current) xr_current = self.gen.decode(c_xa_current, s_xa_current) c_xa = self.gen_test.enc_content(xa) s_xa = self.gen_test.enc_class_model(xa) s_xb = self.gen_test.enc_class_model(xb) xt = self.gen_test.decode(c_xa, s_xb) xr = self.gen_test.decode(c_xa, s_xa) self.train() return xa, xr_current, xt_current, xb, xr, xt def translate_k_shot(self, co_data, cl_data, k): self.eval() xa = co_data[0].cuda() xb = cl_data[0].cuda() c_xa_current = self.gen_test.enc_content(xa) if k == 1: c_xa_current = self.gen_test.enc_content(xa) s_xb_current = self.gen_test.enc_class_model(xb) xt_current = self.gen_test.decode(c_xa_current, s_xb_current) else: s_xb_current_before = self.gen_test.enc_class_model(xb) s_xb_current_after = s_xb_current_before.squeeze(-1).permute(1, 2, 0) s_xb_current_pool = torch.nn.functional.avg_pool1d( s_xb_current_after, k) s_xb_current = s_xb_current_pool.permute(2, 0, 1).unsqueeze(-1) xt_current = self.gen_test.decode(c_xa_current, s_xb_current) return xt_current def compute_k_style(self, style_batch, k): self.eval() style_batch = style_batch.cuda() s_xb_before = self.gen_test.enc_class_model(style_batch) s_xb_after = s_xb_before.squeeze(-1).permute(1, 2, 0) s_xb_pool = torch.nn.functional.avg_pool1d(s_xb_after, k) s_xb = s_xb_pool.permute(2, 0, 1).unsqueeze(-1) return s_xb def translate_simple(self, content_image, class_code): self.eval() xa = content_image.cuda() s_xb_current = class_code.cuda() c_xa_current = self.gen_test.enc_content(xa) xt_current = self.gen_test.decode(c_xa_current, s_xb_current) return xt_current
5,659
41.238806
85
py
FUNIT
FUNIT-master/train.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import torch import os import sys import argparse import shutil from tensorboardX import SummaryWriter from utils import get_config, get_train_loaders, make_result_folders from utils import write_loss, write_html, write_1images, Timer from trainer import Trainer import torch.backends.cudnn as cudnn # Enable auto-tuner to find the best algorithm to use for your hardware. cudnn.benchmark = True # SAMUEL 18.03.2021 CUDA GPU setting os.environ['CUDA_VISIBLE_DEVICES']='1' # SAMUEL 18.03.2021 CUDA GPU setting parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, default='configs/funit_animals.yaml', help='configuration file for training and testing') parser.add_argument('--output_path', type=str, default='.', help="outputs path") parser.add_argument('--multigpus', action="store_true") parser.add_argument('--batch_size', type=int, default=0) parser.add_argument('--test_batch_size', type=int, default=4) parser.add_argument("--resume", action="store_true") opts = parser.parse_args() # Load experiment setting config = get_config(opts.config) max_iter = config['max_iter'] # Override the batch size if specified. if opts.batch_size != 0: config['batch_size'] = opts.batch_size trainer = Trainer(config) trainer.cuda() if opts.multigpus: ngpus = torch.cuda.device_count() config['gpus'] = ngpus print("Number of GPUs: %d" % ngpus) trainer.model = torch.nn.DataParallel( trainer.model, device_ids=range(ngpus)) else: config['gpus'] = 1 loaders = get_train_loaders(config) train_content_loader = loaders[0] train_class_loader = loaders[1] test_content_loader = loaders[2] test_class_loader = loaders[3] # Setup logger and output folders model_name = os.path.splitext(os.path.basename(opts.config))[0] train_writer = SummaryWriter( os.path.join(opts.output_path + "/logs", model_name)) output_directory = os.path.join(opts.output_path + "/outputs", model_name) checkpoint_directory, image_directory = make_result_folders(output_directory) shutil.copy(opts.config, os.path.join(output_directory, 'config.yaml')) iterations = trainer.resume(checkpoint_directory, hp=config, multigpus=opts.multigpus) if opts.resume else 0 while True: for it, (co_data, cl_data) in enumerate( zip(train_content_loader, train_class_loader)): with Timer("Elapsed time in update: %f"): d_acc = trainer.dis_update(co_data, cl_data, config) g_acc = trainer.gen_update(co_data, cl_data, config, opts.multigpus) torch.cuda.synchronize() print('D acc: %.4f\t G acc: %.4f' % (d_acc, g_acc)) if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) if ((iterations + 1) % config['image_save_iter'] == 0 or ( iterations + 1) % config['image_display_iter'] == 0): if (iterations + 1) % config['image_save_iter'] == 0: key_str = '%08d' % (iterations + 1) write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') else: key_str = 'current' with torch.no_grad(): for t, (val_co_data, val_cl_data) in enumerate( zip(train_content_loader, train_class_loader)): if t >= opts.test_batch_size: break val_image_outputs = trainer.test(val_co_data, val_cl_data, opts.multigpus) write_1images(val_image_outputs, image_directory, 'train_%s_%02d' % (key_str, t)) for t, (test_co_data, test_cl_data) in enumerate( zip(test_content_loader, test_class_loader)): if t >= opts.test_batch_size: break test_image_outputs = trainer.test(test_co_data, test_cl_data, opts.multigpus) write_1images(test_image_outputs, image_directory, 'test_%s_%02d' % (key_str, t)) if (iterations + 1) % config['snapshot_save_iter'] == 0: trainer.save(checkpoint_directory, iterations, opts.multigpus) print('Saved model at iteration %d' % (iterations + 1)) iterations += 1 if iterations >= max_iter: print("Finish Training") sys.exit(0)
5,178
37.93985
78
py
FUNIT
FUNIT-master/trainer.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import copy import os import math import torch import torch.nn as nn import torch.nn.init as init from torch.optim import lr_scheduler from funit_model import FUNITModel def update_average(model_tgt, model_src, beta=0.999): with torch.no_grad(): param_dict_src = dict(model_src.named_parameters()) for p_name, p_tgt in model_tgt.named_parameters(): p_src = param_dict_src[p_name] assert(p_src is not p_tgt) p_tgt.copy_(beta*p_tgt + (1. - beta)*p_src) class Trainer(nn.Module): def __init__(self, cfg): super(Trainer, self).__init__() self.model = FUNITModel(cfg) lr_gen = cfg['lr_gen'] lr_dis = cfg['lr_dis'] dis_params = list(self.model.dis.parameters()) gen_params = list(self.model.gen.parameters()) self.dis_opt = torch.optim.RMSprop( [p for p in dis_params if p.requires_grad], lr=lr_gen, weight_decay=cfg['weight_decay']) self.gen_opt = torch.optim.RMSprop( [p for p in gen_params if p.requires_grad], lr=lr_dis, weight_decay=cfg['weight_decay']) self.dis_scheduler = get_scheduler(self.dis_opt, cfg) self.gen_scheduler = get_scheduler(self.gen_opt, cfg) self.apply(weights_init(cfg['init'])) self.model.gen_test = copy.deepcopy(self.model.gen) def gen_update(self, co_data, cl_data, hp, multigpus): self.gen_opt.zero_grad() al, ad, xr, cr, sr, ac = self.model(co_data, cl_data, hp, 'gen_update') self.loss_gen_total = torch.mean(al) self.loss_gen_recon_x = torch.mean(xr) self.loss_gen_recon_c = torch.mean(cr) self.loss_gen_recon_s = torch.mean(sr) self.loss_gen_adv = torch.mean(ad) self.accuracy_gen_adv = torch.mean(ac) self.gen_opt.step() this_model = self.model.module if multigpus else self.model update_average(this_model.gen_test, this_model.gen) return self.accuracy_gen_adv.item() def dis_update(self, co_data, cl_data, hp): self.dis_opt.zero_grad() al, lfa, lre, reg, acc = self.model(co_data, cl_data, hp, 'dis_update') self.loss_dis_total = torch.mean(al) self.loss_dis_fake_adv = torch.mean(lfa) self.loss_dis_real_adv = torch.mean(lre) self.loss_dis_reg = torch.mean(reg) self.accuracy_dis_adv = torch.mean(acc) self.dis_opt.step() return self.accuracy_dis_adv.item() def test(self, co_data, cl_data, multigpus): this_model = self.model.module if multigpus else self.model return this_model.test(co_data, cl_data) def resume(self, checkpoint_dir, hp, multigpus): this_model = self.model.module if multigpus else self.model last_model_name = get_model_list(checkpoint_dir, "gen") state_dict = torch.load(last_model_name) this_model.gen.load_state_dict(state_dict['gen']) this_model.gen_test.load_state_dict(state_dict['gen_test']) iterations = int(last_model_name[-11:-3]) last_model_name = get_model_list(checkpoint_dir, "dis") state_dict = torch.load(last_model_name) this_model.dis.load_state_dict(state_dict['dis']) state_dict = torch.load(os.path.join(checkpoint_dir, 'optimizer.pt')) self.dis_opt.load_state_dict(state_dict['dis']) self.gen_opt.load_state_dict(state_dict['gen']) self.dis_scheduler = get_scheduler(self.dis_opt, hp, iterations) self.gen_scheduler = get_scheduler(self.gen_opt, hp, iterations) print('Resume from iteration %d' % iterations) return iterations def save(self, snapshot_dir, iterations, multigpus): this_model = self.model.module if multigpus else self.model # Save generators, discriminators, and optimizers gen_name = os.path.join(snapshot_dir, 'gen_%08d.pt' % (iterations + 1)) dis_name = os.path.join(snapshot_dir, 'dis_%08d.pt' % (iterations + 1)) opt_name = os.path.join(snapshot_dir, 'optimizer.pt') torch.save({'gen': this_model.gen.state_dict(), 'gen_test': this_model.gen_test.state_dict()}, gen_name) torch.save({'dis': this_model.dis.state_dict()}, dis_name) torch.save({'gen': self.gen_opt.state_dict(), 'dis': self.dis_opt.state_dict()}, opt_name) def load_ckpt(self, ckpt_name): state_dict = torch.load(ckpt_name) self.model.gen.load_state_dict(state_dict['gen']) self.model.gen_test.load_state_dict(state_dict['gen_test']) def translate(self, co_data, cl_data): return self.model.translate(co_data, cl_data) def translate_k_shot(self, co_data, cl_data, k, mode): return self.model.translate_k_shot(co_data, cl_data, k, mode) def forward(self, *inputs): print('Forward function not implemented.') pass def get_model_list(dirname, key): if os.path.exists(dirname) is False: return None gen_models = [os.path.join(dirname, f) for f in os.listdir(dirname) if os.path.isfile(os.path.join(dirname, f)) and key in f and ".pt" in f] if gen_models is None: return None gen_models.sort() last_model_name = gen_models[-1] return last_model_name def get_scheduler(optimizer, hp, it=-1): if 'lr_policy' not in hp or hp['lr_policy'] == 'constant': scheduler = None # constant scheduler elif hp['lr_policy'] == 'step': scheduler = lr_scheduler.StepLR(optimizer, step_size=hp['step_size'], gamma=hp['gamma'], last_epoch=it) else: return NotImplementedError('%s not implemented', hp['lr_policy']) return scheduler def weights_init(init_type='gaussian'): def init_fun(m): classname = m.__class__.__name__ if (classname.find('Conv') == 0 or classname.find( 'Linear') == 0) and hasattr(m, 'weight'): if init_type == 'gaussian': init.normal_(m.weight.data, 0.0, 0.02) elif init_type == 'xavier': init.xavier_normal_(m.weight.data, gain=math.sqrt(2)) elif init_type == 'kaiming': init.kaiming_normal_(m.weight.data, a=0, mode='fan_in') elif init_type == 'orthogonal': init.orthogonal_(m.weight.data, gain=math.sqrt(2)) elif init_type == 'default': pass else: assert 0, "Unsupported initialization: {}".format(init_type) if hasattr(m, 'bias') and m.bias is not None: init.constant_(m.bias.data, 0.0) return init_fun
6,871
39.662722
79
py
FUNIT
FUNIT-master/blocks.py
""" Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). """ import torch import torch.nn.functional as F from torch import nn class ResBlocks(nn.Module): def __init__(self, num_blocks, dim, norm, activation, pad_type): super(ResBlocks, self).__init__() self.model = [] for i in range(num_blocks): self.model += [ResBlock(dim, norm=norm, activation=activation, pad_type=pad_type)] self.model = nn.Sequential(*self.model) def forward(self, x): return self.model(x) class ResBlock(nn.Module): def __init__(self, dim, norm='in', activation='relu', pad_type='zero'): super(ResBlock, self).__init__() model = [] model += [Conv2dBlock(dim, dim, 3, 1, 1, norm=norm, activation=activation, pad_type=pad_type)] model += [Conv2dBlock(dim, dim, 3, 1, 1, norm=norm, activation='none', pad_type=pad_type)] self.model = nn.Sequential(*model) def forward(self, x): residual = x out = self.model(x) out += residual return out class ActFirstResBlock(nn.Module): def __init__(self, fin, fout, fhid=None, activation='lrelu', norm='none'): super().__init__() self.learned_shortcut = (fin != fout) self.fin = fin self.fout = fout self.fhid = min(fin, fout) if fhid is None else fhid self.conv_0 = Conv2dBlock(self.fin, self.fhid, 3, 1, padding=1, pad_type='reflect', norm=norm, activation=activation, activation_first=True) self.conv_1 = Conv2dBlock(self.fhid, self.fout, 3, 1, padding=1, pad_type='reflect', norm=norm, activation=activation, activation_first=True) if self.learned_shortcut: self.conv_s = Conv2dBlock(self.fin, self.fout, 1, 1, activation='none', use_bias=False) def forward(self, x): x_s = self.conv_s(x) if self.learned_shortcut else x dx = self.conv_0(x) dx = self.conv_1(dx) out = x_s + dx return out class LinearBlock(nn.Module): def __init__(self, in_dim, out_dim, norm='none', activation='relu'): super(LinearBlock, self).__init__() use_bias = True self.fc = nn.Linear(in_dim, out_dim, bias=use_bias) # initialize normalization norm_dim = out_dim if norm == 'bn': self.norm = nn.BatchNorm1d(norm_dim) elif norm == 'in': self.norm = nn.InstanceNorm1d(norm_dim) elif norm == 'none': self.norm = None else: assert 0, "Unsupported normalization: {}".format(norm) # initialize activation if activation == 'relu': self.activation = nn.ReLU(inplace=False) elif activation == 'lrelu': self.activation = nn.LeakyReLU(0.2, inplace=False) elif activation == 'tanh': self.activation = nn.Tanh() elif activation == 'none': self.activation = None else: assert 0, "Unsupported activation: {}".format(activation) def forward(self, x): out = self.fc(x) if self.norm: out = self.norm(out) if self.activation: out = self.activation(out) return out class Conv2dBlock(nn.Module): def __init__(self, in_dim, out_dim, ks, st, padding=0, norm='none', activation='relu', pad_type='zero', use_bias=True, activation_first=False): super(Conv2dBlock, self).__init__() self.use_bias = use_bias self.activation_first = activation_first # initialize padding if pad_type == 'reflect': self.pad = nn.ReflectionPad2d(padding) elif pad_type == 'replicate': self.pad = nn.ReplicationPad2d(padding) elif pad_type == 'zero': self.pad = nn.ZeroPad2d(padding) else: assert 0, "Unsupported padding type: {}".format(pad_type) # initialize normalization norm_dim = out_dim if norm == 'bn': self.norm = nn.BatchNorm2d(norm_dim) elif norm == 'in': self.norm = nn.InstanceNorm2d(norm_dim) elif norm == 'adain': self.norm = AdaptiveInstanceNorm2d(norm_dim) elif norm == 'none': self.norm = None else: assert 0, "Unsupported normalization: {}".format(norm) # initialize activation if activation == 'relu': self.activation = nn.ReLU(inplace=False) elif activation == 'lrelu': self.activation = nn.LeakyReLU(0.2, inplace=False) elif activation == 'tanh': self.activation = nn.Tanh() elif activation == 'none': self.activation = None else: assert 0, "Unsupported activation: {}".format(activation) self.conv = nn.Conv2d(in_dim, out_dim, ks, st, bias=self.use_bias) def forward(self, x): if self.activation_first: if self.activation: x = self.activation(x) x = self.conv(self.pad(x)) if self.norm: x = self.norm(x) else: x = self.conv(self.pad(x)) if self.norm: x = self.norm(x) if self.activation: x = self.activation(x) return x class AdaptiveInstanceNorm2d(nn.Module): def __init__(self, num_features, eps=1e-5, momentum=0.1): super(AdaptiveInstanceNorm2d, self).__init__() self.num_features = num_features self.eps = eps self.momentum = momentum self.weight = None self.bias = None self.register_buffer('running_mean', torch.zeros(num_features)) self.register_buffer('running_var', torch.ones(num_features)) def forward(self, x): assert self.weight is not None and \ self.bias is not None, "Please assign AdaIN weight first" b, c = x.size(0), x.size(1) running_mean = self.running_mean.repeat(b) running_var = self.running_var.repeat(b) x_reshaped = x.contiguous().view(1, b * c, *x.size()[2:]) out = F.batch_norm( x_reshaped, running_mean, running_var, self.weight, self.bias, True, self.momentum, self.eps) return out.view(b, c, *x.size()[2:]) def __repr__(self): return self.__class__.__name__ + '(' + str(self.num_features) + ')'
6,986
34.647959
79
py
SimSiam-91.9-top1-acc-on-CIFAR10
SimSiam-91.9-top1-acc-on-CIFAR10-main/main.py
import argparse import time import math from os import path, makedirs import torch from torch import optim from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from torch.backends import cudnn from torchvision import datasets from torchvision import transforms from simsiam.loader import TwoCropsTransform from simsiam.model_factory import SimSiam from simsiam.criterion import SimSiamLoss from simsiam.validation import KNNValidation parser = argparse.ArgumentParser('arguments for training') parser.add_argument('--data_root', type=str, help='path to dataset directory') parser.add_argument('--exp_dir', type=str, help='path to experiment directory') parser.add_argument('--trial', type=str, default='1', help='trial id') parser.add_argument('--img_dim', default=32, type=int) parser.add_argument('--arch', default='resnet18', help='model name is used for training') parser.add_argument('--feat_dim', default=2048, type=int, help='feature dimension') parser.add_argument('--num_proj_layers', type=int, default=2, help='number of projection layer') parser.add_argument('--batch_size', type=int, default=512, help='batch_size') parser.add_argument('--num_workers', type=int, default=8, help='num of workers to use') parser.add_argument('--epochs', type=int, default=800, help='number of training epochs') parser.add_argument('--gpu', default=None, type=int, help='GPU id to use.') parser.add_argument('--loss_version', default='simplified', type=str, choices=['simplified', 'original'], help='do the same thing but simplified version is much faster. ()') parser.add_argument('--print_freq', default=10, type=int, help='print frequency') parser.add_argument('--eval_freq', default=5, type=int, help='evaluate model frequency') parser.add_argument('--save_freq', default=50, type=int, help='save model frequency') parser.add_argument('--resume', default=None, type=str, help='path to latest checkpoint') parser.add_argument('--learning_rate', type=float, default=0.05, help='learning rate') parser.add_argument('--weight_decay', type=float, default=5e-4, help='weight decay') parser.add_argument('--momentum', type=float, default=0.9, help='momentum') args = parser.parse_args() def main(): if not path.exists(args.exp_dir): makedirs(args.exp_dir) trial_dir = path.join(args.exp_dir, args.trial) logger = SummaryWriter(trial_dir) print(vars(args)) train_transforms = transforms.Compose([ transforms.RandomResizedCrop(args.img_dim, scale=(0.2, 1.)), transforms.RandomHorizontalFlip(), transforms.RandomApply([ transforms.ColorJitter(0.4, 0.4, 0.4, 0.1) # not strengthened ], p=0.8), transforms.RandomGrayscale(p=0.2), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)) ]) train_set = datasets.CIFAR10(root=args.data_root, train=True, download=True, transform=TwoCropsTransform(train_transforms)) train_loader = DataLoader(dataset=train_set, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True, drop_last=True) model = SimSiam(args) optimizer = optim.SGD(model.parameters(), lr=args.learning_rate, momentum=args.momentum, weight_decay=args.weight_decay) criterion = SimSiamLoss(args.loss_version) if args.gpu is not None: torch.cuda.set_device(args.gpu) model = model.cuda(args.gpu) criterion = criterion.cuda(args.gpu) cudnn.benchmark = True start_epoch = 1 if args.resume is not None: if path.isfile(args.resume): start_epoch, model, optimizer = load_checkpoint(model, optimizer, args.resume) print("Loaded checkpoint '{}' (epoch {})" .format(args.resume, start_epoch)) else: print("No checkpoint found at '{}'".format(args.resume)) # routine best_acc = 0.0 validation = KNNValidation(args, model.encoder) for epoch in range(start_epoch, args.epochs+1): adjust_learning_rate(optimizer, epoch, args) print("Training...") # train for one epoch train_loss = train(train_loader, model, criterion, optimizer, epoch, args) logger.add_scalar('Loss/train', train_loss, epoch) if epoch % args.eval_freq == 0: print("Validating...") val_top1_acc = validation.eval() print('Top1: {}'.format(val_top1_acc)) # save the best model if val_top1_acc > best_acc: best_acc = val_top1_acc save_checkpoint(epoch, model, optimizer, best_acc, path.join(trial_dir, '{}_best.pth'.format(args.trial)), 'Saving the best model!') logger.add_scalar('Acc/val_top1', val_top1_acc, epoch) # save the model if epoch % args.save_freq == 0: save_checkpoint(epoch, model, optimizer, val_top1_acc, path.join(trial_dir, 'ckpt_epoch_{}_{}.pth'.format(epoch, args.trial)), 'Saving...') print('Best accuracy:', best_acc) # save model save_checkpoint(epoch, model, optimizer, val_top1_acc, path.join(trial_dir, '{}_last.pth'.format(args.trial)), 'Saving the model at the last epoch.') def train(train_loader, model, criterion, optimizer, epoch, args): batch_time = AverageMeter('Time', ':6.3f') losses = AverageMeter('Loss', ':.4e') progress = ProgressMeter( len(train_loader), [batch_time, losses], prefix="Epoch: [{}]".format(epoch)) # switch to train mode model.train() end = time.time() for i, (images, _) in enumerate(train_loader): if args.gpu is not None: images[0] = images[0].cuda(args.gpu, non_blocking=True) images[1] = images[1].cuda(args.gpu, non_blocking=True) # compute output outs = model(im_aug1=images[0], im_aug2=images[1]) loss = criterion(outs['z1'], outs['z2'], outs['p1'], outs['p2']) # compute gradient and do SGD step optimizer.zero_grad() loss.backward() optimizer.step() # measure elapsed time losses.update(loss.item(), images[0].size(0)) batch_time.update(time.time() - end) end = time.time() if i % args.print_freq == 0: progress.display(i) return losses.avg def adjust_learning_rate(optimizer, epoch, args): """Decay the learning rate based on schedule""" lr = args.learning_rate # cosine lr schedule lr *= 0.5 * (1. + math.cos(math.pi * epoch / args.epochs)) for param_group in optimizer.param_groups: param_group['lr'] = lr class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self, name, fmt=':f'): self.name = name self.fmt = fmt self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def __str__(self): fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})' return fmtstr.format(**self.__dict__) class ProgressMeter(object): def __init__(self, num_batches, meters, prefix=""): self.batch_fmtstr = self._get_batch_fmtstr(num_batches) self.meters = meters self.prefix = prefix def display(self, batch): entries = [self.prefix + self.batch_fmtstr.format(batch)] entries += [str(meter) for meter in self.meters] print('\t'.join(entries)) def _get_batch_fmtstr(self, num_batches): num_digits = len(str(num_batches // 1)) fmt = '{:' + str(num_digits) + 'd}' return '[' + fmt + '/' + fmt.format(num_batches) + ']' def save_checkpoint(epoch, model, optimizer, acc, filename, msg): state = { 'epoch': epoch, 'arch': args.arch, 'state_dict': model.state_dict(), 'optimizer': optimizer.state_dict(), 'top1_acc': acc } torch.save(state, filename) print(msg) def load_checkpoint(model, optimizer, filename): checkpoint = torch.load(filename, map_location='cuda:0') start_epoch = checkpoint['epoch'] model.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) return start_epoch, model, optimizer if __name__ == '__main__': main()
9,087
33.687023
99
py
SimSiam-91.9-top1-acc-on-CIFAR10
SimSiam-91.9-top1-acc-on-CIFAR10-main/main_lincls.py
#!/usr/bin/env python # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import argparse import builtins import os import random import shutil import time import warnings import torch import torch.nn as nn import torch.nn.parallel import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.optim import torch.multiprocessing as mp from torch.utils.data import DataLoader import torch.utils.data.distributed import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models from simsiam.resnet_cifar import ResNet18, ResNet34, ResNet50, ResNet101, ResNet152 from PIL import Image model_names = sorted(name for name in models.__dict__ if name.islower() and not name.startswith("__") and callable(models.__dict__[name])) parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') parser.add_argument('data', metavar='DIR', help='path to dataset') parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet50', choices=model_names, help='model architecture: ' + ' | '.join(model_names) + ' (default: resnet50)') parser.add_argument('--num_cls', default=10, type=int, metavar='N', help='number of classes in dataset (output dimention of models)') parser.add_argument('-j', '--workers', default=8, type=int, metavar='N', help='number of data loading workers (default: 32)') parser.add_argument('--epochs', default=100, type=int, metavar='N', help='number of total epochs to run') parser.add_argument('--start_epoch', default=0, type=int, metavar='N', help='manual epoch number (useful on restarts)') parser.add_argument('-b', '--batch_size', default=256, type=int, metavar='N', help='mini-batch size (default: 256), this is the total ' 'batch size of all GPUs on the current node when ' 'using Data Parallel or Distributed Data Parallel') parser.add_argument('--lr', '--learning_rate', default=30., type=float, metavar='LR', help='initial learning rate', dest='lr') parser.add_argument('--schedule', default=[60, 80], nargs='*', type=int, help='learning rate schedule (when to drop lr by a ratio)') parser.add_argument('--momentum', default=0.9, type=float, metavar='M', help='momentum') parser.add_argument('--wd', '--weight_decay', default=0., type=float, metavar='W', help='weight decay (default: 0.)', dest='weight_decay') parser.add_argument('-p', '--print-freq', default=500, type=int, metavar='N', help='print frequency (default: 10)') parser.add_argument('--resume', default='', type=str, metavar='PATH', help='path to latest checkpoint (default: none)') parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true', help='evaluate model on validation set') parser.add_argument('--world-size', default=-1, type=int, help='number of nodes for distributed training') parser.add_argument('--rank', default=-1, type=int, help='node rank for distributed training') parser.add_argument('--dist_url', default='tcp://224.66.41.62:23456', type=str, help='url used to set up distributed training') parser.add_argument('--dist-backend', default='nccl', type=str, help='distributed backend') parser.add_argument('--seed', default=None, type=int, help='seed for initializing training. ') parser.add_argument('--gpu', default=None, type=int, help='GPU id to use.') parser.add_argument('--multiprocessing-distributed', action='store_true', help='Use multi-processing distributed training to launch ' 'N processes per node, which has N GPUs. This is the ' 'fastest way to use PyTorch for either single node or ' 'multi node data parallel training') parser.add_argument('--pretrained', default='', type=str, help='path to pretrained checkpoint') def get_backbone(backbone_name, num_cls=10): models = {'resnet18': ResNet18(low_dim=num_cls), 'resnet34': ResNet34(low_dim=num_cls), 'resnet50': ResNet50(low_dim=num_cls), 'resnet101': ResNet101(low_dim=num_cls), 'resnet152': ResNet152(low_dim=num_cls)} return models[backbone_name] best_acc1 = 0 def main(): args = parser.parse_args() print(vars(args)) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') if args.gpu is not None: warnings.warn('You have chosen a specific GPU. This will completely ' 'disable data parallelism.') if args.dist_url == "env://" and args.world_size == -1: args.world_size = int(os.environ["WORLD_SIZE"]) args.distributed = args.world_size > 1 or args.multiprocessing_distributed ngpus_per_node = torch.cuda.device_count() if args.multiprocessing_distributed: # Since we have ngpus_per_node processes per node, the total world_size # needs to be adjusted accordingly args.world_size = ngpus_per_node * args.world_size # Use torch.multiprocessing.spawn to launch distributed processes: the # main_worker process function mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args)) else: # Simply call main_worker function main_worker(args.gpu, ngpus_per_node, args) def main_worker(gpu, ngpus_per_node, args): global best_acc1 args.gpu = gpu # suppress printing if not master if args.multiprocessing_distributed and args.gpu != 0: def print_pass(*args): pass builtins.print = print_pass if args.gpu is not None: print("Use GPU: {} for training".format(args.gpu)) if args.distributed: if args.dist_url == "env://" and args.rank == -1: args.rank = int(os.environ["RANK"]) if args.multiprocessing_distributed: # For multiprocessing distributed training, rank needs to be the # global rank among all the processes args.rank = args.rank * ngpus_per_node + gpu dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size, rank=args.rank) # create model print("=> creating model '{}'".format(args.arch)) # model = models.__dict__[args.arch]() model = get_backbone(args.arch, args.num_cls) # freeze all layers but the last fc for name, param in model.named_parameters(): if name not in ['fc.weight', 'fc.bias']: param.requires_grad = False # init the fc layer model.fc.weight.data.normal_(mean=0.0, std=0.01) model.fc.bias.data.zero_() # load from pre-trained, before DistributedDataParallel constructor if args.pretrained: if os.path.isfile(args.pretrained): print("=> loading checkpoint '{}'".format(args.pretrained)) checkpoint = torch.load(args.pretrained, map_location="cpu") state_dict = checkpoint['state_dict'] new_state_dict = dict() for old_key, value in state_dict.items(): if old_key.startswith('backbone') and 'fc' not in old_key: new_key = old_key.replace('backbone.', '') new_state_dict[new_key] = value args.start_epoch = 0 msg = model.load_state_dict(new_state_dict, strict=False) assert set(msg.missing_keys) == {"fc.weight", "fc.bias"} print("=> loaded pre-trained model '{}'".format(args.pretrained)) else: print("=> no checkpoint found at '{}'".format(args.pretrained)) if args.distributed: # For multiprocessing distributed, DistributedDataParallel constructor # should always set the single device scope, otherwise, # DistributedDataParallel will use all available devices. if args.gpu is not None: torch.cuda.set_device(args.gpu) model.cuda(args.gpu) # When using a single GPU per process and per # DistributedDataParallel, we need to divide the batch size # ourselves based on the total number of GPUs we have args.batch_size = int(args.batch_size / ngpus_per_node) args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node) model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) else: model.cuda() # DistributedDataParallel will divide and allocate batch_size to all # available GPUs if device_ids are not set model = torch.nn.parallel.DistributedDataParallel(model) elif args.gpu is not None: torch.cuda.set_device(args.gpu) model = model.cuda(args.gpu) else: # DataParallel will divide and allocate batch_size to all available GPUs if args.arch.startswith('alexnet') or args.arch.startswith('vgg'): model.features = torch.nn.DataParallel(model.features) model.cuda() else: model = torch.nn.DataParallel(model).cuda() # define loss function (criterion) and optimizer criterion = nn.CrossEntropyLoss().cuda(args.gpu) # optimize only the linear classifier parameters = list(filter(lambda p: p.requires_grad, model.parameters())) assert len(parameters) == 2 # fc.weight, fc.bias optimizer = torch.optim.SGD(parameters, args.lr, momentum=args.momentum, weight_decay=args.weight_decay) # optionally resume from a checkpoint if args.resume: if os.path.isfile(args.resume): print("=> loading checkpoint '{}'".format(args.resume)) if args.gpu is None: checkpoint = torch.load(args.resume) else: # Map model to be loaded to specified single gpu. loc = 'cuda:{}'.format(args.gpu) checkpoint = torch.load(args.resume, map_location=loc) args.start_epoch = checkpoint['epoch'] best_acc1 = checkpoint['best_acc1'] if args.gpu is not None: # best_acc1 may be from a checkpoint from a different GPU best_acc1 = best_acc1.to(args.gpu) model.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) print("=> loaded checkpoint '{}' (epoch {})" .format(args.resume, checkpoint['epoch'])) else: print("=> no checkpoint found at '{}'".format(args.resume)) cudnn.benchmark = True # Data loading code transform_train = transforms.Compose([ transforms.RandomResizedCrop(32, scale=(0.8, 1.0), ratio=(3.0 / 4.0, 4.0 / 3.0), interpolation=Image.BICUBIC), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)) ]) transform_test = transforms.Compose([ transforms.Resize(int(32 * (8 / 7)), interpolation=Image.BICUBIC), transforms.CenterCrop(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)) ]) trainset = datasets.CIFAR10(args.data, train=True, transform=transform_train) valset = datasets.CIFAR10(args.data, train=False, transform=transform_test) if args.distributed: train_sampler = torch.utils.data.distributed.DistributedSampler(trainset) else: train_sampler = None train_loader = DataLoader(trainset, batch_size=args.batch_size, shuffle=(train_sampler is None), num_workers=args.workers, sampler=train_sampler, pin_memory=True, drop_last=True) val_loader = DataLoader(valset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=True, drop_last=True) if args.evaluate: validate(val_loader, model, criterion, args) return for epoch in range(args.start_epoch, args.epochs): if args.distributed: train_sampler.set_epoch(epoch) adjust_learning_rate(optimizer, epoch, args) # train for one epoch train(train_loader, model, criterion, optimizer, epoch, args) # evaluate on validation set acc1 = validate(val_loader, model, criterion, args) # remember best acc@1 and save checkpoint is_best = acc1 > best_acc1 best_acc1 = max(acc1, best_acc1) if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): save_checkpoint({ 'epoch': epoch + 1, 'arch': args.arch, 'state_dict': model.state_dict(), 'best_acc1': best_acc1, 'optimizer': optimizer.state_dict(), }, is_best) if epoch == args.start_epoch: sanity_check(model.state_dict(), args.pretrained) print('Best acc:', best_acc1) def train(train_loader, model, criterion, optimizer, epoch, args): batch_time = AverageMeter('Time', ':6.3f') data_time = AverageMeter('Data', ':6.3f') losses = AverageMeter('Loss', ':.4e') top1 = AverageMeter('Acc@1', ':6.2f') top5 = AverageMeter('Acc@5', ':6.2f') progress = ProgressMeter( len(train_loader), [batch_time, data_time, losses, top1, top5], prefix="Epoch: [{}]".format(epoch)) """ Switch to eval mode: Under the protocol of linear classification on frozen features/models, it is not legitimate to change any part of the pre-trained model. BatchNorm in train mode may revise running mean/std (even if it receives no gradient), which are part of the model parameters too. """ model.eval() end = time.time() for i, (images, target) in enumerate(train_loader): # measure data loading time data_time.update(time.time() - end) if args.gpu is not None: images = images.cuda(args.gpu, non_blocking=True) target = target.cuda(args.gpu, non_blocking=True) # compute output output = model(images) loss = criterion(output, target) # measure accuracy and record loss acc1, acc5 = accuracy(output, target, topk=(1, 5)) losses.update(loss.item(), images.size(0)) top1.update(acc1[0], images.size(0)) top5.update(acc5[0], images.size(0)) # compute gradient and do SGD step optimizer.zero_grad() loss.backward() optimizer.step() # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % args.print_freq == 0: progress.display(i) def validate(val_loader, model, criterion, args): batch_time = AverageMeter('Time', ':6.3f') losses = AverageMeter('Loss', ':.4e') top1 = AverageMeter('Acc@1', ':6.2f') top5 = AverageMeter('Acc@5', ':6.2f') progress = ProgressMeter( len(val_loader), [batch_time, losses, top1, top5], prefix='Test: ') # switch to evaluate mode model.eval() with torch.no_grad(): end = time.time() for i, (images, target) in enumerate(val_loader): if args.gpu is not None: images = images.cuda(args.gpu, non_blocking=True) target = target.cuda(args.gpu, non_blocking=True) # compute output output = model(images) loss = criterion(output, target) # measure accuracy and record loss acc1, acc5 = accuracy(output, target, topk=(1, 5)) losses.update(loss.item(), images.size(0)) top1.update(acc1[0], images.size(0)) top5.update(acc5[0], images.size(0)) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % args.print_freq == 0: progress.display(i) # TODO: this should also be done with the ProgressMeter print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}' .format(top1=top1, top5=top5)) return top1.avg def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'): torch.save(state, filename) if is_best: shutil.copyfile(filename, 'model_best.pth.tar') def sanity_check(state_dict, pretrained_weights): """ Linear classifier should not change any weights other than the linear layer. This sanity check asserts nothing wrong happens (e.g., BN stats updated). """ print("=> loading '{}' for sanity check".format(pretrained_weights)) checkpoint = torch.load(pretrained_weights, map_location="cpu") state_dict_pre = checkpoint['state_dict'] for k in list(state_dict.keys()): # only ignore fc layer if 'fc.weight' in k or 'fc.bias' in k: continue k_pre = 'backbone.' + k[len('module.'):] \ if k.startswith('module.') else 'backbone.' + k assert ((state_dict[k].cpu() == state_dict_pre[k_pre]).all()), \ '{} is changed in linear classifier training.'.format(k) print("=> sanity check passed.") class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self, name, fmt=':f'): self.name = name self.fmt = fmt self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def __str__(self): fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})' return fmtstr.format(**self.__dict__) class ProgressMeter(object): def __init__(self, num_batches, meters, prefix=""): self.batch_fmtstr = self._get_batch_fmtstr(num_batches) self.meters = meters self.prefix = prefix def display(self, batch): entries = [self.prefix + self.batch_fmtstr.format(batch)] entries += [str(meter) for meter in self.meters] print('\t'.join(entries)) def _get_batch_fmtstr(self, num_batches): num_digits = len(str(num_batches // 1)) fmt = '{:' + str(num_digits) + 'd}' return '[' + fmt + '/' + fmt.format(num_batches) + ']' def adjust_learning_rate(optimizer, epoch, args): """Decay the learning rate based on schedule""" lr = args.lr for milestone in args.schedule: lr *= 0.1 if epoch >= milestone else 1. for param_group in optimizer.param_groups: param_group['lr'] = lr def accuracy(output, target, topk=(1,)): """Computes the accuracy over the k top predictions for the specified values of k""" with torch.no_grad(): maxk = max(topk) batch_size = target.size(0) _, pred = output.topk(maxk, 1, True, True) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) res = [] for k in topk: correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True) res.append(correct_k.mul_(100.0 / batch_size)) return res if __name__ == '__main__': main()
20,587
38.066414
95
py
SimSiam-91.9-top1-acc-on-CIFAR10
SimSiam-91.9-top1-acc-on-CIFAR10-main/simsiam/model_factory.py
from torch import nn from .resnet_cifar import ResNet18, ResNet34, ResNet50, ResNet101, ResNet152 class projection_MLP(nn.Module): def __init__(self, in_dim, out_dim, num_layers=2): super().__init__() hidden_dim = out_dim self.num_layers = num_layers self.layer1 = nn.Sequential( nn.Linear(in_dim, hidden_dim), nn.BatchNorm1d(hidden_dim), nn.ReLU(inplace=True) ) self.layer2 = nn.Sequential( nn.Linear(hidden_dim, hidden_dim), nn.BatchNorm1d(hidden_dim), nn.ReLU(inplace=True) ) self.layer3 = nn.Sequential( nn.Linear(hidden_dim, out_dim), nn.BatchNorm1d(out_dim, affine=False) # Page:5, Paragraph:2 ) def forward(self, x): if self.num_layers == 2: x = self.layer1(x) x = self.layer3(x) elif self.num_layers == 3: x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) return x class prediction_MLP(nn.Module): def __init__(self, in_dim=2048): super().__init__() out_dim = in_dim hidden_dim = int(out_dim / 4) self.layer1 = nn.Sequential( nn.Linear(in_dim, hidden_dim), nn.BatchNorm1d(hidden_dim), nn.ReLU(inplace=True) ) self.layer2 = nn.Linear(hidden_dim, out_dim) def forward(self, x): x = self.layer1(x) x = self.layer2(x) return x class SimSiam(nn.Module): def __init__(self, args): super(SimSiam, self).__init__() self.backbone = SimSiam.get_backbone(args.arch) out_dim = self.backbone.fc.weight.shape[1] self.backbone.fc = nn.Identity() self.projector = projection_MLP(out_dim, args.feat_dim, args.num_proj_layers) self.encoder = nn.Sequential( self.backbone, self.projector ) self.predictor = prediction_MLP(args.feat_dim) @staticmethod def get_backbone(backbone_name): return {'resnet18': ResNet18(), 'resnet34': ResNet34(), 'resnet50': ResNet50(), 'resnet101': ResNet101(), 'resnet152': ResNet152()}[backbone_name] def forward(self, im_aug1, im_aug2): z1 = self.encoder(im_aug1) z2 = self.encoder(im_aug2) p1 = self.predictor(z1) p2 = self.predictor(z2) return {'z1': z1, 'z2': z2, 'p1': p1, 'p2': p2}
2,575
24.76
76
py
SimSiam-91.9-top1-acc-on-CIFAR10
SimSiam-91.9-top1-acc-on-CIFAR10-main/simsiam/validation.py
# https://github.com/zhirongw/lemniscate.pytorch/blob/master/test.py import torch from torch.utils.data import DataLoader from torchvision import transforms from torchvision import datasets from torch import nn class KNNValidation(object): def __init__(self, args, model, K=1): self.model = model self.device = torch.device('cuda' if next(model.parameters()).is_cuda else 'cpu') self.args = args self.K = K base_transforms = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) train_dataset = datasets.CIFAR10(root=args.data_root, train=True, download=True, transform=base_transforms) self.train_dataloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=True) val_dataset = datasets.CIFAR10(root=args.data_root, train=False, download=True, transform=base_transforms) self.val_dataloader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=True) def _topk_retrieval(self): """Extract features from validation split and search on train split features.""" n_data = self.train_dataloader.dataset.data.shape[0] feat_dim = self.args.feat_dim self.model.eval() if str(self.device) == 'cuda': torch.cuda.empty_cache() train_features = torch.zeros([feat_dim, n_data], device=self.device) with torch.no_grad(): for batch_idx, (inputs, _) in enumerate(self.train_dataloader): inputs = inputs.to(self.device) batch_size = inputs.size(0) # forward features = self.model(inputs) features = nn.functional.normalize(features) train_features[:, batch_idx * batch_size:batch_idx * batch_size + batch_size] = features.data.t() train_labels = torch.LongTensor(self.train_dataloader.dataset.targets).cuda() total = 0 correct = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(self.val_dataloader): targets = targets.cuda(non_blocking=True) batch_size = inputs.size(0) features = self.model(inputs.to(self.device)) dist = torch.mm(features, train_features) yd, yi = dist.topk(self.K, dim=1, largest=True, sorted=True) candidates = train_labels.view(1, -1).expand(batch_size, -1) retrieval = torch.gather(candidates, 1, yi) retrieval = retrieval.narrow(1, 0, 1).clone().view(-1) total += targets.size(0) correct += retrieval.eq(targets.data).sum().item() top1 = correct / total return top1 def eval(self): return self._topk_retrieval()
3,662
38.815217
113
py
SimSiam-91.9-top1-acc-on-CIFAR10
SimSiam-91.9-top1-acc-on-CIFAR10-main/simsiam/resnet_cifar.py
'''ResNet in PyTorch. For Pre-activation ResNet, see 'preact_resnet.py'. Reference: [1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun Deep Residual Learning for Image Recognition. arXiv:1512.03385 ''' import torch import torch.nn as nn import torch.nn.functional as F # from lib.normalize import Normalize from torch.autograd import Variable class BasicBlock(nn.Module): expansion = 1 def __init__(self, in_planes, planes, stride=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.shortcut = nn.Sequential() if stride != 1 or in_planes != self.expansion*planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(self.expansion*planes) ) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) out += self.shortcut(x) out = F.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, in_planes, planes, stride=1): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, self.expansion*planes, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(self.expansion*planes) self.shortcut = nn.Sequential() if stride != 1 or in_planes != self.expansion*planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(self.expansion*planes) ) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = F.relu(self.bn2(self.conv2(out))) out = self.bn3(self.conv3(out)) out += self.shortcut(x) out = F.relu(out) return out class ResNet(nn.Module): def __init__(self, block, num_blocks, low_dim=128): super(ResNet, self).__init__() self.in_planes = 64 self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(64) self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) self.fc = nn.Linear(512*block.expansion, low_dim) # self.l2norm = Normalize(2) def _make_layer(self, block, planes, num_blocks, stride): strides = [stride] + [1]*(num_blocks-1) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, stride)) self.in_planes = planes * block.expansion return nn.Sequential(*layers) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = self.layer4(out) out = F.avg_pool2d(out, 4) out = out.view(out.size(0), -1) out = self.fc(out) # out = self.l2norm(out) return out def ResNet18(low_dim=128): return ResNet(BasicBlock, [2,2,2,2], low_dim) def ResNet34(low_dim=128): return ResNet(BasicBlock, [3,4,6,3], low_dim) def ResNet50(low_dim=128): return ResNet(Bottleneck, [3,4,6,3], low_dim) def ResNet101(low_dim=128): return ResNet(Bottleneck, [3,4,23,3], low_dim) def ResNet152(low_dim=128): return ResNet(Bottleneck, [3,8,36,3], low_dim) def test(): net = ResNet18() y = net(Variable(torch.randn(1,3,32,32))) print(y.size()) # test()
4,245
32.433071
102
py
SimSiam-91.9-top1-acc-on-CIFAR10
SimSiam-91.9-top1-acc-on-CIFAR10-main/simsiam/criterion.py
from torch import nn class SimSiamLoss(nn.Module): def __init__(self, version='simplified'): super().__init__() self.ver = version def asymmetric_loss(self, p, z): if self.ver == 'original': z = z.detach() # stop gradient p = nn.functional.normalize(p, dim=1) z = nn.functional.normalize(z, dim=1) return -(p * z).sum(dim=1).mean() elif self.ver == 'simplified': z = z.detach() # stop gradient return - nn.functional.cosine_similarity(p, z, dim=-1).mean() def forward(self, z1, z2, p1, p2): loss1 = self.asymmetric_loss(p1, z2) loss2 = self.asymmetric_loss(p2, z1) return 0.5 * loss1 + 0.5 * loss2
751
24.066667
73
py
mIOHMM
mIOHMM-main/src/utils.py
import numpy as np import pickle import torch def save_pickle(data, filename): with open(filename, "wb") as f: pickle.dump(data, f, pickle.HIGHEST_PROTOCOL) def load_pickle(filepath): with open(filepath, "rb") as handle: data = pickle.load(handle) return data def normalize(A, axis=None): Z = torch.sum(A, axis=axis, keepdims=True) idx = np.where(Z == 0) Z[idx] = 1 return A / Z def normalize_exp(log_P, axis=None): a, _ = torch.max(log_P, keepdims=True, axis=axis) P = normalize(torch.exp(log_P - a), axis=axis) return P
586
19.964286
53
py
mIOHMM
mIOHMM-main/src/piomhmm.py
from scipy.special import gamma as gamma_fn from sklearn.cluster import KMeans from src.utils import normalize_exp import math import numpy as np import pickle import torch torch.set_default_dtype(torch.float64) class mHMM: def __init__( self, data, ins=None, K=2, k=5, TM=None, OM=None, full_cov=False, io=True, state_io=False, personalized_io=False, personalized=False, eps=0, min_var=1e-6, device="cpu", priorV=False, priorMu=False, sample_num=1, alpha=10.0, beta=5.0, UT=False, var_fill=0.5, VI_diag=False, lr=0.001, ): """ personalized input-output hidden markov model. This class of models considers patient observations that are modeled by several possible factors, which are turned on or off using flags. The most complete version of the model is x_i,t | z_i,t = k, d_i,t ~ N(mu_k + R_i + (V_k + M_i)*D_i,t, sigma_k) where x_i,t is the observed data, z_i,t is the latent state, d_i,t is the observed drug information, R_i is a personalized state effect, V_k is a state-based drug effect, M_i is a personalized drug effect, and sigma_k is the covariance. :param data: an n x t x d matrix of clinical observations :param ins: an n x t matrix of input/drug information (note that this is assumed to be univariate) :param k: number of latent states :param TM: the time mask, indicates trailing zeros in the observation array :param OM: the observation mask, indicates missing observations within the time series (i.e. missed visit) :param full_cov: flag indicating if a full covariance matrix should be used, alternatively a diagonal covariance is used :param io: flag indicating if the model is an input-output HMM; drugs should not be none if io=True :param state_io: flag indicating if input-output effects should be a function of state, if io=True and state_io=False, V_k = V for all k. This flag should not be True if io=False :param personalized_io: flag indicating if input-output effects should be a function of patient (i.e. M_i is 'turned on'). This flag should not be True if io=False :param personalized: flag indicating if personalized state effects should be applied (i.e. R_i is 'turned on'). :param eps: prevent division by zero :param min_var: set a minimum allowable variance :param device: either cpu or cuda :param priorV: indicates if priors should be used for the state- and personalized-drug effects :param priorMu: indicates if prios should be used for the state-means :param sample_num: number of saples used in MC sampling; only 1 sample is currently supported :param alpha: parameter of the inverse gamma distribution used as prior for V_k and M_i :param beta: parameter of the inverse gamma distribution used as prior for V_k and M_i :param UT: parameter to enforce an upper triangular structure for the transition matrix :param var_fill: parameter to specify initial guess for variance :param VI_diag: flag to indicate whether or not the variational distributions should have a diagonal covariance structure """ # number of HMM mixtures self.K = K # number of latent states per HMM self.k_per_hmm = k # number of latent states self.k = k * K # flag to indicate whether or not to use a full covariance matrix (alternative is diagonal) self.full_cov = full_cov # flag to indicate whether or not the model is input-output self.io = io # flag to indicate if io effects should be a function of state self.state_io = state_io # flag to indicate if io effects should be personalized self.perso_io = personalized_io # flag to indicate if personalized (non-io) effects should be included self.perso = personalized # flag to indicate whether or not to use GPU self.device = device self.tensor_constructor = ( torch.cuda.DoubleTensor if device == "cuda" and torch.cuda.is_available() else torch.Tensor ) # flag to indicate whether or not to have a prior on V self.priorV = priorV # flag to indicate whether or not to have a prior on mu self.priorMu = priorMu # store the parameters of the IG prior self.alpha = torch.tensor( [alpha], dtype=torch.float64, requires_grad=False, device=self.device ).float() self.beta = torch.tensor( [beta], dtype=torch.float64, requires_grad=False, device=self.device ).float() # flag to indicate upper triangular structure for transition matrix self.ut = UT # flag to indicate whether to use diagonal covariance for variational distribution self.VI_diag = VI_diag # store the data used in analysis self.data = data.to(device=self.device) # n x t x d self.n, self.t, self.d = self.data.shape # store likelihood (this is the objective if no personalized effects are used) self.ll = [] if self.perso_io and self.perso: # case with both personalized state and medication effects self.elbo = [] # objective self.mu_hat = torch.zeros( self.n, self.d, requires_grad=True, device=self.device ) # vector for optimization of covariance which is mapped into lower triangular cholesky factor if self.VI_diag: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * self.d), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril_vec else: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril_vec self.nu_hat = torch.zeros( self.n, self.d, requires_grad=True, device=self.device ) # vector for optimization of covariance which is mapped into lower triangular cholesky factor if self.VI_diag: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * self.d), requires_grad=True, device=self.device, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril else: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril self.optimizer = torch.optim.Adam( [self.mu_hat, self.tril_vec, self.nu_hat, self.tril], lr=lr ) elif self.perso_io: # case with personalized medication effects self.elbo = [] # objective self.mu_hat = torch.zeros( self.n, self.d, requires_grad=True, device=self.device ) # vector for optimization of covariance which is mapped into lower triangular cholesky factor if self.VI_diag: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * self.d), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril_vec else: self.tril_vec = torch.tensor( 0.1 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril_vec self.optimizer = torch.optim.Adam([self.mu_hat, self.tril_vec], lr=lr) elif self.perso: # case with personalized state effects self.elbo = [] # objective self.nu_hat = torch.zeros( self.n, self.d, requires_grad=True, device=self.device ) # vector for optimization of covariance which is mapped into lower triangular cholesky factor if self.VI_diag: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * self.d), requires_grad=True, device=self.device, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril else: self.tril = torch.tensor( 0.1 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril self.optimizer = torch.optim.Adam([self.nu_hat, self.tril], lr=lr, eps=1e-4) # store the inputs used in analysis if self.io: self.ins = ins.to(self.device) # n x t x 1 # store the time mask if TM is None: self.tm = torch.ones( self.n, self.t, requires_grad=False, device=self.device ) else: self.tm = TM.to(self.device) # n x t # store the observation mask if OM is None: self.om = torch.ones( self.n, self.t, requires_grad=False, device=self.device ) else: self.om = OM.to(self.device) # n x t self.om_is = torch.ones( (self.k, self.k, self.n, self.t - 1), requires_grad=False, device=device, dtype=torch.bool, ) for i in range(self.k): self.om_is[i] = self.om[:, 1:].unsqueeze(0).repeat(self.k, 1, 1) self.eps = eps self.min_var = min_var self.ini_var = var_fill self.sample_num = sample_num def initialize_model(self, km_init=True): """ Initializes the parameters of the PIOHMM model; internal method km_init: flag to indicate if kmeans should be used to initialize the state means """ # All implementations of the model have the parameter set {mu, var, pi, A} if km_init: # initialize the means using kmeans kmeans = KMeans(n_clusters=self.k, init="random").fit( torch.reshape(self.data[:, :, :].cpu(), [self.n * self.t, self.d]) ) mu = torch.tensor( kmeans.cluster_centers_, requires_grad=False, device=self.device ).float() else: # choose k initial points from data to initialize means idxs = torch.from_numpy(np.random.choice(self.n, self.k, replace=False)) mu = self.data[idxs, 0, :] if self.full_cov: # create k random symmetric positive definite d x d matrices R = 0.1 * torch.rand(self.k, self.d, self.d, requires_grad=False) var = torch.stack( [ 0.5 * (R[i, :, :].squeeze() + torch.t(R[i, :, :].squeeze())) + self.d * torch.eye(self.d) for i in range(self.k) ] ).to(self.device) else: var = self.tensor_constructor(self.k, self.d, device=self.device).fill_( self.ini_var ) # uniform prior pi = torch.empty(self.k, requires_grad=False, device=self.device).fill_( 1.0 / self.k ) # transition matrix if self.ut: # create UT matrices for each HMM A = [] for _ in range(self.K): A.append( torch.triu( torch.stack( [ 1.0 / (self.k_per_hmm - i) * torch.ones( self.k_per_hmm, requires_grad=False, device=self.device, ) for i in range(self.k_per_hmm) ] ) ) ) # create the corresponding mask new_triangular = torch.block_diag( *[ torch.triu( torch.ones( self.k_per_hmm, self.k_per_hmm, requires_grad=False, device=self.device, ) ) for _ in range(self.K) ] ) else: # create non-UT matrices for each HMM A = [] for _ in range(self.K): A.append( torch.stack( [ 1.0 / self.k_per_hmm * torch.ones( self.k_per_hmm, requires_grad=False, device=self.device ) for _ in range(self.k_per_hmm) ] ) ) # create the corresponding mask new_triangular = torch.block_diag( *[ torch.ones( self.k_per_hmm, self.k_per_hmm, requires_grad=False, device=self.device, ) for _ in range(self.K) ] ) self.new_triangular = new_triangular # obtain a block diagonal transition matrix A = torch.block_diag(*A) # for calculation stability A += self.eps params = {"mu": mu, "var": var, "pi": pi, "A": A} # input transformation matrix if self.io: if self.state_io: V = torch.zeros(self.k, self.d, requires_grad=False, device=self.device) else: V = torch.zeros(self.d, requires_grad=False, device=self.device) params["V"] = V # variational parameters if self.perso_io: # transformation matrix prior noise # initialize using the mean of the IG distribution mnoise = torch.tensor([0.5], device=self.device) params["mnoise"] = mnoise if self.priorV: vnoise = torch.tensor([1.0], device=self.device) params["vnoise"] = vnoise if self.perso: nnoise = torch.tensor([0.5], device=self.device) params["nnoise"] = nnoise if self.priorMu: munoise = torch.tensor([1.0], device=self.device) params["munoise"] = munoise return params def batch_mahalanobis(self, L, x, check=True): """ Computes the squared Mahalanobis distance :math:`\mathbf{x}^\top\mathbf{M}^{-1}\mathbf{x}` for a factored :math:`\mathbf{M} = \mathbf{L}\mathbf{L}^\top`. internal method Accepts batches for both L and x. """ flat_L = L.unsqueeze(0).reshape((-1,) + L.shape[-2:]) L_inv = ( torch.stack([torch.inverse(Li.t()) for Li in flat_L]) .view(L.shape) .to(self.device) ) batch_val = L_inv.shape[0] if check: return ( ( torch.stack( [ x[i, :, :].unsqueeze(-1) * L_inv[i, :, :] for i in range(batch_val) ] ) ) .sum(-2) .pow(2.0) .sum(-1) ) else: return ( ( torch.stack( [ x[i, :].unsqueeze(-1) * L_inv[i, :, :] for i in range(batch_val) ] ) ) .sum(-2) .pow(2.0) .sum(-1) ) def batch_diag(self, bmat): """ Returns the diagonals of a batch of square matrices; internal method """ return bmat.reshape(bmat.shape[:-2] + (-1,))[..., :: bmat.size(-1) + 1] def log_gaussian(self, params, m_sample=None, n_sample=None): """ Returns the density of the model data given the current parameters; internal method :param params: set of model parameters :param m_sample: current sample of m_i, only applicable for perso_io=True :param n_sample: current sample of r_i, only applicable for perso=True :return: log likelihood at each time point for each possible cluster component, k x n x t """ # unpack params mu = params["mu"] var = params["var"] log_norm_constant = self.d * torch.log( 2 * torch.tensor(math.pi, device=self.device) ) if self.full_cov: try: # This try statement helps catch issues related to singular covariance, which can be an issue that is difficult to trace L = torch.linalg.cholesky(var) except: print(var) print(mu) print(params["A"]) print(params["V"]) with open("miohmm_var.pkl", "wb") as handle: pickle.dump(var, handle) r = self.data[None, :, :, :] - mu[:, None, None, :] if self.io: V = params["V"] if self.state_io: r = r - V[:, None, None, :] * self.ins[None, :, :, None] else: r = r - V * self.ins[:, :, None] if self.perso_io: r = r - m_sample[None, :, None, :] * self.ins[None, :, :, None] if self.perso: r = r - n_sample[None, :, None, :] md = self.batch_mahalanobis(L, r) log_det = 2 * self.batch_diag(L).abs().log().sum(-1).to(self.device) log_p = -0.5 * (md + log_norm_constant + log_det[:, None, None]) else: r = self.data[None, :, :, :] - mu[:, None, None, :] if self.io: V = params["V"] if self.state_io: r = r - V[:, None, None, :] * self.ins[None, :, :, None] else: r = r - V * self.ins[:, :, None] if self.perso_io: r = r - m_sample[None, :, None, :] * self.ins[:, :, None] if self.perso: r = r - n_sample[None, :, None, :] r = r ** 2 log_p = -0.5 * (var.log()[:, None, None, :] + r / var[:, None, None, :]) log_p = log_p + log_norm_constant log_p = log_p.sum(-1) return log_p def log_gaussian_prior(self, rv, mu, L): """ Returns the probability of random varaible rv with mean mu and variance var; does not support full covariance structure; internal method :param rv d :param mu d :param L cholesky matrix for covariance of RV :return: log probability """ d = np.shape(rv)[0] log_norm_constant = ( -0.5 * d * torch.log(2 * torch.tensor(math.pi, device=self.device)) ) r = rv - mu md = self.batch_mahalanobis(L, r, check=False) log_det = self.batch_diag(L).abs().log().sum(-1).to(self.device) log_p = -0.5 * md + log_norm_constant - log_det return log_p def log_ig(self, noise): """ Returns the probability of the inverse gamma prior; internal method :return: """ log_ig = ( self.alpha * torch.log(self.beta) - torch.log(gamma_fn(self.alpha.cpu())).to(self.device) - (self.alpha + 1.0) * torch.log(noise) - self.beta / noise ) return log_ig def get_likelihoods(self, params, log=True, m_sample=None, n_sample=None): """ :param log: flag to indicate if likelihood should be returned in log domain; internal method :return likelihoods: (k x n x t) """ log_likelihoods = self.log_gaussian(params, m_sample, n_sample) if not log: log_likelihoods.exp_() # multiply the liklihoods by the observation mask return (log_likelihoods * self.om[None, :, :]).to(self.device) def get_exp_data(self, mu, var, V=None, m_sample=None, n_sample=None): """ Function to calculate the expectation of the conditional log-likelihood with respect to the variational approximation q(M|X); internal method :return: expectation of the conditional log-likelihood wrt the variational approximation """ if self.full_cov: L = torch.linalg.cholesky(var) else: L = torch.zeros(self.k, self.d, self.d) for i in range(self.k): L[i, :, :] = torch.diag(torch.sqrt(var[i, :])) r = self.data[None, :, :, :] - mu[:, None, None, :] if self.io: r = r - V[:, None, None, :] * self.ins[None, :, :, None] if self.perso_io: r = r - m_sample[None, :, None, :] * self.ins[None, :, :, None] if self.perso: r = r - n_sample[None, :, None, :] const = self.d * torch.log( 2 * torch.tensor(math.pi, device=self.device) ) # scalar logdet = 2 * self.batch_diag(L).abs().log().sum(-1).to(self.device) # k md1 = self.batch_mahalanobis(L, r) # k x n x t out = -0.5 * (const + logdet[:, None, None] + md1) return out def get_exp_M(self, mnoise): """ Function to calculate the expectation of the prior with respect to the variational approximation q(M|X); internal method :return: expectation of the prior on M wrt the variational approximation """ out = ( -self.d * self.n / 2 * torch.log(2 * torch.tensor(math.pi, device=self.device) * mnoise) - ( 1 / mnoise / 2 * torch.einsum("kij, kij -> k", [self.L_hat, self.L_hat]) ).sum() - ( 1 / mnoise / 2 * torch.einsum("ij,ij->i", [self.mu_hat, self.mu_hat]) ).sum() ) return out def get_exp_V(self, V, vnoise): """ Function to calculate the expectation of the prior with respect to the variational approximation q(M|X); internal method :return: expectation of the prior on M wrt the variational approximation """ out = ( -self.d * self.k / 2 * torch.log(2 * torch.tensor(math.pi, device=self.device) * vnoise) - (1 / vnoise / 2 * torch.einsum("ij,ij->i", [V, V])).sum() ) return out def get_exp_Mtilde(self, nnoise): """ Function to calculate the expectation of the prior with respect to the variational approximation q(M|X); internal method :return: expectation of the prior on M wrt the variational approximation """ out = ( -self.d * self.n / 2 * torch.log(2 * torch.tensor(math.pi, device=self.device) * nnoise) - ( 1 / nnoise / 2 * torch.einsum("kij, kij -> k", [self.N_hat, self.N_hat]) ).sum() - ( 1 / nnoise / 2 * torch.einsum("ij,ij->i", [self.nu_hat, self.nu_hat]) ).sum() ) return out def exp_log_joint(self, params, e_out, samples): """ Function to calculate the expectation of the joint likelihood with respect to the variational approximation; internal method :return: log joint likelihood """ # unpack parameters pi = params["pi"] A = params["A"] logA = A.log() logA[torch.isinf(logA)] = 0 mu = params["mu"] var = params["var"] m_sample = samples["m_sample"] n_sample = samples["n_sample"] gamma = e_out["gamma"] xi = e_out["xi"] if self.io: V = params["V"] lj = ( (gamma[:, :, 0].exp() * pi[:, None].log()).sum() + (xi.exp() * logA[:, :, None, None] * self.om[None, None, :, 1:]).sum() + ( self.get_exp_data( mu, var, V=V, m_sample=m_sample, n_sample=n_sample ) * gamma.exp() * self.om[None, :, :] ).sum() ) else: lj = ( (gamma[:, :, 0].exp() * pi[:, None].log()).sum() + (xi.exp() * logA[:, :, None, None] * self.om[None, None, :, 1:]).sum() + ( self.get_exp_data(mu, var, m_sample=m_sample, n_sample=n_sample) * gamma.exp() * self.om[None, :, :] ).sum() ) if self.perso: nnoise = params["nnoise"] lj = lj + self.get_exp_Mtilde(nnoise) if self.perso_io: mnoise = params["mnoise"] lj = lj + self.get_exp_M(mnoise) if self.priorV: vnoise = params["vnoise"] mnoise = params["mnoise"] V = params["V"] lj = ( lj + self.log_ig(vnoise) + self.log_ig(mnoise) + self.get_exp_V(V, vnoise) ) if self.priorMu: munoise = params["munoise"] nnoise = params["nnoise"] lj = ( lj + self.log_ig(munoise) + self.log_ig(nnoise) + self.get_exp_V(mu, munoise) ) return lj def entropy(self, e_out): """ Function to calculate the entropy; internal method :return: """ gamma = e_out["gamma"] xi = e_out["xi"] gamma_sum = gamma[:, :, 0].exp() * gamma[:, :, 0] * self.om[None, :, 0] if self.ut: xi_sum = 0 for i in range(self.n): for j in range(1, self.t): xi_sum = ( xi_sum + ( torch.triu(xi[:, :, i, j - 1]).exp() * torch.triu(xi[:, :, i, j - 1]) * self.om[None, None, i, j] ).sum() ) xi_sum = ( xi_sum - (xi.exp() * gamma[:, None, :, :-1] * self.om[None, None, :, 1:]).sum() ) else: xi_sum = ( xi.exp() * (xi - gamma[:, None, :, :-1]) * self.om[None, None, :, 1:] ).sum() et = -gamma_sum.sum() - xi_sum if self.perso_io: logdet = 2 * self.batch_diag(self.L_hat).abs().log().sum(-1) diffe = ( 0.5 * (logdet + self.d * np.log(2 * torch.tensor(math.pi)) + self.d).sum() ) et = et + diffe if self.perso: logdet = 2 * self.batch_diag(self.N_hat).abs().log().sum(-1) diffe = ( 0.5 * (logdet + self.d * np.log(2 * torch.tensor(math.pi)) + self.d).sum() ) et = et + diffe return et def variational_obj(self, params, e_out, samples): """ Function to calculate the elbo using the expectation of the joint likelihood and the entropy; internal method :return: """ obj1 = -self.exp_log_joint(params, e_out, samples) obj2 = -self.entropy(e_out) self.elbo.append((obj1 + obj2).item()) return obj1 + obj2 def baseline_variational_obj(self, params, e_out, samples): """ Function to calculate the elbo when only one time point has been observed; internal method """ # unpack parameters pi = params["pi"] A = params["A"] logA = A.log() logA[torch.isinf(logA)] = 0 mu = params["mu"] var = params["var"] m_sample = samples["m_sample"] n_sample = samples["n_sample"] gamma = e_out["gamma"] if self.io: V = params["V"] lj = (gamma[:, :, 0].exp() * pi[:, None].log()).sum() + ( self.get_exp_data(mu, var, V=V, m_sample=m_sample, n_sample=n_sample) * gamma.exp() * self.om[None, :, :] ).sum() else: lj = (gamma[:, :, 0].exp() * pi[:, None].log()).sum() + ( self.get_exp_data(mu, var, m_sample=m_sample, n_sample=n_sample) * gamma.exp() * self.om[None, :, :] ).sum() if self.perso: nnoise = params["nnoise"] lj = lj + self.get_exp_Mtilde(nnoise) if self.perso_io: mnoise = params["mnoise"] lj = lj + self.get_exp_M(mnoise) if self.priorV: vnoise = params["vnoise"] mnoise = params["mnoise"] V = params["V"] lj = ( lj + self.log_ig(vnoise) + self.log_ig(mnoise) + self.get_exp_V(V, vnoise) ) if self.priorMu: munoise = params["munoise"] nnoise = params["nnoise"] lj = ( lj + self.log_ig(munoise) + self.log_ig(nnoise) + self.get_exp_V(mu, munoise) ) gamma_sum = gamma[:, :, 0].exp() * gamma[:, :, 0] * self.om[None, :, 0] et = -gamma_sum.sum() if self.perso_io: logdet = 2 * self.batch_diag(self.L_hat).abs().log().sum(-1) diffe = ( 0.5 * (logdet + self.d * np.log(2 * torch.tensor(math.pi)) + self.d).sum() ) et = et + diffe if self.perso: logdet = 2 * self.batch_diag(self.N_hat).abs().log().sum(-1) diffe = ( 0.5 * (logdet + self.d * np.log(2 * torch.tensor(math.pi)) + self.d).sum() ) et = et + diffe self.elbo.append((-lj + -et).item()) return -lj - et def forward(self, likelihood, params): """ Calculate the forward pass of the EM algorithm for the HMM (Baum-Welch); internal method :param likelihood: log-likelihood of the data for the current parameters :param params: current model parameters :return: k x n x t log alpha's and the n x t scaling factors Note this implementation uses the rescaled 'alpha-hats' """ # unpack params pi = params["pi"] A = params["A"] logA = A.log() alpha = torch.zeros(self.k, self.n, self.t, device=self.device) scaling_factor = torch.zeros(self.n, self.t, device=self.device) a = pi[:, None].log() + likelihood[:, :, 0] scaling_factor[:, 0] = torch.logsumexp(a, dim=0) alpha[:, :, 0] = a - scaling_factor[:, 0] for i in range(1, self.t): asample = alpha[ :, :, i - 1 ] # this is the previous time point alpha, we need this for the recursion # we'll use the log-sum-exp trick for stable calculation a = likelihood[:, :, i] + torch.logsumexp( asample[:, None, :] + logA[:, :, None], dim=0 ) scaling_factor[:, i] = torch.logsumexp(a, dim=0) alpha[:, :, i] = a - scaling_factor[:, i] # multiply the final results with the time mask to reset missing values to zero alpha = alpha * self.tm[None, :, :] scaling_factor = scaling_factor * self.tm return ( alpha, scaling_factor, ) # note that this is log alpha and log scaling factor def backward(self, likelihood, params, scaling_factor): """ Calaculate the backward pass of the EM algorithm for the HMM (Baum-Welch); internal method :param likelihood: log-likelihood of the data for the current parameters :param params: current model parameters :param scaling_factor: scaling factors calculated during the forward pass; required for numerical stability :return: k x n x t log beta's Note this implementation uses the rescaled 'beta-hats' """ # unpack params logA = params["A"].log() beta = torch.zeros(self.k, self.n, self.t, device=self.device) for i in range(self.t - 2, -1, -1): bsample = beta[ :, :, i + 1 ] # this is the next time point beta, we need this for the recusion # we'll use the log-sum-exp trick for stable calculation b = torch.logsumexp( bsample[None, :, :] + logA[:, :, None] + likelihood[None, :, :, i + 1], dim=1, ) tmi = self.tm[:, i + 1] beta[:, :, i] = (b - scaling_factor[:, i + 1]) * tmi[None, :] return beta # note that this is log beta def e_step(self, params, fixSample=False): """ 'expectation step' for the EM algorithm (Baum-Welch); internal method :return: updates gamma, xi and the log-likelihood """ # default setting is to assume no personalized effects m_sample = None n_sample = None if self.perso_io: # sample from the variational approximation of M_i if fixSample: m_sample = self.mu_hat else: e_sample = torch.randn(self.n, self.d, device=self.device) m_sample = ( torch.einsum("ijk,ik->ij", [self.L_hat, e_sample]) + self.mu_hat ) if self.perso: if fixSample: n_sample = self.nu_hat else: e_sample = torch.randn(self.n, self.d, device=self.device) n_sample = ( torch.einsum("ijk,ik->ij", [self.N_hat, e_sample]) + self.nu_hat ) likelihood = self.get_likelihoods(params, m_sample=m_sample, n_sample=n_sample) alpha, scaling_factor = self.forward(likelihood, params) # NB: the exponentiated alpha sum over the first dimension should be one beta = self.backward(likelihood, params, scaling_factor) # the expontiated beta sum over the first dimension should be numerically well-behaved gamma = alpha + beta # note this is log gamma logA = params["A"].log() xi = ( alpha[:, None, :, :-1] + beta[None, :, :, 1:] + likelihood[None, :, :, 1:] + logA[:, :, None, None] - scaling_factor[None, None, :, 1:] ) # note this is log xi pX = scaling_factor.sum() e_out = {"xi": xi, "gamma": gamma, "pX": pX} samples = {"m_sample": m_sample, "n_sample": n_sample} if self.perso_io or self.perso: self.optimizer.zero_grad() self.variational_obj(params, e_out, samples).backward(retain_graph=True) self.optimizer.step() if self.perso_io: # update the variational parameters mu_hat and L_hat using gradient descent if self.VI_diag: self.L_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril_vec else: self.L_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril_vec if self.perso: if self.VI_diag: self.N_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril else: self.N_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril return e_out, params, samples def m_step(self, e_out, params, samples): """ fixed point equation for updating 'theta' :return: updates mu_k, sigma_k, V, A, pi """ with torch.no_grad(): # un-pack parameters gamma = e_out["gamma"] xi = e_out["xi"] var = params["var"] if self.io: V = params["V"] if self.priorV: vnoise = params["vnoise"] if self.perso_io: m_sample = samples["m_sample"] if self.perso: n_sample = samples["n_sample"] if self.priorMu: munoise = params["munoise"] # compute `N_k` the proxy "number of points" assigned to each distribution. # gamma is k x n x t N_k1 = ((gamma[:, :, 0].exp()) * self.om[None, :, 0]).sum(1) N_k = ((gamma.exp()) * self.om[None, :, :]).sum(-1).sum(-1) # get the means by taking the weighted combination of points r = self.data if self.io: if self.state_io: r = r - V[:, None, None, :] * self.ins[None, :, :, None] else: r = r - V * self.ins[:, :, None] if self.perso_io: r = r - m_sample[None, :, None, :] * self.ins[None, :, :, None] if self.perso: r = r - n_sample[None, :, None, :] else: if self.perso: r = r - n_sample[:, None, :] if self.priorMu: if self.full_cov: if self.state_io: num = torch.einsum( "ijk,ijkl->il", [ gamma.exp() * self.om[None, :, :], r * self.om[:, :, None], ], ) else: num = torch.einsum( "ijk,jkl->il", [ gamma.exp() * self.om[None, :, :], r * self.om[:, :, None], ], ) denom = ( torch.einsum( "i, ijk->ijk", [ torch.sum(gamma.exp() * self.om[None, :, :], (1, 2)), torch.stack( [ torch.eye(self.d, device=self.device) for _ in range(self.k) ] ), ], ) + var / munoise ) MU_LU = torch.lu(denom) mu = torch.lu_solve(num, *MU_LU) else: if self.state_io: mu = torch.einsum( "ijk,ijkl->il", [ gamma.exp() * self.om[None, :, :], r * self.om[:, :, None], ], ) / (torch.sum(gamma.exp(), (1, 2)) + var / munoise) else: mu = torch.einsum( "ijk,jkl->il", [ gamma.exp() * self.om[None, :, :], r * self.om[:, :, None], ], ) / (torch.sum(gamma.exp(), (1, 2))[:, None] + var / munoise) else: if self.state_io: mu = torch.einsum( "ijk,ijkl->il", [gamma.exp() * self.om[None, :, :], r * self.om[:, :, None]], ) mu = mu / (N_k[:, None] + self.eps) else: mu = torch.einsum( "ijk,jkl->il", [gamma.exp() * self.om[None, :, :], r * self.om[:, :, None]], ) mu = mu / (N_k[:, None] + self.eps) # update the matrix which tansforms the drug information if self.io: r = self.data - mu[:, None, None, :] if self.perso_io: r = r - m_sample[None, :, None, :] * self.ins[None, :, :, None] if self.perso: r = r - n_sample[None, :, None, :] if self.priorV: num = torch.einsum( "ijk,ijkl->il", [ gamma.exp() * self.om[None, :, :], r * self.ins[None, :, :, None] * self.om[None, :, :, None], ], ) denom = ( torch.einsum( "i,ijk->ijk", [ torch.sum( gamma.exp() * self.ins[None, :, :] ** 2 * self.om[None, :, :], (1, 2), ), torch.stack( [ torch.eye(self.d, device=self.device) for _ in range(self.k) ] ), ], ) + var / vnoise ) V_LU = torch.lu(denom) V = torch.lu_solve(num[:, :, None], *V_LU).squeeze() else: if self.state_io: V = torch.einsum( "ijk,ijkl->il", [ gamma.exp() * self.om[None, :, :], r * self.ins[None, :, :, None] * self.om[None, :, :, None], ], ) denom = torch.sum( gamma.exp() * self.ins[None, :, :] ** 2 * self.om[None, :, :], (1, 2), ) V = V / denom[:, None] else: V = torch.einsum( "ijk,ijkl->l", [ gamma.exp() * self.om[None, :, :], r * self.ins[:, :, None] * self.om[None, :, :, None], ], ) V = V / torch.sum( ((gamma.exp()) * self.tm[None, :, :]) * ((self.ins[None, :, :] * self.om[None, :, :]) ** 2) ) # compute the diagonal covar. matrix, by taking a weighted combination of # the each point's square distance from the mean r = self.data - mu[:, None, None] if self.io: if self.state_io: r = r - V[:, None, None, :] * self.ins[None, :, :, None] else: r = r - V[None, None, None, :] * self.ins[None, :, :, None] if self.perso_io: r = r - m_sample[None, :, None, :] * self.ins[None, :, :, None] if self.perso: r = r - n_sample[None, :, None, :] r = r * self.om[None, :, :, None] if self.full_cov: if self.perso_io: var = ( (gamma[:, :, :, None, None].exp()) * self.om[None, :, :, None, None] ) * (torch.einsum("ijkl,ijkm->ijklm", [r, r])) else: var = ( (gamma[:, :, :, None, None].exp()) * self.om[None, :, :, None, None] ) * torch.einsum("ijkl,ijkm->ijklm", [r, r]) var = var.sum(1).sum(1) / (N_k[:, None, None] + self.eps) # add variance ridge to prevent non psd covariance matrices var = torch.stack( [ var[i, :, :] + 1e-4 * torch.eye(self.d).to(self.device) for i in range(self.k) ] ) else: var = torch.einsum( "ijk,ijkl->il", [(gamma.exp()) * self.om[None, :, :], r ** 2] ) var = var / (N_k[:, None] + self.eps) var = torch.clamp(var, min=self.min_var) if self.perso_io: # compute the prior mnoise if self.priorV: mnoise = ( 1 / (2 * self.alpha + 2 + self.n * self.d) * ( 2 * self.beta + (torch.einsum("ij,ij->i", [m_sample, m_sample])).sum() ) ) else: mnoise = ( 1 / self.n / self.d * (torch.einsum("ij,ij->i", [m_sample, m_sample])).sum() ) if self.priorV: vnoise = ( 1 / (2 * self.alpha + 2 + self.d * self.k) * (2 * self.beta + (torch.einsum("ij, ij ->i", [V, V])).sum()) ) # CHECK DERIVATION HERE if self.perso: if self.priorMu: nnoise = ( 1 / (2 * self.alpha + 2 + self.n * self.d) * ( 2 * self.beta + (torch.einsum("ij,ij->i", [n_sample, n_sample])).sum() ) ) else: nnoise = ( 1 / self.n / self.d * (torch.einsum("ij,ij->i", [n_sample, n_sample])).sum() ) if self.priorMu: munoise = ( 1 / (2 * self.alpha + 2 + self.d * self.k) * (2 * self.beta + (torch.einsum("ij, ij-> i", [mu, mu])).sum()) ) # recompute the mixing probabilities pi = N_k1 / N_k1.sum() + self.eps # recompute the transition matrix # # xi is k x k x n x t - 1 logA = torch.logsumexp( xi.masked_fill(self.om_is == False, -1e18).reshape( self.k, self.k, self.n * (self.t - 1) ), dim=-1, ) A = normalize_exp(logA, axis=1) + self.eps * self.new_triangular params = { "mu": mu.to(self.device), "var": var.to(self.device), "pi": pi.to(self.device), "A": A.to(self.device), } if self.io: params["V"] = V.to(self.device) if self.perso_io: params["mnoise"] = mnoise.to(self.device) if self.perso: params["nnoise"] = nnoise.to(self.device) if self.priorV: params["vnoise"] = vnoise.to(self.device) if self.priorMu: params["munoise"] = munoise.to(self.device) return params def learn_model( self, num_iter=1000, use_cc=False, cc=1e-6, intermediate_save=True, load_model=False, model_name=None, ): """ function to learn the parameters of the PIOHMM :param num_iter: number of steps for the learning procedure :param use_cc: flag to indicate if a convergence criteria should be used :param cc: tolerance for the convergence criteria :param intermediate_save: flag to indicate if parameters should be saved during training :param load_model: flag to indicate if parameters should be loaded from a saved model :param model_name: file name of model to be loaded :return: depends on the type of model and will include parameters, variational parameters, liklihood, elbo """ if load_model: load_params = torch.load(model_name) A = load_params["A"] mu = load_params["mu"] var = load_params["var"] pi = load_params["pi"] V = load_params["V"] params = {"mu": mu, "var": var, "pi": pi, "A": A, "V": V} # variational parameters if self.perso_io: # transformation matrix prior noise # initialize using the mean of the IG distribution mnoise = torch.tensor([1.5], device=self.device) params["mnoise"] = mnoise if self.perso: nnoise = torch.tensor([0.5], device=self.device) params["nnoise"] = nnoise if self.priorMu: munoise = torch.tensor([1.0], device=self.device) params["munoise"] = munoise else: params = self.initialize_model() prev_cost = float("inf") for _ in range(num_iter): if _ % 500 == 0: print("Iteration ", _, flush=True) if intermediate_save: if _ % 500 == 0: print("Iteration ", _) if self.device[:4] == "cuda": print(torch.cuda.get_device_name(0)) print("Memory Usage:") print( "Allocated:", round(torch.cuda.memory_allocated(0) / 1024 ** 3, 1), "GB", ) print( "Cached: ", round(torch.cuda.memory_cached(0) / 1024 ** 3, 1), "GB", ) torch.save( { "params": params, "elbo": self.elbo, "entropy": self.ent, "exp_ll": self.ell, "log_prob": self.ll, "mi": self.mu_hat, "Li": self.L_hat, }, "../results/PD_HMM_Model_iter" + str(_) + "_k" + str(self.k) + ".pkl", ) # e-step, calculate the 'responsibilities' e_out, params, samples = self.e_step(params) # compute the cost and check for convergence obj = e_out["pX"].item() self.ll.append(obj) if use_cc: diff = prev_cost - obj if np.abs(diff) < cc: print("Breaking ", _) break prev_cost = obj # m-step, update the parameters params = self.m_step(e_out, params, samples) params["A"] = params["A"] * self.new_triangular if self.perso_io and self.perso: return ( params, e_out, self.ll, self.elbo, self.mu_hat, self.L_hat, self.nu_hat, self.N_hat, ) elif self.perso: return params, e_out, self.ll, self.elbo, self.nu_hat, self.N_hat elif self.perso_io: return params, e_out, self.ll, self.elbo, self.mu_hat, self.L_hat else: return params, e_out, self.ll def est_test_pX(self, params): """ calaculate the marginal probability of the observed data """ likelihood = self.get_likelihoods(params, fixSample=True) alpha, scaling_factor = self.forward(likelihood, params) return scaling_factor.sum() def learn_baseline_vi_params(self, params, num_iter=1000, intermediate_save=False): """ function to learn personalized parameters using only baseline data and fixed estimates of 'theta' :param params: fixed estimates of 'theta' :param num_iter: number of steps for the learning procedure :param intermediate_save: flag to indicate if parameters should be saved during training :return inferred parameters """ for _ in range(num_iter): # default setting is to assume no personalized effects m_sample = None n_sample = None if self.perso_io: e_sample = torch.randn(self.n, self.d, device=self.device) m_sample = ( torch.einsum("ijk,ik->ij", [self.L_hat, e_sample]) + self.mu_hat ) if self.perso: e_sample = torch.randn(self.n, self.d, device=self.device) n_sample = ( torch.einsum("ijk,ik->ij", [self.N_hat, e_sample]) + self.nu_hat ) likelihood = self.get_likelihoods( params, m_sample=m_sample, n_sample=n_sample ) pi = params["pi"] gamma = pi[:, None, None].log() + likelihood e_out = {"gamma": gamma} samples = {"m_sample": m_sample, "n_sample": n_sample} if self.perso_io or self.perso: self.optimizer.zero_grad() self.baseline_variational_obj(params, e_out, samples).backward( retain_graph=True ) self.optimizer.step() if self.perso_io: # update the variational parameters mu_hat and L_hat using gradient descent if self.VI_diag: self.L_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril_vec else: self.L_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril_vec if self.perso: if self.VI_diag: self.N_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril else: self.N_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril if self.perso_io and self.perso: return ( params, e_out, self.ll, self.elbo, self.mu_hat, self.L_hat, self.nu_hat, self.N_hat, ) elif self.perso: return params, e_out, self.ll, self.elbo, self.nu_hat, self.N_hat elif self.perso_io: return params, e_out, self.ll, self.elbo, self.mu_hat, self.L_hat else: return params, e_out, self.ll def learn_vi_params(self, params, num_iter=1000, intermediate_save=False): """ function to learn personalized parameters using fixed estimates of 'theta' :param params: fixed estimates of 'theta' :param num_iter: number of steps for the learning procedure :param intermediate_save: flag to indicate if parameters should be saved during training :return inferred parameters """ for _ in range(num_iter): if _ % 50 == 0: print("Iteration ", _, flush=True) if intermediate_save: if _ % 500 == 0: print("Iteration ", _) if self.device[:4] == "cuda": print(torch.cuda.get_device_name(0)) print("Memory Usage:") print( "Allocated:", round(torch.cuda.memory_allocated(0) / 1024 ** 3, 1), "GB", ) print( "Cached: ", round(torch.cuda.memory_cached(0) / 1024 ** 3, 1), "GB", ) torch.save( { "params": params, "elbo": self.elbo, "entropy": self.ent, "exp_ll": self.ell, "log_prob": self.ll, "mi": self.mu_hat, "Li": self.L_hat, }, "../results/PD_HMM_Model_iter" + str(_) + "_k" + str(self.k) + ".pkl", ) # e-step, calculate the 'responsibilities' e_out, params, samples = self.e_step(params) if self.perso_io and self.perso: return ( params, e_out, self.ll, self.elbo, self.mu_hat, self.L_hat, self.nu_hat, self.N_hat, ) elif self.perso: return params, e_out, self.ll, self.elbo, self.nu_hat, self.N_hat elif self.perso_io: return params, e_out, self.ll, self.elbo, self.mu_hat, self.L_hat else: return params, e_out, self.ll def calc_pX( self, params, num_samples=1, importance_sampling=False, mu_hat=None, nu_hat=None, L_hat=None, N_hat=None, fixSample=False, ): """ Function to calculate the test log likelihood :param params: fixed estimate for 'theta' :param num_samples: number of samples for importance sampling :param importance_sampling: flag to indicate if importance sampling should be use :param mu_hat: mean of the variational distribution for personalized medication effects :param nu_hat: mean of the variational distrbution for personalized state effects :param L_hat: cholesky factor for covariance matrix of the variational distribution for personalized medication effects :param N_hat: cholesky factor for covariance matric of the variational distribution for personalized state effects :param fixSample: flag to indicate if only the variational mean should be used :return: test log likelihood """ if importance_sampling: px = torch.zeros(num_samples, self.n) for i in range(num_samples): if self.perso_io: e_sample = torch.randn(self.n, self.d, device=self.device) m_sample = torch.einsum("ijk,ik->ij", [L_hat, e_sample]) + mu_hat L_prior = torch.stack( [ ( params["mnoise"].sqrt() * torch.eye(self.d).to(self.device) ) for _ in range(self.n) ] ).to(self.device) sample_weight_m = self.log_gaussian_prior( m_sample, torch.zeros(m_sample.shape).to(self.device), L_prior ) - self.log_gaussian_prior(m_sample, mu_hat, L_hat) else: m_sample = None sample_weight_m = 0 if self.perso: e_sample = torch.randn(self.n, self.d, device=self.device) n_sample = torch.einsum("ijk,ik->ij", [N_hat, e_sample]) + nu_hat N_prior = torch.stack( [ (params["nnoise"].sqrt() * torch.eye(self.d)).to( self.device ) for _ in range(self.n) ] ).to(self.device) sample_weight_n = self.log_gaussian_prior( n_sample, torch.zeros(n_sample.shape).to(self.device), N_prior ) - self.log_gaussian_prior(n_sample, nu_hat, N_hat) else: n_sample = None sample_weight_n = 0 likelihood = self.get_likelihoods( params, m_sample=m_sample, n_sample=n_sample ) alpha, scaling_factor = self.forward(likelihood, params) # print((scaling_factor*sample_weight[:, None]).sum()) px[i, :] = scaling_factor.sum(-1) + sample_weight_m + sample_weight_n out = torch.logsumexp(px, 0) - np.log(num_samples) elif fixSample: likelihood = self.get_likelihoods(params, m_sample=mu_hat, n_sample=nu_hat) alpha, scaling_factor = self.forward(likelihood, params) out = scaling_factor else: px = torch.zeros(num_samples, self.n) for i in range(num_samples): if self.perso_io: mnoise = params["mnoise"] # sample from the sampled from the MLE params m_sample = mnoise.sqrt() * torch.randn( self.n, self.d, device=self.device ) else: m_sample = None if self.perso: nnoise = params["nnoise"] n_sample = nnoise.sqrt() * torch.randn( self.n, self.d, device=self.device ) else: n_sample = None likelihood = self.get_likelihoods( params, m_sample=m_sample, n_sample=n_sample ) alpha, scaling_factor = self.forward(likelihood, params) px[i, :] = scaling_factor.sum(-1) out = torch.logsumexp(px, 0) - np.log(num_samples) return out def predict_sequence(self, params, m_sample=None, n_sample=None): """ function to apply viterbi algorithm :param params: fixed estimates of the model parameters :param m_sample: value to use for the personalized medication effects :param n_sample: value to use for the personalized state effects :return mps: most probable sequence n x t """ likelihood = self.get_likelihoods(params, m_sample=m_sample, n_sample=n_sample) mps = self.viterbi(likelihood, params) return mps def viterbi(self, likelihood, params): """ apply the viterbi algorithm to find the most probable sequence per patient; internal method omega is the maximimum joint probability of the previous data and latent states ; the last value is the joint distribution of the most probable path :return: """ omega = torch.zeros(self.k, self.n, self.t).to(self.device) psi = torch.zeros(self.k, self.n, self.t).to(self.device) mps = torch.zeros(self.n, self.t).to(self.device) logA = params["A"].log() pi = params["pi"] omega[:, :, 0] = pi[:, None].log() + likelihood[:, :, 0] for i in range(1, self.t): inner_max, psi[:, :, i] = torch.max( logA[:, :, None] + omega[:, None, :, i - 1], dim=0 ) omega[:, :, i] = likelihood[:, :, i] + inner_max mps[:, -1] = torch.argmax(omega[:, :, -1], dim=0) val, _ = torch.max(omega[:, :, -1], dim=0) for i in range(self.t - 2, -1, -1): psi_sample = psi[:, :, i + 1] mps[:, i] = torch.gather(psi_sample, 0, mps[:, i + 1].long().unsqueeze(0)) return mps def change_data( self, data, ins=None, OM=None, TM=None, reset_VI=True, params=[], lr=0.001 ): """ Replace model dataset :param data: new dataset to use; additional parameters are corresponding dataset features / descriptors :return: none, updates to model params only """ self.data = data.to(self.device) self.n = data.shape[0] self.t = data.shape[1] # store the inputs used in analysis if self.io: self.ins = ins.to(self.device) # n x t x 1 # store the time mask if TM is None: self.tm = torch.ones( self.n, self.t, requires_grad=False, device=self.device ) else: self.tm = TM.to(self.device) # n x t # store the observation mask if OM is None: self.om = torch.ones( self.n, self.t, requires_grad=False, device=self.device ) else: self.om = OM.to(self.device) # n x t if reset_VI: if self.perso_io and self.perso: self.elbo = [] mnoise = params["mnoise"] mi_numpy = np.sqrt(mnoise.cpu().numpy()) * np.random.randn( self.n, self.d ) if self.device[:4] == "cuda": self.mu_hat = ( torch.from_numpy(mi_numpy) .float() .cuda() .to(self.device) .requires_grad_() ) else: self.mu_hat = torch.from_numpy(mi_numpy).float().requires_grad_() if self.VI_diag: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * self.d), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril_vec else: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril_vec nnoise = params["nnoise"] ni_numpy = np.sqrt(nnoise.cpu().numpy()) * np.random.randn( self.n, self.d ) if self.device[:4] == "cuda": self.nu_hat = ( torch.from_numpy(ni_numpy) .float() .cuda() .to(self.device) .requires_grad_() ) else: self.nu_hat = torch.from_numpy(ni_numpy).float().requires_grad_() if self.VI_diag: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * self.d), device=self.device, requires_grad=True, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril else: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril self.optimizer = torch.optim.Adam( [self.mu_hat, self.tril_vec, self.nu_hat, self.tril], lr=lr ) elif self.perso_io: self.elbo = [] mnoise = params["mnoise"] mi_numpy = np.sqrt(mnoise.cpu().numpy()) * np.random.randn( self.n, self.d ) if self.device[:4] == "cuda": self.mu_hat = ( torch.from_numpy(mi_numpy) .float() .cuda() .to(self.device) .requires_grad_() ) else: self.mu_hat = torch.from_numpy(mi_numpy).float().requires_grad_() if self.VI_diag: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * self.d), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril_vec else: self.tril_vec = torch.tensor( 0.01 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.L_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.L_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril_vec self.optimizer = torch.optim.Adam([self.mu_hat, self.tril_vec], lr=lr) elif self.perso: self.elbo = [] nnoise = params["nnoise"] ni_numpy = np.sqrt(nnoise.cpu().numpy()) * np.random.randn( self.n, self.d ) if self.device[:4] == "cuda": self.nu_hat = ( torch.from_numpy(ni_numpy) .float() .cuda() .to(self.device) .requires_grad_() ) else: self.nu_hat = torch.from_numpy(ni_numpy).float().requires_grad_() if self.VI_diag: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * self.d), device=self.device, requires_grad=True, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.stack([torch.eye(self.d) for _ in range(self.n)]) == 1 ] = self.tril else: self.tril = torch.tensor( 0.01 * np.random.randn(self.n * int(0.5 * self.d * (self.d + 1))), requires_grad=True, device=self.device, dtype=torch.float64, ) self.N_hat = torch.zeros(self.n, self.d, self.d, device=self.device) self.N_hat[ torch.tril(torch.ones(self.n, self.d, self.d)) == 1 ] = self.tril self.optimizer = torch.optim.Adam( [self.nu_hat, self.tril], lr=lr, eps=1e-4 ) def forward_pred(self, params, m_sample=None, n_sample=None): """ function to forecast one-step-ahead :return: osapd: one-step-ahead predictive density bs: belief state lpe: log probability evidence """ pi = params["pi"] A = params["A"] likelihood = self.get_likelihoods(params, m_sample=m_sample, n_sample=n_sample) alpha, scaling_factor = self.forward(likelihood, params) osapd = torch.zeros(self.k, self.n, self.t + 1).to( self.device ) # one-step-ahead predictive density # there is no data yet at t=0, use pi osapd[:, :, 0] = pi[:, None].log() osapd[:, :, 1:] = torch.logsumexp( A[:, :, None, None].log() + alpha[:, None, :, :], dim=0 ) bs = likelihood + osapd # belief state lpe = torch.logsumexp(bs, dim=0) # log probability evidence bs = (bs - lpe[None, :, :]).exp() osapd = osapd.exp() # return values in probability space return osapd, bs, lpe def forward_sample(self, prob, ns=100): """ prob: k x n x t 'one-step-ahead predictive density' p(z_it=j | x_i1, ... x_it-1) return: samples from the one-step-ahed predictive density """ vals = torch.zeros(ns, self.n, self.t - 1, self.d) for i in range(self.t - 1): for j in range(self.n): m = torch.distributions.categorical.Categorical(prob[:, j, i]) for k in range(ns): draw = m.sample() mvn = torch.distributions.multivariate_normal.MultivariateNormal( self.mu[draw, :] + self.V[draw, :] * self.ins[j, i + 1] + self.mu_hat[j, :] * self.ins[j, i + 1], covariance_matrix=self.var[draw, :, :] + torch.mm(self.L_hat[j, :, :], self.L_hat[j, :, :].t()) * self.ins[j, i + 1], ) vals[k, j, i, :] = mvn.sample() return vals def load_model(self, filename, cpu=True): """ function to specifically add the variational model parameters because they are properties of the model; note that variable names have been assumed here and the save file needs to be formatted accordingly :param filename: name of filename containing model :return: none """ if cpu: trained_model = torch.load(filename, map_location=torch.device("cpu")) else: trained_model = torch.load(filename) ### IMPORTANT #### # Note that this is currently not setup to continue training. tril_vec needs to be populated and have # requires_grad = True to be able to continue training; this function is only to load in a model. Additional # functionality is required to continue training if self.perso_io: self.mu_hat = trained_model["Mi"].to(self.device) self.L_hat = trained_model["Li"].to(self.device) if self.perso: self.nu_hat = trained_model["ni"].to(self.device) self.N_hat = trained_model["Ni"].to(self.device) def baseline_risk(self, params, ns=500, type="sample", m_sample=None): """ Determine probabilities of state assignment at 1- and 2-years from baseline; sample the observed data using those probabilities :return: """ # unpack params pi = params["pi"] A = params["A"] mu = params["mu"] var = params["var"] if self.io: V = params["V"] mnoise = params["mnoise"] if type == "sample": sample_1year = torch.zeros(ns, self.n, self.d) sample_2year = torch.zeros(ns, self.n, self.d) for k in range(ns): if self.perso_io: if m_sample is None: m_sample = mnoise.sqrt() * torch.randn(size=(self.n, self.d)) likelihood = self.get_likelihoods( params, log=False, m_sample=m_sample ) # k x n x t # NB: one of the pi elements is zero so we don't work directly in the log space p_z1 = ( pi[:, None] * likelihood[:, :, 0].squeeze() / (pi[:, None] * likelihood[:, :, 0].squeeze()).sum(0) ) p_z6month = ( A[:, :, None] * (A[:, :, None] * p_z1[:, None, :]).sum(0) ).sum(0) p_z1year = p_z1 for i in range(4): p_z1year = (A[:, :, None] * p_z1year[:, None, :]).sum(0) p_z2year = p_z1year for i in range(4): p_z2year = (A[:, :, None] * p_z2year[:, None, :]).sum(0) for j in range(self.n): # check for underflow issues if np.isnan(p_z1year[:, j].detach().numpy()).all(): p_z1year[:, j] = torch.zeros(self.k) p_z1year[-1, j] = 1 p_z2year[:, j] = p_z1year[:, j] m1 = torch.distributions.categorical.Categorical(p_z1year[:, j]) m2 = torch.distributions.categorical.Categorical(p_z2year[:, j]) draw = m1.sample() if self.io: mvn = torch.distributions.multivariate_normal.MultivariateNormal( mu[draw, :] + (V[draw, :] + m_sample[j, :]) * self.ins[j, 4], covariance_matrix=var[draw, :, :], ) else: mvn = torch.distributions.multivariate_normal.MultivariateNormal( mu[draw, :], covariance_matrix=var[draw, :, :] ) sample_1year[k, j, :] = mvn.sample() draw = m2.sample() if self.io: mvn = torch.distributions.multivariate_normal.MultivariateNormal( mu[draw, :] + (V[draw, :] + m_sample[j, :]) * self.ins[j, 8], covariance_matrix=var[draw, :, :], ) else: mvn = torch.distributions.multivariate_normal.MultivariateNormal( mu[draw, :], covariance_matrix=var[draw, :, :] ) sample_2year[k, j, :] = mvn.sample() elif type == "mean": sample_1year = torch.zeros(self.n, self.d) sample_2year = torch.zeros(self.n, self.d) if self.perso_io: m_sample = mnoise.sqrt() * torch.randn(size=(self.n, self.d)) likelihood = self.get_likelihoods( params, log=False, m_sample=m_sample ) # k x n x t # NB: one of the pi elements is zero so we don't work directly in the log space p_z1 = ( pi[:, None] * likelihood[:, :, 0].squeeze() / (pi[:, None] * likelihood[:, :, 0].squeeze()).sum(0) ) # print('Check sum:', p_z1.sum(0)) p_z6month = (A[:, :, None] * (A[:, :, None] * p_z1[:, None, :]).sum(0)).sum( 0 ) p_z1year = p_z1 for i in range(4): p_z1year = (A[:, :, None] * p_z1year[:, None, :]).sum(0) p_z2year = p_z1year for i in range(4): p_z2year = (A[:, :, None] * p_z2year[:, None, :]).sum(0) meds1 = self.ins[:, 4] meds2 = self.ins[:, 8] sample_1year = ( torch.einsum("ij, ik->jk", [p_z1year, mu]) + torch.einsum("ij, ik->jk", [p_z1year, V]) * meds1[:, None] ) sample_2year = ( torch.einsum("ij, ik->jk", [p_z2year, mu]) + torch.einsum("ij, ik->jk", [p_z2year, V]) * meds2[:, None] ) else: sample_1year = torch.zeros(self.n, self.d) sample_2year = torch.zeros(self.n, self.d) if self.perso_io: m_sample = mnoise.sqrt() * torch.randn(size=(self.n, self.d)) likelihood = self.get_likelihoods( params, log=False, m_sample=m_sample ) # k x n x t # NB: one of the pi elements is zero so we don't work directly in the log space p_z1 = ( pi[:, None] * likelihood[:, :, 0].squeeze() / (pi[:, None] * likelihood[:, :, 0].squeeze()).sum(0) ) p_z6month = (A[:, :, None] * (A[:, :, None] * p_z1[:, None, :]).sum(0)).sum( 0 ) p_z1year = p_z1 for i in range(4): p_z1year = (A[:, :, None] * p_z1year[:, None, :]).sum(0) p_z2year = p_z1year for i in range(4): p_z2year = (A[:, :, None] * p_z2year[:, None, :]).sum(0) idx1 = torch.argmax(p_z1year, dim=0) idx2 = torch.argmax(p_z2year, dim=0) for j in range(self.n): # check for underflow issues if np.isnan(p_z1year[:, j]).all(): p_z1year[:, j] = torch.zeros(self.k) p_z1year[-1, j] = 1 p_z2year[:, j] = p_z1year[:, j] sample_1year[j, :] = mu[idx1[j]] + V[idx1[j]] * self.ins[j, 4] sample_2year[j, :] = mu[idx2[j]] + V[idx2[j]] * self.ins[j, 8] return p_z1year, p_z2year, sample_1year, sample_2year, p_z1, p_z6month if __name__ == "__main__": # Note that this is also contained in the jupyter notebook for more integrated visuals n = 200 # number of samples d = 1 # dimensionality of observations t = 30 # number of time steps k = 2 # number of states K = 2 # number of HMM mixtures A_1 = torch.tensor([[0.8, 0.2], [0.2, 0.8]]) # transition matrix A_2 = torch.tensor([[0.2, 0.8], [0.8, 0.2]]) # transition matrix A = torch.block_diag(*[A_1, A_2]) pi = torch.ones(k * K) / (k * K) # initial state distribution mu = torch.tensor([0.0, 2.0, 0.0, 2.0]) # state means var = torch.tensor([0.1, 0.1, 0.1, 0.1]) # state covariance b = 1 # limit of the uniform distribution to specify personalized state effects X = torch.zeros((n, t, d)) Z = torch.zeros((n, t), dtype=torch.long) for i in range(n): for j in range(t): if j == 0: Z[i, j] = torch.multinomial(pi, num_samples=1).byte() m_dist = torch.distributions.normal.Normal( mu.index_select(0, Z[i, j]), var.index_select(0, Z[i, j]) ) X[i, j, :] = m_dist.sample() else: Z[i, j] = torch.multinomial(A[Z[i, j - 1], :], num_samples=1) m_dist = torch.distributions.normal.Normal( mu.index_select(0, Z[i, j]), var.index_select(0, Z[i, j]) ) X[i, j, :] = m_dist.sample() X_hat = torch.zeros(n, t, d) l = 1.0 # lengthscale for the SE kernel s = 0.1 # sigma^2 for the SE kernel # build covariance matrix var_x = torch.zeros(t, t) t_vec = torch.range(0, t) for j in range(t): for jj in range(t): r = (t_vec[j] - t_vec[jj]) ** 2 var_x[j, jj] = 1 / s * torch.exp(-r / (2 * l)) L = torch.cholesky(var_x) b_stor = torch.zeros(n) for i in range(n): e = torch.randn(t) b_stor[i] = 2 * b * torch.rand(1) - b X_hat[i, :, :] = ( torch.einsum("ik,k->i", [L, e])[None, :, None] + X[i, :, :] + b_stor[i] * torch.ones(1, t, 1) ) # fit a personalized hmm piohmm = mHMM( X_hat, k=k, K=K, full_cov=False, priorV=False, io=False, personalized=True, personalized_io=False, state_io=False, UT=False, device="cpu", eps=1e-18, priorMu=True, var_fill=0.5, ) piohmm_params, _, _, elbo, b_hat, _ = piohmm.learn_model( num_iter=10000, intermediate_save=False ) piohmm_mps, _, _ = piohmm.predict_sequence(piohmm_params, n_sample=b_hat) piohmm_xhat = np.zeros((n, t)) piohmm_xvar = np.zeros((n, t)) for i in range(n): for j in range(t): idx = np.where(piohmm_mps[i, j].numpy() == np.arange(k))[0][0] piohmm_xhat[i, j] = ( piohmm_params["mu"][idx].numpy() + b_hat[i].detach().numpy() ) piohmm_xvar[i, j] = 2 * np.sqrt(piohmm_params["var"][idx].numpy()) torch.save( "model_results.pkl", {"params": piohmm_params, "mps": piohmm_mps, "piohmm_xhat": piohmm_xhat}, )
89,160
37.867044
136
py
mIOHMM
mIOHMM-main/experiments/synthetic.py
from src.piomhmm import mHMM from src.utils import save_pickle import matplotlib.pyplot as plt import numpy as np import torch def pred(model_name, model, params, b_hat): model_mps = model.predict_sequence(params, n_sample=b_hat) xhat = np.zeros((n, t)) xvar = np.zeros((n, t)) for i in range(n): for j in range(t): idx = np.where(model_mps[i, j].cpu().numpy() == np.arange(model.k))[0][0] if model_name in ["HMM", "mHMM"]: xhat[i, j] = params["mu"][idx].cpu().numpy() else: xhat[i, j] = ( params["mu"][idx].cpu().numpy() + b_hat[i].cpu().detach().numpy() ) xvar[i, j] = 2 * np.sqrt(params["var"][idx].cpu().numpy()) return {"xhat": xhat, "xvar": xvar} torch.manual_seed(0) torch.set_default_dtype(torch.float64) torch.set_printoptions(precision=2) device = "cpu" # DATA GENERATION n = 200 # number of samples d = 1 # dimensionality of observations t = 30 # number of time steps k = 2 # number of states K = 2 # number of HMM mixtures # set parameters A_1 = torch.tensor([[0.8, 0.2], [0.2, 0.8]]) # transition matrix A_2 = torch.tensor([[0.2, 0.8], [0.8, 0.2]]) # transition matrix A = torch.block_diag(*[A_1, A_2]) pi = torch.ones(k * K) / (k * K) # initial state distribution mu = torch.tensor([0.0, 2.0, 0.0, 2.0]) # state means var = torch.tensor([0.1, 0.1, 0.1, 0.1]) # state covariance b = 1.0 # specify the range of a uniform distribution over personalized state effects, e.g. r_i ~ Unif[-b, b] # simulate the model X = torch.zeros((n, t, d)) Z = torch.zeros((n, t), dtype=torch.long) for i in range(n): for j in range(t): if j == 0: Z[i, j] = torch.multinomial(pi, num_samples=1).byte() # D[i, j] = torch.rand(1) m_dist = torch.distributions.normal.Normal( mu.index_select(0, Z[i, j]), var.index_select(0, Z[i, j]) ) X[i, j, :] = m_dist.sample() else: Z[i, j] = torch.multinomial(A[Z[i, j - 1], :], num_samples=1) # D[i, j] = torch.rand(1) m_dist = torch.distributions.normal.Normal( mu.index_select(0, Z[i, j]), var.index_select(0, Z[i, j]) ) X[i, j, :] = m_dist.sample() # add noise X_hat = torch.zeros(n, t, d) l = 1.0 # lengthscale for the SE kernel s = 0.1 # sigma^2 for the SE kernel # build covariance matrix var_x = torch.zeros(t, t) t_vec = torch.range(0, t) for j in range(t): for jj in range(t): r = (t_vec[j] - t_vec[jj]) ** 2 var_x[j, jj] = s * torch.exp(-r / (2 * l ** 2)) L = torch.cholesky(var_x) b_stor = torch.zeros(n) for i in range(n): e = torch.randn(t) b_stor[i] = 2 * b * torch.rand(1) - b X_hat[i, :, :] = ( torch.einsum("ik,k->i", [L, e])[None, :, None] + X[i, :, :] + b_stor[i] * torch.ones(1, t, 1) ) # plot a number of samples generated fig, axs = plt.subplots(3, 5, dpi=200) fig.set_size_inches(12, 5) for i, ax in enumerate(axs.flatten()): ax.plot(X[i, :].numpy(), label="$x_i$") ax.plot(X_hat[i, :].numpy(), label="$\hat{x}_i$") ax.set_title("Sample " + str(i)) fig.tight_layout() ax.legend( loc="lower center", bbox_to_anchor=(-2.2, -0.95), fancybox=True, shadow=True, ncol=5 ) # FITTING MODELS print("fitting standard HMM...") hmm = mHMM( X_hat, k=k, K=1, full_cov=False, priorV=False, io=False, personalized=False, personalized_io=False, state_io=False, device=device, eps=1e-18, ) hmm_params, _, ll_hmm = hmm.learn_model(num_iter=10000, intermediate_save=False) print("fitting standard mHMM...") mhmm = mHMM( X_hat, k=k, K=K, full_cov=False, priorV=False, io=False, personalized=False, personalized_io=False, state_io=False, device=device, eps=1e-18, ) mhmm_params, _, ll_mhmm = mhmm.learn_model(num_iter=10000, intermediate_save=False) print("fitting personalized HMM...") phmm = mHMM( X_hat, k=k, K=1, full_cov=False, priorV=False, io=False, personalized=True, personalized_io=False, state_io=False, device=device, eps=1e-18, ) phmm_params, _, _, elbo_phmm, b_hat_phmm, _ = phmm.learn_model( num_iter=10000, intermediate_save=False ) print("fitting personalized mHMM...") mphmm = mHMM( X_hat, k=2, K=2, full_cov=False, priorV=False, io=False, personalized=True, personalized_io=False, state_io=False, device=device, eps=1e-18, ) mphmm_params, _, _, elbo_mphmm, b_hat_mphmm, _ = mphmm.learn_model( num_iter=10000, intermediate_save=False ) # RESULTS # save outputs outputs = { "HMM": {"model": hmm, "params": hmm_params, "b_hat": None}, "mHMM": {"model": mhmm, "params": mhmm_params, "b_hat": None}, "PHMM": {"model": phmm, "params": phmm_params, "b_hat": b_hat_phmm}, "mPHMM": {"model": mphmm, "params": mphmm_params, "b_hat": b_hat_mphmm}, } preds = {} for model_name in outputs: model = outputs[model_name]["model"] params = outputs[model_name]["params"] b_hat = outputs[model_name]["b_hat"] preds[model_name] = pred(model_name, model, params, b_hat) save_pickle({"outputs": outputs, "preds": preds}, "outputs/synthetic_all.pkl") # figure fig, axs = plt.subplots(3, 4, dpi=200) fig.set_size_inches(12, 8) for j in range(3): for i in range(4): ax = axs[j][i] ax.plot(X_hat[i, :].numpy(), "k:", label="$\hat{x}_i$") if j == 0: xhat = preds["HMM"]["xhat"] xvar = preds["HMM"]["xvar"] ax.plot(xhat[i, :], label="HMM $\mu_k \pm 2\sigma_{k,i}$") ax.fill_between( np.arange(t), xhat[i, :] - xvar[i, :], xhat[i, :] + xvar[i, :], alpha=0.5, ) xhat = preds["PHMM"]["xhat"] xvar = preds["PHMM"]["xvar"] ax.plot(xhat[i, :], label="PHMM $(\mu_k + r^{(i)}) \pm 2\sigma_{k,i}$") ax.fill_between( np.arange(t), xhat[i, :] - xvar[i, :], xhat[i, :] + xvar[i, :], alpha=0.5, ) elif j == 1: xhat = preds["mHMM"]["xhat"] xvar = preds["mHMM"]["xvar"] ax.plot(xhat[i, :], label="mHMM $\mu_k \pm 2\sigma_{k,i}$") ax.fill_between( np.arange(t), xhat[i, :] - xvar[i, :], xhat[i, :] + xvar[i, :], alpha=0.5, ) xhat = preds["mPHMM"]["xhat"] xvar = preds["mPHMM"]["xvar"] ax.plot(xhat[i, :], label="mPHMM $(\mu_k + r_i) \pm 2\sigma_{k,i}$") ax.fill_between( np.arange(t), xhat[i, :] - xvar[i, :], xhat[i, :] + xvar[i, :], alpha=0.5, ) else: xhat = preds["PHMM"]["xhat"] xvar = preds["PHMM"]["xvar"] ax.plot(xhat[i, :], label="PHMM $(\mu_k + r_i) \pm 2\sigma_{k,i}$") ax.fill_between( np.arange(t), xhat[i, :] - xvar[i, :], xhat[i, :] + xvar[i, :], alpha=0.5, ) xhat = preds["mPHMM"]["xhat"] xvar = preds["mPHMM"]["xvar"] ax.plot(xhat[i, :], label="mPHMM $(\mu_k + r_i) \pm 2\sigma_{k,i}$") ax.fill_between( np.arange(t), xhat[i, :] - xvar[i, :], xhat[i, :] + xvar[i, :], alpha=0.5, ) ax.set_xlabel("Time") ax.set_xticks([]) ax.set_xticklabels([]) ax.set_title("Sample " + str(i)) ax.set_ylim([-4, 4]) ax.legend( loc="lower left", bbox_to_anchor=(-2.7, -0.42), fancybox=True, shadow=True, ncol=4, ) fig.tight_layout() fig.subplots_adjust(hspace=0.7) fig.savefig( "outputs/synthetic_x_hats.png", dpi=400, facecolor="w", edgecolor="w", orientation="portrait", bbox_inches="tight", pad_inches=0, metadata={"Creator": None, "Producer": None, "CreationDate": None}, )
8,235
28
110
py
mIOHMM
mIOHMM-main/experiments/real.py
from src.piomhmm import mHMM import numpy as np import random import torch from src.utils import save_pickle, load_pickle RANDOM_SEED = 0 torch.manual_seed(RANDOM_SEED) random.seed(RANDOM_SEED) np.random.seed(RANDOM_SEED) torch.set_default_dtype(torch.float64) torch.set_printoptions(precision=2) def preprocess(x, d): # don't include samples which only have one measurement, i.e. aren't time series remove_idx = np.where(np.sum(~np.isnan(x[:, :, 0]), axis=1) == 1) x = np.delete(x, remove_idx, 0) d = np.delete(d, remove_idx, 0) # set any LEDD values greater than 5000 to 620 and rescale d[d > 5000] = 620 d[np.isnan(d)] = 0 d = d / np.max(d) # get time and observation masks N, T, D, = x.shape time_mask = np.ones((N, T)) for i in range(N): ind = np.where(~np.isnan(x[i, :, 0]))[0][-1] + 1 time_mask[i, ind:] = 0 missing_mask = (~np.isnan(x[:, :, 0])).astype(float) x[np.isnan(x)] = 0 # convert everything to tensors X = torch.Tensor(x).float() D = torch.Tensor(d).float() TM = torch.Tensor(time_mask).float() OM = torch.Tensor(missing_mask).bool() return X, D, TM, OM # data data = load_pickle("processed/data_for_PIOHMM.pkl") X_train, D_train, TM_train, OM_train = preprocess(data["x_train"], data["train_med"]) X_test, D_test, TM_test, OM_test = preprocess(data["x_test"], data["test_med"]) # experiment setting device = "cpu" k = 8 # number of hidden states num_iter_train = 10000 num_iter_test = 5000 for K in [1, 2, 3, 4, 5]: # mIOHMM print("fitting mIOHMM...", flush=True) model = mHMM( X_train, ins=D_train, k=k, K=K, TM=TM_train, OM=OM_train, full_cov=False, io=True, personalized=False, personalized_io=False, state_io=True, UT=True, device=device, eps=1e-18, ) params_hat, e_out, ll = model.learn_model( num_iter=num_iter_train, intermediate_save=False ) training_pX = model.calc_pX(params_hat) mIOHMM_model = { "params": params_hat, "e_out": e_out, "ll": ll, "training_pX": training_pX, } print("learning vi params ...\n", flush=True) model.change_data( X_test, ins=D_test, TM=TM_test, OM=OM_test, reset_VI=True, params=params_hat ) params_hat, e_out_test, ll_test = model.learn_vi_params( params_hat, num_iter=num_iter_test ) test_pX = model.calc_pX(params_hat) mIOHMM_model["params_test"] = params_hat mIOHMM_model["e_out_test"] = e_out_test mIOHMM_model["ll_test"] = ll_test mIOHMM_model["test_pX"] = test_pX save_pickle( mIOHMM_model, "models/mIOHMM_" + str(K) + ".pkl", )
2,760
26.61
85
py
MostAccurableMNIST_keras
MostAccurableMNIST_keras-master/DeepCNN.py
import numpy as np from keras.models import Sequential from keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, pooling, Input from keras.layers.convolutional import Conv2D, ZeroPadding2D from keras.layers.pooling import MaxPooling2D from keras.utils import np_utils from keras.datasets import mnist from keras.optimizers import Adam, Adagrad, RMSprop, Adadelta np.random.seed(777) # for reproducibility (X_train, y_train), (X_test, y_test) = mnist.load_data() X_train = X_train.reshape(X_train.shape[0], 28,28,1) X_test = X_test.reshape(X_test.shape[0], 28,28,1) X_train = X_train.astype('float32') X_test = X_test.astype('float32') X_train /= 255 X_test /= 255 Y_train = np_utils.to_categorical(y_train, 10) Y_test = np_utils.to_categorical(y_test, 10) cnn = Sequential() cnn.add(ZeroPadding2D((2, 2), input_shape=(28, 28, 1))) cnn.add(Conv2D(64, (5, 5), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(ZeroPadding2D((2, 2))) cnn.add(Conv2D(128, (5, 5), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(MaxPooling2D(strides=(2, 2))) cnn.add(ZeroPadding2D((2, 2))) cnn.add(Conv2D(256, (5, 5), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(ZeroPadding2D((1, 1))) cnn.add(Conv2D(256, (3, 3), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(MaxPooling2D(strides=(2, 2))) cnn.add(Dropout(0.2)) cnn.add(ZeroPadding2D((1, 1))) cnn.add(Conv2D(512, (3, 3), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(Dropout(0.2)) cnn.add(ZeroPadding2D((1, 1))) cnn.add(Conv2D(512, (3, 3), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(MaxPooling2D(strides=(2, 2))) cnn.add(ZeroPadding2D((1, 1))) cnn.add(Conv2D(1024, (3, 3), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(Dropout(0.2)) cnn.add(ZeroPadding2D((1, 1))) cnn.add(Conv2D(1024, (3, 3), kernel_initializer='he_normal')) cnn.add(Activation('relu')) cnn.add(MaxPooling2D(strides=(2, 2))) cnn.add(Flatten()) cnn.add(Dropout(0.5)) cnn.add(Dense(2048, activation="relu", kernel_initializer='he_normal')) cnn.add(Dense(128, activation="relu", kernel_initializer='he_normal')) cnn.add(Dense(10, activation="softmax")) cnn.summary() opt = Adagrad(lr=0.001, epsilon=1e-8, decay=0.) cnn.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy']) cnn.fit(X_train, Y_train, batch_size=64, shuffle=True, epochs=50, validation_split=0.1) score = cnn.evaluate(X_test, Y_test) print(cnn.metrics_names) print(score) f = open("./saved/MNIST_DeepCNN_model.json", 'w') f.write(cnn.to_json()) f.close() cnn.save_weights('./saved/MNIST_DeepCNN_weight.h5')
2,673
32.012346
87
py
EDGY
EDGY-master/DDF.py
import hydra import hydra.utils as utils import json from pathlib import Path import torch import numpy as np import librosa from tqdm import tqdm import pyloudnorm from preprocess import preemphasis from model import Encoder, Decoder @hydra.main(config_path="Training/VQ-VAE/Configuration_files/DDF.yaml") def DDF(cfg): filter_list_path = Path(utils.to_absolute_path(cfg.filter_list)) with open(filter_list_path) as file: filter_list = json.load(file) in_dir = Path(utils.to_absolute_path(cfg.in_dir)) out_dir = Path(utils.to_absolute_path(cfg.out_dir)) out_dir.mkdir(exist_ok=True, parents=True) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") encoder = Encoder(**cfg.model.encoder) decoder = Decoder(**cfg.model.decoder) encoder.to(device) decoder.to(device) print("Load checkpoint from: {}:".format(cfg.checkpoint)) checkpoint_path = utils.to_absolute_path(cfg.checkpoint) checkpoint = torch.load(checkpoint_path, map_location=lambda storage, loc: storage) encoder.load_state_dict(checkpoint["encoder"]) decoder.load_state_dict(checkpoint["decoder"]) encoder.eval() decoder.eval() meter = pyloudnorm.Meter(cfg.preprocessing.sr) #--------------------------------------- if cfg.privacy_preference == "Low": for wav_path, speaker_id, out_filename in tqdm(filter_list): wav_path = in_dir / wav_path # librosa.load (it will return audio time series, and its sampling rate) wav, _ = librosa.load(wav_path.with_suffix(".wav"), sr=cfg.preprocessing.sr) ref_loudness = meter.integrated_loudness(wav) wav = wav / np.abs(wav).max() * 0.999 path = out_dir / out_filename # to return raw recording in mel-spectrogram without any filtering if cfg.output_type == "Embedding": mel = librosa.feature.melspectrogram( preemphasis(wav, cfg.preprocessing.preemph), sr=cfg.preprocessing.sr, n_fft=cfg.preprocessing.n_fft, n_mels=cfg.preprocessing.n_mels, hop_length=cfg.preprocessing.hop_length, win_length=cfg.preprocessing.win_length, fmin=cfg.preprocessing.fmin, power=1) logmel = librosa.amplitude_to_db(mel, top_db=cfg.preprocessing.top_db) logmel = logmel / cfg.preprocessing.top_db + 1 mel = torch.FloatTensor(logmel).squeeze().to(device).numpy() np.savetxt(path.with_suffix(".mel.txt"), mel) # to return raw recording in waveform without any filtering if cfg.output_type == "Recording": librosa.output.write_wav(path.with_suffix(".wav"), wav.astype(np.float32), sr=cfg.preprocessing.sr) #--------------------------------------- if cfg.privacy_preference == "Moderate": dataset_path = Path(utils.to_absolute_path("Training/Datasets")) / cfg.dataset.path with open(dataset_path / "speakers.json") as file: speakers = sorted(json.load(file)) for wav_path, speaker_id, out_filename in tqdm(filter_list): wav_path = in_dir / wav_path wav, _ = librosa.load( wav_path.with_suffix(".wav"), sr=cfg.preprocessing.sr) ref_loudness = meter.integrated_loudness(wav) wav = wav / np.abs(wav).max() * 0.999 mel = librosa.feature.melspectrogram( preemphasis(wav, cfg.preprocessing.preemph), sr=cfg.preprocessing.sr, n_fft=cfg.preprocessing.n_fft, n_mels=cfg.preprocessing.n_mels, hop_length=cfg.preprocessing.hop_length, win_length=cfg.preprocessing.win_length, fmin=cfg.preprocessing.fmin, power=1) logmel = librosa.amplitude_to_db(mel, top_db=cfg.preprocessing.top_db) logmel = logmel / cfg.preprocessing.top_db + 1 mel = torch.FloatTensor(logmel).unsqueeze(0).to(device) speaker = torch.LongTensor([speakers.index(speaker_id)]).to(device) path = out_dir / out_filename if cfg.output_type == "Recording": with torch.no_grad(): vq, _ = encoder.encode(mel) output = decoder.generate(vq, speaker) output_loudness = meter.integrated_loudness(output) output = pyloudnorm.normalize.loudness(output, output_loudness, ref_loudness) librosa.output.write_wav(path.with_suffix(".wav"), output.astype(np.float32), sr=cfg.preprocessing.sr) if cfg.output_type == "Embedding": with torch.no_grad(): vq, _ = encoder.encode(mel) speaker = decoder.speaker(speaker) vq = vq.squeeze().to(device).numpy() speaker = speaker.squeeze().to(device).numpy() np.savetxt(path.with_suffix(".vq.txt"), vq) np.savetxt(path.with_suffix(".speaker.txt"), speaker) #--------------------------------------- if cfg.privacy_preference == "High": dataset_path = Path(utils.to_absolute_path("Training/Datasets")) / cfg.dataset.path with open(dataset_path / "speakers.json") as file: speakers = sorted(json.load(file)) for wav_path, speaker_id, out_filename in tqdm(filter_list): wav_path = in_dir / wav_path wav, _ = librosa.load( wav_path.with_suffix(".wav"),sr=cfg.preprocessing.sr) ref_loudness = meter.integrated_loudness(wav) wav = wav / np.abs(wav).max() * 0.999 mel = librosa.feature.melspectrogram( preemphasis(wav, cfg.preprocessing.preemph), sr=cfg.preprocessing.sr, n_fft=cfg.preprocessing.n_fft, n_mels=cfg.preprocessing.n_mels, hop_length=cfg.preprocessing.hop_length, win_length=cfg.preprocessing.win_length, fmin=cfg.preprocessing.fmin, power=1) logmel = librosa.amplitude_to_db(mel, top_db=cfg.preprocessing.top_db) logmel = logmel / cfg.preprocessing.top_db + 1 mel = torch.FloatTensor(logmel).unsqueeze(0).to(device) speaker = torch.LongTensor([speakers.index(speaker_id)]).to(device) path = out_dir / out_filename if cfg.output_type == "Recording": with torch.no_grad(): vq, _ = encoder.encode(mel) output = decoder.generate(vq, speaker) output_loudness = meter.integrated_loudness(output) output = pyloudnorm.normalize.loudness(output, output_loudness, ref_loudness) librosa.output.write_wav(path.with_suffix(".wav"), output.astype(np.float32), sr=cfg.preprocessing.sr) if cfg.output_type == "Embedding": with torch.no_grad(): vq, _ = encoder.encode(mel) vq = vq.squeeze().cpu().numpy() np.savetxt(path.with_suffix(".vq.txt"), vq) if __name__ == "__main__": DDF()
6,913
43.320513
111
py
EDGY
EDGY-master/Training/VQ-VAE/dataset.py
import numpy as np import torch from torch.utils.data import Dataset import json from random import randint from pathlib import Path class SpeechDataset(Dataset): def __init__(self, root, hop_length, sr, sample_frames): self.root = Path(root) self.hop_length = hop_length self.sample_frames = sample_frames with open(self.root / "speakers.json") as file: self.speakers = sorted(json.load(file)) min_duration = (sample_frames + 2) * hop_length / sr with open(self.root / "train.json") as file: metadata = json.load(file) self.metadata = [ Path(out_path) for _, _, duration, out_path in metadata if duration > min_duration ] def __len__(self): return len(self.metadata) def __getitem__(self, index): path = self.metadata[index] path = self.root.parent / path audio = np.load(path.with_suffix(".wav.npy")) mel = np.load(path.with_suffix(".mel.npy")) pos = randint(1, mel.shape[-1] - self.sample_frames - 2) mel = mel[:, pos - 1:pos + self.sample_frames + 1] audio = audio[pos * self.hop_length:(pos + self.sample_frames) * self.hop_length + 1] speaker = self.speakers.index(path.parts[-2]) return torch.LongTensor(audio), torch.FloatTensor(mel), speaker
1,381
31.139535
93
py
EDGY
EDGY-master/Training/VQ-VAE/models.py
import torch import torch.nn as nn import torch.nn.functional as F from torch.distributions import Categorical from tqdm import tqdm import numpy as np from preprocess import mulaw_decode def get_gru_cell(gru): gru_cell = nn.GRUCell(gru.input_size, gru.hidden_size) gru_cell.weight_hh.data = gru.weight_hh_l0.data gru_cell.weight_ih.data = gru.weight_ih_l0.data gru_cell.bias_hh.data = gru.bias_hh_l0.data gru_cell.bias_ih.data = gru.bias_ih_l0.data return gru_cell class Encoder(nn.Module): def __init__(self, in_channels, channels, n_embeddings, embedding_dim, jitter=0): super(Encoder, self).__init__() self.encoder = nn.Sequential( nn.Conv1d(in_channels, channels, 3, 1, 0, bias=False), nn.BatchNorm1d(channels), nn.ReLU(True), nn.Conv1d(channels, channels, 3, 1, 1, bias=False), nn.BatchNorm1d(channels), nn.ReLU(True), nn.Conv1d(channels, channels, 4, 2, 1, bias=False), nn.BatchNorm1d(channels), nn.ReLU(True), nn.Conv1d(channels, channels, 3, 1, 1, bias=False), nn.BatchNorm1d(channels), nn.ReLU(True), nn.Conv1d(channels, channels, 3, 1, 1, bias=False), nn.BatchNorm1d(channels), nn.ReLU(True), nn.Conv1d(channels, embedding_dim, 1) ) self.codebook = VQEmbeddingEMA(n_embeddings, embedding_dim) self.jitter = Jitter(jitter) def forward(self, mels): z = self.encoder(mels) z, loss, perplexity = self.codebook(z.transpose(1, 2)) z = self.jitter(z) return z, loss, perplexity def encode(self, mel): z = self.encoder(mel) z, indices = self.codebook.encode(z.transpose(1, 2)) return z, indices class Jitter(nn.Module): def __init__(self, p): super().__init__() self.p = p prob = torch.Tensor([p / 2, 1 - p, p / 2]) self.register_buffer("prob", prob) def forward(self, x): if not self.training or self.p == 0: return x else: batch_size, sample_size, channels = x.size() dist = Categorical(self.prob) index = dist.sample(torch.Size([batch_size, sample_size])) - 1 index[:, 0].clamp_(0, 1) index[:, -1].clamp_(-1, 0) index += torch.arange(sample_size, device=x.device) x = torch.gather(x, 1, index.unsqueeze(-1).expand(-1, -1, channels)) return x class VQEmbeddingEMA(nn.Module): def __init__(self, n_embeddings, embedding_dim, commitment_cost=0.25, decay=0.999, epsilon=1e-5): super(VQEmbeddingEMA, self).__init__() self.commitment_cost = commitment_cost self.decay = decay self.epsilon = epsilon init_bound = 1 / 512 embedding = torch.Tensor(n_embeddings, embedding_dim) embedding.uniform_(-init_bound, init_bound) self.register_buffer("embedding", embedding) self.register_buffer("ema_count", torch.zeros(n_embeddings)) self.register_buffer("ema_weight", self.embedding.clone()) def encode(self, x): M, D = self.embedding.size() x_flat = x.detach().reshape(-1, D) distances = torch.addmm(torch.sum(self.embedding ** 2, dim=1) + torch.sum(x_flat ** 2, dim=1, keepdim=True), x_flat, self.embedding.t(), alpha=-2.0, beta=1.0) indices = torch.argmin(distances.float(), dim=-1) quantized = F.embedding(indices, self.embedding) quantized = quantized.view_as(x) return quantized, indices def forward(self, x): M, D = self.embedding.size() x_flat = x.detach().reshape(-1, D) distances = torch.addmm(torch.sum(self.embedding ** 2, dim=1) + torch.sum(x_flat ** 2, dim=1, keepdim=True), x_flat, self.embedding.t(), alpha=-2.0, beta=1.0) indices = torch.argmin(distances.float(), dim=-1) encodings = F.one_hot(indices, M).float() quantized = F.embedding(indices, self.embedding) quantized = quantized.view_as(x) if self.training: self.ema_count = self.decay * self.ema_count + (1 - self.decay) * torch.sum(encodings, dim=0) n = torch.sum(self.ema_count) self.ema_count = (self.ema_count + self.epsilon) / (n + M * self.epsilon) * n dw = torch.matmul(encodings.t(), x_flat) self.ema_weight = self.decay * self.ema_weight + (1 - self.decay) * dw self.embedding = self.ema_weight / self.ema_count.unsqueeze(-1) e_latent_loss = F.mse_loss(x, quantized.detach()) loss = self.commitment_cost * e_latent_loss quantized = x + (quantized - x).detach() avg_probs = torch.mean(encodings, dim=0) perplexity = torch.exp(-torch.sum(avg_probs * torch.log(avg_probs + 1e-10))) return quantized, loss, perplexity class Decoder(nn.Module): def __init__(self, in_channels, n_speakers, speaker_embedding_dim, conditioning_channels, mu_embedding_dim, rnn_channels, fc_channels, bits, hop_length): super().__init__() self.rnn_channels = rnn_channels self.quantization_channels = 2**bits self.hop_length = hop_length self.speaker_embedding = nn.Embedding(n_speakers, speaker_embedding_dim) self.rnn1 = nn.GRU(in_channels + speaker_embedding_dim, conditioning_channels, num_layers=2, batch_first=True, bidirectional=True) self.mu_embedding = nn.Embedding(self.quantization_channels, mu_embedding_dim) self.rnn2 = nn.GRU(mu_embedding_dim + 2*conditioning_channels, rnn_channels, batch_first=True) self.fc1 = nn.Linear(rnn_channels, fc_channels) self.fc2 = nn.Linear(fc_channels, self.quantization_channels) def forward(self, x, z, speakers): z = F.interpolate(z.transpose(1, 2), scale_factor=2) z = z.transpose(1, 2) speakers = self.speaker_embedding(speakers) speakers = speakers.unsqueeze(1).expand(-1, z.size(1), -1) z = torch.cat((z, speakers), dim=-1) z, _ = self.rnn1(z) z = F.interpolate(z.transpose(1, 2), scale_factor=self.hop_length) z = z.transpose(1, 2) x = self.mu_embedding(x) x, _ = self.rnn2(torch.cat((x, z), dim=2)) x = F.relu(self.fc1(x)) x = self.fc2(x) return x def generate(self, z, speaker): output = [] cell = get_gru_cell(self.rnn2) z = F.interpolate(z.transpose(1, 2), scale_factor=2) z = z.transpose(1, 2) speaker = self.speaker_embedding(speaker) speaker = speaker.unsqueeze(1).expand(-1, z.size(1), -1) z = torch.cat((z, speaker), dim=-1) z, _ = self.rnn1(z) z = F.interpolate(z.transpose(1, 2), scale_factor=self.hop_length) z = z.transpose(1, 2) batch_size, sample_size, _ = z.size() h = torch.zeros(batch_size, self.rnn_channels, device=z.device) x = torch.zeros(batch_size, device=z.device).fill_(self.quantization_channels // 2).long() for m in tqdm(torch.unbind(z, dim=1), leave=False): x = self.mu_embedding(x) h = cell(torch.cat((x, m), dim=1), h) x = F.relu(self.fc1(h)) logits = self.fc2(x) dist = Categorical(logits=logits) x = dist.sample() output.append(2 * x.float().item() / (self.quantization_channels - 1.) - 1.) output = np.asarray(output, dtype=np.float64) output = mulaw_decode(output, self.quantization_channels) return output def speaker(self, speaker): speaker = self.speaker_embedding(speaker) return speaker
7,998
35.861751
105
py
EDGY
EDGY-master/Training/VQ-VAE/train_VQ.py
import hydra from hydra import utils from itertools import chain from pathlib import Path from tqdm import tqdm import apex.amp as amp import torch import torch.nn.functional as F import torch.optim as optim from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from dataset import SpeechDataset from models import Encoder, Decoder def save_checkpoint(encoder, decoder, optimizer, amp, scheduler, step, checkpoint_dir): checkpoint_state = { "encoder": encoder.state_dict(), "decoder": decoder.state_dict(), "optimizer": optimizer.state_dict(), "amp": amp.state_dict(), "scheduler": scheduler.state_dict(), "step": step} checkpoint_dir.mkdir(exist_ok=True, parents=True) checkpoint_path = checkpoint_dir / "model.ckpt-{}.pt".format(step) torch.save(checkpoint_state, checkpoint_path) print("Saved checkpoint: {}".format(checkpoint_path.stem)) @hydra.main(config_path="Training/VQ-VAE/Configuration_files/training.yaml") def train_model(cfg): tensorboard_path = Path(utils.to_absolute_path("tensorboard")) / cfg.checkpoint_dir checkpoint_dir = Path(utils.to_absolute_path(cfg.checkpoint_dir)) writer = SummaryWriter(tensorboard_path) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") encoder = Encoder(**cfg.model.encoder) decoder = Decoder(**cfg.model.decoder) encoder.to(device) decoder.to(device) optimizer = optim.Adam( chain(encoder.parameters(), decoder.parameters()), lr=cfg.training.optimizer.lr) [encoder, decoder], optimizer = amp.initialize([encoder, decoder], optimizer, opt_level="O1") scheduler = optim.lr_scheduler.MultiStepLR( optimizer, milestones=cfg.training.scheduler.milestones, gamma=cfg.training.scheduler.gamma) if cfg.resume: print("Resume checkpoint from: {}:".format(cfg.resume)) resume_path = utils.to_absolute_path(cfg.resume) checkpoint = torch.load(resume_path, map_location=lambda storage, loc: storage) encoder.load_state_dict(checkpoint["encoder"]) decoder.load_state_dict(checkpoint["decoder"]) optimizer.load_state_dict(checkpoint["optimizer"]) amp.load_state_dict(checkpoint["amp"]) scheduler.load_state_dict(checkpoint["scheduler"]) global_step = checkpoint["step"] else: global_step = 0 root_path = Path(utils.to_absolute_path("datasets")) / cfg.dataset.path dataset = SpeechDataset( root=root_path, hop_length=cfg.preprocessing.hop_length, sr=cfg.preprocessing.sr, sample_frames=cfg.training.sample_frames) dataloader = DataLoader( dataset, batch_size=cfg.training.batch_size, shuffle=True, num_workers=cfg.training.n_workers, pin_memory=True, drop_last=True) n_epochs = cfg.training.n_steps // len(dataloader) + 1 start_epoch = global_step // len(dataloader) + 1 for epoch in range(start_epoch, n_epochs + 1): average_recon_loss = average_vq_loss = average_perplexity = 0 for i, (audio, mels, speakers) in enumerate(tqdm(dataloader), 1): audio, mels, speakers = audio.to(device), mels.to(device), speakers.to(device) optimizer.zero_grad() z, vq_loss, perplexity = encoder(mels) output = decoder(audio[:, :-1], z, speakers) recon_loss = F.cross_entropy(output.transpose(1, 2), audio[:, 1:]) loss = recon_loss + vq_loss with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), 1) optimizer.step() scheduler.step() average_recon_loss += (recon_loss.item() - average_recon_loss) / i average_vq_loss += (vq_loss.item() - average_vq_loss) / i average_perplexity += (perplexity.item() - average_perplexity) / i global_step += 1 if global_step % cfg.training.checkpoint_interval == 0: save_checkpoint( encoder, decoder, optimizer, amp, scheduler, global_step, checkpoint_dir) writer.add_scalar("recon_loss/train", average_recon_loss, global_step) writer.add_scalar("vq_loss/train", average_vq_loss, global_step) writer.add_scalar("average_perplexity", average_perplexity, global_step) print("epoch:{}, recon loss:{:.2E}, vq loss:{:.2E}, perpexlity:{:.3f}" .format(epoch, average_recon_loss, average_vq_loss, average_perplexity)) if __name__ == "__main__": train_model()
4,714
37.647541
97
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/incremental_train_and_eval_AMR_LF.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * cur_features = [] ref_features = [] old_scores = [] new_scores = [] def get_ref_features(self, inputs, outputs): global ref_features ref_features = inputs[0] def get_cur_features(self, inputs, outputs): global cur_features cur_features = inputs[0] def get_old_scores_before_scale(self, inputs, outputs): global old_scores old_scores = outputs def get_new_scores_before_scale(self, inputs, outputs): global new_scores new_scores = outputs def incremental_train_and_eval_AMR_LF(epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iteration, \ lamda, \ dist, K, lw_mr, \ weight_per_class=None, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = Y_train #trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, # shuffle=True, num_workers=2) #testset.test_data = X_valid.astype('uint8') #testset.test_labels = Y_valid #testloader = torch.utils.data.DataLoader(testset, batch_size=100, # shuffle=False, num_workers=2) #print('Max and Min of train labels: {}, {}'.format(min(Y_train), max(Y_train))) #print('Max and Min of valid labels: {}, {}'.format(min(Y_valid), max(Y_valid))) if iteration > start_iteration: ref_model.eval() num_old_classes = ref_model.fc.out_features handle_ref_features = ref_model.fc.register_forward_hook(get_ref_features) handle_cur_features = tg_model.fc.register_forward_hook(get_cur_features) handle_old_scores_bs = tg_model.fc.fc1.register_forward_hook(get_old_scores_before_scale) handle_new_scores_bs = tg_model.fc.fc2.register_forward_hook(get_new_scores_before_scale) for epoch in range(epochs): #train tg_model.train() train_loss = 0 train_loss1 = 0 train_loss2 = 0 train_loss3 = 0 correct = 0 total = 0 tg_lr_scheduler.step() print('\nEpoch: %d, LR: ' % epoch, end='') print(tg_lr_scheduler.get_lr()) for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.to(device), targets.to(device) tg_optimizer.zero_grad() outputs = tg_model(inputs) if iteration == start_iteration: loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) else: ref_outputs = ref_model(inputs) loss1 = nn.CosineEmbeddingLoss()(cur_features, ref_features.detach(), \ torch.ones(inputs.shape[0]).to(device)) * lamda loss2 = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) ################################################# #scores before scale, [-1, 1] outputs_bs = torch.cat((old_scores, new_scores), dim=1) assert(outputs_bs.size()==outputs.size()) #print("outputs_bs:", outputs_bs.size(), outputs_bs) #print("targets:", targets.size(), targets) #get groud truth scores gt_index = torch.zeros(outputs_bs.size()).to(device) gt_index = gt_index.scatter(1, targets.view(-1,1), 1).ge(0.5) gt_scores = outputs_bs.masked_select(gt_index) #print("gt_index:", gt_index.size(), gt_index) #print("gt_scores:", gt_scores.size(), gt_scores) #get top-K scores on none gt classes none_gt_index = torch.zeros(outputs_bs.size()).to(device) none_gt_index = none_gt_index.scatter(1, targets.view(-1,1), 1).le(0.5) none_gt_scores = outputs_bs.masked_select(none_gt_index).reshape((outputs_bs.size(0), outputs.size(1)-1)) #print("none_gt_index:", none_gt_index.size(), none_gt_index) #print("none_gt_scores:", none_gt_scores.size(), none_gt_scores) hard_scores = none_gt_scores.topk(K, dim=1)[0] #print("hard_scores:", hard_scores.size(), hard_scores) #the index of hard samples, i.e., samples of old classes hard_index = targets.lt(num_old_classes) hard_num = torch.nonzero(hard_index).size(0) #print("hard examples size: ", hard_num) if hard_num > 0: gt_scores = gt_scores[hard_index].view(-1, 1).repeat(1, K) hard_scores = hard_scores[hard_index] assert(gt_scores.size() == hard_scores.size()) assert(gt_scores.size(0) == hard_num) #print("hard example gt scores: ", gt_scores.size(), gt_scores) #print("hard example max novel scores: ", hard_scores.size(), hard_scores) loss3 = nn.MarginRankingLoss(margin=dist)(gt_scores.view(-1, 1), \ hard_scores.view(-1, 1), torch.ones(hard_num*K).to(device)) * lw_mr else: loss3 = torch.zeros(1).to(device) ################################################# loss = loss1 + loss2 + loss3 loss.backward() tg_optimizer.step() train_loss += loss.item() if iteration > start_iteration: train_loss1 += loss1.item() train_loss2 += loss2.item() train_loss3 += loss3.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #if iteration == 0: # msg = 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (train_loss/(batch_idx+1), 100.*correct/total, correct, total) #else: # msg = 'Loss1: %.3f Loss2: %.3f Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (loss1.item(), loss2.item(), train_loss/(batch_idx+1), 100.*correct/total, correct, total) #progress_bar(batch_idx, len(trainloader), msg) if iteration == start_iteration: print('Train set: {}, Train Loss: {:.4f} Acc: {:.4f}'.format(\ len(trainloader), train_loss/(batch_idx+1), 100.*correct/total)) else: print('Train set: {}, Train Loss1: {:.4f}, Train Loss2: {:.4f}, Train Loss3: {:.4f},\ Train Loss: {:.4f} Acc: {:.4f}'.format(len(trainloader), \ train_loss1/(batch_idx+1), train_loss2/(batch_idx+1), train_loss3/(batch_idx+1), train_loss/(batch_idx+1), 100.*correct/total)) #eval tg_model.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = tg_model(inputs) loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' # % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) print('Test set: {} Test Loss: {:.4f} Acc: {:.4f}'.format(\ len(testloader), test_loss/(batch_idx+1), 100.*correct/total)) if iteration > start_iteration: print("Removing register_forward_hook") handle_ref_features.remove() handle_cur_features.remove() handle_old_scores_bs.remove() handle_new_scores_bs.remove() return tg_model
8,384
45.071429
121
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/compute_confusion_matrix.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * def compute_confusion_matrix(tg_model, tg_feature_model, class_means, evalloader, print_info=False, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") tg_model.eval() tg_feature_model.eval() #evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) #evalset.test_data = input_data.astype('uint8') #evalset.test_labels = input_labels #evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, # shuffle=False, num_workers=2) correct = 0 correct_icarl = 0 correct_ncm = 0 total = 0 num_classes = tg_model.fc.out_features cm = np.zeros((3, num_classes, num_classes)) all_targets = [] all_predicted = [] all_predicted_icarl = [] all_predicted_ncm = [] with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(evalloader): inputs, targets = inputs.to(device), targets.to(device) total += targets.size(0) all_targets.append(targets) outputs = tg_model(inputs) _, predicted = outputs.max(1) correct += predicted.eq(targets).sum().item() all_predicted.append(predicted) outputs_feature = np.squeeze(tg_feature_model(inputs)) # Compute score for iCaRL sqd_icarl = cdist(class_means[:,:,0].T, outputs_feature, 'sqeuclidean') score_icarl = torch.from_numpy((-sqd_icarl).T).to(device) _, predicted_icarl = score_icarl.max(1) correct_icarl += predicted_icarl.eq(targets).sum().item() all_predicted_icarl.append(predicted_icarl) # Compute score for NCM sqd_ncm = cdist(class_means[:,:,1].T, outputs_feature, 'sqeuclidean') score_ncm = torch.from_numpy((-sqd_ncm).T).to(device) _, predicted_ncm = score_ncm.max(1) correct_ncm += predicted_ncm.eq(targets).sum().item() all_predicted_ncm.append(predicted_ncm) # print(sqd_icarl.shape, score_icarl.shape, predicted_icarl.shape, \ # sqd_ncm.shape, score_ncm.shape, predicted_ncm.shape) cm[0, :, :] = confusion_matrix(np.concatenate(all_targets), np.concatenate(all_predicted)) cm[1, :, :] = confusion_matrix(np.concatenate(all_targets), np.concatenate(all_predicted_icarl)) cm[2, :, :] = confusion_matrix(np.concatenate(all_targets), np.concatenate(all_predicted_ncm)) if print_info: print(" top 1 accuracy CNN :\t\t{:.2f} %".format( 100.*correct/total )) print(" top 1 accuracy iCaRL :\t\t{:.2f} %".format( 100.*correct_icarl/total )) print(" top 1 accuracy NCM :\t\t{:.2f} %".format( 100.*correct_ncm/total )) print(" top 1 accuracy CNN :\t\t{:.2f} %".format( 100.*np.mean(np.diag(cm[0])/np.sum(cm[0],axis=1)) )) print(" top 1 accuracy iCaRL :\t\t{:.2f} %".format( 100.*np.mean(np.diag(cm[1])/np.sum(cm[1],axis=1)) )) print(" top 1 accuracy NCM :\t\t{:.2f} %".format( 100.*np.mean(np.diag(cm[2])/np.sum(cm[2],axis=1)) )) return cm
3,725
43.357143
122
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/incremental_train_and_eval_MS.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * def get_old_scores_before_scale(self, inputs, outputs): global old_scores old_scores = outputs def get_new_scores_before_scale(self, inputs, outputs): global new_scores new_scores = outputs def incremental_train_and_eval_MS(epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iteration, \ lw_ms, \ fix_bn=False, weight_per_class=None, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = Y_train #trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, # shuffle=True, num_workers=2) #testset.test_data = X_valid.astype('uint8') #testset.test_labels = Y_valid #testloader = torch.utils.data.DataLoader(testset, batch_size=100, # shuffle=False, num_workers=2) #print('Max and Min of train labels: {}, {}'.format(min(Y_train), max(Y_train))) #print('Max and Min of valid labels: {}, {}'.format(min(Y_valid), max(Y_valid))) if iteration > start_iteration: ref_model.eval() num_old_classes = ref_model.fc.out_features handle_old_scores_bs = tg_model.fc.fc1.register_forward_hook(get_old_scores_before_scale) handle_new_scores_bs = tg_model.fc.fc2.register_forward_hook(get_new_scores_before_scale) for epoch in range(epochs): #train tg_model.train() if fix_bn: for m in tg_model.modules(): if isinstance(m, nn.BatchNorm2d): m.eval() #m.weight.requires_grad = False #m.bias.requires_grad = False train_loss = 0 train_loss1 = 0 train_loss2 = 0 correct = 0 total = 0 tg_lr_scheduler.step() print('\nEpoch: %d, LR: ' % epoch, end='') print(tg_lr_scheduler.get_lr()) for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.to(device), targets.to(device) tg_optimizer.zero_grad() outputs = tg_model(inputs) if iteration == start_iteration: loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) else: ref_outputs = ref_model(inputs) #loss1 = nn.KLDivLoss()(F.log_softmax(outputs[:,:num_old_classes]/T, dim=1), \ # F.softmax(ref_outputs.detach()/T, dim=1)) * T * T * beta * num_old_classes ref_scores = ref_outputs.detach() / ref_model.fc.sigma.detach() loss1 = nn.MSELoss()(old_scores, ref_scores.detach()) * lw_ms * num_old_classes loss2 = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) loss = loss1 + loss2 loss.backward() tg_optimizer.step() train_loss += loss.item() if iteration > start_iteration: train_loss1 += loss1.item() train_loss2 += loss2.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #if iteration == 0: # msg = 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (train_loss/(batch_idx+1), 100.*correct/total, correct, total) #else: # msg = 'Loss1: %.3f Loss2: %.3f Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (loss1.item(), loss2.item(), train_loss/(batch_idx+1), 100.*correct/total, correct, total) #progress_bar(batch_idx, len(trainloader), msg) if iteration == start_iteration: print('Train set: {}, Train Loss: {:.4f} Acc: {:.4f}'.format(\ len(trainloader), train_loss/(batch_idx+1), 100.*correct/total)) else: print('Train set: {}, Train Loss1: {:.4f}, Train Loss2: {:.4f},\ Train Loss: {:.4f} Acc: {:.4f}'.format(len(trainloader), \ train_loss1/(batch_idx+1), train_loss2/(batch_idx+1), train_loss/(batch_idx+1), 100.*correct/total)) #eval tg_model.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = tg_model(inputs) loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' # % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) print('Test set: {} Test Loss: {:.4f} Acc: {:.4f}'.format(\ len(testloader), test_loss/(batch_idx+1), 100.*correct/total)) if iteration > start_iteration: print("Removing register_forward_hook") handle_old_scores_bs.remove() handle_new_scores_bs.remove() return tg_model
5,759
41.666667
107
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/incremental_train_and_eval.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * def incremental_train_and_eval(epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iteration, \ T, beta, \ fix_bn=False, weight_per_class=None, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = Y_train #trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, # shuffle=True, num_workers=2) #testset.test_data = X_valid.astype('uint8') #testset.test_labels = Y_valid #testloader = torch.utils.data.DataLoader(testset, batch_size=100, # shuffle=False, num_workers=2) #print('Max and Min of train labels: {}, {}'.format(min(Y_train), max(Y_train))) #print('Max and Min of valid labels: {}, {}'.format(min(Y_valid), max(Y_valid))) if iteration > start_iteration: ref_model.eval() num_old_classes = ref_model.fc.out_features for epoch in range(epochs): #train tg_model.train() if fix_bn: for m in tg_model.modules(): if isinstance(m, nn.BatchNorm2d): m.eval() #m.weight.requires_grad = False #m.bias.requires_grad = False train_loss = 0 train_loss1 = 0 train_loss2 = 0 correct = 0 total = 0 tg_lr_scheduler.step() print('\nEpoch: %d, LR: ' % epoch, end='') print(tg_lr_scheduler.get_lr()) for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.to(device), targets.to(device) tg_optimizer.zero_grad() outputs = tg_model(inputs) if iteration == start_iteration: loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) else: ref_outputs = ref_model(inputs) loss1 = nn.KLDivLoss()(F.log_softmax(outputs[:,:num_old_classes]/T, dim=1), \ F.softmax(ref_outputs.detach()/T, dim=1)) * T * T * beta * num_old_classes loss2 = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) loss = loss1 + loss2 loss.backward() tg_optimizer.step() train_loss += loss.item() if iteration > start_iteration: train_loss1 += loss1.item() train_loss2 += loss2.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #if iteration == 0: # msg = 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (train_loss/(batch_idx+1), 100.*correct/total, correct, total) #else: # msg = 'Loss1: %.3f Loss2: %.3f Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (loss1.item(), loss2.item(), train_loss/(batch_idx+1), 100.*correct/total, correct, total) #progress_bar(batch_idx, len(trainloader), msg) if iteration == start_iteration: print('Train set: {}, Train Loss: {:.4f} Acc: {:.4f}'.format(\ len(trainloader), train_loss/(batch_idx+1), 100.*correct/total)) else: print('Train set: {}, Train Loss1: {:.4f}, Train Loss2: {:.4f},\ Train Loss: {:.4f} Acc: {:.4f}'.format(len(trainloader), \ train_loss1/(batch_idx+1), train_loss2/(batch_idx+1), train_loss/(batch_idx+1), 100.*correct/total)) #eval tg_model.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = tg_model(inputs) loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' # % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) print('Test set: {} Test Loss: {:.4f} Acc: {:.4f}'.format(\ len(testloader), test_loss/(batch_idx+1), 100.*correct/total)) return tg_model
5,014
41.5
107
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/compute_accuracy.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * def compute_accuracy(tg_model, tg_feature_model, class_means, evalloader, scale=None, print_info=True, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") tg_model.eval() tg_feature_model.eval() #evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) #evalset.test_data = input_data.astype('uint8') #evalset.test_labels = input_labels #evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, # shuffle=False, num_workers=2) correct = 0 correct_icarl = 0 correct_ncm = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(evalloader): inputs, targets = inputs.to(device), targets.to(device) total += targets.size(0) outputs = tg_model(inputs) outputs = F.softmax(outputs, dim=1) if scale is not None: assert(scale.shape[0] == 1) assert(outputs.shape[1] == scale.shape[1]) outputs = outputs / scale.repeat(outputs.shape[0], 1).type(torch.FloatTensor).to(device) _, predicted = outputs.max(1) correct += predicted.eq(targets).sum().item() outputs_feature = np.squeeze(tg_feature_model(inputs)) # Compute score for iCaRL sqd_icarl = cdist(class_means[:,:,0].T, outputs_feature, 'sqeuclidean') score_icarl = torch.from_numpy((-sqd_icarl).T).to(device) _, predicted_icarl = score_icarl.max(1) correct_icarl += predicted_icarl.eq(targets).sum().item() # Compute score for NCM sqd_ncm = cdist(class_means[:,:,1].T, outputs_feature, 'sqeuclidean') score_ncm = torch.from_numpy((-sqd_ncm).T).to(device) _, predicted_ncm = score_ncm.max(1) correct_ncm += predicted_ncm.eq(targets).sum().item() # print(sqd_icarl.shape, score_icarl.shape, predicted_icarl.shape, \ # sqd_ncm.shape, score_ncm.shape, predicted_ncm.shape) if print_info: print(" top 1 accuracy CNN :\t\t{:.2f} %".format(100.*correct/total)) print(" top 1 accuracy iCaRL :\t\t{:.2f} %".format(100.*correct_icarl/total)) print(" top 1 accuracy NCM :\t\t{:.2f} %".format(100.*correct_ncm/total)) cnn_acc = 100.*correct/total icarl_acc = 100.*correct_icarl/total ncm_acc = 100.*correct_ncm/total return [cnn_acc, icarl_acc, ncm_acc]
3,097
40.306667
116
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/compute_features.py
#!/usr/bin/env python # coding=utf-8 #!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * def compute_features(tg_feature_model, evalloader, num_samples, num_features, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") tg_feature_model.eval() #evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) #evalset.test_data = input_data.astype('uint8') #evalset.test_labels = np.zeros(input_data.shape[0]) #evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, # shuffle=False, num_workers=2) features = np.zeros([num_samples, num_features]) start_idx = 0 with torch.no_grad(): for inputs, targets in evalloader: inputs = inputs.to(device) features[start_idx:start_idx+inputs.shape[0], :] = np.squeeze(tg_feature_model(inputs)) start_idx = start_idx+inputs.shape[0] assert(start_idx==num_samples) return features
1,503
33.181818
99
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/incremental_train_and_eval_LF.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * cur_features = [] ref_features = [] def get_ref_features(self, inputs, outputs): global ref_features ref_features = inputs[0] def get_cur_features(self, inputs, outputs): global cur_features cur_features = inputs[0] def incremental_train_and_eval_LF(epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iteration, \ lamda, \ fix_bn=False, weight_per_class=None, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = Y_train #trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, # shuffle=True, num_workers=2) #testset.test_data = X_valid.astype('uint8') #testset.test_labels = Y_valid #testloader = torch.utils.data.DataLoader(testset, batch_size=100, # shuffle=False, num_workers=2) #print('Max and Min of train labels: {}, {}'.format(min(Y_train), max(Y_train))) #print('Max and Min of valid labels: {}, {}'.format(min(Y_valid), max(Y_valid))) if iteration > start_iteration: ref_model.eval() handle_ref_features = ref_model.fc.register_forward_hook(get_ref_features) handle_cur_features = tg_model.fc.register_forward_hook(get_cur_features) for epoch in range(epochs): #train tg_model.train() if fix_bn: for m in tg_model.modules(): if isinstance(m, nn.BatchNorm2d): m.eval() #m.weight.requires_grad = False #m.bias.requires_grad = False train_loss = 0 train_loss1 = 0 train_loss2 = 0 correct = 0 total = 0 tg_lr_scheduler.step() print('\nEpoch: %d, LR: ' % epoch, end='') print(tg_lr_scheduler.get_lr()) for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.to(device), targets.to(device) tg_optimizer.zero_grad() outputs = tg_model(inputs) if iteration == start_iteration: loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) else: ref_outputs = ref_model(inputs) loss1 = nn.CosineEmbeddingLoss()(cur_features, ref_features.detach(), \ torch.ones(inputs.shape[0]).to(device)) * lamda loss2 = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) loss = loss1 + loss2 loss.backward() tg_optimizer.step() train_loss += loss.item() if iteration > start_iteration: train_loss1 += loss1.item() train_loss2 += loss2.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #if iteration == 0: # msg = 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (train_loss/(batch_idx+1), 100.*correct/total, correct, total) #else: # msg = 'Loss1: %.3f Loss2: %.3f Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (loss1.item(), loss2.item(), train_loss/(batch_idx+1), 100.*correct/total, correct, total) #progress_bar(batch_idx, len(trainloader), msg) if iteration == start_iteration: print('Train set: {}, Train Loss: {:.4f} Acc: {:.4f}'.format(\ len(trainloader), train_loss/(batch_idx+1), 100.*correct/total)) else: print('Train set: {}, Train Loss1: {:.4f}, Train Loss2: {:.4f},\ Train Loss: {:.4f} Acc: {:.4f}'.format(len(trainloader), \ train_loss1/(batch_idx+1), train_loss2/(batch_idx+1), train_loss/(batch_idx+1), 100.*correct/total)) #eval tg_model.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = tg_model(inputs) loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' # % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) print('Test set: {} Test Loss: {:.4f} Acc: {:.4f}'.format(\ len(testloader), test_loss/(batch_idx+1), 100.*correct/total)) if iteration > start_iteration: print("Removing register_forward_hook") handle_ref_features.remove() handle_cur_features.remove() return tg_model
5,489
39.970149
107
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/utils_incremental/incremental_train_and_eval_MR_LF.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import copy import argparse from PIL import Image from scipy.spatial.distance import cdist from sklearn.metrics import confusion_matrix from utils_pytorch import * cur_features = [] ref_features = [] old_scores = [] new_scores = [] def get_ref_features(self, inputs, outputs): global ref_features ref_features = inputs[0] def get_cur_features(self, inputs, outputs): global cur_features cur_features = inputs[0] def get_old_scores_before_scale(self, inputs, outputs): global old_scores old_scores = outputs def get_new_scores_before_scale(self, inputs, outputs): global new_scores new_scores = outputs def incremental_train_and_eval_MR_LF(epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iteration, \ lamda, \ dist, K, lw_mr, \ fix_bn=False, weight_per_class=None, device=None): if device is None: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = Y_train #trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, # shuffle=True, num_workers=2) #testset.test_data = X_valid.astype('uint8') #testset.test_labels = Y_valid #testloader = torch.utils.data.DataLoader(testset, batch_size=100, # shuffle=False, num_workers=2) #print('Max and Min of train labels: {}, {}'.format(min(Y_train), max(Y_train))) #print('Max and Min of valid labels: {}, {}'.format(min(Y_valid), max(Y_valid))) if iteration > start_iteration: ref_model.eval() num_old_classes = ref_model.fc.out_features handle_ref_features = ref_model.fc.register_forward_hook(get_ref_features) handle_cur_features = tg_model.fc.register_forward_hook(get_cur_features) handle_old_scores_bs = tg_model.fc.fc1.register_forward_hook(get_old_scores_before_scale) handle_new_scores_bs = tg_model.fc.fc2.register_forward_hook(get_new_scores_before_scale) for epoch in range(epochs): #train tg_model.train() if fix_bn: for m in tg_model.modules(): if isinstance(m, nn.BatchNorm2d): m.eval() #m.weight.requires_grad = False #m.bias.requires_grad = False train_loss = 0 train_loss1 = 0 train_loss2 = 0 train_loss3 = 0 correct = 0 total = 0 tg_lr_scheduler.step() print('\nEpoch: %d, LR: ' % epoch, end='') print(tg_lr_scheduler.get_lr()) for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.to(device), targets.to(device) tg_optimizer.zero_grad() outputs = tg_model(inputs) if iteration == start_iteration: loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) else: ref_outputs = ref_model(inputs) loss1 = nn.CosineEmbeddingLoss()(cur_features, ref_features.detach(), \ torch.ones(inputs.shape[0]).to(device)) * lamda loss2 = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) ################################################# #scores before scale, [-1, 1] outputs_bs = torch.cat((old_scores, new_scores), dim=1) #print(tg_model.fc.fc1.in_features, tg_model.fc.fc1.out_features) #print(tg_model.fc.fc2.in_features, tg_model.fc.fc2.out_features) #print(old_scores.size(), new_scores.size(), outputs_bs.size(), outputs.size()) assert(outputs_bs.size()==outputs.size()) #get groud truth scores gt_index = torch.zeros(outputs_bs.size()).to(device) gt_index = gt_index.scatter(1, targets.view(-1,1), 1).ge(0.5) gt_scores = outputs_bs.masked_select(gt_index) #get top-K scores on novel classes max_novel_scores = outputs_bs[:, num_old_classes:].topk(K, dim=1)[0] #the index of hard samples, i.e., samples of old classes hard_index = targets.lt(num_old_classes) hard_num = torch.nonzero(hard_index).size(0) #print("hard examples size: ", hard_num) if hard_num > 0: gt_scores = gt_scores[hard_index].view(-1, 1).repeat(1, K) max_novel_scores = max_novel_scores[hard_index] assert(gt_scores.size() == max_novel_scores.size()) assert(gt_scores.size(0) == hard_num) #print("hard example gt scores: ", gt_scores.size(), gt_scores) #print("hard example max novel scores: ", max_novel_scores.size(), max_novel_scores) loss3 = nn.MarginRankingLoss(margin=dist)(gt_scores.view(-1, 1), \ max_novel_scores.view(-1, 1), torch.ones(hard_num*K).to(device)) * lw_mr else: loss3 = torch.zeros(1).to(device) ################################################# loss = loss1 + loss2 + loss3 loss.backward() tg_optimizer.step() train_loss += loss.item() if iteration > start_iteration: train_loss1 += loss1.item() train_loss2 += loss2.item() train_loss3 += loss3.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #if iteration == 0: # msg = 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (train_loss/(batch_idx+1), 100.*correct/total, correct, total) #else: # msg = 'Loss1: %.3f Loss2: %.3f Loss: %.3f | Acc: %.3f%% (%d/%d)' % \ # (loss1.item(), loss2.item(), train_loss/(batch_idx+1), 100.*correct/total, correct, total) #progress_bar(batch_idx, len(trainloader), msg) if iteration == start_iteration: print('Train set: {}, Train Loss: {:.4f} Acc: {:.4f}'.format(\ len(trainloader), train_loss/(batch_idx+1), 100.*correct/total)) else: print('Train set: {}, Train Loss1: {:.4f}, Train Loss2: {:.4f}, Train Loss3: {:.4f},\ Train Loss: {:.4f} Acc: {:.4f}'.format(len(trainloader), \ train_loss1/(batch_idx+1), train_loss2/(batch_idx+1), train_loss3/(batch_idx+1), train_loss/(batch_idx+1), 100.*correct/total)) #eval tg_model.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = tg_model(inputs) loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() #progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' # % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) print('Test set: {} Test Loss: {:.4f} Acc: {:.4f}'.format(\ len(testloader), test_loss/(batch_idx+1), 100.*correct/total)) if iteration > start_iteration: print("Removing register_forward_hook") handle_ref_features.remove() handle_cur_features.remove() handle_old_scores_bs.remove() handle_new_scores_bs.remove() return tg_model
8,171
44.149171
107
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/gen_imagenet_subset.py
#!/usr/bin/env python # coding=utf-8 import argparse import os import random import shutil import time import warnings import numpy as np import torch import torch.nn as nn import torch.nn.parallel import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.optim import torch.utils.data import torch.utils.data.distributed import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models data_dir = 'data/imagenet/data/' # Data loading code traindir = os.path.join(data_dir, 'train') train_dataset = datasets.ImageFolder(traindir, None) classes = train_dataset.classes print("the number of total classes: {}".format(len(classes))) seed = 1993 np.random.seed(seed) subset_num = 100 subset_classes = np.random.choice(classes, subset_num, replace=False) print("the number of subset classes: {}".format(len(subset_classes))) print(subset_classes) des_root_dir = 'data/seed_{}_subset_{}_imagenet/data/'.format(seed, subset_num) if not os.path.exists(des_root_dir): os.makedirs(des_root_dir) phase_list = ['train', 'val'] for phase in phase_list: if not os.path.exists(os.path.join(des_root_dir, phase)): os.mkdir(os.path.join(des_root_dir, phase)) for sc in subset_classes: src_dir = os.path.join(data_dir, phase, sc) des_dir = os.path.join(des_root_dir, phase, sc) cmd = "cp -r {} {}".format(src_dir, des_dir) print(cmd) os.system(cmd) print("Hello World")
1,499
26.777778
79
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/resnet.py
import torch.nn as nn import math import torch.utils.model_zoo as model_zoo __all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152'] model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', } def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers, num_classes=1000): self.inplanes = 64 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) self.avgpool = nn.AvgPool2d(7, stride=1) self.fc = nn.Linear(512 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x def resnet18(pretrained=False, **kwargs): """Constructs a ResNet-18 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet18'])) return model def resnet34(pretrained=False, **kwargs): """Constructs a ResNet-34 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet34'])) return model def resnet50(pretrained=False, **kwargs): """Constructs a ResNet-50 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet50'])) return model def resnet101(pretrained=False, **kwargs): """Constructs a ResNet-101 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet101'])) return model def resnet152(pretrained=False, **kwargs): """Constructs a ResNet-152 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet152'])) return model
6,582
29.906103
90
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/utils_pytorch.py
#!/usr/bin/env python # coding=utf-8 from __future__ import print_function, division import torch import torch.nn as nn import torch.nn.init as init from collections import OrderedDict import numpy as np import os import os.path as osp import sys import time import math import subprocess try: import cPickle as pickle except: import pickle def savepickle(data, file_path): mkdir_p(osp.dirname(file_path), delete=False) print('pickle into', file_path) with open(file_path, 'wb') as f: pickle.dump(data, f, pickle.HIGHEST_PROTOCOL) def unpickle(file_path): with open(file_path, 'rb') as f: data = pickle.load(f) return data def mkdir_p(path, delete=False, print_info=True): if path == '': return if delete: subprocess.call(('rm -r ' + path).split()) if not osp.exists(path): if print_info: print('mkdir -p ' + path) subprocess.call(('mkdir -p ' + path).split()) def get_mean_and_std(dataset): '''Compute the mean and std value of dataset.''' dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2) mean = torch.zeros(3) std = torch.zeros(3) print('==> Computing mean and std..') for inputs, targets in dataloader: for i in range(3): mean[i] += inputs[:,i,:,:].mean() std[i] += inputs[:,i,:,:].std() mean.div_(len(dataset)) std.div_(len(dataset)) return mean, std def init_params(net): '''Init layer parameters.''' for m in net.modules(): if isinstance(m, nn.Conv2d): init.kaiming_normal_(m.weight, mode='fan_out') if m.bias: init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d): init.constant_(m.weight, 1) init.constant_(m.bias, 0) elif isinstance(m, nn.Linear): init.normal_(m.weight, std=1e-3) if m.bias is not None: init.constant_(m.bias, 0) _, term_width = os.popen('stty size', 'r').read().split() term_width = int(term_width) TOTAL_BAR_LENGTH = 65. last_time = time.time() begin_time = last_time def progress_bar(current, total, msg=None): global last_time, begin_time if current == 0: begin_time = time.time() # Reset for new bar. cur_len = int(TOTAL_BAR_LENGTH*current/total) rest_len = int(TOTAL_BAR_LENGTH - cur_len) - 1 sys.stdout.write(' [') for i in range(cur_len): sys.stdout.write('=') sys.stdout.write('>') for i in range(rest_len): sys.stdout.write('.') sys.stdout.write(']') cur_time = time.time() step_time = cur_time - last_time last_time = cur_time tot_time = cur_time - begin_time L = [] L.append(' Step: %s' % format_time(step_time)) L.append(' | Tot: %s' % format_time(tot_time)) if msg: L.append(' | ' + msg) msg = ''.join(L) sys.stdout.write(msg) for i in range(term_width-int(TOTAL_BAR_LENGTH)-len(msg)-3): sys.stdout.write(' ') # Go back to the center of the bar. for i in range(term_width-int(TOTAL_BAR_LENGTH/2)+2): sys.stdout.write('\b') sys.stdout.write(' %d/%d ' % (current+1, total)) if current < total-1: sys.stdout.write('\r') else: sys.stdout.write('\n') sys.stdout.flush() def format_time(seconds): days = int(seconds / 3600/24) seconds = seconds - days*3600*24 hours = int(seconds / 3600) seconds = seconds - hours*3600 minutes = int(seconds / 60) seconds = seconds - minutes*60 secondsf = int(seconds) seconds = seconds - secondsf millis = int(seconds*1000) f = '' i = 1 if days > 0: f += str(days) + 'D' i += 1 if hours > 0 and i <= 2: f += str(hours) + 'h' i += 1 if minutes > 0 and i <= 2: f += str(minutes) + 'm' i += 1 if secondsf > 0 and i <= 2: f += str(secondsf) + 's' i += 1 if millis > 0 and i <= 2: f += str(millis) + 'ms' i += 1 if f == '': f = '0ms' return f
4,102
26.172185
96
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/eval_cumul_acc.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle from scipy.spatial.distance import cdist import utils_pytorch from utils_imagenet.utils_dataset import split_images_labels from utils_imagenet.utils_dataset import merge_images_labels from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--datadir', default='data/seed_1993_subset_100_imagenet/data', type=str) parser.add_argument('--num_classes', default=100, type=int) parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--ckp_prefix', \ default='checkpoint/class_incremental_imagenet_nb_cl_fg_50_nb_cl_10_nb_protos_200_run_0_', \ type=str) parser.add_argument('--order', \ default='./checkpoint/seed_1993_subset_100_imagenet_order_run_0.pkl', \ type=str) parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') args = parser.parse_args() print(args) order = utils_pytorch.unpickle(args.order) order_list = list(order) # transform_test = transforms.Compose([ # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), # ]) # evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) # input_data = evalset.test_data # input_labels = evalset.test_labels # map_input_labels = np.array([order_list.index(i) for i in input_labels]) valdir = os.path.join(args.datadir, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) evalset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) input_data, input_labels = split_images_labels(evalset.imgs) map_input_labels = np.array([order_list.index(i) for i in input_labels]) # evalset.test_labels = map_input_labels # evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, # shuffle=False, num_workers=2) cnn_cumul_acc = [] icarl_cumul_acc = [] ncm_cumul_acc = [] num_classes = [] nb_cl = args.nb_cl start_iter = int(args.nb_cl_fg/nb_cl)-1 for iteration in range(start_iter, int(args.num_classes/nb_cl)): # print("###########################################################") # print("For iteration {}".format(iteration)) # print("###########################################################") ckp_name = '{}iteration_{}_model.pth'.format(args.ckp_prefix, iteration) class_means_name = '{}iteration_{}_class_means.pth'.format(args.ckp_prefix, iteration) if not os.path.exists(ckp_name): break tg_model = torch.load(ckp_name) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) class_means = torch.load(class_means_name) current_means = class_means[:, order[:(iteration+1)*nb_cl]] indices = np.array([i in range(0, (iteration+1)*nb_cl) for i in map_input_labels]) # evalset.test_data = input_data[indices] # evalset.test_labels = map_input_labels[indices] # print('Max and Min of valid labels: {}, {}'.format(min(evalset.test_labels), max(evalset.test_labels))) current_eval_set = merge_images_labels(input_data[indices], map_input_labels[indices]) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, shuffle=False, num_workers=8, pin_memory=True) print("###########################################################") acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader, print_info=True) print("###########################################################") cnn_cumul_acc.append(acc[0]) icarl_cumul_acc.append(acc[1]) ncm_cumul_acc.append(acc[2]) num_classes.append((iteration+1)*nb_cl) print("###########################################################") print(' CNN acc: \t iCaRL acc \t NCM acc') print("###########################################################") for i in range(len(cnn_cumul_acc)): print("{:.2f} ".format(cnn_cumul_acc[i]), end='') print("[{:.2f}] ".format(np.mean(cnn_cumul_acc[-1])), end='') print("[{:.2f}] ".format(np.mean(cnn_cumul_acc)), end='') print("[{:.2f}] ".format(np.sum(np.array(cnn_cumul_acc)*np.array(num_classes)) / np.sum(num_classes)), end='') print("") for i in range(len(icarl_cumul_acc)): print("{:.2f} ".format(icarl_cumul_acc[i]), end='') print("[{:.2f}] ".format(np.mean(icarl_cumul_acc[-1])), end='') print("[{:.2f}] ".format(np.mean(icarl_cumul_acc)), end='') print("[{:.2f}] ".format(np.sum(np.array(icarl_cumul_acc)*np.array(num_classes)) / np.sum(num_classes)), end='') print("") for i in range(len(cnn_cumul_acc)): print("{:.2f} ".format(ncm_cumul_acc[i]), end='') print("[{:.2f}] ".format(np.mean(ncm_cumul_acc[-1])), end='') print("[{:.2f}] ".format(np.mean(ncm_cumul_acc)), end='') print("[{:.2f}] ".format(np.sum(np.array(ncm_cumul_acc)*np.array(num_classes)) / np.sum(num_classes)), end='') print("") print("###########################################################") print("") print('Number of classes', num_classes) print("###########################################################") print("Final acc on all classes") print("CNN:{:.2f}\t iCaRL:{:.2f}\t NCM:{:.2f}".format(cnn_cumul_acc[-1], icarl_cumul_acc[-1], ncm_cumul_acc[-1])) print("###########################################################") print("Average acc in each phase") print("CNN:{:.2f}\t iCaRL:{:.2f}\t NCM:{:.2f}".format(np.mean(cnn_cumul_acc), np.mean(icarl_cumul_acc), np.mean(ncm_cumul_acc))) print("###########################################################") print("Weighted average acc in each phase") print("CNN:{:.2f}\t iCaRL:{:.2f}\t NCM:{:.2f}".format( np.sum(np.array(cnn_cumul_acc)*np.array(num_classes)) / np.sum(num_classes), np.sum(np.array(icarl_cumul_acc)*np.array(num_classes)) / np.sum(num_classes), np.sum(np.array(ncm_cumul_acc)*np.array(num_classes)) / np.sum(num_classes) )) print("###########################################################")
6,830
45.469388
128
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/class_incremental_imagenet.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle import utils_pytorch from utils_imagenet.utils_dataset import split_images_labels from utils_imagenet.utils_dataset import merge_images_labels from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix from utils_incremental.incremental_train_and_eval import incremental_train_and_eval ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--dataset', default='seed_1993_subset_100_imagenet', type=str) parser.add_argument('--datadir', default='data/seed_1993_subset_100_imagenet/data', type=str) parser.add_argument('--num_classes', default=100, type=int) parser.add_argument('--num_workers', default=16, type=int, \ help='the number of workers for loading data') parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--nb_protos', default=20, type=int, \ help='Number of prototypes per class at the end') parser.add_argument('--nb_runs', default=1, type=int, \ help='Number of runs (random ordering of classes at each run)') parser.add_argument('--ckp_prefix', default=os.path.basename(sys.argv[0])[:-3], type=str, \ help='Checkpoint prefix') parser.add_argument('--epochs', default=90, type=int, \ help='Epochs') parser.add_argument('--T', default=2, type=float, \ help='Temporature for distialltion') parser.add_argument('--beta', default=0.25, type=float, \ help='Beta for distialltion') parser.add_argument('--resume', action='store_true', \ help='resume from checkpoint') parser.add_argument('--fix_budget', action='store_true', \ help='fix budget') parser.add_argument('--rs_ratio', default=0, type=float, \ help='The ratio for resample') parser.add_argument('--random_seed', default=1993, type=int, \ help='random seed') args = parser.parse_args() ######################################## assert(args.nb_cl_fg % args.nb_cl == 0) assert(args.nb_cl_fg >= args.nb_cl) train_batch_size = 128 # Batch size for train test_batch_size = 50 # Batch size for test eval_batch_size = 128 # Batch size for eval base_lr = 0.1 # Initial learning rate lr_strat = [30, 60] # Epochs where learning rate gets decreased lr_factor = 0.1 # Learning rate decrease factor custom_weight_decay = 1e-4 # Weight Decay custom_momentum = 0.9 # Momentum args.ckp_prefix = '{}_nb_cl_fg_{}_nb_cl_{}_nb_protos_{}'.format(args.ckp_prefix, args.nb_cl_fg, args.nb_cl, args.nb_protos) np.random.seed(args.random_seed) # Fix the random seed print(args) ######################################## device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #transform_train = transforms.Compose([ # transforms.RandomCrop(32, padding=4), # transforms.RandomHorizontalFlip(), # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), #]) #transform_test = transforms.Compose([ # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), #]) #trainset = torchvision.datasets.CIFAR100(root='./data', train=True, # download=True, transform=transform_train) #testset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=True, transform=transform_test) #evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) # Data loading code traindir = os.path.join(args.datadir, 'train') valdir = os.path.join(args.datadir, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) trainset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) testset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) evalset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) # Initialization dictionary_size = 1500 top1_acc_list_cumul = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) top1_acc_list_ori = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) #X_train_total = np.array(trainset.train_data) #Y_train_total = np.array(trainset.train_labels) #X_valid_total = np.array(testset.test_data) #Y_valid_total = np.array(testset.test_labels) X_train_total, Y_train_total = split_images_labels(trainset.imgs) X_valid_total, Y_valid_total = split_images_labels(testset.imgs) # Launch the different runs for iteration_total in range(args.nb_runs): # Select the order for the class learning order_name = "./checkpoint/seed_{}_{}_order_run_{}.pkl".format(args.random_seed, args.dataset, iteration_total) print("Order name:{}".format(order_name)) if os.path.exists(order_name): print("Loading orders") order = utils_pytorch.unpickle(order_name) else: print("Generating orders") order = np.arange(args.num_classes) np.random.shuffle(order) utils_pytorch.savepickle(order, order_name) order_list = list(order) print(order_list) # Initialization of the variables for this run X_valid_cumuls = [] X_protoset_cumuls = [] X_train_cumuls = [] Y_valid_cumuls = [] Y_protoset_cumuls = [] Y_train_cumuls = [] alpha_dr_herding = np.zeros((int(args.num_classes/args.nb_cl),dictionary_size,args.nb_cl),np.float32) # The following contains all the training samples of the different classes # because we want to compare our method with the theoretical case where all the training samples are stored # prototypes = np.zeros((args.num_classes,dictionary_size,X_train_total.shape[1],X_train_total.shape[2],X_train_total.shape[3])) prototypes = [[] for i in range(args.num_classes)] for orde in range(args.num_classes): prototypes[orde] = X_train_total[np.where(Y_train_total==order[orde])] prototypes = np.array(prototypes) start_iter = int(args.nb_cl_fg/args.nb_cl)-1 for iteration in range(start_iter, int(args.num_classes/args.nb_cl)): #init model if iteration == start_iter: ############################################################ last_iter = 0 ############################################################ tg_model = models.resnet18(num_classes=args.nb_cl_fg) ref_model = None else: ############################################################ last_iter = iteration ############################################################ #increment classes ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features new_fc = nn.Linear(in_features, out_features+args.nb_cl) new_fc.weight.data[:out_features] = tg_model.fc.weight.data new_fc.bias.data[:out_features] = tg_model.fc.bias.data tg_model.fc = new_fc # Prepare the training data for the current batch of classes actual_cl = order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] indices_train_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_train_total]) indices_test_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_valid_total]) X_train = X_train_total[indices_train_10] X_valid = X_valid_total[indices_test_10] X_valid_cumuls.append(X_valid) X_train_cumuls.append(X_train) X_valid_cumul = np.concatenate(X_valid_cumuls) X_train_cumul = np.concatenate(X_train_cumuls) Y_train = Y_train_total[indices_train_10] Y_valid = Y_valid_total[indices_test_10] Y_valid_cumuls.append(Y_valid) Y_train_cumuls.append(Y_train) Y_valid_cumul = np.concatenate(Y_valid_cumuls) Y_train_cumul = np.concatenate(Y_train_cumuls) # Add the stored exemplars to the training data if iteration == start_iter: X_valid_ori = X_valid Y_valid_ori = Y_valid else: X_protoset = np.concatenate(X_protoset_cumuls) Y_protoset = np.concatenate(Y_protoset_cumuls) if args.rs_ratio > 0: #1/rs_ratio = (len(X_train)+len(X_protoset)*scale_factor)/(len(X_protoset)*scale_factor) scale_factor = (len(X_train) * args.rs_ratio) / (len(X_protoset) * (1 - args.rs_ratio)) rs_sample_weights = np.concatenate((np.ones(len(X_train)), np.ones(len(X_protoset))*scale_factor)) #number of samples per epoch #rs_num_samples = len(X_train) + len(X_protoset) rs_num_samples = int(len(X_train) / (1 - args.rs_ratio)) print("X_train:{}, X_protoset:{}, rs_num_samples:{}".format(len(X_train), len(X_protoset), rs_num_samples)) X_train = np.concatenate((X_train,X_protoset),axis=0) Y_train = np.concatenate((Y_train,Y_protoset)) # Launch the training loop print('Batch of classes number {0} arrives ...'.format(iteration+1)) map_Y_train = np.array([order_list.index(i) for i in Y_train]) map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) ############################################################ #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = map_Y_train current_train_imgs = merge_images_labels(X_train, map_Y_train) trainset.imgs = trainset.samples = current_train_imgs if iteration > start_iter and args.rs_ratio > 0 and scale_factor > 1: print("Weights from sampling:", rs_sample_weights) index1 = np.where(rs_sample_weights>1)[0] index2 = np.where(map_Y_train<iteration*args.nb_cl)[0] assert((index1==index2).all()) train_sampler = torch.utils.data.sampler.WeightedRandomSampler(rs_sample_weights, rs_num_samples) #trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ # shuffle=False, sampler=train_sampler, num_workers=2) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ shuffle=False, sampler=train_sampler, num_workers=args.num_workers, pin_memory=True) else: #trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, # shuffle=True, num_workers=2) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) #testset.test_data = X_valid_cumul.astype('uint8') #testset.test_labels = map_Y_valid_cumul current_test_images = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) testset.imgs = testset.samples = current_test_images testloader = torch.utils.data.DataLoader(testset, batch_size=test_batch_size, shuffle=False, num_workers=2) print('Max and Min of train labels: {}, {}'.format(min(map_Y_train), max(map_Y_train))) print('Max and Min of valid labels: {}, {}'.format(min(map_Y_valid_cumul), max(map_Y_valid_cumul))) ############################################################## ckp_name = './checkpoint/{}_run_{}_iteration_{}_model.pth'.format(args.ckp_prefix, iteration_total, iteration) print('ckp_name', ckp_name) if args.resume and os.path.exists(ckp_name): print("###############################") print("Loading models from checkpoint") tg_model = torch.load(ckp_name) print("###############################") else: tg_params = tg_model.parameters() tg_model = tg_model.to(device) if iteration > start_iter: ref_model = ref_model.to(device) tg_optimizer = optim.SGD(tg_params, lr=base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_lr_scheduler = lr_scheduler.MultiStepLR(tg_optimizer, milestones=lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ args.T, args.beta) if not os.path.isdir('checkpoint'): os.mkdir('checkpoint') torch.save(tg_model, ckp_name) ### Exemplars if args.fix_budget: nb_protos_cl = int(np.ceil(args.nb_protos*args.num_classes*1.0/args.nb_cl/(iteration+1))) else: nb_protos_cl = args.nb_protos tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features # Herding print('Updating exemplar set...') for iter_dico in range(last_iter*args.nb_cl, (iteration+1)*args.nb_cl): # Possible exemplars in the feature space and projected on the L2 sphere # evalset.test_data = prototypes[iter_dico].astype('uint8') # evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(prototypes[iter_dico], np.zeros(len(prototypes[iter_dico]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) num_samples = len(prototypes[iter_dico]) mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Herding procedure : ranking of the potential exemplars mu = np.mean(D,axis=1) index1 = int(iter_dico/args.nb_cl) index2 = iter_dico % args.nb_cl alpha_dr_herding[index1,:,index2] = alpha_dr_herding[index1,:,index2]*0 w_t = mu iter_herding = 0 iter_herding_eff = 0 while not(np.sum(alpha_dr_herding[index1,:,index2]!=0)==min(nb_protos_cl,500)) and iter_herding_eff<1000: tmp_t = np.dot(w_t,D) ind_max = np.argmax(tmp_t) iter_herding_eff += 1 if alpha_dr_herding[index1,ind_max,index2] == 0: alpha_dr_herding[index1,ind_max,index2] = 1+iter_herding iter_herding += 1 w_t = w_t+mu-D[:,ind_max] # Prepare the protoset X_protoset_cumuls = [] Y_protoset_cumuls = [] # Class means for iCaRL and NCM + Storing the selected exemplars in the protoset print('Computing mean-of_exemplars and theoretical mean...') # class_means = np.zeros((64,100,2)) class_means = np.zeros((num_features, args.num_classes, 2)) for iteration2 in range(iteration+1): for iter_dico in range(args.nb_cl): current_cl = order[range(iteration2*args.nb_cl,(iteration2+1)*args.nb_cl)] # Collect data in the feature space for each class # evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico].astype('uint8') # evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(prototypes[iteration2*args.nb_cl+iter_dico], \ np.zeros(len(prototypes[iteration2*args.nb_cl+iter_dico]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) num_samples = len(prototypes[iteration2*args.nb_cl+iter_dico]) mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Flipped version also # evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico][:,:,:,::-1].astype('uint8') # evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, # shuffle=False, num_workers=2) # mapped_prototypes2 = compute_features(tg_feature_model, evalloader, num_samples, num_features) # D2 = mapped_prototypes2.T # D2 = D2/np.linalg.norm(D2,axis=0) D2 = D # iCaRL alph = alpha_dr_herding[iteration2,:,iter_dico] assert((alph[num_samples:]==0).all()) alph = alph[:num_samples] alph = (alph>0)*(alph<nb_protos_cl+1)*1. # X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico,np.where(alph==1)[0]]) X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico][np.where(alph==1)[0]]) Y_protoset_cumuls.append(order[iteration2*args.nb_cl+iter_dico]*np.ones(len(np.where(alph==1)[0]))) alph = alph/np.sum(alph) class_means[:,current_cl[iter_dico],0] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],0] /= np.linalg.norm(class_means[:,current_cl[iter_dico],0]) # Normal NCM # alph = np.ones(dictionary_size)/dictionary_size alph = np.ones(num_samples)/num_samples class_means[:,current_cl[iter_dico],1] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],1] /= np.linalg.norm(class_means[:,current_cl[iter_dico],1]) torch.save(class_means, \ './checkpoint/{}_run_{}_iteration_{}_class_means.pth'.format(args.ckp_prefix,iteration_total, iteration)) current_means = class_means[:, order[range(0,(iteration+1)*args.nb_cl)]] ############################################################## # Calculate validation error of model on the first nb_cl classes: map_Y_valid_ori = np.array([order_list.index(i) for i in Y_valid_ori]) print('Computing accuracy on the original batch of classes...') # evalset.test_data = X_valid_ori.astype('uint8') # evalset.test_labels = map_Y_valid_ori current_eval_set = merge_images_labels(X_valid_ori, map_Y_valid_ori) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) ori_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_ori[iteration, :, iteration_total] = np.array(ori_acc).T ############################################################## # Calculate validation error of model on the cumul of classes: map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) print('Computing cumulative accuracy...') # evalset.test_data = X_valid_cumul.astype('uint8') # evalset.test_labels = map_Y_valid_cumul current_eval_set = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) cumul_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_cumul[iteration, :, iteration_total] = np.array(cumul_acc).T ############################################################## # Calculate confusion matrix print('Computing confusion matrix...') cm = compute_confusion_matrix(tg_model, tg_feature_model, current_means, evalloader) cm_name = './checkpoint/{}_run_{}_iteration_{}_confusion_matrix.pth'.format(args.ckp_prefix,iteration_total, iteration) with open(cm_name, 'wb') as f: pickle.dump(cm, f, 2) #for reading with Python 2 ############################################################## # Final save of the data torch.save(top1_acc_list_ori, \ './checkpoint/{}_run_{}_top1_acc_list_ori.pth'.format(args.ckp_prefix, iteration_total)) torch.save(top1_acc_list_cumul, \ './checkpoint/{}_run_{}_top1_acc_list_cumul.pth'.format(args.ckp_prefix, iteration_total))
22,015
52.307506
132
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/modified_resnet.py
import torch.nn as nn import math import torch.utils.model_zoo as model_zoo import modified_linear def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None, last=False): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride self.last = last def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual if not self.last: #remove ReLU in the last layer out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers, num_classes=1000): self.inplanes = 64 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2, last_phase=True) self.avgpool = nn.AvgPool2d(7, stride=1) self.fc = modified_linear.CosineLinear(512 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) def _make_layer(self, block, planes, blocks, stride=1, last_phase=False): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion if last_phase: for i in range(1, blocks-1): layers.append(block(self.inplanes, planes)) layers.append(block(self.inplanes, planes, last=True)) else: for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x def resnet18(pretrained=False, **kwargs): """Constructs a ResNet-18 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs) return model
3,850
32.198276
88
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/cbf_class_incremental_cosine_imagenet.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle import math import modified_resnet import modified_linear import utils_pytorch from utils_imagenet.utils_dataset import split_images_labels from utils_imagenet.utils_dataset import merge_images_labels from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix from utils_incremental.incremental_train_and_eval import incremental_train_and_eval from utils_incremental.incremental_train_and_eval_MS import incremental_train_and_eval_MS from utils_incremental.incremental_train_and_eval_LF import incremental_train_and_eval_LF from utils_incremental.incremental_train_and_eval_MR_LF import incremental_train_and_eval_MR_LF from utils_incremental.incremental_train_and_eval_AMR_LF import incremental_train_and_eval_AMR_LF ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--dataset', default='seed_1993_subset_100_imagenet', type=str) parser.add_argument('--datadir', default='data/seed_1993_subset_100_imagenet/data', type=str) parser.add_argument('--num_classes', default=100, type=int) parser.add_argument('--num_workers', default=16, type=int, \ help='the number of workers for loading data') parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--nb_protos', default=20, type=int, \ help='Number of prototypes per class at the end') parser.add_argument('--nb_runs', default=1, type=int, \ help='Number of runs (random ordering of classes at each run)') parser.add_argument('--ckp_prefix', default=os.path.basename(sys.argv[0])[:-3], type=str, \ help='Checkpoint prefix') parser.add_argument('--epochs', default=90, type=int, \ help='Epochs') parser.add_argument('--T', default=2, type=float, \ help='Temporature for distialltion') parser.add_argument('--beta', default=0.25, type=float, \ help='Beta for distialltion') parser.add_argument('--resume', action='store_true', \ help='resume from checkpoint') parser.add_argument('--fix_budget', action='store_true', \ help='fix budget') ######################################## parser.add_argument('--mimic_score', action='store_true', \ help='To mimic scores for cosine embedding') parser.add_argument('--lw_ms', default=1, type=float, \ help='loss weight for mimicking score') ######################################## #improved class incremental learning parser.add_argument('--rs_ratio', default=0, type=float, \ help='The ratio for resample') parser.add_argument('--imprint_weights', action='store_true', \ help='Imprint the weights for novel classes') parser.add_argument('--less_forget', action='store_true', \ help='Less forgetful') parser.add_argument('--lamda', default=5, type=float, \ help='Lamda for LF') parser.add_argument('--adapt_lamda', action='store_true', \ help='Adaptively change lamda') parser.add_argument('--mr_loss', action='store_true', \ help='Margin ranking loss v1') parser.add_argument('--amr_loss', action='store_true', \ help='Margin ranking loss v2') parser.add_argument('--dist', default=0.5, type=float, \ help='Dist for MarginRankingLoss') parser.add_argument('--K', default=2, type=int, \ help='K for MarginRankingLoss') parser.add_argument('--lw_mr', default=1, type=float, \ help='loss weight for margin ranking loss') ######################################## parser.add_argument('--random_seed', default=1993, type=int, \ help='random seed') ######################################## parser.add_argument('--cb_finetune', action='store_true', \ help='class balance finetune') parser.add_argument('--ft_epochs', default=20, type=int, \ help='Epochs for class balance finetune') parser.add_argument('--ft_base_lr', default=0.01, type=float, \ help='Base learning rate for class balance finetune') parser.add_argument('--ft_lr_strat', default=[10], type=int, nargs='+', \ help='Lr_strat for class balance finetune') parser.add_argument('--ft_flag', default=2, type=int, \ help='Flag for class balance finetune') args = parser.parse_args() ######################################## assert(args.nb_cl_fg % args.nb_cl == 0) assert(args.nb_cl_fg >= args.nb_cl) train_batch_size = 128 # Batch size for train test_batch_size = 50 # Batch size for test eval_batch_size = 128 # Batch size for eval base_lr = 0.1 # Initial learning rate lr_strat = [30, 60] # Epochs where learning rate gets decreased lr_factor = 0.1 # Learning rate decrease factor custom_weight_decay = 1e-4 # Weight Decay custom_momentum = 0.9 # Momentum args.ckp_prefix = '{}_nb_cl_fg_{}_nb_cl_{}_nb_protos_{}'.format(args.ckp_prefix, args.nb_cl_fg, args.nb_cl, args.nb_protos) np.random.seed(args.random_seed) # Fix the random seed print(args) ######################################## device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #transform_train = transforms.Compose([ # transforms.RandomCrop(32, padding=4), # transforms.RandomHorizontalFlip(), # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), #]) #transform_test = transforms.Compose([ # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), #]) #trainset = torchvision.datasets.CIFAR100(root='./data', train=True, # download=True, transform=transform_train) #testset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=True, transform=transform_test) #evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) # Data loading code traindir = os.path.join(args.datadir, 'train') valdir = os.path.join(args.datadir, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) trainset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) testset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) evalset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) # Initialization dictionary_size = 1500 top1_acc_list_cumul = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) top1_acc_list_ori = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) #X_train_total = np.array(trainset.train_data) #Y_train_total = np.array(trainset.train_labels) #X_valid_total = np.array(testset.test_data) #Y_valid_total = np.array(testset.test_labels) X_train_total, Y_train_total = split_images_labels(trainset.imgs) X_valid_total, Y_valid_total = split_images_labels(testset.imgs) # Launch the different runs for iteration_total in range(args.nb_runs): # Select the order for the class learning order_name = "./checkpoint/seed_{}_{}_order_run_{}.pkl".format(args.random_seed, args.dataset, iteration_total) print("Order name:{}".format(order_name)) if os.path.exists(order_name): print("Loading orders") order = utils_pytorch.unpickle(order_name) else: print("Generating orders") order = np.arange(args.num_classes) np.random.shuffle(order) utils_pytorch.savepickle(order, order_name) order_list = list(order) print(order_list) # Initialization of the variables for this run X_valid_cumuls = [] X_protoset_cumuls = [] X_train_cumuls = [] Y_valid_cumuls = [] Y_protoset_cumuls = [] Y_train_cumuls = [] alpha_dr_herding = np.zeros((int(args.num_classes/args.nb_cl),dictionary_size,args.nb_cl),np.float32) # The following contains all the training samples of the different classes # because we want to compare our method with the theoretical case where all the training samples are stored # prototypes = np.zeros((args.num_classes,dictionary_size,X_train_total.shape[1],X_train_total.shape[2],X_train_total.shape[3])) prototypes = [[] for i in range(args.num_classes)] for orde in range(args.num_classes): prototypes[orde] = X_train_total[np.where(Y_train_total==order[orde])] prototypes = np.array(prototypes) start_iter = int(args.nb_cl_fg/args.nb_cl)-1 for iteration in range(start_iter, int(args.num_classes/args.nb_cl)): #init model if iteration == start_iter: ############################################################ last_iter = 0 ############################################################ tg_model = modified_resnet.resnet18(num_classes=args.nb_cl_fg) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features print("in_features:", in_features, "out_features:", out_features) ref_model = None elif iteration == start_iter+1: ############################################################ last_iter = iteration ############################################################ #increment classes ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features print("in_features:", in_features, "out_features:", out_features) new_fc = modified_linear.SplitCosineLinear(in_features, out_features, args.nb_cl) new_fc.fc1.weight.data = tg_model.fc.weight.data new_fc.sigma.data = tg_model.fc.sigma.data tg_model.fc = new_fc lamda_mult = out_features*1.0 / args.nb_cl else: ############################################################ last_iter = iteration ############################################################ ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features1 = tg_model.fc.fc1.out_features out_features2 = tg_model.fc.fc2.out_features print("in_features:", in_features, "out_features1:", \ out_features1, "out_features2:", out_features2) new_fc = modified_linear.SplitCosineLinear(in_features, out_features1+out_features2, args.nb_cl) new_fc.fc1.weight.data[:out_features1] = tg_model.fc.fc1.weight.data new_fc.fc1.weight.data[out_features1:] = tg_model.fc.fc2.weight.data new_fc.sigma.data = tg_model.fc.sigma.data tg_model.fc = new_fc lamda_mult = (out_features1+out_features2)*1.0 / (args.nb_cl) if iteration > start_iter and args.less_forget and args.adapt_lamda: #cur_lamda = lamda_base * sqrt(num_old_classes/num_new_classes) cur_lamda = args.lamda * math.sqrt(lamda_mult) else: cur_lamda = args.lamda if iteration > start_iter and args.less_forget: print("###############################") print("Lamda for less forget is set to ", cur_lamda) print("###############################") # Prepare the training data for the current batch of classes actual_cl = order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] indices_train_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_train_total]) indices_test_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_valid_total]) X_train = X_train_total[indices_train_10] X_valid = X_valid_total[indices_test_10] X_valid_cumuls.append(X_valid) X_train_cumuls.append(X_train) X_valid_cumul = np.concatenate(X_valid_cumuls) X_train_cumul = np.concatenate(X_train_cumuls) Y_train = Y_train_total[indices_train_10] Y_valid = Y_valid_total[indices_test_10] Y_valid_cumuls.append(Y_valid) Y_train_cumuls.append(Y_train) Y_valid_cumul = np.concatenate(Y_valid_cumuls) Y_train_cumul = np.concatenate(Y_train_cumuls) # Add the stored exemplars to the training data if iteration == start_iter: X_valid_ori = X_valid Y_valid_ori = Y_valid else: X_protoset = np.concatenate(X_protoset_cumuls) Y_protoset = np.concatenate(Y_protoset_cumuls) if args.rs_ratio > 0: #1/rs_ratio = (len(X_train)+len(X_protoset)*scale_factor)/(len(X_protoset)*scale_factor) scale_factor = (len(X_train) * args.rs_ratio) / (len(X_protoset) * (1 - args.rs_ratio)) rs_sample_weights = np.concatenate((np.ones(len(X_train)), np.ones(len(X_protoset))*scale_factor)) #number of samples per epoch #rs_num_samples = len(X_train) + len(X_protoset) rs_num_samples = int(len(X_train) / (1 - args.rs_ratio)) print("X_train:{}, X_protoset:{}, rs_num_samples:{}".format(len(X_train), len(X_protoset), rs_num_samples)) X_train = np.concatenate((X_train,X_protoset),axis=0) Y_train = np.concatenate((Y_train,Y_protoset)) # Launch the training loop print('Batch of classes number {0} arrives ...'.format(iteration+1)) map_Y_train = np.array([order_list.index(i) for i in Y_train]) map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) #imprint weights if iteration > start_iter and args.imprint_weights: #input: tg_model, X_train, map_Y_train #class_start = iteration*nb_cl class_end = (iteration+1)*nb_cl print("Imprint weights") ######################################### #compute the average norm of old embdding old_embedding_norm = tg_model.fc.fc1.weight.data.norm(dim=1, keepdim=True) average_old_embedding_norm = torch.mean(old_embedding_norm, dim=0).to('cpu').type(torch.DoubleTensor) ######################################### tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features novel_embedding = torch.zeros((args.nb_cl, num_features)) for cls_idx in range(iteration*args.nb_cl, (iteration+1)*args.nb_cl): cls_indices = np.array([i == cls_idx for i in map_Y_train]) assert(len(np.where(cls_indices==1)[0])<=dictionary_size) #evalset.test_data = X_train[cls_indices].astype('uint8') #evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(X_train[cls_indices], np.zeros(len(X_train[cls_indices]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = len(X_train[cls_indices]) cls_features = compute_features(tg_feature_model, evalloader, num_samples, num_features) #cls_features = cls_features.T #cls_features = cls_features / np.linalg.norm(cls_features,axis=0) #cls_embedding = np.mean(cls_features, axis=1) norm_features = F.normalize(torch.from_numpy(cls_features), p=2, dim=1) cls_embedding = torch.mean(norm_features, dim=0) #novel_embedding[cls_idx-iteration*args.nb_cl] = cls_embedding novel_embedding[cls_idx-iteration*args.nb_cl] = F.normalize(cls_embedding, p=2, dim=0) * average_old_embedding_norm tg_model.to(device) #torch.save(tg_model, "tg_model_before_imprint_weights.pth") tg_model.fc.fc2.weight.data = novel_embedding.to(device) #torch.save(tg_model, "tg_model_after_imprint_weights.pth") ############################################################ #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = map_Y_train current_train_imgs = merge_images_labels(X_train, map_Y_train) trainset.imgs = trainset.samples = current_train_imgs if iteration > start_iter and args.rs_ratio > 0 and scale_factor > 1: print("Weights from sampling:", rs_sample_weights) index1 = np.where(rs_sample_weights>1)[0] index2 = np.where(map_Y_train<iteration*args.nb_cl)[0] assert((index1==index2).all()) train_sampler = torch.utils.data.sampler.WeightedRandomSampler(rs_sample_weights, rs_num_samples) #trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ # shuffle=False, sampler=train_sampler, num_workers=2) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ shuffle=False, sampler=train_sampler, num_workers=args.num_workers, pin_memory=True) else: #trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, # shuffle=True, num_workers=2) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) #testset.test_data = X_valid_cumul.astype('uint8') #testset.test_labels = map_Y_valid_cumul current_test_imgs = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) testset.imgs = testset.samples = current_test_imgs testloader = torch.utils.data.DataLoader(testset, batch_size=test_batch_size, shuffle=False, num_workers=2) print('Max and Min of train labels: {}, {}'.format(min(map_Y_train), max(map_Y_train))) print('Max and Min of valid labels: {}, {}'.format(min(map_Y_valid_cumul), max(map_Y_valid_cumul))) ############################################################## ckp_name = './checkpoint/{}_run_{}_iteration_{}_model.pth'.format(args.ckp_prefix, iteration_total, iteration) print('ckp_name', ckp_name) if args.resume and os.path.exists(ckp_name): print("###############################") print("Loading models from checkpoint") tg_model = torch.load(ckp_name) print("###############################") else: ############################### if iteration > start_iter and args.less_forget: #fix the embedding of old classes ignored_params = list(map(id, tg_model.fc.fc1.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, \ tg_model.parameters()) tg_params =[{'params': base_params, 'lr': base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': 0, 'weight_decay': 0}] else: tg_params = tg_model.parameters() ############################### tg_model = tg_model.to(device) if iteration > start_iter: ref_model = ref_model.to(device) tg_optimizer = optim.SGD(tg_params, lr=base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_lr_scheduler = lr_scheduler.MultiStepLR(tg_optimizer, milestones=lr_strat, gamma=lr_factor) ############################### if args.less_forget and args.mr_loss: print("incremental_train_and_eval_MR_LF") tg_model = incremental_train_and_eval_MR_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr) elif args.less_forget and args.amr_loss: print("incremental_train_and_eval_AMR_LF") tg_model = incremental_train_and_eval_AMR_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr) else: if args.less_forget: print("incremental_train_and_eval_LF") tg_model = incremental_train_and_eval_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda) else: if args.mimic_score: print("incremental_train_and_eval_MS") tg_model = incremental_train_and_eval_MS(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, args.lw_ms) else: print("incremental_train_and_eval") tg_model = incremental_train_and_eval(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, args.T, args.beta) if not os.path.isdir('checkpoint'): os.mkdir('checkpoint') torch.save(tg_model, ckp_name) ### Exemplars if args.fix_budget: nb_protos_cl = int(np.ceil(args.nb_protos*args.num_classes*1.0/args.nb_cl/(iteration+1))) else: nb_protos_cl = args.nb_protos tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features # Herding print('Updating exemplar set...') for iter_dico in range(last_iter*args.nb_cl, (iteration+1)*args.nb_cl): # Possible exemplars in the feature space and projected on the L2 sphere # evalset.test_data = prototypes[iter_dico].astype('uint8') # evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(prototypes[iter_dico], np.zeros(len(prototypes[iter_dico]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) num_samples = len(prototypes[iter_dico]) mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Herding procedure : ranking of the potential exemplars mu = np.mean(D,axis=1) index1 = int(iter_dico/args.nb_cl) index2 = iter_dico % args.nb_cl alpha_dr_herding[index1,:,index2] = alpha_dr_herding[index1,:,index2]*0 w_t = mu iter_herding = 0 iter_herding_eff = 0 while not(np.sum(alpha_dr_herding[index1,:,index2]!=0)==min(nb_protos_cl,500)) and iter_herding_eff<1000: tmp_t = np.dot(w_t,D) ind_max = np.argmax(tmp_t) iter_herding_eff += 1 if alpha_dr_herding[index1,ind_max,index2] == 0: alpha_dr_herding[index1,ind_max,index2] = 1+iter_herding iter_herding += 1 w_t = w_t+mu-D[:,ind_max] # Prepare the protoset X_protoset_cumuls = [] Y_protoset_cumuls = [] # Class means for iCaRL and NCM + Storing the selected exemplars in the protoset print('Computing mean-of_exemplars and theoretical mean...') # class_means = np.zeros((64,100,2)) class_means = np.zeros((num_features, args.num_classes, 2)) for iteration2 in range(iteration+1): for iter_dico in range(args.nb_cl): current_cl = order[range(iteration2*args.nb_cl,(iteration2+1)*args.nb_cl)] # Collect data in the feature space for each class # evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico].astype('uint8') # evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(prototypes[iteration2*args.nb_cl+iter_dico], \ np.zeros(len(prototypes[iteration2*args.nb_cl+iter_dico]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) num_samples = len(prototypes[iteration2*args.nb_cl+iter_dico]) mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Flipped version also # evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico][:,:,:,::-1].astype('uint8') # evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, # shuffle=False, num_workers=2) # mapped_prototypes2 = compute_features(tg_feature_model, evalloader, num_samples, num_features) # D2 = mapped_prototypes2.T # D2 = D2/np.linalg.norm(D2,axis=0) D2 = D # iCaRL alph = alpha_dr_herding[iteration2,:,iter_dico] assert((alph[num_samples:]==0).all()) alph = alph[:num_samples] alph = (alph>0)*(alph<nb_protos_cl+1)*1. # X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico,np.where(alph==1)[0]]) X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico][np.where(alph==1)[0]]) Y_protoset_cumuls.append(order[iteration2*args.nb_cl+iter_dico]*np.ones(len(np.where(alph==1)[0]))) alph = alph/np.sum(alph) class_means[:,current_cl[iter_dico],0] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],0] /= np.linalg.norm(class_means[:,current_cl[iter_dico],0]) # Normal NCM # alph = np.ones(dictionary_size)/dictionary_size alph = np.ones(num_samples)/num_samples class_means[:,current_cl[iter_dico],1] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],1] /= np.linalg.norm(class_means[:,current_cl[iter_dico],1]) # torch.save(class_means, \ # './checkpoint/{}_run_{}_iteration_{}_class_means.pth'.format(args.ckp_prefix,iteration_total, iteration)) class_means_name = './checkpoint/{}_run_{}_iteration_{}_class_means.pth'.format(args.ckp_prefix,iteration_total, iteration) torch.save(class_means, class_means_name) current_means = class_means[:, order[range(0,(iteration+1)*args.nb_cl)]] ############################################################## # Calculate validation error of model on the first nb_cl classes: map_Y_valid_ori = np.array([order_list.index(i) for i in Y_valid_ori]) print('Computing accuracy on the original batch of classes...') # evalset.test_data = X_valid_ori.astype('uint8') # evalset.test_labels = map_Y_valid_ori current_eval_set = merge_images_labels(X_valid_ori, map_Y_valid_ori) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) ori_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_ori[iteration, :, iteration_total] = np.array(ori_acc).T ############################################################## # Calculate validation error of model on the cumul of classes: map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) print('Computing cumulative accuracy...') # evalset.test_data = X_valid_cumul.astype('uint8') # evalset.test_labels = map_Y_valid_cumul current_eval_set = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) cumul_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_cumul[iteration, :, iteration_total] = np.array(cumul_acc).T ############################################################## # Calculate confusion matrix # print('Computing confusion matrix...') # cm = compute_confusion_matrix(tg_model, tg_feature_model, current_means, evalloader) # cm_name = './checkpoint/{}_run_{}_iteration_{}_confusion_matrix.pth'.format(args.ckp_prefix,iteration_total, iteration) # with open(cm_name, 'wb') as f: # pickle.dump(cm, f, 2) #for reading with Python 2 ############################################################## if iteration == start_iter and args.cb_finetune: #for the convenience of evaluation torch.save(tg_model, ckp_name.replace("/checkpoint/", "/checkpoint/AFTER_CBF_")) torch.save(class_means, class_means_name.replace("/checkpoint/", "/checkpoint/AFTER_CBF_")) if iteration > start_iter and args.cb_finetune: # Class balance finetuning on the protoset print("###############################") print("Class balance finetuning on the protoset") print("###############################") map_Y_protoset_cumuls = np.array([order_list.index(i) for i in np.concatenate(Y_protoset_cumuls)]) current_train_imgs = merge_images_labels(np.concatenate(X_protoset_cumuls), map_Y_protoset_cumuls) trainset.imgs = trainset.samples = current_train_imgs trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) print('Min and Max of train labels: {}, {}'.format(min(map_Y_protoset_cumuls), max(map_Y_protoset_cumuls))) ############################### print('Computing accuracy on the protoset...') current_eval_set = merge_images_labels(np.concatenate(X_protoset_cumuls), map_Y_protoset_cumuls) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) cbf_proto_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) ############################### print('Computing accuracy on the old protoset...') indices = np.array([i in range(0, iteration*args.nb_cl) for i in map_Y_protoset_cumuls]) current_eval_set = merge_images_labels(np.concatenate(X_protoset_cumuls)[indices], map_Y_protoset_cumuls[indices]) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) print('Min and Max of eval labels: {}, {}'.format(min(map_Y_protoset_cumuls[indices]), max(map_Y_protoset_cumuls[indices]))) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) cbf_proto_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) ############################### print('Computing accuracy on the new protoset...') indices = np.array([i in range(iteration*args.nb_cl, (iteration+1)*args.nb_cl) for i in map_Y_protoset_cumuls]) current_eval_set = merge_images_labels(np.concatenate(X_protoset_cumuls)[indices], map_Y_protoset_cumuls[indices]) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) print('Min and Max of eval labels: {}, {}'.format(min(map_Y_protoset_cumuls[indices]), max(map_Y_protoset_cumuls[indices]))) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) cbf_proto_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) ############################################################## #tg_params = tg_model.parameters() if args.ft_flag == 0: #everything is not updated ignored_params = list(map(id, tg_model.fc.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, tg_model.parameters()) tg_params =[{'params': base_params, 'lr': 0, 'weight_decay': 0}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': 0, 'weight_decay': 0}, \ {'params': tg_model.fc.fc2.parameters(), 'lr': 0, 'weight_decay': 0}] fix_bn_flag = True tg_model = tg_model.to(device) ref_model = ref_model.to(device) tg_ft_optimizer = optim.SGD(tg_params, lr=args.ft_base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_ft_lr_scheduler = lr_scheduler.MultiStepLR(tg_ft_optimizer, milestones=args.ft_lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval_MR_LF(args.ft_epochs, tg_model, ref_model, tg_ft_optimizer, tg_ft_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr, \ fix_bn=fix_bn_flag) elif args.ft_flag == 1: #only the novel embeddings are updated with the feature extractor fixed ignored_params = list(map(id, tg_model.fc.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, tg_model.parameters()) tg_params =[{'params': base_params, 'lr': 0, 'weight_decay': 0}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': 0, 'weight_decay': 0}, \ {'params': tg_model.fc.fc2.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}] fix_bn_flag = True tg_model = tg_model.to(device) ref_model = ref_model.to(device) tg_ft_optimizer = optim.SGD(tg_params, lr=args.ft_base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_ft_lr_scheduler = lr_scheduler.MultiStepLR(tg_ft_optimizer, milestones=args.ft_lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval_MR_LF(args.ft_epochs, tg_model, ref_model, tg_ft_optimizer, tg_ft_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr, \ fix_bn=fix_bn_flag) elif args.ft_flag == 2: #both the old and novel embeddings are updated with the feature extractor fixed ignored_params = list(map(id, tg_model.fc.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, tg_model.parameters()) tg_params =[{'params': base_params, 'lr': 0, 'weight_decay': 0}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc2.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}] fix_bn_flag = True tg_model = tg_model.to(device) ref_model = ref_model.to(device) tg_ft_optimizer = optim.SGD(tg_params, lr=args.ft_base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_ft_lr_scheduler = lr_scheduler.MultiStepLR(tg_ft_optimizer, milestones=args.ft_lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval_MR_LF(args.ft_epochs, tg_model, ref_model, tg_ft_optimizer, tg_ft_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr, \ fix_bn=fix_bn_flag) elif args.ft_flag == 3: #everything is updated ignored_params = list(map(id, tg_model.fc.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, tg_model.parameters()) tg_params =[{'params': base_params, 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc2.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}] fix_bn_flag = False tg_model = tg_model.to(device) ref_model = ref_model.to(device) tg_ft_optimizer = optim.SGD(tg_params, lr=args.ft_base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_ft_lr_scheduler = lr_scheduler.MultiStepLR(tg_ft_optimizer, milestones=args.ft_lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval_MR_LF(args.ft_epochs, tg_model, ref_model, tg_ft_optimizer, tg_ft_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr, \ fix_bn=fix_bn_flag) #both the old and novel embeddings are updated with the feature extractor fixed #the MR loss is removed in the CBF (ft_flag=4) and removed in all the training (ft_flag=5) #the differences lie in the models for CBF elif args.ft_flag == 4 or args.ft_flag == 5: ignored_params = list(map(id, tg_model.fc.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, tg_model.parameters()) tg_params =[{'params': base_params, 'lr': 0, 'weight_decay': 0}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc2.parameters(), 'lr': args.ft_base_lr, 'weight_decay': custom_weight_decay}] fix_bn_flag = True tg_model = tg_model.to(device) ref_model = ref_model.to(device) tg_ft_optimizer = optim.SGD(tg_params, lr=args.ft_base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_ft_lr_scheduler = lr_scheduler.MultiStepLR(tg_ft_optimizer, milestones=args.ft_lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval_LF(args.ft_epochs, tg_model, ref_model, tg_ft_optimizer, tg_ft_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ fix_bn=fix_bn_flag) else: print("Unknown ft_flag") sys.exit() ############################### # tg_model = tg_model.to(device) # ref_model = ref_model.to(device) # tg_ft_optimizer = optim.SGD(tg_params, lr=args.ft_base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) # tg_ft_lr_scheduler = lr_scheduler.MultiStepLR(tg_ft_optimizer, milestones=args.ft_lr_strat, gamma=lr_factor) # tg_model = incremental_train_and_eval_MR_LF(args.ft_epochs, tg_model, ref_model, tg_ft_optimizer, tg_ft_lr_scheduler, \ # trainloader, testloader, \ # iteration, start_iter, \ # cur_lamda, \ # args.dist, args.K, args.lw_mr, \ # fix_bn=fix_bn_flag) torch.save(tg_model, ckp_name.replace("/checkpoint/", "/checkpoint/AFTER_CBF_")) torch.save(class_means, class_means_name.replace("/checkpoint/", "/checkpoint/AFTER_CBF_")) ############################################################### ############################### print('Computing accuracy on the protoset...') current_eval_set = merge_images_labels(np.concatenate(X_protoset_cumuls), map_Y_protoset_cumuls) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) cbf_proto_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) ############################### print('Computing accuracy on the old protoset...') indices = np.array([i in range(0, iteration*args.nb_cl) for i in map_Y_protoset_cumuls]) current_eval_set = merge_images_labels(np.concatenate(X_protoset_cumuls)[indices], map_Y_protoset_cumuls[indices]) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) print('Min and Max of eval labels: {}, {}'.format(min(map_Y_protoset_cumuls[indices]), max(map_Y_protoset_cumuls[indices]))) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) cbf_proto_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) ############################### print('Computing accuracy on the new protoset...') indices = np.array([i in range(iteration*args.nb_cl, (iteration+1)*args.nb_cl) for i in map_Y_protoset_cumuls]) current_eval_set = merge_images_labels(np.concatenate(X_protoset_cumuls)[indices], map_Y_protoset_cumuls[indices]) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) print('Min and Max of eval labels: {}, {}'.format(min(map_Y_protoset_cumuls[indices]), max(map_Y_protoset_cumuls[indices]))) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) cbf_proto_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) ############################### ############################################################## # Calculate validation error of model on the first nb_cl classes: map_Y_valid_ori = np.array([order_list.index(i) for i in Y_valid_ori]) print('Computing accuracy on the original batch of classes...') # evalset.test_data = X_valid_ori.astype('uint8') # evalset.test_labels = map_Y_valid_ori current_eval_set = merge_images_labels(X_valid_ori, map_Y_valid_ori) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) ori_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_ori[iteration, :, iteration_total] = np.array(ori_acc).T ############################################################## # Calculate validation error of model on the cumul of classes: map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) print('Computing cumulative accuracy...') # evalset.test_data = X_valid_cumul.astype('uint8') # evalset.test_labels = map_Y_valid_cumul current_eval_set = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) cumul_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_cumul[iteration, :, iteration_total] = np.array(cumul_acc).T ############################################################## # Final save of the data torch.save(top1_acc_list_ori, \ './checkpoint/{}_run_{}_top1_acc_list_ori.pth'.format(args.ckp_prefix, iteration_total)) torch.save(top1_acc_list_cumul, \ './checkpoint/{}_run_{}_top1_acc_list_cumul.pth'.format(args.ckp_prefix, iteration_total))
47,336
60.08
136
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/class_incremental_cosine_imagenet.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle import math import modified_resnet import modified_linear import utils_pytorch from utils_imagenet.utils_dataset import split_images_labels from utils_imagenet.utils_dataset import merge_images_labels from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix from utils_incremental.incremental_train_and_eval import incremental_train_and_eval from utils_incremental.incremental_train_and_eval_MS import incremental_train_and_eval_MS from utils_incremental.incremental_train_and_eval_LF import incremental_train_and_eval_LF from utils_incremental.incremental_train_and_eval_MR_LF import incremental_train_and_eval_MR_LF from utils_incremental.incremental_train_and_eval_AMR_LF import incremental_train_and_eval_AMR_LF ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--dataset', default='seed_1993_subset_100_imagenet', type=str) parser.add_argument('--datadir', default='data/seed_1993_subset_100_imagenet/data', type=str) parser.add_argument('--num_classes', default=100, type=int) parser.add_argument('--num_workers', default=16, type=int, \ help='the number of workers for loading data') parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--nb_protos', default=20, type=int, \ help='Number of prototypes per class at the end') parser.add_argument('--nb_runs', default=1, type=int, \ help='Number of runs (random ordering of classes at each run)') parser.add_argument('--ckp_prefix', default=os.path.basename(sys.argv[0])[:-3], type=str, \ help='Checkpoint prefix') parser.add_argument('--epochs', default=90, type=int, \ help='Epochs') parser.add_argument('--T', default=2, type=float, \ help='Temporature for distialltion') parser.add_argument('--beta', default=0.25, type=float, \ help='Beta for distialltion') parser.add_argument('--resume', action='store_true', \ help='resume from checkpoint') parser.add_argument('--fix_budget', action='store_true', \ help='fix budget') ######################################## parser.add_argument('--mimic_score', action='store_true', \ help='To mimic scores for cosine embedding') parser.add_argument('--lw_ms', default=1, type=float, \ help='loss weight for mimicking score') ######################################## #improved class incremental learning parser.add_argument('--rs_ratio', default=0, type=float, \ help='The ratio for resample') parser.add_argument('--imprint_weights', action='store_true', \ help='Imprint the weights for novel classes') parser.add_argument('--less_forget', action='store_true', \ help='Less forgetful') parser.add_argument('--lamda', default=5, type=float, \ help='Lamda for LF') parser.add_argument('--adapt_lamda', action='store_true', \ help='Adaptively change lamda') parser.add_argument('--mr_loss', action='store_true', \ help='Margin ranking loss v1') parser.add_argument('--amr_loss', action='store_true', \ help='Margin ranking loss v2') parser.add_argument('--dist', default=0.5, type=float, \ help='Dist for MarginRankingLoss') parser.add_argument('--K', default=2, type=int, \ help='K for MarginRankingLoss') parser.add_argument('--lw_mr', default=1, type=float, \ help='loss weight for margin ranking loss') ######################################## parser.add_argument('--random_seed', default=1993, type=int, \ help='random seed') args = parser.parse_args() ######################################## assert(args.nb_cl_fg % args.nb_cl == 0) assert(args.nb_cl_fg >= args.nb_cl) train_batch_size = 128 # Batch size for train test_batch_size = 50 # Batch size for test eval_batch_size = 128 # Batch size for eval base_lr = 0.1 # Initial learning rate lr_strat = [30, 60] # Epochs where learning rate gets decreased lr_factor = 0.1 # Learning rate decrease factor custom_weight_decay = 1e-4 # Weight Decay custom_momentum = 0.9 # Momentum args.ckp_prefix = '{}_nb_cl_fg_{}_nb_cl_{}_nb_protos_{}'.format(args.ckp_prefix, args.nb_cl_fg, args.nb_cl, args.nb_protos) np.random.seed(args.random_seed) # Fix the random seed print(args) ######################################## device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") #transform_train = transforms.Compose([ # transforms.RandomCrop(32, padding=4), # transforms.RandomHorizontalFlip(), # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), #]) #transform_test = transforms.Compose([ # transforms.ToTensor(), # transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), #]) #trainset = torchvision.datasets.CIFAR100(root='./data', train=True, # download=True, transform=transform_train) #testset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=True, transform=transform_test) #evalset = torchvision.datasets.CIFAR100(root='./data', train=False, # download=False, transform=transform_test) # Data loading code traindir = os.path.join(args.datadir, 'train') valdir = os.path.join(args.datadir, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) trainset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) testset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) evalset = datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])) # Initialization dictionary_size = 1500 top1_acc_list_cumul = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) top1_acc_list_ori = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) #X_train_total = np.array(trainset.train_data) #Y_train_total = np.array(trainset.train_labels) #X_valid_total = np.array(testset.test_data) #Y_valid_total = np.array(testset.test_labels) X_train_total, Y_train_total = split_images_labels(trainset.imgs) X_valid_total, Y_valid_total = split_images_labels(testset.imgs) # Launch the different runs for iteration_total in range(args.nb_runs): # Select the order for the class learning order_name = "./checkpoint/seed_{}_{}_order_run_{}.pkl".format(args.random_seed, args.dataset, iteration_total) print("Order name:{}".format(order_name)) if os.path.exists(order_name): print("Loading orders") order = utils_pytorch.unpickle(order_name) else: print("Generating orders") order = np.arange(args.num_classes) np.random.shuffle(order) utils_pytorch.savepickle(order, order_name) order_list = list(order) print(order_list) # Initialization of the variables for this run X_valid_cumuls = [] X_protoset_cumuls = [] X_train_cumuls = [] Y_valid_cumuls = [] Y_protoset_cumuls = [] Y_train_cumuls = [] alpha_dr_herding = np.zeros((int(args.num_classes/args.nb_cl),dictionary_size,args.nb_cl),np.float32) # The following contains all the training samples of the different classes # because we want to compare our method with the theoretical case where all the training samples are stored # prototypes = np.zeros((args.num_classes,dictionary_size,X_train_total.shape[1],X_train_total.shape[2],X_train_total.shape[3])) prototypes = [[] for i in range(args.num_classes)] for orde in range(args.num_classes): prototypes[orde] = X_train_total[np.where(Y_train_total==order[orde])] prototypes = np.array(prototypes) start_iter = int(args.nb_cl_fg/args.nb_cl)-1 for iteration in range(start_iter, int(args.num_classes/args.nb_cl)): #init model if iteration == start_iter: ############################################################ last_iter = 0 ############################################################ tg_model = modified_resnet.resnet18(num_classes=args.nb_cl_fg) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features print("in_features:", in_features, "out_features:", out_features) ref_model = None elif iteration == start_iter+1: ############################################################ last_iter = iteration ############################################################ #increment classes ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features print("in_features:", in_features, "out_features:", out_features) new_fc = modified_linear.SplitCosineLinear(in_features, out_features, args.nb_cl) new_fc.fc1.weight.data = tg_model.fc.weight.data new_fc.sigma.data = tg_model.fc.sigma.data tg_model.fc = new_fc lamda_mult = out_features*1.0 / args.nb_cl else: ############################################################ last_iter = iteration ############################################################ ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features1 = tg_model.fc.fc1.out_features out_features2 = tg_model.fc.fc2.out_features print("in_features:", in_features, "out_features1:", \ out_features1, "out_features2:", out_features2) new_fc = modified_linear.SplitCosineLinear(in_features, out_features1+out_features2, args.nb_cl) new_fc.fc1.weight.data[:out_features1] = tg_model.fc.fc1.weight.data new_fc.fc1.weight.data[out_features1:] = tg_model.fc.fc2.weight.data new_fc.sigma.data = tg_model.fc.sigma.data tg_model.fc = new_fc lamda_mult = (out_features1+out_features2)*1.0 / (args.nb_cl) if iteration > start_iter and args.less_forget and args.adapt_lamda: #cur_lamda = lamda_base * sqrt(num_old_classes/num_new_classes) cur_lamda = args.lamda * math.sqrt(lamda_mult) else: cur_lamda = args.lamda if iteration > start_iter and args.less_forget: print("###############################") print("Lamda for less forget is set to ", cur_lamda) print("###############################") # Prepare the training data for the current batch of classes actual_cl = order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] indices_train_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_train_total]) indices_test_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_valid_total]) X_train = X_train_total[indices_train_10] X_valid = X_valid_total[indices_test_10] X_valid_cumuls.append(X_valid) X_train_cumuls.append(X_train) X_valid_cumul = np.concatenate(X_valid_cumuls) X_train_cumul = np.concatenate(X_train_cumuls) Y_train = Y_train_total[indices_train_10] Y_valid = Y_valid_total[indices_test_10] Y_valid_cumuls.append(Y_valid) Y_train_cumuls.append(Y_train) Y_valid_cumul = np.concatenate(Y_valid_cumuls) Y_train_cumul = np.concatenate(Y_train_cumuls) # Add the stored exemplars to the training data if iteration == start_iter: X_valid_ori = X_valid Y_valid_ori = Y_valid else: X_protoset = np.concatenate(X_protoset_cumuls) Y_protoset = np.concatenate(Y_protoset_cumuls) if args.rs_ratio > 0: #1/rs_ratio = (len(X_train)+len(X_protoset)*scale_factor)/(len(X_protoset)*scale_factor) scale_factor = (len(X_train) * args.rs_ratio) / (len(X_protoset) * (1 - args.rs_ratio)) rs_sample_weights = np.concatenate((np.ones(len(X_train)), np.ones(len(X_protoset))*scale_factor)) #number of samples per epoch #rs_num_samples = len(X_train) + len(X_protoset) rs_num_samples = int(len(X_train) / (1 - args.rs_ratio)) print("X_train:{}, X_protoset:{}, rs_num_samples:{}".format(len(X_train), len(X_protoset), rs_num_samples)) X_train = np.concatenate((X_train,X_protoset),axis=0) Y_train = np.concatenate((Y_train,Y_protoset)) # Launch the training loop print('Batch of classes number {0} arrives ...'.format(iteration+1)) map_Y_train = np.array([order_list.index(i) for i in Y_train]) map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) #imprint weights if iteration > start_iter and args.imprint_weights: #input: tg_model, X_train, map_Y_train #class_start = iteration*nb_cl class_end = (iteration+1)*nb_cl print("Imprint weights") ######################################### #compute the average norm of old embdding old_embedding_norm = tg_model.fc.fc1.weight.data.norm(dim=1, keepdim=True) average_old_embedding_norm = torch.mean(old_embedding_norm, dim=0).to('cpu').type(torch.DoubleTensor) ######################################### tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features novel_embedding = torch.zeros((args.nb_cl, num_features)) for cls_idx in range(iteration*args.nb_cl, (iteration+1)*args.nb_cl): cls_indices = np.array([i == cls_idx for i in map_Y_train]) assert(len(np.where(cls_indices==1)[0])<=dictionary_size) #evalset.test_data = X_train[cls_indices].astype('uint8') #evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(X_train[cls_indices], np.zeros(len(X_train[cls_indices]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = len(X_train[cls_indices]) cls_features = compute_features(tg_feature_model, evalloader, num_samples, num_features) #cls_features = cls_features.T #cls_features = cls_features / np.linalg.norm(cls_features,axis=0) #cls_embedding = np.mean(cls_features, axis=1) norm_features = F.normalize(torch.from_numpy(cls_features), p=2, dim=1) cls_embedding = torch.mean(norm_features, dim=0) #novel_embedding[cls_idx-iteration*args.nb_cl] = cls_embedding novel_embedding[cls_idx-iteration*args.nb_cl] = F.normalize(cls_embedding, p=2, dim=0) * average_old_embedding_norm tg_model.to(device) #torch.save(tg_model, "tg_model_before_imprint_weights.pth") tg_model.fc.fc2.weight.data = novel_embedding.to(device) #torch.save(tg_model, "tg_model_after_imprint_weights.pth") ############################################################ #trainset.train_data = X_train.astype('uint8') #trainset.train_labels = map_Y_train current_train_imgs = merge_images_labels(X_train, map_Y_train) trainset.imgs = trainset.samples = current_train_imgs if iteration > start_iter and args.rs_ratio > 0 and scale_factor > 1: print("Weights from sampling:", rs_sample_weights) index1 = np.where(rs_sample_weights>1)[0] index2 = np.where(map_Y_train<iteration*args.nb_cl)[0] assert((index1==index2).all()) train_sampler = torch.utils.data.sampler.WeightedRandomSampler(rs_sample_weights, rs_num_samples) #trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ # shuffle=False, sampler=train_sampler, num_workers=2) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ shuffle=False, sampler=train_sampler, num_workers=args.num_workers, pin_memory=True) else: #trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, # shuffle=True, num_workers=2) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) #testset.test_data = X_valid_cumul.astype('uint8') #testset.test_labels = map_Y_valid_cumul current_test_imgs = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) testset.imgs = testset.samples = current_test_imgs testloader = torch.utils.data.DataLoader(testset, batch_size=test_batch_size, shuffle=False, num_workers=2) print('Max and Min of train labels: {}, {}'.format(min(map_Y_train), max(map_Y_train))) print('Max and Min of valid labels: {}, {}'.format(min(map_Y_valid_cumul), max(map_Y_valid_cumul))) ############################################################## ckp_name = './checkpoint/{}_run_{}_iteration_{}_model.pth'.format(args.ckp_prefix, iteration_total, iteration) print('ckp_name', ckp_name) if args.resume and os.path.exists(ckp_name): print("###############################") print("Loading models from checkpoint") tg_model = torch.load(ckp_name) print("###############################") else: ############################### if iteration > start_iter and args.less_forget: #fix the embedding of old classes ignored_params = list(map(id, tg_model.fc.fc1.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, \ tg_model.parameters()) tg_params =[{'params': base_params, 'lr': base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': 0, 'weight_decay': 0}] else: tg_params = tg_model.parameters() ############################### tg_model = tg_model.to(device) if iteration > start_iter: ref_model = ref_model.to(device) tg_optimizer = optim.SGD(tg_params, lr=base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_lr_scheduler = lr_scheduler.MultiStepLR(tg_optimizer, milestones=lr_strat, gamma=lr_factor) ############################### if args.less_forget and args.mr_loss: print("incremental_train_and_eval_MR_LF") tg_model = incremental_train_and_eval_MR_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr) elif args.less_forget and args.amr_loss: print("incremental_train_and_eval_AMR_LF") tg_model = incremental_train_and_eval_AMR_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr) else: if args.less_forget: print("incremental_train_and_eval_LF") tg_model = incremental_train_and_eval_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda) else: if args.mimic_score: print("incremental_train_and_eval_MS") tg_model = incremental_train_and_eval_MS(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, args.lw_ms) else: print("incremental_train_and_eval") tg_model = incremental_train_and_eval(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, args.T, args.beta) if not os.path.isdir('checkpoint'): os.mkdir('checkpoint') torch.save(tg_model, ckp_name) ### Exemplars if args.fix_budget: nb_protos_cl = int(np.ceil(args.nb_protos*args.num_classes*1.0/args.nb_cl/(iteration+1))) else: nb_protos_cl = args.nb_protos tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features # Herding print('Updating exemplar set...') for iter_dico in range(last_iter*args.nb_cl, (iteration+1)*args.nb_cl): # Possible exemplars in the feature space and projected on the L2 sphere # evalset.test_data = prototypes[iter_dico].astype('uint8') # evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(prototypes[iter_dico], np.zeros(len(prototypes[iter_dico]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) num_samples = len(prototypes[iter_dico]) mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Herding procedure : ranking of the potential exemplars mu = np.mean(D,axis=1) index1 = int(iter_dico/args.nb_cl) index2 = iter_dico % args.nb_cl alpha_dr_herding[index1,:,index2] = alpha_dr_herding[index1,:,index2]*0 w_t = mu iter_herding = 0 iter_herding_eff = 0 while not(np.sum(alpha_dr_herding[index1,:,index2]!=0)==min(nb_protos_cl,500)) and iter_herding_eff<1000: tmp_t = np.dot(w_t,D) ind_max = np.argmax(tmp_t) iter_herding_eff += 1 if alpha_dr_herding[index1,ind_max,index2] == 0: alpha_dr_herding[index1,ind_max,index2] = 1+iter_herding iter_herding += 1 w_t = w_t+mu-D[:,ind_max] # Prepare the protoset X_protoset_cumuls = [] Y_protoset_cumuls = [] # Class means for iCaRL and NCM + Storing the selected exemplars in the protoset print('Computing mean-of_exemplars and theoretical mean...') # class_means = np.zeros((64,100,2)) class_means = np.zeros((num_features, args.num_classes, 2)) for iteration2 in range(iteration+1): for iter_dico in range(args.nb_cl): current_cl = order[range(iteration2*args.nb_cl,(iteration2+1)*args.nb_cl)] # Collect data in the feature space for each class # evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico].astype('uint8') # evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels current_eval_set = merge_images_labels(prototypes[iteration2*args.nb_cl+iter_dico], \ np.zeros(len(prototypes[iteration2*args.nb_cl+iter_dico]))) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) num_samples = len(prototypes[iteration2*args.nb_cl+iter_dico]) mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Flipped version also # evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico][:,:,:,::-1].astype('uint8') # evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, # shuffle=False, num_workers=2) # mapped_prototypes2 = compute_features(tg_feature_model, evalloader, num_samples, num_features) # D2 = mapped_prototypes2.T # D2 = D2/np.linalg.norm(D2,axis=0) D2 = D # iCaRL alph = alpha_dr_herding[iteration2,:,iter_dico] assert((alph[num_samples:]==0).all()) alph = alph[:num_samples] alph = (alph>0)*(alph<nb_protos_cl+1)*1. # X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico,np.where(alph==1)[0]]) X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico][np.where(alph==1)[0]]) Y_protoset_cumuls.append(order[iteration2*args.nb_cl+iter_dico]*np.ones(len(np.where(alph==1)[0]))) alph = alph/np.sum(alph) class_means[:,current_cl[iter_dico],0] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],0] /= np.linalg.norm(class_means[:,current_cl[iter_dico],0]) # Normal NCM # alph = np.ones(dictionary_size)/dictionary_size alph = np.ones(num_samples)/num_samples class_means[:,current_cl[iter_dico],1] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],1] /= np.linalg.norm(class_means[:,current_cl[iter_dico],1]) torch.save(class_means, \ './checkpoint/{}_run_{}_iteration_{}_class_means.pth'.format(args.ckp_prefix,iteration_total, iteration)) current_means = class_means[:, order[range(0,(iteration+1)*args.nb_cl)]] ############################################################## # Calculate validation error of model on the first nb_cl classes: map_Y_valid_ori = np.array([order_list.index(i) for i in Y_valid_ori]) print('Computing accuracy on the original batch of classes...') # evalset.test_data = X_valid_ori.astype('uint8') # evalset.test_labels = map_Y_valid_ori current_eval_set = merge_images_labels(X_valid_ori, map_Y_valid_ori) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) ori_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_ori[iteration, :, iteration_total] = np.array(ori_acc).T ############################################################## # Calculate validation error of model on the cumul of classes: map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) print('Computing cumulative accuracy...') # evalset.test_data = X_valid_cumul.astype('uint8') # evalset.test_labels = map_Y_valid_cumul current_eval_set = merge_images_labels(X_valid_cumul, map_Y_valid_cumul) evalset.imgs = evalset.samples = current_eval_set evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) cumul_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_cumul[iteration, :, iteration_total] = np.array(cumul_acc).T ############################################################## # Calculate confusion matrix # print('Computing confusion matrix...') # cm = compute_confusion_matrix(tg_model, tg_feature_model, current_means, evalloader) # cm_name = './checkpoint/{}_run_{}_iteration_{}_confusion_matrix.pth'.format(args.ckp_prefix,iteration_total, iteration) # with open(cm_name, 'wb') as f: # pickle.dump(cm, f, 2) #for reading with Python 2 ############################################################## # Final save of the data torch.save(top1_acc_list_ori, \ './checkpoint/{}_run_{}_top1_acc_list_ori.pth'.format(args.ckp_prefix, iteration_total)) torch.save(top1_acc_list_cumul, \ './checkpoint/{}_run_{}_top1_acc_list_cumul.pth'.format(args.ckp_prefix, iteration_total))
30,418
53.809009
132
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/gen_resized_imagenet.py
#!/usr/bin/env python # coding=utf-8 import argparse import os import random import shutil import time import warnings import numpy as np import torch import torch.nn as nn import torch.nn.parallel import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.optim import torch.utils.data import torch.utils.data.distributed import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models from PIL import Image src_root_dir = 'data/imagenet/data/' des_root_dir = 'data/imagenet_resized_256/data/' if not os.path.exists(des_root_dir): os.makedirs(des_root_dir) phase_list = ['train', 'val'] for phase in phase_list: if not os.path.exists(os.path.join(des_root_dir, phase)): os.mkdir(os.path.join(des_root_dir, phase)) data_dir = os.path.join(src_root_dir, phase) tg_dataset = datasets.ImageFolder(data_dir) for cls_name in tg_dataset.classes: if not os.path.exists(os.path.join(des_root_dir, phase, cls_name)): os.mkdir(os.path.join(des_root_dir, phase, cls_name)) cnt = 0 for item in tg_dataset.imgs: img_path = item[0] img = Image.open(img_path) img = img.convert('RGB') save_path = img_path.replace('imagenet', 'imagenet_resized_256') resized_img = img.resize((256,256), Image.BILINEAR) resized_img.save(save_path) cnt = cnt+1 if cnt % 1000 == 0: print(cnt, save_path) print("Hello World")
1,507
28
75
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/modified_linear.py
import math import torch from torch.nn.parameter import Parameter from torch.nn import functional as F from torch.nn import Module class CosineLinear(Module): def __init__(self, in_features, out_features, sigma=True): super(CosineLinear, self).__init__() self.in_features = in_features self.out_features = out_features self.weight = Parameter(torch.Tensor(out_features, in_features)) if sigma: self.sigma = Parameter(torch.Tensor(1)) else: self.register_parameter('sigma', None) self.reset_parameters() def reset_parameters(self): stdv = 1. / math.sqrt(self.weight.size(1)) self.weight.data.uniform_(-stdv, stdv) if self.sigma is not None: self.sigma.data.fill_(1) #for initializaiton of sigma def forward(self, input): #w_norm = self.weight.data.norm(dim=1, keepdim=True) #w_norm = w_norm.expand_as(self.weight).add_(self.epsilon) #x_norm = input.data.norm(dim=1, keepdim=True) #x_norm = x_norm.expand_as(input).add_(self.epsilon) #w = self.weight.div(w_norm) #x = input.div(x_norm) out = F.linear(F.normalize(input, p=2,dim=1), \ F.normalize(self.weight, p=2, dim=1)) if self.sigma is not None: out = self.sigma * out return out class SplitCosineLinear(Module): #consists of two fc layers and concatenate their outputs def __init__(self, in_features, out_features1, out_features2, sigma=True): super(SplitCosineLinear, self).__init__() self.in_features = in_features self.out_features = out_features1 + out_features2 self.fc1 = CosineLinear(in_features, out_features1, False) self.fc2 = CosineLinear(in_features, out_features2, False) if sigma: self.sigma = Parameter(torch.Tensor(1)) self.sigma.data.fill_(1) else: self.register_parameter('sigma', None) def forward(self, x): out1 = self.fc1(x) out2 = self.fc2(x) out = torch.cat((out1, out2), dim=1) #concatenate along the channel if self.sigma is not None: out = self.sigma * out return out
2,235
36.898305
78
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/utils_imagenet/train_and_eval.py
import argparse import os import shutil import time import torch import torch.nn as nn import torch.nn.parallel import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.optim import torch.utils.data import torch.utils.data.distributed import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models from .utils_train import * def train_and_eval(epochs, start_epoch, model, optimizer, lr_scheduler, \ train_loader, val_loader, gpu=None): for epoch in range(start_epoch, epochs): #adjust_learning_rate(optimizer, epoch) lr_scheduler.step() print('\nEpoch: %d, LR: ' % epoch, end='') print(lr_scheduler.get_lr()) # train for one epoch train(train_loader, model, optimizer, epoch, gpu) # evaluate on validation set validate(val_loader, model, gpu) return model def train(train_loader, model, optimizer, epoch, gpu=None): batch_time = AverageMeter() data_time = AverageMeter() losses = AverageMeter() top1 = AverageMeter() top5 = AverageMeter() # switch to train mode model.train() criterion = nn.CrossEntropyLoss().cuda(gpu) end = time.time() for i, (input, target) in enumerate(train_loader): # measure data loading time data_time.update(time.time() - end) if gpu is not None: input = input.cuda(gpu, non_blocking=True) target = target.cuda(gpu, non_blocking=True) # compute output output = model(input) loss = criterion(output, target) # measure accuracy and record loss prec1, prec5 = accuracy(output, target, topk=(1, 5)) losses.update(loss.item(), input.size(0)) top1.update(prec1[0], input.size(0)) top5.update(prec5[0], input.size(0)) # compute gradient and do SGD step optimizer.zero_grad() loss.backward() optimizer.step() # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % 10 == 0: print('Epoch: [{0}][{1}/{2}]\t' 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 'Data {data_time.val:.3f} ({data_time.avg:.3f})\t' 'Loss {loss.val:.4f} ({loss.avg:.4f})\t' 'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t' 'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format( epoch, i, len(train_loader), batch_time=batch_time, data_time=data_time, loss=losses, top1=top1, top5=top5))
2,629
30.309524
75
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/imagenet-class-incremental/utils_imagenet/utils_train.py
import argparse import os import shutil import time import torch import torch.nn as nn import torch.nn.parallel import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.optim import torch.utils.data import torch.utils.data.distributed import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models def validate(val_loader, model, gpu=None): batch_time = AverageMeter() losses = AverageMeter() top1 = AverageMeter() top5 = AverageMeter() # switch to evaluate mode model.eval() criterion = nn.CrossEntropyLoss().cuda(gpu) with torch.no_grad(): end = time.time() for i, (input, target) in enumerate(val_loader): if gpu is not None: input = input.cuda(gpu, non_blocking=True) target = target.cuda(gpu, non_blocking=True) # compute output output = model(input) loss = criterion(output, target) # measure accuracy and record loss prec1, prec5 = accuracy(output, target, topk=(1, 5)) losses.update(loss.item(), input.size(0)) top1.update(prec1[0], input.size(0)) top5.update(prec5[0], input.size(0)) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % 10 == 0: print('Test: [{0}/{1}]\t' 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 'Loss {loss.val:.4f} ({loss.avg:.4f})\t' 'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t' 'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format( i, len(val_loader), batch_time=batch_time, loss=losses, top1=top1, top5=top5)) print(' * Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f}' .format(top1=top1, top5=top5)) return top1.avg class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def accuracy(output, target, topk=(1,)): """Computes the precision@k for the specified values of k""" with torch.no_grad(): maxk = max(topk) batch_size = target.size(0) _, pred = output.topk(maxk, 1, True, True) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) res = [] for k in topk: correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) res.append(correct_k.mul_(100.0 / batch_size)) return res
2,897
29.505263
78
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/cifar100-class-incremental/utils_pytorch.py
#!/usr/bin/env python # coding=utf-8 from __future__ import print_function, division import torch import torch.nn as nn import torch.nn.init as init from collections import OrderedDict import numpy as np import os import os.path as osp import sys import time import math import subprocess try: import cPickle as pickle except: import pickle def savepickle(data, file_path): mkdir_p(osp.dirname(file_path), delete=False) print('pickle into', file_path) with open(file_path, 'wb') as f: pickle.dump(data, f, pickle.HIGHEST_PROTOCOL) def unpickle(file_path): with open(file_path, 'rb') as f: data = pickle.load(f) return data def mkdir_p(path, delete=False, print_info=True): if path == '': return if delete: subprocess.call(('rm -r ' + path).split()) if not osp.exists(path): if print_info: print('mkdir -p ' + path) subprocess.call(('mkdir -p ' + path).split()) def get_mean_and_std(dataset): '''Compute the mean and std value of dataset.''' dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2) mean = torch.zeros(3) std = torch.zeros(3) print('==> Computing mean and std..') for inputs, targets in dataloader: for i in range(3): mean[i] += inputs[:,i,:,:].mean() std[i] += inputs[:,i,:,:].std() mean.div_(len(dataset)) std.div_(len(dataset)) return mean, std def init_params(net): '''Init layer parameters.''' for m in net.modules(): if isinstance(m, nn.Conv2d): init.kaiming_normal_(m.weight, mode='fan_out') if m.bias: init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d): init.constant_(m.weight, 1) init.constant_(m.bias, 0) elif isinstance(m, nn.Linear): init.normal_(m.weight, std=1e-3) if m.bias is not None: init.constant_(m.bias, 0) _, term_width = os.popen('stty size', 'r').read().split() term_width = int(term_width) TOTAL_BAR_LENGTH = 65. last_time = time.time() begin_time = last_time def progress_bar(current, total, msg=None): global last_time, begin_time if current == 0: begin_time = time.time() # Reset for new bar. cur_len = int(TOTAL_BAR_LENGTH*current/total) rest_len = int(TOTAL_BAR_LENGTH - cur_len) - 1 sys.stdout.write(' [') for i in range(cur_len): sys.stdout.write('=') sys.stdout.write('>') for i in range(rest_len): sys.stdout.write('.') sys.stdout.write(']') cur_time = time.time() step_time = cur_time - last_time last_time = cur_time tot_time = cur_time - begin_time L = [] L.append(' Step: %s' % format_time(step_time)) L.append(' | Tot: %s' % format_time(tot_time)) if msg: L.append(' | ' + msg) msg = ''.join(L) sys.stdout.write(msg) for i in range(term_width-int(TOTAL_BAR_LENGTH)-len(msg)-3): sys.stdout.write(' ') # Go back to the center of the bar. for i in range(term_width-int(TOTAL_BAR_LENGTH/2)+2): sys.stdout.write('\b') sys.stdout.write(' %d/%d ' % (current+1, total)) if current < total-1: sys.stdout.write('\r') else: sys.stdout.write('\n') sys.stdout.flush() def format_time(seconds): days = int(seconds / 3600/24) seconds = seconds - days*3600*24 hours = int(seconds / 3600) seconds = seconds - hours*3600 minutes = int(seconds / 60) seconds = seconds - minutes*60 secondsf = int(seconds) seconds = seconds - secondsf millis = int(seconds*1000) f = '' i = 1 if days > 0: f += str(days) + 'D' i += 1 if hours > 0 and i <= 2: f += str(hours) + 'h' i += 1 if minutes > 0 and i <= 2: f += str(minutes) + 'm' i += 1 if secondsf > 0 and i <= 2: f += str(secondsf) + 's' i += 1 if millis > 0 and i <= 2: f += str(millis) + 'ms' i += 1 if f == '': f = '0ms' return f
4,102
26.172185
96
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/cifar100-class-incremental/eval_cumul_acc.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle from scipy.spatial.distance import cdist import modified_resnet_cifar import modified_linear import utils_pytorch from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--ckp_prefix', \ default='checkpoint/class_incremental_cifar100_nb_cl_10_nb_protos_20_run_0_', \ type=str) parser.add_argument('--order', \ default='./checkpoint/cifar100_order_run_0.pkl', \ type=str) parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') args = parser.parse_args() print(args) order = utils_pytorch.unpickle(args.order) order_list = list(order) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), ]) evalset = torchvision.datasets.CIFAR100(root='./data', train=False, download=False, transform=transform_test) input_data = evalset.test_data input_labels = evalset.test_labels map_input_labels = np.array([order_list.index(i) for i in input_labels]) #evalset.test_labels = map_input_labels #evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, # shuffle=False, num_workers=2) cnn_cumul_acc = [] icarl_cumul_acc = [] ncm_cumul_acc = [] num_classes = [] nb_cl = args.nb_cl start_iter = int(args.nb_cl_fg/nb_cl)-1 for iteration in range(start_iter, int(100/nb_cl)): #print("###########################################################") #print("For iteration {}".format(iteration)) #print("###########################################################") ckp_name = '{}iteration_{}_model.pth'.format(args.ckp_prefix, iteration) class_means_name = '{}iteration_{}_class_means.pth'.format(args.ckp_prefix, iteration) tg_model = torch.load(ckp_name) tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) class_means = torch.load(class_means_name) current_means = class_means[:, order[:(iteration+1)*nb_cl]] indices = np.array([i in range(0, (iteration+1)*nb_cl) for i in map_input_labels]) evalset.test_data = input_data[indices] evalset.test_labels = map_input_labels[indices] #print('Max and Min of valid labels: {}, {}'.format(min(evalset.test_labels), max(evalset.test_labels))) evalloader = torch.utils.data.DataLoader(evalset, batch_size=128, shuffle=False, num_workers=2) acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader, print_info=False) cnn_cumul_acc.append(acc[0]) icarl_cumul_acc.append(acc[1]) ncm_cumul_acc.append(acc[2]) num_classes.append((iteration+1)*nb_cl) print("###########################################################") print(' CNN acc: \t iCaRL acc \t NCM acc') print("###########################################################") for i in range(len(cnn_cumul_acc)): print("{:.2f} ".format(cnn_cumul_acc[i]), end='') print("[{:.2f}] ".format(np.mean(cnn_cumul_acc[-1])), end='') print("[{:.2f}] ".format(np.mean(cnn_cumul_acc)), end='') print("[{:.2f}] ".format(np.sum(np.array(cnn_cumul_acc)*np.array(num_classes)) / np.sum(num_classes)), end='') print("") for i in range(len(icarl_cumul_acc)): print("{:.2f} ".format(icarl_cumul_acc[i]), end='') print("[{:.2f}] ".format(np.mean(icarl_cumul_acc[-1])), end='') print("[{:.2f}] ".format(np.mean(icarl_cumul_acc)), end='') print("[{:.2f}] ".format(np.sum(np.array(icarl_cumul_acc)*np.array(num_classes)) / np.sum(num_classes)), end='') print("") for i in range(len(cnn_cumul_acc)): print("{:.2f} ".format(ncm_cumul_acc[i]), end='') print("[{:.2f}] ".format(np.mean(ncm_cumul_acc[-1])), end='') print("[{:.2f}] ".format(np.mean(ncm_cumul_acc)), end='') print("[{:.2f}] ".format(np.sum(np.array(ncm_cumul_acc)*np.array(num_classes)) / np.sum(num_classes)), end='') print("") print("###########################################################") print("") print('Number of classes', num_classes) print("###########################################################") print("Final acc on all classes") print("CNN:{:.2f}\t iCaRL:{:.2f}\t NCM:{:.2f}".format(cnn_cumul_acc[-1], icarl_cumul_acc[-1], ncm_cumul_acc[-1])) print("###########################################################") print("Average acc in each phase") print("CNN:{:.2f}\t iCaRL:{:.2f}\t NCM:{:.2f}".format(np.mean(cnn_cumul_acc), np.mean(icarl_cumul_acc), np.mean(ncm_cumul_acc))) print("###########################################################") print("Weighted average acc in each phase") print("CNN:{:.2f}\t iCaRL:{:.2f}\t NCM:{:.2f}".format( np.sum(np.array(cnn_cumul_acc)*np.array(num_classes)) / np.sum(num_classes), np.sum(np.array(icarl_cumul_acc)*np.array(num_classes)) / np.sum(num_classes), np.sum(np.array(ncm_cumul_acc)*np.array(num_classes)) / np.sum(num_classes) )) print("###########################################################")
5,687
43.4375
128
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/cifar100-class-incremental/class_incremental_cifar100.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle import resnet_cifar import utils_pytorch from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix from utils_incremental.incremental_train_and_eval import incremental_train_and_eval ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--dataset', default='cifar100', type=str) parser.add_argument('--num_classes', default=100, type=int) parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--nb_protos', default=20, type=int, \ help='Number of prototypes per class at the end') parser.add_argument('--nb_runs', default=1, type=int, \ help='Number of runs (random ordering of classes at each run)') parser.add_argument('--ckp_prefix', default=os.path.basename(sys.argv[0])[:-3], type=str, \ help='Checkpoint prefix') parser.add_argument('--epochs', default=160, type=int, \ help='Epochs') parser.add_argument('--T', default=2, type=float, \ help='Temperature for distialltion') parser.add_argument('--beta', default=0.25, type=float, \ help='Beta for distialltion') parser.add_argument('--resume', action='store_true', \ help='resume from checkpoint') parser.add_argument('--fix_budget', action='store_true', \ help='fix budget') parser.add_argument('--rs_ratio', default=0, type=float, \ help='The ratio for resample') parser.add_argument('--random_seed', default=1993, type=int, \ help='random seed') args = parser.parse_args() ######################################## assert(args.nb_cl_fg % args.nb_cl == 0) assert(args.nb_cl_fg >= args.nb_cl) train_batch_size = 128 # Batch size for train test_batch_size = 100 # Batch size for test eval_batch_size = 128 # Batch size for eval base_lr = 0.1 # Initial learning rate lr_strat = [80, 120] # Epochs where learning rate gets decreased lr_factor = 0.1 # Learning rate decrease factor custom_weight_decay = 5e-4 # Weight Decay custom_momentum = 0.9 # Momentum args.ckp_prefix = '{}_nb_cl_fg_{}_nb_cl_{}_nb_protos_{}'.format(args.ckp_prefix, args.nb_cl_fg, args.nb_cl, args.nb_protos) np.random.seed(args.random_seed) # Fix the random seed print(args) ######################################## device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), ]) trainset = torchvision.datasets.CIFAR100(root='./data', train=True, download=True, transform=transform_train) testset = torchvision.datasets.CIFAR100(root='./data', train=False, download=True, transform=transform_test) evalset = torchvision.datasets.CIFAR100(root='./data', train=False, download=False, transform=transform_test) # Initialization dictionary_size = 500 top1_acc_list_cumul = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) top1_acc_list_ori = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) X_train_total = np.array(trainset.train_data) Y_train_total = np.array(trainset.train_labels) X_valid_total = np.array(testset.test_data) Y_valid_total = np.array(testset.test_labels) # Launch the different runs for iteration_total in range(args.nb_runs): # Select the order for the class learning order_name = "./checkpoint/seed_{}_{}_order_run_{}.pkl".format(args.random_seed, args.dataset, iteration_total) print("Order name:{}".format(order_name)) if os.path.exists(order_name): print("Loading orders") order = utils_pytorch.unpickle(order_name) else: print("Generating orders") order = np.arange(args.num_classes) np.random.shuffle(order) utils_pytorch.savepickle(order, order_name) order_list = list(order) print(order_list) # Initialization of the variables for this run X_valid_cumuls = [] X_protoset_cumuls = [] X_train_cumuls = [] Y_valid_cumuls = [] Y_protoset_cumuls = [] Y_train_cumuls = [] alpha_dr_herding = np.zeros((int(args.num_classes/args.nb_cl),dictionary_size,args.nb_cl),np.float32) # The following contains all the training samples of the different classes # because we want to compare our method with the theoretical case where all the training samples are stored prototypes = np.zeros((args.num_classes,dictionary_size,X_train_total.shape[1],X_train_total.shape[2],X_train_total.shape[3])) for orde in range(args.num_classes): prototypes[orde,:,:,:,:] = X_train_total[np.where(Y_train_total==order[orde])] start_iter = int(args.nb_cl_fg/args.nb_cl)-1 for iteration in range(start_iter, int(args.num_classes/args.nb_cl)): #init model if iteration == start_iter: ############################################################ last_iter = 0 ############################################################ tg_model = resnet_cifar.resnet32(num_classes=args.nb_cl_fg) ref_model = None else: ############################################################ last_iter = iteration ############################################################ #increment classes ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features new_fc = nn.Linear(in_features, out_features+args.nb_cl) new_fc.weight.data[:out_features] = tg_model.fc.weight.data new_fc.bias.data[:out_features] = tg_model.fc.bias.data tg_model.fc = new_fc # Prepare the training data for the current batch of classes actual_cl = order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] indices_train_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_train_total]) indices_test_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_valid_total]) X_train = X_train_total[indices_train_10] X_valid = X_valid_total[indices_test_10] X_valid_cumuls.append(X_valid) X_train_cumuls.append(X_train) X_valid_cumul = np.concatenate(X_valid_cumuls) X_train_cumul = np.concatenate(X_train_cumuls) Y_train = Y_train_total[indices_train_10] Y_valid = Y_valid_total[indices_test_10] Y_valid_cumuls.append(Y_valid) Y_train_cumuls.append(Y_train) Y_valid_cumul = np.concatenate(Y_valid_cumuls) Y_train_cumul = np.concatenate(Y_train_cumuls) # Add the stored exemplars to the training data if iteration == start_iter: X_valid_ori = X_valid Y_valid_ori = Y_valid else: X_protoset = np.concatenate(X_protoset_cumuls) Y_protoset = np.concatenate(Y_protoset_cumuls) if args.rs_ratio > 0: #1/rs_ratio = (len(X_train)+len(X_protoset)*scale_factor)/(len(X_protoset)*scale_factor) scale_factor = (len(X_train) * args.rs_ratio) / (len(X_protoset) * (1 - args.rs_ratio)) rs_sample_weights = np.concatenate((np.ones(len(X_train)), np.ones(len(X_protoset))*scale_factor)) #number of samples per epoch, undersample on the new classes #rs_num_samples = len(X_train) + len(X_protoset) rs_num_samples = int(len(X_train) / (1 - args.rs_ratio)) print("X_train:{}, X_protoset:{}, rs_num_samples:{}".format(len(X_train), len(X_protoset), rs_num_samples)) X_train = np.concatenate((X_train,X_protoset),axis=0) Y_train = np.concatenate((Y_train,Y_protoset)) # Launch the training loop print('Batch of classes number {0} arrives ...'.format(iteration+1)) map_Y_train = np.array([order_list.index(i) for i in Y_train]) map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) ############################################################ trainset.train_data = X_train.astype('uint8') trainset.train_labels = map_Y_train if iteration > start_iter and args.rs_ratio > 0 and scale_factor > 1: print("Weights from sampling:", rs_sample_weights) index1 = np.where(rs_sample_weights>1)[0] index2 = np.where(map_Y_train<iteration*args.nb_cl)[0] assert((index1==index2).all()) train_sampler = torch.utils.data.sampler.WeightedRandomSampler(rs_sample_weights, rs_num_samples) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ shuffle=False, sampler=train_sampler, num_workers=2) else: trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=2) testset.test_data = X_valid_cumul.astype('uint8') testset.test_labels = map_Y_valid_cumul testloader = torch.utils.data.DataLoader(testset, batch_size=test_batch_size, shuffle=False, num_workers=2) print('Max and Min of train labels: {}, {}'.format(min(map_Y_train), max(map_Y_train))) print('Max and Min of valid labels: {}, {}'.format(min(map_Y_valid_cumul), max(map_Y_valid_cumul))) ############################################################## ckp_name = './checkpoint/{}_run_{}_iteration_{}_model.pth'.format(args.ckp_prefix, iteration_total, iteration) print('ckp_name', ckp_name) if args.resume and os.path.exists(ckp_name): print("###############################") print("Loading models from checkpoint") tg_model = torch.load(ckp_name) print("###############################") else: tg_params = tg_model.parameters() tg_model = tg_model.to(device) if iteration > start_iter: ref_model = ref_model.to(device) tg_optimizer = optim.SGD(tg_params, lr=base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_lr_scheduler = lr_scheduler.MultiStepLR(tg_optimizer, milestones=lr_strat, gamma=lr_factor) tg_model = incremental_train_and_eval(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ args.T, args.beta) if not os.path.isdir('checkpoint'): os.mkdir('checkpoint') torch.save(tg_model, ckp_name) ### Exemplars if args.fix_budget: nb_protos_cl = int(np.ceil(args.nb_protos*100./args.nb_cl/(iteration+1))) else: nb_protos_cl = args.nb_protos tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features # Herding print('Updating exemplar set...') for iter_dico in range(last_iter*args.nb_cl, (iteration+1)*args.nb_cl): # Possible exemplars in the feature space and projected on the L2 sphere evalset.test_data = prototypes[iter_dico].astype('uint8') evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = evalset.test_data.shape[0] mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Herding procedure : ranking of the potential exemplars mu = np.mean(D,axis=1) index1 = int(iter_dico/args.nb_cl) index2 = iter_dico % args.nb_cl alpha_dr_herding[index1,:,index2] = alpha_dr_herding[index1,:,index2]*0 w_t = mu iter_herding = 0 iter_herding_eff = 0 while not(np.sum(alpha_dr_herding[index1,:,index2]!=0)==min(nb_protos_cl,500)) and iter_herding_eff<1000: tmp_t = np.dot(w_t,D) ind_max = np.argmax(tmp_t) iter_herding_eff += 1 if alpha_dr_herding[index1,ind_max,index2] == 0: alpha_dr_herding[index1,ind_max,index2] = 1+iter_herding iter_herding += 1 w_t = w_t+mu-D[:,ind_max] # Prepare the protoset X_protoset_cumuls = [] Y_protoset_cumuls = [] # Class means for iCaRL and NCM + Storing the selected exemplars in the protoset print('Computing mean-of_exemplars and theoretical mean...') class_means = np.zeros((64,100,2)) for iteration2 in range(iteration+1): for iter_dico in range(args.nb_cl): current_cl = order[range(iteration2*args.nb_cl,(iteration2+1)*args.nb_cl)] # Collect data in the feature space for each class evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico].astype('uint8') evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = evalset.test_data.shape[0] mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Flipped version also evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico][:,:,:,::-1].astype('uint8') evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) mapped_prototypes2 = compute_features(tg_feature_model, evalloader, num_samples, num_features) D2 = mapped_prototypes2.T D2 = D2/np.linalg.norm(D2,axis=0) # iCaRL alph = alpha_dr_herding[iteration2,:,iter_dico] alph = (alph>0)*(alph<nb_protos_cl+1)*1. X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico,np.where(alph==1)[0]]) Y_protoset_cumuls.append(order[iteration2*args.nb_cl+iter_dico]*np.ones(len(np.where(alph==1)[0]))) alph = alph/np.sum(alph) class_means[:,current_cl[iter_dico],0] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],0] /= np.linalg.norm(class_means[:,current_cl[iter_dico],0]) # Normal NCM alph = np.ones(dictionary_size)/dictionary_size class_means[:,current_cl[iter_dico],1] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],1] /= np.linalg.norm(class_means[:,current_cl[iter_dico],1]) torch.save(class_means, \ './checkpoint/{}_run_{}_iteration_{}_class_means.pth'.format(args.ckp_prefix,iteration_total, iteration)) current_means = class_means[:, order[range(0,(iteration+1)*args.nb_cl)]] ############################################################## # Calculate validation error of model on the first nb_cl classes: map_Y_valid_ori = np.array([order_list.index(i) for i in Y_valid_ori]) print('Computing accuracy on the original batch of classes...') evalset.test_data = X_valid_ori.astype('uint8') evalset.test_labels = map_Y_valid_ori evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) ori_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_ori[iteration, :, iteration_total] = np.array(ori_acc).T ############################################################## # Calculate validation error of model on the cumul of classes: map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) print('Computing cumulative accuracy...') evalset.test_data = X_valid_cumul.astype('uint8') evalset.test_labels = map_Y_valid_cumul evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) cumul_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_cumul[iteration, :, iteration_total] = np.array(cumul_acc).T ############################################################## # Calculate confusion matrix print('Computing confusion matrix...') cm = compute_confusion_matrix(tg_model, tg_feature_model, current_means, evalloader) cm_name = './checkpoint/{}_run_{}_iteration_{}_confusion_matrix.pth'.format(args.ckp_prefix,iteration_total, iteration) with open(cm_name, 'wb') as f: pickle.dump(cm, f, 2) #for reading with Python 2 ############################################################## # Final save of the data torch.save(top1_acc_list_ori, \ './checkpoint/{}_run_{}_top1_acc_list_ori.pth'.format(args.ckp_prefix, iteration_total)) torch.save(top1_acc_list_cumul, \ './checkpoint/{}_run_{}_top1_acc_list_cumul.pth'.format(args.ckp_prefix, iteration_total))
18,765
51.565826
130
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/cifar100-class-incremental/resnet_cifar.py
import torch.nn as nn import math import torch.utils.model_zoo as model_zoo def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers, num_classes=10): self.inplanes = 16 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(16) self.relu = nn.ReLU(inplace=True) self.layer1 = self._make_layer(block, 16, layers[0]) self.layer2 = self._make_layer(block, 32, layers[1], stride=2) self.layer3 = self._make_layer(block, 64, layers[2], stride=2) self.avgpool = nn.AvgPool2d(8, stride=1) self.fc = nn.Linear(64 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x def resnet20(pretrained=False, **kwargs): n = 3 model = ResNet(BasicBlock, [n, n, n], **kwargs) return model def resnet32(pretrained=False, **kwargs): n = 5 model = ResNet(BasicBlock, [n, n, n], **kwargs) return model def resnet56(pretrained=False, **kwargs): n = 9 model = ResNet(Bottleneck, [n, n, n], **kwargs) return model
4,525
29.375839
90
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/cifar100-class-incremental/modified_resnet_cifar.py
#remove ReLU in the last layer, and use cosine layer to replace nn.Linear import torch.nn as nn import math import torch.utils.model_zoo as model_zoo import modified_linear def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None, last=False): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride self.last = last def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual if not self.last: #remove ReLU in the last layer out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers, num_classes=10): self.inplanes = 16 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(16) self.relu = nn.ReLU(inplace=True) self.layer1 = self._make_layer(block, 16, layers[0]) self.layer2 = self._make_layer(block, 32, layers[1], stride=2) self.layer3 = self._make_layer(block, 64, layers[2], stride=2, last_phase=True) self.avgpool = nn.AvgPool2d(8, stride=1) self.fc = modified_linear.CosineLinear(64 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) def _make_layer(self, block, planes, blocks, stride=1, last_phase=False): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion if last_phase: for i in range(1, blocks-1): layers.append(block(self.inplanes, planes)) layers.append(block(self.inplanes, planes, last=True)) else: for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x def resnet20(pretrained=False, **kwargs): n = 3 model = ResNet(BasicBlock, [n, n, n], **kwargs) return model def resnet32(pretrained=False, **kwargs): n = 5 model = ResNet(BasicBlock, [n, n, n], **kwargs) return model
3,716
31.893805
87
py
CVPR19_Incremental_Learning
CVPR19_Incremental_Learning-master/cifar100-class-incremental/class_incremental_cosine_cifar100.py
#!/usr/bin/env python # coding=utf-8 import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim import lr_scheduler import torchvision from torchvision import datasets, models, transforms from torch.autograd import Variable import numpy as np import time import os import sys import copy import argparse from PIL import Image try: import cPickle as pickle except: import pickle import math import modified_resnet_cifar import modified_linear import utils_pytorch from utils_incremental.compute_features import compute_features from utils_incremental.compute_accuracy import compute_accuracy from utils_incremental.compute_confusion_matrix import compute_confusion_matrix from utils_incremental.incremental_train_and_eval import incremental_train_and_eval from utils_incremental.incremental_train_and_eval_MS import incremental_train_and_eval_MS from utils_incremental.incremental_train_and_eval_LF import incremental_train_and_eval_LF from utils_incremental.incremental_train_and_eval_MR_LF import incremental_train_and_eval_MR_LF from utils_incremental.incremental_train_and_eval_AMR_LF import incremental_train_and_eval_AMR_LF ######### Modifiable Settings ########## parser = argparse.ArgumentParser() parser.add_argument('--dataset', default='cifar100', type=str) parser.add_argument('--num_classes', default=100, type=int) parser.add_argument('--nb_cl_fg', default=50, type=int, \ help='the number of classes in first group') parser.add_argument('--nb_cl', default=10, type=int, \ help='Classes per group') parser.add_argument('--nb_protos', default=20, type=int, \ help='Number of prototypes per class at the end') parser.add_argument('--nb_runs', default=1, type=int, \ help='Number of runs (random ordering of classes at each run)') parser.add_argument('--ckp_prefix', default=os.path.basename(sys.argv[0])[:-3], type=str, \ help='Checkpoint prefix') parser.add_argument('--epochs', default=160, type=int, \ help='Epochs') parser.add_argument('--T', default=2, type=float, \ help='Temporature for distialltion') parser.add_argument('--beta', default=0.25, type=float, \ help='Beta for distialltion') parser.add_argument('--resume', action='store_true', \ help='resume from checkpoint') parser.add_argument('--fix_budget', action='store_true', \ help='fix budget') ######################################## parser.add_argument('--mimic_score', action='store_true', \ help='To mimic scores for cosine embedding') parser.add_argument('--lw_ms', default=1, type=float, \ help='loss weight for mimicking score') ######################################## #improved class incremental learning parser.add_argument('--rs_ratio', default=0, type=float, \ help='The ratio for resample') parser.add_argument('--imprint_weights', action='store_true', \ help='Imprint the weights for novel classes') parser.add_argument('--less_forget', action='store_true', \ help='Less forgetful') parser.add_argument('--lamda', default=5, type=float, \ help='Lamda for LF') parser.add_argument('--adapt_lamda', action='store_true', \ help='Adaptively change lamda') parser.add_argument('--mr_loss', action='store_true', \ help='Margin ranking loss v1') parser.add_argument('--amr_loss', action='store_true', \ help='Margin ranking loss v2') parser.add_argument('--dist', default=0.5, type=float, \ help='Dist for MarginRankingLoss') parser.add_argument('--K', default=2, type=int, \ help='K for MarginRankingLoss') parser.add_argument('--lw_mr', default=1, type=float, \ help='loss weight for margin ranking loss') ######################################## parser.add_argument('--random_seed', default=1993, type=int, \ help='random seed') args = parser.parse_args() ######################################## assert(args.nb_cl_fg % args.nb_cl == 0) assert(args.nb_cl_fg >= args.nb_cl) train_batch_size = 128 # Batch size for train test_batch_size = 100 # Batch size for test eval_batch_size = 128 # Batch size for eval base_lr = 0.1 # Initial learning rate lr_strat = [80, 120] # Epochs where learning rate gets decreased lr_factor = 0.1 # Learning rate decrease factor custom_weight_decay = 5e-4 # Weight Decay custom_momentum = 0.9 # Momentum args.ckp_prefix = '{}_nb_cl_fg_{}_nb_cl_{}_nb_protos_{}'.format(args.ckp_prefix, args.nb_cl_fg, args.nb_cl, args.nb_protos) np.random.seed(args.random_seed) # Fix the random seed print(args) ######################################## device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)), ]) trainset = torchvision.datasets.CIFAR100(root='./data', train=True, download=True, transform=transform_train) testset = torchvision.datasets.CIFAR100(root='./data', train=False, download=True, transform=transform_test) evalset = torchvision.datasets.CIFAR100(root='./data', train=False, download=False, transform=transform_test) # Initialization dictionary_size = 500 top1_acc_list_cumul = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) top1_acc_list_ori = np.zeros((int(args.num_classes/args.nb_cl),3,args.nb_runs)) X_train_total = np.array(trainset.train_data) Y_train_total = np.array(trainset.train_labels) X_valid_total = np.array(testset.test_data) Y_valid_total = np.array(testset.test_labels) # Launch the different runs for iteration_total in range(args.nb_runs): # Select the order for the class learning order_name = "./checkpoint/seed_{}_{}_order_run_{}.pkl".format(args.random_seed, args.dataset, iteration_total) print("Order name:{}".format(order_name)) if os.path.exists(order_name): print("Loading orders") order = utils_pytorch.unpickle(order_name) else: print("Generating orders") order = np.arange(args.num_classes) np.random.shuffle(order) utils_pytorch.savepickle(order, order_name) order_list = list(order) print(order_list) # Initialization of the variables for this run X_valid_cumuls = [] X_protoset_cumuls = [] X_train_cumuls = [] Y_valid_cumuls = [] Y_protoset_cumuls = [] Y_train_cumuls = [] alpha_dr_herding = np.zeros((int(args.num_classes/args.nb_cl),dictionary_size,args.nb_cl),np.float32) # The following contains all the training samples of the different classes # because we want to compare our method with the theoretical case where all the training samples are stored prototypes = np.zeros((args.num_classes,dictionary_size,X_train_total.shape[1],X_train_total.shape[2],X_train_total.shape[3])) for orde in range(args.num_classes): prototypes[orde,:,:,:,:] = X_train_total[np.where(Y_train_total==order[orde])] start_iter = int(args.nb_cl_fg/args.nb_cl)-1 for iteration in range(start_iter, int(args.num_classes/args.nb_cl)): #init model if iteration == start_iter: ############################################################ last_iter = 0 ############################################################ tg_model = modified_resnet_cifar.resnet32(num_classes=args.nb_cl_fg) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features print("in_features:", in_features, "out_features:", out_features) ref_model = None elif iteration == start_iter+1: ############################################################ last_iter = iteration ############################################################ #increment classes ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features = tg_model.fc.out_features print("in_features:", in_features, "out_features:", out_features) new_fc = modified_linear.SplitCosineLinear(in_features, out_features, args.nb_cl) new_fc.fc1.weight.data = tg_model.fc.weight.data new_fc.sigma.data = tg_model.fc.sigma.data tg_model.fc = new_fc lamda_mult = out_features*1.0 / args.nb_cl else: ############################################################ last_iter = iteration ############################################################ ref_model = copy.deepcopy(tg_model) in_features = tg_model.fc.in_features out_features1 = tg_model.fc.fc1.out_features out_features2 = tg_model.fc.fc2.out_features print("in_features:", in_features, "out_features1:", \ out_features1, "out_features2:", out_features2) new_fc = modified_linear.SplitCosineLinear(in_features, out_features1+out_features2, args.nb_cl) new_fc.fc1.weight.data[:out_features1] = tg_model.fc.fc1.weight.data new_fc.fc1.weight.data[out_features1:] = tg_model.fc.fc2.weight.data new_fc.sigma.data = tg_model.fc.sigma.data tg_model.fc = new_fc lamda_mult = (out_features1+out_features2)*1.0 / (args.nb_cl) if iteration > start_iter and args.less_forget and args.adapt_lamda: #cur_lamda = lamda_base * sqrt(num_old_classes/num_new_classes) cur_lamda = args.lamda * math.sqrt(lamda_mult) else: cur_lamda = args.lamda if iteration > start_iter and args.less_forget: print("###############################") print("Lamda for less forget is set to ", cur_lamda) print("###############################") # Prepare the training data for the current batch of classes actual_cl = order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] indices_train_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_train_total]) indices_test_10 = np.array([i in order[range(last_iter*args.nb_cl,(iteration+1)*args.nb_cl)] for i in Y_valid_total]) X_train = X_train_total[indices_train_10] X_valid = X_valid_total[indices_test_10] X_valid_cumuls.append(X_valid) X_train_cumuls.append(X_train) X_valid_cumul = np.concatenate(X_valid_cumuls) X_train_cumul = np.concatenate(X_train_cumuls) Y_train = Y_train_total[indices_train_10] Y_valid = Y_valid_total[indices_test_10] Y_valid_cumuls.append(Y_valid) Y_train_cumuls.append(Y_train) Y_valid_cumul = np.concatenate(Y_valid_cumuls) Y_train_cumul = np.concatenate(Y_train_cumuls) # Add the stored exemplars to the training data if iteration == start_iter: X_valid_ori = X_valid Y_valid_ori = Y_valid else: X_protoset = np.concatenate(X_protoset_cumuls) Y_protoset = np.concatenate(Y_protoset_cumuls) if args.rs_ratio > 0: #1/rs_ratio = (len(X_train)+len(X_protoset)*scale_factor)/(len(X_protoset)*scale_factor) scale_factor = (len(X_train) * args.rs_ratio) / (len(X_protoset) * (1 - args.rs_ratio)) rs_sample_weights = np.concatenate((np.ones(len(X_train)), np.ones(len(X_protoset))*scale_factor)) #number of samples per epoch, undersample on the new classes #rs_num_samples = len(X_train) + len(X_protoset) rs_num_samples = int(len(X_train) / (1 - args.rs_ratio)) print("X_train:{}, X_protoset:{}, rs_num_samples:{}".format(len(X_train), len(X_protoset), rs_num_samples)) X_train = np.concatenate((X_train,X_protoset),axis=0) Y_train = np.concatenate((Y_train,Y_protoset)) # Launch the training loop print('Batch of classes number {0} arrives ...'.format(iteration+1)) map_Y_train = np.array([order_list.index(i) for i in Y_train]) map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) #imprint weights if iteration > start_iter and args.imprint_weights: #input: tg_model, X_train, map_Y_train #class_start = iteration*nb_cl class_end = (iteration+1)*nb_cl print("Imprint weights") ######################################### #compute the average norm of old embdding old_embedding_norm = tg_model.fc.fc1.weight.data.norm(dim=1, keepdim=True) average_old_embedding_norm = torch.mean(old_embedding_norm, dim=0).to('cpu').type(torch.DoubleTensor) ######################################### tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features novel_embedding = torch.zeros((args.nb_cl, num_features)) for cls_idx in range(iteration*args.nb_cl, (iteration+1)*args.nb_cl): cls_indices = np.array([i == cls_idx for i in map_Y_train]) assert(len(np.where(cls_indices==1)[0])==dictionary_size) evalset.test_data = X_train[cls_indices].astype('uint8') evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = evalset.test_data.shape[0] cls_features = compute_features(tg_feature_model, evalloader, num_samples, num_features) #cls_features = cls_features.T #cls_features = cls_features / np.linalg.norm(cls_features,axis=0) #cls_embedding = np.mean(cls_features, axis=1) norm_features = F.normalize(torch.from_numpy(cls_features), p=2, dim=1) cls_embedding = torch.mean(norm_features, dim=0) #novel_embedding[cls_idx-iteration*args.nb_cl] = cls_embedding novel_embedding[cls_idx-iteration*args.nb_cl] = F.normalize(cls_embedding, p=2, dim=0) * average_old_embedding_norm tg_model.to(device) #torch.save(tg_model, "tg_model_before_imprint_weights.pth") tg_model.fc.fc2.weight.data = novel_embedding.to(device) #torch.save(tg_model, "tg_model_after_imprint_weights.pth") ############################################################ trainset.train_data = X_train.astype('uint8') trainset.train_labels = map_Y_train if iteration > start_iter and args.rs_ratio > 0 and scale_factor > 1: print("Weights from sampling:", rs_sample_weights) index1 = np.where(rs_sample_weights>1)[0] index2 = np.where(map_Y_train<iteration*args.nb_cl)[0] assert((index1==index2).all()) train_sampler = torch.utils.data.sampler.WeightedRandomSampler(rs_sample_weights, rs_num_samples) trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, \ shuffle=False, sampler=train_sampler, num_workers=2) else: trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=2) testset.test_data = X_valid_cumul.astype('uint8') testset.test_labels = map_Y_valid_cumul testloader = torch.utils.data.DataLoader(testset, batch_size=test_batch_size, shuffle=False, num_workers=2) print('Max and Min of train labels: {}, {}'.format(min(map_Y_train), max(map_Y_train))) print('Max and Min of valid labels: {}, {}'.format(min(map_Y_valid_cumul), max(map_Y_valid_cumul))) ############################################################## ckp_name = './checkpoint/{}_run_{}_iteration_{}_model.pth'.format(args.ckp_prefix, iteration_total, iteration) print('ckp_name', ckp_name) if args.resume and os.path.exists(ckp_name): print("###############################") print("Loading models from checkpoint") tg_model = torch.load(ckp_name) print("###############################") else: ############################### if iteration > start_iter and args.less_forget: #fix the embedding of old classes ignored_params = list(map(id, tg_model.fc.fc1.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, \ tg_model.parameters()) tg_params =[{'params': base_params, 'lr': base_lr, 'weight_decay': custom_weight_decay}, \ {'params': tg_model.fc.fc1.parameters(), 'lr': 0, 'weight_decay': 0}] else: tg_params = tg_model.parameters() ############################### tg_model = tg_model.to(device) if iteration > start_iter: ref_model = ref_model.to(device) tg_optimizer = optim.SGD(tg_params, lr=base_lr, momentum=custom_momentum, weight_decay=custom_weight_decay) tg_lr_scheduler = lr_scheduler.MultiStepLR(tg_optimizer, milestones=lr_strat, gamma=lr_factor) ############################### if args.less_forget and args.mr_loss: print("incremental_train_and_eval_MR_LF") tg_model = incremental_train_and_eval_MR_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr) elif args.less_forget and args.amr_loss: print("incremental_train_and_eval_AMR_LF") tg_model = incremental_train_and_eval_AMR_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda, \ args.dist, args.K, args.lw_mr) else: if args.less_forget: print("incremental_train_and_eval_LF") tg_model = incremental_train_and_eval_LF(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, \ cur_lamda) else: if args.mimic_score: print("incremental_train_and_eval_MS") tg_model = incremental_train_and_eval_MS(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, args.lw_ms) else: print("incremental_train_and_eval") tg_model = incremental_train_and_eval(args.epochs, tg_model, ref_model, tg_optimizer, tg_lr_scheduler, \ trainloader, testloader, \ iteration, start_iter, args.T, args.beta) if not os.path.isdir('checkpoint'): os.mkdir('checkpoint') torch.save(tg_model, ckp_name) ### Exemplars if args.fix_budget: nb_protos_cl = int(np.ceil(args.nb_protos*100./args.nb_cl/(iteration+1))) else: nb_protos_cl = args.nb_protos tg_feature_model = nn.Sequential(*list(tg_model.children())[:-1]) num_features = tg_model.fc.in_features # Herding print('Updating exemplar set...') for iter_dico in range(last_iter*args.nb_cl, (iteration+1)*args.nb_cl): # Possible exemplars in the feature space and projected on the L2 sphere evalset.test_data = prototypes[iter_dico].astype('uint8') evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = evalset.test_data.shape[0] mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Herding procedure : ranking of the potential exemplars mu = np.mean(D,axis=1) index1 = int(iter_dico/args.nb_cl) index2 = iter_dico % args.nb_cl alpha_dr_herding[index1,:,index2] = alpha_dr_herding[index1,:,index2]*0 w_t = mu iter_herding = 0 iter_herding_eff = 0 while not(np.sum(alpha_dr_herding[index1,:,index2]!=0)==min(nb_protos_cl,500)) and iter_herding_eff<1000: tmp_t = np.dot(w_t,D) ind_max = np.argmax(tmp_t) iter_herding_eff += 1 if alpha_dr_herding[index1,ind_max,index2] == 0: alpha_dr_herding[index1,ind_max,index2] = 1+iter_herding iter_herding += 1 w_t = w_t+mu-D[:,ind_max] # Prepare the protoset X_protoset_cumuls = [] Y_protoset_cumuls = [] # Class means for iCaRL and NCM + Storing the selected exemplars in the protoset print('Computing mean-of_exemplars and theoretical mean...') class_means = np.zeros((64,100,2)) for iteration2 in range(iteration+1): for iter_dico in range(args.nb_cl): current_cl = order[range(iteration2*args.nb_cl,(iteration2+1)*args.nb_cl)] # Collect data in the feature space for each class evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico].astype('uint8') evalset.test_labels = np.zeros(evalset.test_data.shape[0]) #zero labels evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) num_samples = evalset.test_data.shape[0] mapped_prototypes = compute_features(tg_feature_model, evalloader, num_samples, num_features) D = mapped_prototypes.T D = D/np.linalg.norm(D,axis=0) # Flipped version also evalset.test_data = prototypes[iteration2*args.nb_cl+iter_dico][:,:,:,::-1].astype('uint8') evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) mapped_prototypes2 = compute_features(tg_feature_model, evalloader, num_samples, num_features) D2 = mapped_prototypes2.T D2 = D2/np.linalg.norm(D2,axis=0) # iCaRL alph = alpha_dr_herding[iteration2,:,iter_dico] alph = (alph>0)*(alph<nb_protos_cl+1)*1. X_protoset_cumuls.append(prototypes[iteration2*args.nb_cl+iter_dico,np.where(alph==1)[0]]) Y_protoset_cumuls.append(order[iteration2*args.nb_cl+iter_dico]*np.ones(len(np.where(alph==1)[0]))) alph = alph/np.sum(alph) class_means[:,current_cl[iter_dico],0] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],0] /= np.linalg.norm(class_means[:,current_cl[iter_dico],0]) # Normal NCM alph = np.ones(dictionary_size)/dictionary_size class_means[:,current_cl[iter_dico],1] = (np.dot(D,alph)+np.dot(D2,alph))/2 class_means[:,current_cl[iter_dico],1] /= np.linalg.norm(class_means[:,current_cl[iter_dico],1]) torch.save(class_means, \ './checkpoint/{}_run_{}_iteration_{}_class_means.pth'.format(args.ckp_prefix,iteration_total, iteration)) current_means = class_means[:, order[range(0,(iteration+1)*args.nb_cl)]] ############################################################## # Calculate validation error of model on the first nb_cl classes: map_Y_valid_ori = np.array([order_list.index(i) for i in Y_valid_ori]) print('Computing accuracy on the original batch of classes...') evalset.test_data = X_valid_ori.astype('uint8') evalset.test_labels = map_Y_valid_ori evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) ori_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_ori[iteration, :, iteration_total] = np.array(ori_acc).T ############################################################## # Calculate validation error of model on the cumul of classes: map_Y_valid_cumul = np.array([order_list.index(i) for i in Y_valid_cumul]) print('Computing cumulative accuracy...') evalset.test_data = X_valid_cumul.astype('uint8') evalset.test_labels = map_Y_valid_cumul evalloader = torch.utils.data.DataLoader(evalset, batch_size=eval_batch_size, shuffle=False, num_workers=2) cumul_acc = compute_accuracy(tg_model, tg_feature_model, current_means, evalloader) top1_acc_list_cumul[iteration, :, iteration_total] = np.array(cumul_acc).T ############################################################## # Calculate confusion matrix print('Computing confusion matrix...') cm = compute_confusion_matrix(tg_model, tg_feature_model, current_means, evalloader) cm_name = './checkpoint/{}_run_{}_iteration_{}_confusion_matrix.pth'.format(args.ckp_prefix,iteration_total, iteration) with open(cm_name, 'wb') as f: pickle.dump(cm, f, 2) #for reading with Python 2 ############################################################## # Final save of the data torch.save(top1_acc_list_ori, \ './checkpoint/{}_run_{}_top1_acc_list_ori.pth'.format(args.ckp_prefix, iteration_total)) torch.save(top1_acc_list_cumul, \ './checkpoint/{}_run_{}_top1_acc_list_cumul.pth'.format(args.ckp_prefix, iteration_total))
26,967
53.370968
131
py