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
DaVinci
DaVinci-main/taming/modules/vqvae/quantize.py
import torch import torch.nn as nn import torch.nn.functional as F import numpy as np from torch import einsum from einops import rearrange class VectorQuantizer(nn.Module): """ see https://github.com/MishaLaskin/vqvae/blob/d761a999e2267766400dc646d82d3ac3657771d4/models/quantizer.py ____________________________________________ Discretization bottleneck part of the VQ-VAE. Inputs: - n_e : number of embeddings - e_dim : dimension of embedding - beta : commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2 _____________________________________________ """ # NOTE: this class contains a bug regarding beta; see VectorQuantizer2 for # a fix and use legacy=False to apply that fix. VectorQuantizer2 can be # used wherever VectorQuantizer has been used before and is additionally # more efficient. def __init__(self, n_e, e_dim, beta): super(VectorQuantizer, self).__init__() self.n_e = n_e self.e_dim = e_dim self.beta = beta self.embedding = nn.Embedding(self.n_e, self.e_dim) self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e) def forward(self, z): """ Inputs the output of the encoder network z and maps it to a discrete one-hot vector that is the index of the closest embedding vector e_j z (continuous) -> z_q (discrete) z.shape = (batch, channel, height, width) quantization pipeline: 1. get encoder input (B,C,H,W) 2. flatten input to (B*H*W,C) """ # reshape z -> (batch, height, width, channel) and flatten z = z.permute(0, 2, 3, 1).contiguous() z_flattened = z.view(-1, self.e_dim) # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \ torch.sum(self.embedding.weight**2, dim=1) - 2 * \ torch.matmul(z_flattened, self.embedding.weight.t()) ## could possible replace this here # #\start... # find closest encodings min_encoding_indices = torch.argmin(d, dim=1).unsqueeze(1) min_encodings = torch.zeros( min_encoding_indices.shape[0], self.n_e).to(z) min_encodings.scatter_(1, min_encoding_indices, 1) # dtype min encodings: torch.float32 # min_encodings shape: torch.Size([2048, 512]) # min_encoding_indices.shape: torch.Size([2048, 1]) # get quantized latent vectors z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape) #.........\end # with: # .........\start #min_encoding_indices = torch.argmin(d, dim=1) #z_q = self.embedding(min_encoding_indices) # ......\end......... (TODO) # compute loss for embedding loss = torch.mean((z_q.detach()-z)**2) + self.beta * \ torch.mean((z_q - z.detach()) ** 2) # preserve gradients z_q = z + (z_q - z).detach() # perplexity e_mean = torch.mean(min_encodings, dim=0) perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10))) # reshape back to match original input shape z_q = z_q.permute(0, 3, 1, 2).contiguous() return z_q, loss, (perplexity, min_encodings, min_encoding_indices) def get_codebook_entry(self, indices, shape): # shape specifying (batch, height, width, channel) # TODO: check for more easy handling with nn.Embedding min_encodings = torch.zeros(indices.shape[0], self.n_e).to(indices) min_encodings.scatter_(1, indices[:,None], 1) # get quantized latent vectors z_q = torch.matmul(min_encodings.float(), self.embedding.weight) if shape is not None: z_q = z_q.view(shape) # reshape back to match original input shape z_q = z_q.permute(0, 3, 1, 2).contiguous() return z_q class GumbelQuantize(nn.Module): """ credit to @karpathy: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py (thanks!) Gumbel Softmax trick quantizer Categorical Reparameterization with Gumbel-Softmax, Jang et al. 2016 https://arxiv.org/abs/1611.01144 """ def __init__(self, num_hiddens, embedding_dim, n_embed, straight_through=True, kl_weight=5e-4, temp_init=1.0, use_vqinterface=True, remap=None, unknown_index="random"): super().__init__() self.embedding_dim = embedding_dim self.n_embed = n_embed self.straight_through = straight_through self.temperature = temp_init self.kl_weight = kl_weight self.proj = nn.Conv2d(num_hiddens, n_embed, 1) self.embed = nn.Embedding(n_embed, embedding_dim) self.use_vqinterface = use_vqinterface self.remap = remap if self.remap is not None: self.register_buffer("used", torch.tensor(np.load(self.remap))) self.re_embed = self.used.shape[0] self.unknown_index = unknown_index # "random" or "extra" or integer if self.unknown_index == "extra": self.unknown_index = self.re_embed self.re_embed = self.re_embed+1 print(f"Remapping {self.n_embed} indices to {self.re_embed} indices. " f"Using {self.unknown_index} for unknown indices.") else: self.re_embed = n_embed def remap_to_used(self, inds): ishape = inds.shape assert len(ishape)>1 inds = inds.reshape(ishape[0],-1) used = self.used.to(inds) match = (inds[:,:,None]==used[None,None,...]).long() new = match.argmax(-1) unknown = match.sum(2)<1 if self.unknown_index == "random": new[unknown]=torch.randint(0,self.re_embed,size=new[unknown].shape).to(device=new.device) else: new[unknown] = self.unknown_index return new.reshape(ishape) def unmap_to_all(self, inds): ishape = inds.shape assert len(ishape)>1 inds = inds.reshape(ishape[0],-1) used = self.used.to(inds) if self.re_embed > self.used.shape[0]: # extra token inds[inds>=self.used.shape[0]] = 0 # simply set to zero back=torch.gather(used[None,:][inds.shape[0]*[0],:], 1, inds) return back.reshape(ishape) def forward(self, z, temp=None, return_logits=False): # force hard = True when we are in eval mode, as we must quantize. actually, always true seems to work hard = self.straight_through if self.training else True temp = self.temperature if temp is None else temp logits = self.proj(z) if self.remap is not None: # continue only with used logits full_zeros = torch.zeros_like(logits) logits = logits[:,self.used,...] soft_one_hot = F.gumbel_softmax(logits, tau=temp, dim=1, hard=hard) if self.remap is not None: # go back to all entries but unused set to zero full_zeros[:,self.used,...] = soft_one_hot soft_one_hot = full_zeros z_q = einsum('b n h w, n d -> b d h w', soft_one_hot, self.embed.weight) # + kl divergence to the prior loss qy = F.softmax(logits, dim=1) diff = self.kl_weight * torch.sum(qy * torch.log(qy * self.n_embed + 1e-10), dim=1).mean() ind = soft_one_hot.argmax(dim=1) if self.remap is not None: ind = self.remap_to_used(ind) if self.use_vqinterface: if return_logits: return z_q, diff, (None, None, ind), logits return z_q, diff, (None, None, ind) return z_q, diff, ind def get_codebook_entry(self, indices, shape): b, h, w, c = shape assert b*h*w == indices.shape[0] indices = rearrange(indices, '(b h w) -> b h w', b=b, h=h, w=w) if self.remap is not None: indices = self.unmap_to_all(indices) one_hot = F.one_hot(indices, num_classes=self.n_embed).permute(0, 3, 1, 2).float() z_q = einsum('b n h w, n d -> b d h w', one_hot, self.embed.weight) return z_q class VectorQuantizer2(nn.Module): """ Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly avoids costly matrix multiplications and allows for post-hoc remapping of indices. """ # NOTE: due to a bug the beta term was applied to the wrong term. for # backwards compatibility we use the buggy version by default, but you can # specify legacy=False to fix it. def __init__(self, n_e, e_dim, beta, remap=None, unknown_index="random", sane_index_shape=False, legacy=True): super().__init__() self.n_e = n_e self.e_dim = e_dim self.beta = beta self.legacy = legacy self.embedding = nn.Embedding(self.n_e, self.e_dim) self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e) self.remap = remap if self.remap is not None: self.register_buffer("used", torch.tensor(np.load(self.remap))) self.re_embed = self.used.shape[0] self.unknown_index = unknown_index # "random" or "extra" or integer if self.unknown_index == "extra": self.unknown_index = self.re_embed self.re_embed = self.re_embed+1 print(f"Remapping {self.n_e} indices to {self.re_embed} indices. " f"Using {self.unknown_index} for unknown indices.") else: self.re_embed = n_e self.sane_index_shape = sane_index_shape def remap_to_used(self, inds): ishape = inds.shape assert len(ishape)>1 inds = inds.reshape(ishape[0],-1) used = self.used.to(inds) match = (inds[:,:,None]==used[None,None,...]).long() new = match.argmax(-1) unknown = match.sum(2)<1 if self.unknown_index == "random": new[unknown]=torch.randint(0,self.re_embed,size=new[unknown].shape).to(device=new.device) else: new[unknown] = self.unknown_index return new.reshape(ishape) def unmap_to_all(self, inds): ishape = inds.shape assert len(ishape)>1 inds = inds.reshape(ishape[0],-1) used = self.used.to(inds) if self.re_embed > self.used.shape[0]: # extra token inds[inds>=self.used.shape[0]] = 0 # simply set to zero back=torch.gather(used[None,:][inds.shape[0]*[0],:], 1, inds) return back.reshape(ishape) def forward(self, z, temp=None, rescale_logits=False, return_logits=False): assert temp is None or temp==1.0, "Only for interface compatible with Gumbel" assert rescale_logits==False, "Only for interface compatible with Gumbel" assert return_logits==False, "Only for interface compatible with Gumbel" # reshape z -> (batch, height, width, channel) and flatten z = rearrange(z, 'b c h w -> b h w c').contiguous() z_flattened = z.view(-1, self.e_dim) # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \ torch.sum(self.embedding.weight**2, dim=1) - 2 * \ torch.einsum('bd,dn->bn', z_flattened, rearrange(self.embedding.weight, 'n d -> d n')) min_encoding_indices = torch.argmin(d, dim=1) z_q = self.embedding(min_encoding_indices).view(z.shape) perplexity = None min_encodings = None # compute loss for embedding if not self.legacy: loss = self.beta * torch.mean((z_q.detach()-z)**2) + \ torch.mean((z_q - z.detach()) ** 2) else: loss = torch.mean((z_q.detach()-z)**2) + self.beta * \ torch.mean((z_q - z.detach()) ** 2) # preserve gradients z_q = z + (z_q - z).detach() # reshape back to match original input shape z_q = rearrange(z_q, 'b h w c -> b c h w').contiguous() if self.remap is not None: min_encoding_indices = min_encoding_indices.reshape(z.shape[0],-1) # add batch axis min_encoding_indices = self.remap_to_used(min_encoding_indices) min_encoding_indices = min_encoding_indices.reshape(-1,1) # flatten if self.sane_index_shape: min_encoding_indices = min_encoding_indices.reshape( z_q.shape[0], z_q.shape[2], z_q.shape[3]) return z_q, loss, (perplexity, min_encodings, min_encoding_indices) def get_codebook_entry(self, indices, shape): # shape specifying (batch, height, width, channel) if self.remap is not None: indices = indices.reshape(shape[0],-1) # add batch axis indices = self.unmap_to_all(indices) indices = indices.reshape(-1) # flatten again # get quantized latent vectors z_q = self.embedding(indices) if shape is not None: z_q = z_q.view(shape) # reshape back to match original input shape z_q = z_q.permute(0, 3, 1, 2).contiguous() return z_q
13,259
39.181818
110
py
DaVinci
DaVinci-main/taming/modules/discriminator/model.py
import functools import torch.nn as nn from taming.modules.util import ActNorm def weights_init(m): classname = m.__class__.__name__ if classname.find('Conv') != -1: nn.init.normal_(m.weight.data, 0.0, 0.02) elif classname.find('BatchNorm') != -1: nn.init.normal_(m.weight.data, 1.0, 0.02) nn.init.constant_(m.bias.data, 0) class NLayerDiscriminator(nn.Module): """Defines a PatchGAN discriminator as in Pix2Pix --> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py """ def __init__(self, input_nc=3, ndf=64, n_layers=3, use_actnorm=False): """Construct a PatchGAN discriminator Parameters: input_nc (int) -- the number of channels in input images ndf (int) -- the number of filters in the last conv layer n_layers (int) -- the number of conv layers in the discriminator norm_layer -- normalization layer """ super(NLayerDiscriminator, self).__init__() if not use_actnorm: norm_layer = nn.BatchNorm2d else: norm_layer = ActNorm if type(norm_layer) == functools.partial: # no need to use bias as BatchNorm2d has affine parameters use_bias = norm_layer.func != nn.BatchNorm2d else: use_bias = norm_layer != nn.BatchNorm2d kw = 4 padw = 1 sequence = [nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), nn.LeakyReLU(0.2, True)] nf_mult = 1 nf_mult_prev = 1 for n in range(1, n_layers): # gradually increase the number of filters nf_mult_prev = nf_mult nf_mult = min(2 ** n, 8) sequence += [ nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=2, padding=padw, bias=use_bias), norm_layer(ndf * nf_mult), nn.LeakyReLU(0.2, True) ] nf_mult_prev = nf_mult nf_mult = min(2 ** n_layers, 8) sequence += [ nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias), norm_layer(ndf * nf_mult), nn.LeakyReLU(0.2, True) ] sequence += [ nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)] # output 1 channel prediction map self.main = nn.Sequential(*sequence) def forward(self, input): """Standard forward.""" return self.main(input)
2,550
36.514706
116
py
DaVinci
DaVinci-main/taming/modules/misc/coord.py
import torch class CoordStage(object): def __init__(self, n_embed, down_factor): self.n_embed = n_embed self.down_factor = down_factor def eval(self): return self def encode(self, c): """fake vqmodel interface""" assert 0.0 <= c.min() and c.max() <= 1.0 b,ch,h,w = c.shape assert ch == 1 c = torch.nn.functional.interpolate(c, scale_factor=1/self.down_factor, mode="area") c = c.clamp(0.0, 1.0) c = self.n_embed*c c_quant = c.round() c_ind = c_quant.to(dtype=torch.long) info = None, None, c_ind return c_quant, None, info def decode(self, c): c = c/self.n_embed c = torch.nn.functional.interpolate(c, scale_factor=self.down_factor, mode="nearest") return c
904
27.28125
79
py
DaVinci
DaVinci-main/taming/modules/diffusionmodules/model.py
# pytorch_diffusion + derived encoder decoder import math import torch import torch.nn as nn import numpy as np def get_timestep_embedding(timesteps, embedding_dim): """ This matches the implementation in Denoising Diffusion Probabilistic Models: From Fairseq. Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of "Attention Is All You Need". """ assert len(timesteps.shape) == 1 half_dim = embedding_dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) emb = emb.to(device=timesteps.device) emb = timesteps.float()[:, None] * emb[None, :] emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) if embedding_dim % 2 == 1: # zero pad emb = torch.nn.functional.pad(emb, (0,1,0,0)) return emb def nonlinearity(x): # swish return x*torch.sigmoid(x) def Normalize(in_channels): return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) class Upsample(nn.Module): def __init__(self, in_channels, with_conv): super().__init__() self.with_conv = with_conv if self.with_conv: self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1) def forward(self, x): x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") if self.with_conv: x = self.conv(x) return x class Downsample(nn.Module): def __init__(self, in_channels, with_conv): super().__init__() self.with_conv = with_conv if self.with_conv: # no asymmetric padding in torch conv, must do it ourselves self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0) def forward(self, x): if self.with_conv: pad = (0,1,0,1) x = torch.nn.functional.pad(x, pad, mode="constant", value=0) x = self.conv(x) else: x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) return x class ResnetBlock(nn.Module): def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512): super().__init__() self.in_channels = in_channels out_channels = in_channels if out_channels is None else out_channels self.out_channels = out_channels self.use_conv_shortcut = conv_shortcut self.norm1 = Normalize(in_channels) self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) if temb_channels > 0: self.temb_proj = torch.nn.Linear(temb_channels, out_channels) self.norm2 = Normalize(out_channels) self.dropout = torch.nn.Dropout(dropout) self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1) if self.in_channels != self.out_channels: if self.use_conv_shortcut: self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) else: self.nin_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0) def forward(self, x, temb): h = x h = self.norm1(h) h = nonlinearity(h) h = self.conv1(h) if temb is not None: h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] h = self.norm2(h) h = nonlinearity(h) h = self.dropout(h) h = self.conv2(h) if self.in_channels != self.out_channels: if self.use_conv_shortcut: x = self.conv_shortcut(x) else: x = self.nin_shortcut(x) return x+h class AttnBlock(nn.Module): def __init__(self, in_channels): super().__init__() self.in_channels = in_channels self.norm = Normalize(in_channels) self.q = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) self.k = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) self.v = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) self.proj_out = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) def forward(self, x): h_ = x h_ = self.norm(h_) q = self.q(h_) k = self.k(h_) v = self.v(h_) # compute attention b,c,h,w = q.shape q = q.reshape(b,c,h*w) q = q.permute(0,2,1) # b,hw,c k = k.reshape(b,c,h*w) # b,c,hw w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] w_ = w_ * (int(c)**(-0.5)) w_ = torch.nn.functional.softmax(w_, dim=2) # attend to values v = v.reshape(b,c,h*w) w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] h_ = h_.reshape(b,c,h,w) h_ = self.proj_out(h_) return x+h_ class Model(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, resolution, use_timestep=True): super().__init__() self.ch = ch self.temb_ch = self.ch*4 self.num_resolutions = len(ch_mult) self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels self.use_timestep = use_timestep if self.use_timestep: # timestep embedding self.temb = nn.Module() self.temb.dense = nn.ModuleList([ torch.nn.Linear(self.ch, self.temb_ch), torch.nn.Linear(self.temb_ch, self.temb_ch), ]) # downsampling self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1) curr_res = resolution in_ch_mult = (1,)+tuple(ch_mult) self.down = nn.ModuleList() for i_level in range(self.num_resolutions): block = nn.ModuleList() attn = nn.ModuleList() block_in = ch*in_ch_mult[i_level] block_out = ch*ch_mult[i_level] for i_block in range(self.num_res_blocks): block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out if curr_res in attn_resolutions: attn.append(AttnBlock(block_in)) down = nn.Module() down.block = block down.attn = attn if i_level != self.num_resolutions-1: down.downsample = Downsample(block_in, resamp_with_conv) curr_res = curr_res // 2 self.down.append(down) # middle self.mid = nn.Module() self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) self.mid.attn_1 = AttnBlock(block_in) self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) # upsampling self.up = nn.ModuleList() for i_level in reversed(range(self.num_resolutions)): block = nn.ModuleList() attn = nn.ModuleList() block_out = ch*ch_mult[i_level] skip_in = ch*ch_mult[i_level] for i_block in range(self.num_res_blocks+1): if i_block == self.num_res_blocks: skip_in = ch*in_ch_mult[i_level] block.append(ResnetBlock(in_channels=block_in+skip_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out if curr_res in attn_resolutions: attn.append(AttnBlock(block_in)) up = nn.Module() up.block = block up.attn = attn if i_level != 0: up.upsample = Upsample(block_in, resamp_with_conv) curr_res = curr_res * 2 self.up.insert(0, up) # prepend to get consistent order # end self.norm_out = Normalize(block_in) self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1) def forward(self, x, t=None): #assert x.shape[2] == x.shape[3] == self.resolution if self.use_timestep: # timestep embedding assert t is not None temb = get_timestep_embedding(t, self.ch) temb = self.temb.dense[0](temb) temb = nonlinearity(temb) temb = self.temb.dense[1](temb) else: temb = None # downsampling hs = [self.conv_in(x)] for i_level in range(self.num_resolutions): for i_block in range(self.num_res_blocks): h = self.down[i_level].block[i_block](hs[-1], temb) if len(self.down[i_level].attn) > 0: h = self.down[i_level].attn[i_block](h) hs.append(h) if i_level != self.num_resolutions-1: hs.append(self.down[i_level].downsample(hs[-1])) # middle h = hs[-1] h = self.mid.block_1(h, temb) h = self.mid.attn_1(h) h = self.mid.block_2(h, temb) # upsampling for i_level in reversed(range(self.num_resolutions)): for i_block in range(self.num_res_blocks+1): h = self.up[i_level].block[i_block]( torch.cat([h, hs.pop()], dim=1), temb) if len(self.up[i_level].attn) > 0: h = self.up[i_level].attn[i_block](h) if i_level != 0: h = self.up[i_level].upsample(h) # end h = self.norm_out(h) h = nonlinearity(h) h = self.conv_out(h) return h class Encoder(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, resolution, z_channels, double_z=True, **ignore_kwargs): super().__init__() self.ch = ch self.temb_ch = 0 self.num_resolutions = len(ch_mult) self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels # downsampling self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1) curr_res = resolution in_ch_mult = (1,)+tuple(ch_mult) self.down = nn.ModuleList() for i_level in range(self.num_resolutions): block = nn.ModuleList() attn = nn.ModuleList() block_in = ch*in_ch_mult[i_level] block_out = ch*ch_mult[i_level] for i_block in range(self.num_res_blocks): block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out if curr_res in attn_resolutions: attn.append(AttnBlock(block_in)) down = nn.Module() down.block = block down.attn = attn if i_level != self.num_resolutions-1: down.downsample = Downsample(block_in, resamp_with_conv) curr_res = curr_res // 2 self.down.append(down) # middle self.mid = nn.Module() self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) self.mid.attn_1 = AttnBlock(block_in) self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) # end self.norm_out = Normalize(block_in) self.conv_out = torch.nn.Conv2d(block_in, 2*z_channels if double_z else z_channels, kernel_size=3, stride=1, padding=1) def forward(self, x): #assert x.shape[2] == x.shape[3] == self.resolution, "{}, {}, {}".format(x.shape[2], x.shape[3], self.resolution) # timestep embedding temb = None # downsampling hs = [self.conv_in(x)] for i_level in range(self.num_resolutions): for i_block in range(self.num_res_blocks): h = self.down[i_level].block[i_block](hs[-1], temb) if len(self.down[i_level].attn) > 0: h = self.down[i_level].attn[i_block](h) hs.append(h) if i_level != self.num_resolutions-1: hs.append(self.down[i_level].downsample(hs[-1])) # middle h = hs[-1] h = self.mid.block_1(h, temb) h = self.mid.attn_1(h) h = self.mid.block_2(h, temb) # end h = self.norm_out(h) h = nonlinearity(h) h = self.conv_out(h) return h class Decoder(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, resolution, z_channels, give_pre_end=False, **ignorekwargs): super().__init__() self.ch = ch self.temb_ch = 0 self.num_resolutions = len(ch_mult) self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels self.give_pre_end = give_pre_end # compute in_ch_mult, block_in and curr_res at lowest res in_ch_mult = (1,)+tuple(ch_mult) block_in = ch*ch_mult[self.num_resolutions-1] curr_res = resolution // 2**(self.num_resolutions-1) self.z_shape = (1,z_channels,curr_res,curr_res) print("Working with z of shape {} = {} dimensions.".format( self.z_shape, np.prod(self.z_shape))) # z to block_in self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1) # middle self.mid = nn.Module() self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) self.mid.attn_1 = AttnBlock(block_in) self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) # upsampling self.up = nn.ModuleList() for i_level in reversed(range(self.num_resolutions)): block = nn.ModuleList() attn = nn.ModuleList() block_out = ch*ch_mult[i_level] for i_block in range(self.num_res_blocks+1): block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out if curr_res in attn_resolutions: attn.append(AttnBlock(block_in)) up = nn.Module() up.block = block up.attn = attn if i_level != 0: up.upsample = Upsample(block_in, resamp_with_conv) curr_res = curr_res * 2 self.up.insert(0, up) # prepend to get consistent order # end self.norm_out = Normalize(block_in) self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1) def forward(self, z): #assert z.shape[1:] == self.z_shape[1:] self.last_z_shape = z.shape # timestep embedding temb = None # z to block_in h = self.conv_in(z) # middle h = self.mid.block_1(h, temb) h = self.mid.attn_1(h) h = self.mid.block_2(h, temb) # upsampling for i_level in reversed(range(self.num_resolutions)): for i_block in range(self.num_res_blocks+1): h = self.up[i_level].block[i_block](h, temb) if len(self.up[i_level].attn) > 0: h = self.up[i_level].attn[i_block](h) if i_level != 0: h = self.up[i_level].upsample(h) # end if self.give_pre_end: return h h = self.norm_out(h) h = nonlinearity(h) h = self.conv_out(h) return h class VUNet(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, c_channels, resolution, z_channels, use_timestep=False, **ignore_kwargs): super().__init__() self.ch = ch self.temb_ch = self.ch*4 self.num_resolutions = len(ch_mult) self.num_res_blocks = num_res_blocks self.resolution = resolution self.use_timestep = use_timestep if self.use_timestep: # timestep embedding self.temb = nn.Module() self.temb.dense = nn.ModuleList([ torch.nn.Linear(self.ch, self.temb_ch), torch.nn.Linear(self.temb_ch, self.temb_ch), ]) # downsampling self.conv_in = torch.nn.Conv2d(c_channels, self.ch, kernel_size=3, stride=1, padding=1) curr_res = resolution in_ch_mult = (1,)+tuple(ch_mult) self.down = nn.ModuleList() for i_level in range(self.num_resolutions): block = nn.ModuleList() attn = nn.ModuleList() block_in = ch*in_ch_mult[i_level] block_out = ch*ch_mult[i_level] for i_block in range(self.num_res_blocks): block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out if curr_res in attn_resolutions: attn.append(AttnBlock(block_in)) down = nn.Module() down.block = block down.attn = attn if i_level != self.num_resolutions-1: down.downsample = Downsample(block_in, resamp_with_conv) curr_res = curr_res // 2 self.down.append(down) self.z_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=1, stride=1, padding=0) # middle self.mid = nn.Module() self.mid.block_1 = ResnetBlock(in_channels=2*block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) self.mid.attn_1 = AttnBlock(block_in) self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout) # upsampling self.up = nn.ModuleList() for i_level in reversed(range(self.num_resolutions)): block = nn.ModuleList() attn = nn.ModuleList() block_out = ch*ch_mult[i_level] skip_in = ch*ch_mult[i_level] for i_block in range(self.num_res_blocks+1): if i_block == self.num_res_blocks: skip_in = ch*in_ch_mult[i_level] block.append(ResnetBlock(in_channels=block_in+skip_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out if curr_res in attn_resolutions: attn.append(AttnBlock(block_in)) up = nn.Module() up.block = block up.attn = attn if i_level != 0: up.upsample = Upsample(block_in, resamp_with_conv) curr_res = curr_res * 2 self.up.insert(0, up) # prepend to get consistent order # end self.norm_out = Normalize(block_in) self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1) def forward(self, x, z): #assert x.shape[2] == x.shape[3] == self.resolution if self.use_timestep: # timestep embedding assert t is not None temb = get_timestep_embedding(t, self.ch) temb = self.temb.dense[0](temb) temb = nonlinearity(temb) temb = self.temb.dense[1](temb) else: temb = None # downsampling hs = [self.conv_in(x)] for i_level in range(self.num_resolutions): for i_block in range(self.num_res_blocks): h = self.down[i_level].block[i_block](hs[-1], temb) if len(self.down[i_level].attn) > 0: h = self.down[i_level].attn[i_block](h) hs.append(h) if i_level != self.num_resolutions-1: hs.append(self.down[i_level].downsample(hs[-1])) # middle h = hs[-1] z = self.z_in(z) h = torch.cat((h,z),dim=1) h = self.mid.block_1(h, temb) h = self.mid.attn_1(h) h = self.mid.block_2(h, temb) # upsampling for i_level in reversed(range(self.num_resolutions)): for i_block in range(self.num_res_blocks+1): h = self.up[i_level].block[i_block]( torch.cat([h, hs.pop()], dim=1), temb) if len(self.up[i_level].attn) > 0: h = self.up[i_level].attn[i_block](h) if i_level != 0: h = self.up[i_level].upsample(h) # end h = self.norm_out(h) h = nonlinearity(h) h = self.conv_out(h) return h class SimpleDecoder(nn.Module): def __init__(self, in_channels, out_channels, *args, **kwargs): super().__init__() self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), ResnetBlock(in_channels=in_channels, out_channels=2 * in_channels, temb_channels=0, dropout=0.0), ResnetBlock(in_channels=2 * in_channels, out_channels=4 * in_channels, temb_channels=0, dropout=0.0), ResnetBlock(in_channels=4 * in_channels, out_channels=2 * in_channels, temb_channels=0, dropout=0.0), nn.Conv2d(2*in_channels, in_channels, 1), Upsample(in_channels, with_conv=True)]) # end self.norm_out = Normalize(in_channels) self.conv_out = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) def forward(self, x): for i, layer in enumerate(self.model): if i in [1,2,3]: x = layer(x, None) else: x = layer(x) h = self.norm_out(x) h = nonlinearity(h) x = self.conv_out(h) return x class UpsampleDecoder(nn.Module): def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, ch_mult=(2,2), dropout=0.0): super().__init__() # upsampling self.temb_ch = 0 self.num_resolutions = len(ch_mult) self.num_res_blocks = num_res_blocks block_in = in_channels curr_res = resolution // 2 ** (self.num_resolutions - 1) self.res_blocks = nn.ModuleList() self.upsample_blocks = nn.ModuleList() for i_level in range(self.num_resolutions): res_block = [] block_out = ch * ch_mult[i_level] for i_block in range(self.num_res_blocks + 1): res_block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout)) block_in = block_out self.res_blocks.append(nn.ModuleList(res_block)) if i_level != self.num_resolutions - 1: self.upsample_blocks.append(Upsample(block_in, True)) curr_res = curr_res * 2 # end self.norm_out = Normalize(block_in) self.conv_out = torch.nn.Conv2d(block_in, out_channels, kernel_size=3, stride=1, padding=1) def forward(self, x): # upsampling h = x for k, i_level in enumerate(range(self.num_resolutions)): for i_block in range(self.num_res_blocks + 1): h = self.res_blocks[i_level][i_block](h, None) if i_level != self.num_resolutions - 1: h = self.upsample_blocks[k](h) h = self.norm_out(h) h = nonlinearity(h) h = self.conv_out(h) return h
30,221
37.895753
121
py
DaVinci
DaVinci-main/taming/modules/transformer/mingpt.py
""" taken from: https://github.com/karpathy/minGPT/ GPT model: - the initial stem consists of a combination of token encoding and a positional encoding - the meat of it is a uniform sequence of Transformer blocks - each Transformer is a sequential combination of a 1-hidden-layer MLP block and a self-attention block - all blocks feed into a central residual pathway similar to resnets - the final decoder is a linear projection into a vanilla Softmax classifier """ import math import logging import torch import torch.nn as nn from torch.nn import functional as F from transformers import top_k_top_p_filtering logger = logging.getLogger(__name__) class GPTConfig: """ base GPT config, params common to all GPT versions """ embd_pdrop = 0.1 resid_pdrop = 0.1 attn_pdrop = 0.1 def __init__(self, vocab_size, block_size, **kwargs): self.vocab_size = vocab_size self.block_size = block_size for k,v in kwargs.items(): setattr(self, k, v) class GPT1Config(GPTConfig): """ GPT-1 like network roughly 125M params """ n_layer = 12 n_head = 12 n_embd = 768 class CausalSelfAttention(nn.Module): """ A vanilla multi-head masked self-attention layer with a projection at the end. It is possible to use torch.nn.MultiheadAttention here but I am including an explicit implementation here to show that there is nothing too scary here. """ def __init__(self, config): super().__init__() assert config.n_embd % config.n_head == 0 # key, query, value projections for all heads self.key = nn.Linear(config.n_embd, config.n_embd) self.query = nn.Linear(config.n_embd, config.n_embd) self.value = nn.Linear(config.n_embd, config.n_embd) # regularization self.attn_drop = nn.Dropout(config.attn_pdrop) self.resid_drop = nn.Dropout(config.resid_pdrop) # output projection self.proj = nn.Linear(config.n_embd, config.n_embd) # causal mask to ensure that attention is only applied to the left in the input sequence mask = torch.tril(torch.ones(config.block_size, config.block_size)) if hasattr(config, "n_unmasked"): mask[:config.n_unmasked, :config.n_unmasked] = 1 self.register_buffer("mask", mask.view(1, 1, config.block_size, config.block_size)) self.n_head = config.n_head def forward(self, x, layer_past=None): B, T, C = x.size() # calculate query, key, values for all heads in batch and move head forward to be the batch dim k = self.key(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs) q = self.query(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs) v = self.value(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs) present = torch.stack((k, v)) if layer_past is not None: past_key, past_value = layer_past k = torch.cat((past_key, k), dim=-2) v = torch.cat((past_value, v), dim=-2) # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T) att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1))) if layer_past is None: att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf')) att = F.softmax(att, dim=-1) att = self.attn_drop(att) y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs) y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side # output projection y = self.resid_drop(self.proj(y)) return y, present # TODO: check that this does not break anything class Block(nn.Module): """ an unassuming Transformer block """ def __init__(self, config): super().__init__() self.ln1 = nn.LayerNorm(config.n_embd) self.ln2 = nn.LayerNorm(config.n_embd) self.attn = CausalSelfAttention(config) self.mlp = nn.Sequential( nn.Linear(config.n_embd, 4 * config.n_embd), nn.GELU(), # nice nn.Linear(4 * config.n_embd, config.n_embd), nn.Dropout(config.resid_pdrop), ) def forward(self, x, layer_past=None, return_present=False): # TODO: check that training still works if return_present: assert not self.training # layer past: tuple of length two with B, nh, T, hs attn, present = self.attn(self.ln1(x), layer_past=layer_past) x = x + attn x = x + self.mlp(self.ln2(x)) if layer_past is not None or return_present: return x, present return x class GPT(nn.Module): """ the full GPT language model, with a context size of block_size """ def __init__(self, vocab_size, block_size, n_layer=12, n_head=8, n_embd=256, embd_pdrop=0., resid_pdrop=0., attn_pdrop=0., n_unmasked=0): super().__init__() config = GPTConfig(vocab_size=vocab_size, block_size=block_size, embd_pdrop=embd_pdrop, resid_pdrop=resid_pdrop, attn_pdrop=attn_pdrop, n_layer=n_layer, n_head=n_head, n_embd=n_embd, n_unmasked=n_unmasked) # input embedding stem self.tok_emb = nn.Embedding(config.vocab_size, config.n_embd) self.pos_emb = nn.Parameter(torch.zeros(1, config.block_size, config.n_embd)) self.drop = nn.Dropout(config.embd_pdrop) # transformer self.blocks = nn.Sequential(*[Block(config) for _ in range(config.n_layer)]) # decoder head self.ln_f = nn.LayerNorm(config.n_embd) self.head = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.block_size = config.block_size self.apply(self._init_weights) self.config = config logger.info("number of parameters: %e", sum(p.numel() for p in self.parameters())) def get_block_size(self): return self.block_size def _init_weights(self, module): if isinstance(module, (nn.Linear, nn.Embedding)): module.weight.data.normal_(mean=0.0, std=0.02) if isinstance(module, nn.Linear) and module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.LayerNorm): module.bias.data.zero_() module.weight.data.fill_(1.0) def forward(self, idx, embeddings=None, targets=None): # forward the GPT model token_embeddings = self.tok_emb(idx) # each index maps to a (learnable) vector if embeddings is not None: # prepend explicit embeddings token_embeddings = torch.cat((embeddings, token_embeddings), dim=1) t = token_embeddings.shape[1] assert t <= self.block_size, "Cannot forward, model block size is exhausted." position_embeddings = self.pos_emb[:, :t, :] # each position maps to a (learnable) vector x = self.drop(token_embeddings + position_embeddings) x = self.blocks(x) x = self.ln_f(x) logits = self.head(x) # if we are given some desired targets also calculate the loss loss = None if targets is not None: loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1)) return logits, loss def forward_with_past(self, idx, embeddings=None, targets=None, past=None, past_length=None): # inference only assert not self.training token_embeddings = self.tok_emb(idx) # each index maps to a (learnable) vector if embeddings is not None: # prepend explicit embeddings token_embeddings = torch.cat((embeddings, token_embeddings), dim=1) if past is not None: assert past_length is not None past = torch.cat(past, dim=-2) # n_layer, 2, b, nh, len_past, dim_head past_shape = list(past.shape) expected_shape = [self.config.n_layer, 2, idx.shape[0], self.config.n_head, past_length, self.config.n_embd//self.config.n_head] assert past_shape == expected_shape, f"{past_shape} =/= {expected_shape}" position_embeddings = self.pos_emb[:, past_length, :] # each position maps to a (learnable) vector else: position_embeddings = self.pos_emb[:, :token_embeddings.shape[1], :] x = self.drop(token_embeddings + position_embeddings) presents = [] # accumulate over layers for i, block in enumerate(self.blocks): x, present = block(x, layer_past=past[i, ...] if past is not None else None, return_present=True) presents.append(present) x = self.ln_f(x) logits = self.head(x) # if we are given some desired targets also calculate the loss loss = None if targets is not None: loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1)) return logits, loss, torch.stack(presents) # _, _, n_layer, 2, b, nh, 1, dim_head class DummyGPT(nn.Module): # for debugging def __init__(self, add_value=1): super().__init__() self.add_value = add_value def forward(self, idx): return idx + self.add_value, None class CodeGPT(nn.Module): """Takes in semi-embeddings""" def __init__(self, vocab_size, block_size, in_channels, n_layer=12, n_head=8, n_embd=256, embd_pdrop=0., resid_pdrop=0., attn_pdrop=0., n_unmasked=0): super().__init__() config = GPTConfig(vocab_size=vocab_size, block_size=block_size, embd_pdrop=embd_pdrop, resid_pdrop=resid_pdrop, attn_pdrop=attn_pdrop, n_layer=n_layer, n_head=n_head, n_embd=n_embd, n_unmasked=n_unmasked) # input embedding stem self.tok_emb = nn.Linear(in_channels, config.n_embd) self.pos_emb = nn.Parameter(torch.zeros(1, config.block_size, config.n_embd)) self.drop = nn.Dropout(config.embd_pdrop) # transformer self.blocks = nn.Sequential(*[Block(config) for _ in range(config.n_layer)]) # decoder head self.ln_f = nn.LayerNorm(config.n_embd) self.head = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.block_size = config.block_size self.apply(self._init_weights) self.config = config logger.info("number of parameters: %e", sum(p.numel() for p in self.parameters())) def get_block_size(self): return self.block_size def _init_weights(self, module): if isinstance(module, (nn.Linear, nn.Embedding)): module.weight.data.normal_(mean=0.0, std=0.02) if isinstance(module, nn.Linear) and module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.LayerNorm): module.bias.data.zero_() module.weight.data.fill_(1.0) def forward(self, idx, embeddings=None, targets=None): # forward the GPT model token_embeddings = self.tok_emb(idx) # each index maps to a (learnable) vector if embeddings is not None: # prepend explicit embeddings token_embeddings = torch.cat((embeddings, token_embeddings), dim=1) t = token_embeddings.shape[1] assert t <= self.block_size, "Cannot forward, model block size is exhausted." position_embeddings = self.pos_emb[:, :t, :] # each position maps to a (learnable) vector x = self.drop(token_embeddings + position_embeddings) x = self.blocks(x) x = self.taming_cinln_f(x) logits = self.head(x) # if we are given some desired targets also calculate the loss loss = None if targets is not None: loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1)) return logits, loss #### sampling utils def top_k_logits(logits, k): v, ix = torch.topk(logits, k) out = logits.clone() out[out < v[:, [-1]]] = -float('Inf') return out @torch.no_grad() def sample(model, x, steps, temperature=1.0, sample=False, top_k=None): """ take a conditioning sequence of indices in x (of shape (b,t)) and predict the next token in the sequence, feeding the predictions back into the model each time. Clearly the sampling has quadratic complexity unlike an RNN that is only linear, and has a finite context window of block_size, unlike an RNN that has an infinite context window. """ block_size = model.get_block_size() model.eval() for k in range(steps): x_cond = x if x.size(1) <= block_size else x[:, -block_size:] # crop context if needed logits, _ = model(x_cond) # pluck the logits at the final step and scale by temperature logits = logits[:, -1, :] / temperature # optionally crop probabilities to only the top k options if top_k is not None: logits = top_k_logits(logits, top_k) # apply softmax to convert to probabilities probs = F.softmax(logits, dim=-1) # sample from the distribution or take the most likely if sample: ix = torch.multinomial(probs, num_samples=1) else: _, ix = torch.topk(probs, k=1, dim=-1) # append to the sequence and continue x = torch.cat((x, ix), dim=1) return x @torch.no_grad() def sample_with_past(x, model, steps, temperature=1., sample_logits=True, top_k=None, top_p=None, callback=None): # x is conditioning sample = x cond_len = x.shape[1] past = None for n in range(steps): if callback is not None: callback(n) logits, _, present = model.forward_with_past(x, past=past, past_length=(n+cond_len-1)) if past is None: past = [present] else: past.append(present) logits = logits[:, -1, :] / temperature if top_k is not None: logits = top_k_top_p_filtering(logits, top_k=top_k, top_p=top_p) probs = F.softmax(logits, dim=-1) if not sample_logits: _, x = torch.topk(probs, k=1, dim=-1) else: x = torch.multinomial(probs, num_samples=1) # append to the sequence and continue sample = torch.cat((sample, x), dim=1) del past sample = sample[:, cond_len:] # cut conditioning off return sample #### clustering utils class KMeans(nn.Module): def __init__(self, ncluster=512, nc=3, niter=10): super().__init__() self.ncluster = ncluster self.nc = nc self.niter = niter self.shape = (3,32,32) self.register_buffer("C", torch.zeros(self.ncluster,nc)) self.register_buffer('initialized', torch.tensor(0, dtype=torch.uint8)) def is_initialized(self): return self.initialized.item() == 1 @torch.no_grad() def initialize(self, x): N, D = x.shape assert D == self.nc, D c = x[torch.randperm(N)[:self.ncluster]] # init clusters at random for i in range(self.niter): # assign all pixels to the closest codebook element a = ((x[:, None, :] - c[None, :, :])**2).sum(-1).argmin(1) # move each codebook element to be the mean of the pixels that assigned to it c = torch.stack([x[a==k].mean(0) for k in range(self.ncluster)]) # re-assign any poorly positioned codebook elements nanix = torch.any(torch.isnan(c), dim=1) ndead = nanix.sum().item() print('done step %d/%d, re-initialized %d dead clusters' % (i+1, self.niter, ndead)) c[nanix] = x[torch.randperm(N)[:ndead]] # re-init dead clusters self.C.copy_(c) self.initialized.fill_(1) def forward(self, x, reverse=False, shape=None): if not reverse: # flatten bs,c,h,w = x.shape assert c == self.nc x = x.reshape(bs,c,h*w,1) C = self.C.permute(1,0) C = C.reshape(1,c,1,self.ncluster) a = ((x-C)**2).sum(1).argmin(-1) # bs, h*w indices return a else: # flatten bs, HW = x.shape """ c = self.C.reshape( 1, self.nc, 1, self.ncluster) c = c[bs*[0],:,:,:] c = c[:,:,HW*[0],:] x = x.reshape(bs, 1, HW, 1) x = x[:,3*[0],:,:] x = torch.gather(c, dim=3, index=x) """ x = self.C[x] x = x.permute(0,2,1) shape = shape if shape is not None else self.shape x = x.reshape(bs, *shape) return x
16,836
39.473558
140
py
DaVinci
DaVinci-main/taming/modules/transformer/permuter.py
import torch import torch.nn as nn import numpy as np class AbstractPermuter(nn.Module): def __init__(self, *args, **kwargs): super().__init__() def forward(self, x, reverse=False): raise NotImplementedError class Identity(AbstractPermuter): def __init__(self): super().__init__() def forward(self, x, reverse=False): return x class Subsample(AbstractPermuter): def __init__(self, H, W): super().__init__() C = 1 indices = np.arange(H*W).reshape(C,H,W) while min(H, W) > 1: indices = indices.reshape(C,H//2,2,W//2,2) indices = indices.transpose(0,2,4,1,3) indices = indices.reshape(C*4,H//2, W//2) H = H//2 W = W//2 C = C*4 assert H == W == 1 idx = torch.tensor(indices.ravel()) self.register_buffer('forward_shuffle_idx', nn.Parameter(idx, requires_grad=False)) self.register_buffer('backward_shuffle_idx', nn.Parameter(torch.argsort(idx), requires_grad=False)) def forward(self, x, reverse=False): if not reverse: return x[:, self.forward_shuffle_idx] else: return x[:, self.backward_shuffle_idx] def mortonify(i, j): """(i,j) index to linear morton code""" i = np.uint64(i) j = np.uint64(j) z = np.uint(0) for pos in range(32): z = (z | ((j & (np.uint64(1) << np.uint64(pos))) << np.uint64(pos)) | ((i & (np.uint64(1) << np.uint64(pos))) << np.uint64(pos+1)) ) return z class ZCurve(AbstractPermuter): def __init__(self, H, W): super().__init__() reverseidx = [np.int64(mortonify(i,j)) for i in range(H) for j in range(W)] idx = np.argsort(reverseidx) idx = torch.tensor(idx) reverseidx = torch.tensor(reverseidx) self.register_buffer('forward_shuffle_idx', idx) self.register_buffer('backward_shuffle_idx', reverseidx) def forward(self, x, reverse=False): if not reverse: return x[:, self.forward_shuffle_idx] else: return x[:, self.backward_shuffle_idx] class SpiralOut(AbstractPermuter): def __init__(self, H, W): super().__init__() assert H == W size = W indices = np.arange(size*size).reshape(size,size) i0 = size//2 j0 = size//2-1 i = i0 j = j0 idx = [indices[i0, j0]] step_mult = 0 for c in range(1, size//2+1): step_mult += 1 # steps left for k in range(step_mult): i = i - 1 j = j idx.append(indices[i, j]) # step down for k in range(step_mult): i = i j = j + 1 idx.append(indices[i, j]) step_mult += 1 if c < size//2: # step right for k in range(step_mult): i = i + 1 j = j idx.append(indices[i, j]) # step up for k in range(step_mult): i = i j = j - 1 idx.append(indices[i, j]) else: # end reached for k in range(step_mult-1): i = i + 1 idx.append(indices[i, j]) assert len(idx) == size*size idx = torch.tensor(idx) self.register_buffer('forward_shuffle_idx', idx) self.register_buffer('backward_shuffle_idx', torch.argsort(idx)) def forward(self, x, reverse=False): if not reverse: return x[:, self.forward_shuffle_idx] else: return x[:, self.backward_shuffle_idx] class SpiralIn(AbstractPermuter): def __init__(self, H, W): super().__init__() assert H == W size = W indices = np.arange(size*size).reshape(size,size) i0 = size//2 j0 = size//2-1 i = i0 j = j0 idx = [indices[i0, j0]] step_mult = 0 for c in range(1, size//2+1): step_mult += 1 # steps left for k in range(step_mult): i = i - 1 j = j idx.append(indices[i, j]) # step down for k in range(step_mult): i = i j = j + 1 idx.append(indices[i, j]) step_mult += 1 if c < size//2: # step right for k in range(step_mult): i = i + 1 j = j idx.append(indices[i, j]) # step up for k in range(step_mult): i = i j = j - 1 idx.append(indices[i, j]) else: # end reached for k in range(step_mult-1): i = i + 1 idx.append(indices[i, j]) assert len(idx) == size*size idx = idx[::-1] idx = torch.tensor(idx) self.register_buffer('forward_shuffle_idx', idx) self.register_buffer('backward_shuffle_idx', torch.argsort(idx)) def forward(self, x, reverse=False): if not reverse: return x[:, self.forward_shuffle_idx] else: return x[:, self.backward_shuffle_idx] class Random(nn.Module): def __init__(self, H, W): super().__init__() indices = np.random.RandomState(1).permutation(H*W) idx = torch.tensor(indices.ravel()) self.register_buffer('forward_shuffle_idx', idx) self.register_buffer('backward_shuffle_idx', torch.argsort(idx)) def forward(self, x, reverse=False): if not reverse: return x[:, self.forward_shuffle_idx] else: return x[:, self.backward_shuffle_idx] class AlternateParsing(AbstractPermuter): def __init__(self, H, W): super().__init__() indices = np.arange(W*H).reshape(H,W) for i in range(1, H, 2): indices[i, :] = indices[i, ::-1] idx = indices.flatten() assert len(idx) == H*W idx = torch.tensor(idx) self.register_buffer('forward_shuffle_idx', idx) self.register_buffer('backward_shuffle_idx', torch.argsort(idx)) def forward(self, x, reverse=False): if not reverse: return x[:, self.forward_shuffle_idx] else: return x[:, self.backward_shuffle_idx] if __name__ == "__main__": p0 = AlternateParsing(16, 16) print(p0.forward_shuffle_idx) print(p0.backward_shuffle_idx) x = torch.randint(0, 768, size=(11, 256)) y = p0(x) xre = p0(y, reverse=True) assert torch.equal(x, xre) p1 = SpiralOut(2, 2) print(p1.forward_shuffle_idx) print(p1.backward_shuffle_idx)
7,093
27.48996
83
py
DaVinci
DaVinci-main/taming/modules/losses/lpips.py
"""Stripped version of https://github.com/richzhang/PerceptualSimilarity/tree/master/models""" import torch import torch.nn as nn from torchvision import models from collections import namedtuple from taming.util import get_ckpt_path class LPIPS(nn.Module): # Learned perceptual metric def __init__(self, use_dropout=True): super().__init__() self.scaling_layer = ScalingLayer() self.chns = [64, 128, 256, 512, 512] # vg16 features self.net = vgg16(pretrained=True, requires_grad=False) self.lin0 = NetLinLayer(self.chns[0], use_dropout=use_dropout) self.lin1 = NetLinLayer(self.chns[1], use_dropout=use_dropout) self.lin2 = NetLinLayer(self.chns[2], use_dropout=use_dropout) self.lin3 = NetLinLayer(self.chns[3], use_dropout=use_dropout) self.lin4 = NetLinLayer(self.chns[4], use_dropout=use_dropout) self.load_from_pretrained() for param in self.parameters(): param.requires_grad = False def load_from_pretrained(self, name="vgg_lpips"): ckpt = get_ckpt_path(name, "taming/modules/autoencoder/lpips") self.load_state_dict(torch.load(ckpt, map_location=torch.device("cpu")), strict=False) print("loaded pretrained LPIPS loss from {}".format(ckpt)) @classmethod def from_pretrained(cls, name="vgg_lpips"): if name is not "vgg_lpips": raise NotImplementedError model = cls() ckpt = get_ckpt_path(name) model.load_state_dict(torch.load(ckpt, map_location=torch.device("cpu")), strict=False) return model def forward(self, input, target): in0_input, in1_input = (self.scaling_layer(input), self.scaling_layer(target)) outs0, outs1 = self.net(in0_input), self.net(in1_input) feats0, feats1, diffs = {}, {}, {} lins = [self.lin0, self.lin1, self.lin2, self.lin3, self.lin4] for kk in range(len(self.chns)): feats0[kk], feats1[kk] = normalize_tensor(outs0[kk]), normalize_tensor(outs1[kk]) diffs[kk] = (feats0[kk] - feats1[kk]) ** 2 res = [spatial_average(lins[kk].model(diffs[kk]), keepdim=True) for kk in range(len(self.chns))] val = res[0] for l in range(1, len(self.chns)): val += res[l] return val class ScalingLayer(nn.Module): def __init__(self): super(ScalingLayer, self).__init__() self.register_buffer('shift', torch.Tensor([-.030, -.088, -.188])[None, :, None, None]) self.register_buffer('scale', torch.Tensor([.458, .448, .450])[None, :, None, None]) def forward(self, inp): return (inp - self.shift) / self.scale class NetLinLayer(nn.Module): """ A single linear layer which does a 1x1 conv """ def __init__(self, chn_in, chn_out=1, use_dropout=False): super(NetLinLayer, self).__init__() layers = [nn.Dropout(), ] if (use_dropout) else [] layers += [nn.Conv2d(chn_in, chn_out, 1, stride=1, padding=0, bias=False), ] self.model = nn.Sequential(*layers) class vgg16(torch.nn.Module): def __init__(self, requires_grad=False, pretrained=True): super(vgg16, self).__init__() vgg_pretrained_features = models.vgg16(pretrained=pretrained).features self.slice1 = torch.nn.Sequential() self.slice2 = torch.nn.Sequential() self.slice3 = torch.nn.Sequential() self.slice4 = torch.nn.Sequential() self.slice5 = torch.nn.Sequential() self.N_slices = 5 for x in range(4): self.slice1.add_module(str(x), vgg_pretrained_features[x]) for x in range(4, 9): self.slice2.add_module(str(x), vgg_pretrained_features[x]) for x in range(9, 16): self.slice3.add_module(str(x), vgg_pretrained_features[x]) for x in range(16, 23): self.slice4.add_module(str(x), vgg_pretrained_features[x]) for x in range(23, 30): self.slice5.add_module(str(x), vgg_pretrained_features[x]) if not requires_grad: for param in self.parameters(): param.requires_grad = False def forward(self, X): h = self.slice1(X) h_relu1_2 = h h = self.slice2(h) h_relu2_2 = h h = self.slice3(h) h_relu3_3 = h h = self.slice4(h) h_relu4_3 = h h = self.slice5(h) h_relu5_3 = h vgg_outputs = namedtuple("VggOutputs", ['relu1_2', 'relu2_2', 'relu3_3', 'relu4_3', 'relu5_3']) out = vgg_outputs(h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3) return out def normalize_tensor(x,eps=1e-10): norm_factor = torch.sqrt(torch.sum(x**2,dim=1,keepdim=True)) return x/(norm_factor+eps) def spatial_average(x, keepdim=True): return x.mean([2,3],keepdim=keepdim)
4,836
38.008065
104
py
DaVinci
DaVinci-main/taming/modules/losses/segmentation.py
import torch.nn as nn import torch.nn.functional as F class BCELoss(nn.Module): def forward(self, prediction, target): loss = F.binary_cross_entropy_with_logits(prediction,target) return loss, {} class BCELossWithQuant(nn.Module): def __init__(self, codebook_weight=1.): super().__init__() self.codebook_weight = codebook_weight def forward(self, qloss, target, prediction, split): bce_loss = F.binary_cross_entropy_with_logits(prediction,target) loss = bce_loss + self.codebook_weight*qloss return loss, {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/bce_loss".format(split): bce_loss.detach().mean(), "{}/quant_loss".format(split): qloss.detach().mean() }
816
34.521739
82
py
DaVinci
DaVinci-main/taming/modules/losses/vqperceptual.py
import torch import torch.nn as nn import torch.nn.functional as F from taming.modules.losses.lpips import LPIPS from taming.modules.discriminator.model import NLayerDiscriminator, weights_init class DummyLoss(nn.Module): def __init__(self): super().__init__() def adopt_weight(weight, global_step, threshold=0, value=0.): if global_step < threshold: weight = value return weight def hinge_d_loss(logits_real, logits_fake): loss_real = torch.mean(F.relu(1. - logits_real)) loss_fake = torch.mean(F.relu(1. + logits_fake)) d_loss = 0.5 * (loss_real + loss_fake) return d_loss def vanilla_d_loss(logits_real, logits_fake): d_loss = 0.5 * ( torch.mean(torch.nn.functional.softplus(-logits_real)) + torch.mean(torch.nn.functional.softplus(logits_fake))) return d_loss class VQLPIPSWithDiscriminator(nn.Module): def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0, disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0, perceptual_weight=1.0, use_actnorm=False, disc_conditional=False, disc_ndf=64, disc_loss="hinge"): super().__init__() assert disc_loss in ["hinge", "vanilla"] self.codebook_weight = codebook_weight self.pixel_weight = pixelloss_weight self.perceptual_loss = LPIPS().eval() self.perceptual_weight = perceptual_weight self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels, n_layers=disc_num_layers, use_actnorm=use_actnorm, ndf=disc_ndf ).apply(weights_init) self.discriminator_iter_start = disc_start if disc_loss == "hinge": self.disc_loss = hinge_d_loss elif disc_loss == "vanilla": self.disc_loss = vanilla_d_loss else: raise ValueError(f"Unknown GAN loss '{disc_loss}'.") print(f"VQLPIPSWithDiscriminator running with {disc_loss} loss.") self.disc_factor = disc_factor self.discriminator_weight = disc_weight self.disc_conditional = disc_conditional def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None): if last_layer is not None: nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0] g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0] else: nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0] g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0] d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4) d_weight = torch.clamp(d_weight, 0.0, 1e4).detach() d_weight = d_weight * self.discriminator_weight return d_weight def forward(self, codebook_loss, inputs, reconstructions, optimizer_idx, global_step, last_layer=None, cond=None, split="train"): rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous()) if self.perceptual_weight > 0: p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous()) rec_loss = rec_loss + self.perceptual_weight * p_loss else: p_loss = torch.tensor([0.0]) nll_loss = rec_loss #nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] nll_loss = torch.mean(nll_loss) # now the GAN part if optimizer_idx == 0: # generator update if cond is None: assert not self.disc_conditional logits_fake = self.discriminator(reconstructions.contiguous()) else: assert self.disc_conditional logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1)) g_loss = -torch.mean(logits_fake) try: d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer) except RuntimeError: assert not self.training d_weight = torch.tensor(0.0) disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start) loss = nll_loss + d_weight * disc_factor * g_loss + self.codebook_weight * codebook_loss.mean() log = {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/quant_loss".format(split): codebook_loss.detach().mean(), "{}/nll_loss".format(split): nll_loss.detach().mean(), "{}/rec_loss".format(split): rec_loss.detach().mean(), "{}/p_loss".format(split): p_loss.detach().mean(), "{}/d_weight".format(split): d_weight.detach(), "{}/disc_factor".format(split): torch.tensor(disc_factor), "{}/g_loss".format(split): g_loss.detach().mean(), } return loss, log if optimizer_idx == 1: # second pass for discriminator update if cond is None: logits_real = self.discriminator(inputs.contiguous().detach()) logits_fake = self.discriminator(reconstructions.contiguous().detach()) else: logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1)) logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1)) disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start) d_loss = disc_factor * self.disc_loss(logits_real, logits_fake) log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(), "{}/logits_real".format(split): logits_real.detach().mean(), "{}/logits_fake".format(split): logits_fake.detach().mean() } return d_loss, log
6,179
44.109489
113
py
DaVinci
DaVinci-main/taming/models/vqgan.py
import torch import torch.nn.functional as F import pytorch_lightning as pl from taming.main import instantiate_from_config from taming.modules.diffusionmodules.model import Encoder, Decoder from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer from taming.modules.vqvae.quantize import GumbelQuantize class VQModel(pl.LightningModule): def __init__(self, ddconfig, lossconfig, n_embed, embed_dim, ckpt_path=None, ignore_keys=[], image_key="image", colorize_nlabels=None, monitor=None, remap=None, sane_index_shape=False, # tell vector quantizer to return indices as bhw ): super().__init__() self.image_key = image_key self.encoder = Encoder(**ddconfig) self.decoder = Decoder(**ddconfig) self.loss = instantiate_from_config(lossconfig) self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25, remap=remap, sane_index_shape=sane_index_shape) self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1) self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) if ckpt_path is not None: self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) self.image_key = image_key if colorize_nlabels is not None: assert type(colorize_nlabels)==int self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) if monitor is not None: self.monitor = monitor def init_from_ckpt(self, path, ignore_keys=list()): sd = torch.load(path, map_location="cpu")["state_dict"] keys = list(sd.keys()) for k in keys: for ik in ignore_keys: if k.startswith(ik): print("Deleting key {} from state_dict.".format(k)) del sd[k] self.load_state_dict(sd, strict=False) print(f"Restored from {path}") def encode(self, x): h = self.encoder(x).float() h = self.quant_conv(h).float() quant, emb_loss, info = self.quantize(h) return quant, emb_loss, info def decode(self, quant): quant = self.post_quant_conv(quant) dec = self.decoder(quant) return dec def decode_code(self, code_b): quant_b = self.quantize.embed_code(code_b) dec = self.decode(quant_b) return dec def forward(self, input): quant, diff, _ = self.encode(input) dec = self.decode(quant) return dec, diff def get_input(self, batch, k): x = batch[k] if len(x.shape) == 3: x = x[..., None] x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format) return x.float() def training_step(self, batch, batch_idx, optimizer_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x) if optimizer_idx == 0: # autoencode aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step, last_layer=self.get_last_layer(), split="train") self.log("train/aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True) return aeloss if optimizer_idx == 1: # discriminator discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step, last_layer=self.get_last_layer(), split="train") self.log("train/discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True) return discloss def validation_step(self, batch, batch_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x) aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0, self.global_step, last_layer=self.get_last_layer(), split="val") discloss, log_dict_disc = self.loss(qloss, x, xrec, 1, self.global_step, last_layer=self.get_last_layer(), split="val") rec_loss = log_dict_ae["val/rec_loss"] self.log("val/rec_loss", rec_loss, prog_bar=True, logger=True, on_step=True, on_epoch=True, sync_dist=True) self.log("val/aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True, sync_dist=True) self.log_dict(log_dict_ae) self.log_dict(log_dict_disc) return self.log_dict def configure_optimizers(self): lr = self.learning_rate opt_ae = torch.optim.Adam(list(self.encoder.parameters())+ list(self.decoder.parameters())+ list(self.quantize.parameters())+ list(self.quant_conv.parameters())+ list(self.post_quant_conv.parameters()), lr=lr, betas=(0.5, 0.9)) opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), lr=lr, betas=(0.5, 0.9)) return [opt_ae, opt_disc], [] def get_last_layer(self): return self.decoder.conv_out.weight def log_images(self, batch, **kwargs): log = dict() x = self.get_input(batch, self.image_key) x = x.to(self.device) xrec, _ = self(x) if x.shape[1] > 3: # colorize with random projection assert xrec.shape[1] > 3 x = self.to_rgb(x) xrec = self.to_rgb(xrec) log["inputs"] = x log["reconstructions"] = xrec return log def to_rgb(self, x): assert self.image_key == "segmentation" if not hasattr(self, "colorize"): self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) x = F.conv2d(x, weight=self.colorize) x = 2.*(x-x.min())/(x.max()-x.min()) - 1. return x class VQSegmentationModel(VQModel): def __init__(self, n_labels, *args, **kwargs): super().__init__(*args, **kwargs) self.register_buffer("colorize", torch.randn(3, n_labels, 1, 1)) def configure_optimizers(self): lr = self.learning_rate opt_ae = torch.optim.Adam(list(self.encoder.parameters())+ list(self.decoder.parameters())+ list(self.quantize.parameters())+ list(self.quant_conv.parameters())+ list(self.post_quant_conv.parameters()), lr=lr, betas=(0.5, 0.9)) return opt_ae def training_step(self, batch, batch_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x) aeloss, log_dict_ae = self.loss(qloss, x, xrec, split="train") self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True) return aeloss def validation_step(self, batch, batch_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x) aeloss, log_dict_ae = self.loss(qloss, x, xrec, split="val") self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True) total_loss = log_dict_ae["val/total_loss"] self.log("val/total_loss", total_loss, prog_bar=True, logger=True, on_step=True, on_epoch=True, sync_dist=True) return aeloss @torch.no_grad() def log_images(self, batch, **kwargs): log = dict() x = self.get_input(batch, self.image_key) x = x.to(self.device) xrec, _ = self(x) if x.shape[1] > 3: # colorize with random projection assert xrec.shape[1] > 3 # convert logits to indices xrec = torch.argmax(xrec, dim=1, keepdim=True) xrec = F.one_hot(xrec, num_classes=x.shape[1]) xrec = xrec.squeeze(1).permute(0, 3, 1, 2).float() x = self.to_rgb(x) xrec = self.to_rgb(xrec) log["inputs"] = x log["reconstructions"] = xrec return log class VQNoDiscModel(VQModel): def __init__(self, ddconfig, lossconfig, n_embed, embed_dim, ckpt_path=None, ignore_keys=[], image_key="image", colorize_nlabels=None ): super().__init__(ddconfig=ddconfig, lossconfig=lossconfig, n_embed=n_embed, embed_dim=embed_dim, ckpt_path=ckpt_path, ignore_keys=ignore_keys, image_key=image_key, colorize_nlabels=colorize_nlabels) def training_step(self, batch, batch_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x) # autoencode aeloss, log_dict_ae = self.loss(qloss, x, xrec, self.global_step, split="train") output = pl.TrainResult(minimize=aeloss) output.log("train/aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) output.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True) return output def validation_step(self, batch, batch_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x) aeloss, log_dict_ae = self.loss(qloss, x, xrec, self.global_step, split="val") rec_loss = log_dict_ae["val/rec_loss"] output = pl.EvalResult(checkpoint_on=rec_loss) output.log("val/rec_loss", rec_loss, prog_bar=True, logger=True, on_step=True, on_epoch=True) output.log("val/aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) output.log_dict(log_dict_ae) return output def configure_optimizers(self): optimizer = torch.optim.Adam(list(self.encoder.parameters())+ list(self.decoder.parameters())+ list(self.quantize.parameters())+ list(self.quant_conv.parameters())+ list(self.post_quant_conv.parameters()), lr=self.learning_rate, betas=(0.5, 0.9)) return optimizer class GumbelVQ(VQModel): def __init__(self, ddconfig, lossconfig, n_embed, embed_dim, temperature_scheduler_config, ckpt_path=None, ignore_keys=[], image_key="image", colorize_nlabels=None, monitor=None, kl_weight=1e-8, remap=None, ): z_channels = ddconfig["z_channels"] super().__init__(ddconfig, lossconfig, n_embed, embed_dim, ckpt_path=None, ignore_keys=ignore_keys, image_key=image_key, colorize_nlabels=colorize_nlabels, monitor=monitor, ) self.loss.n_classes = n_embed self.vocab_size = n_embed self.quantize = GumbelQuantize(z_channels, embed_dim, n_embed=n_embed, kl_weight=kl_weight, temp_init=1.0, remap=remap) self.temperature_scheduler = instantiate_from_config(temperature_scheduler_config) # annealing of temp if ckpt_path is not None: self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) def temperature_scheduling(self): self.quantize.temperature = self.temperature_scheduler(self.global_step) def encode_to_prequant(self, x): h = self.encoder(x) h = self.quant_conv(h) return h def decode_code(self, code_b): raise NotImplementedError def training_step(self, batch, batch_idx, optimizer_idx): self.temperature_scheduling() x = self.get_input(batch, self.image_key) xrec, qloss = self(x) if optimizer_idx == 0: # autoencode aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step, last_layer=self.get_last_layer(), split="train") self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True) self.log("temperature", self.quantize.temperature, prog_bar=False, logger=True, on_step=True, on_epoch=True) return aeloss if optimizer_idx == 1: # discriminator discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step, last_layer=self.get_last_layer(), split="train") self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True) return discloss def validation_step(self, batch, batch_idx): x = self.get_input(batch, self.image_key) xrec, qloss = self(x, return_pred_indices=True) aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0, self.global_step, last_layer=self.get_last_layer(), split="val") discloss, log_dict_disc = self.loss(qloss, x, xrec, 1, self.global_step, last_layer=self.get_last_layer(), split="val") rec_loss = log_dict_ae["val/rec_loss"] self.log("val/rec_loss", rec_loss, prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True) self.log("val/aeloss", aeloss, prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True) self.log_dict(log_dict_ae) self.log_dict(log_dict_disc) return self.log_dict def log_images(self, batch, **kwargs): log = dict() x = self.get_input(batch, self.image_key) x = x.to(self.device) # encode h = self.encoder(x) h = self.quant_conv(h) quant, _, _ = self.quantize(h) # decode x_rec = self.decode(quant) log["inputs"] = x log["reconstructions"] = x_rec return log
14,908
39.958791
120
py
DaVinci
DaVinci-main/taming/models/cond_transformer.py
import os, math import torch import torch.nn.functional as F import pytorch_lightning as pl from taming.main import instantiate_from_config from taming.modules.util import SOSProvider def disabled_train(self, mode=True): """Overwrite model.train with this function to make sure train/eval mode does not change anymore.""" return self class Net2NetTransformer(pl.LightningModule): def __init__(self, transformer_config, first_stage_config, cond_stage_config, permuter_config=None, ckpt_path=None, ignore_keys=[], first_stage_key="image", cond_stage_key="depth", downsample_cond_size=-1, pkeep=1.0, sos_token=0, unconditional=False, ): super().__init__() self.be_unconditional = unconditional self.sos_token = sos_token self.first_stage_key = first_stage_key self.cond_stage_key = cond_stage_key self.init_first_stage_from_ckpt(first_stage_config) self.init_cond_stage_from_ckpt(cond_stage_config) if permuter_config is None: permuter_config = {"target": "taming.modules.transformer.permuter.Identity"} self.permuter = instantiate_from_config(config=permuter_config) self.transformer = instantiate_from_config(config=transformer_config) if ckpt_path is not None: self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) self.downsample_cond_size = downsample_cond_size self.pkeep = pkeep def init_from_ckpt(self, path, ignore_keys=list()): sd = torch.load(path, map_location="cpu")["state_dict"] for k in sd.keys(): for ik in ignore_keys: if k.startswith(ik): self.print("Deleting key {} from state_dict.".format(k)) del sd[k] self.load_state_dict(sd, strict=False) print(f"Restored from {path}") def init_first_stage_from_ckpt(self, config): model = instantiate_from_config(config) model = model.eval() model.train = disabled_train self.first_stage_model = model def init_cond_stage_from_ckpt(self, config): if config == "__is_first_stage__": print("Using first stage also as cond stage.") self.cond_stage_model = self.first_stage_model elif config == "__is_unconditional__" or self.be_unconditional: print(f"Using no cond stage. Assuming the training is intended to be unconditional. " f"Prepending {self.sos_token} as a sos token.") self.be_unconditional = True self.cond_stage_key = self.first_stage_key self.cond_stage_model = SOSProvider(self.sos_token) else: model = instantiate_from_config(config) model = model.eval() model.train = disabled_train self.cond_stage_model = model def forward(self, x, c): # one step to produce the logits _, z_indices = self.encode_to_z(x) _, c_indices = self.encode_to_c(c) if self.training and self.pkeep < 1.0: mask = torch.bernoulli(self.pkeep*torch.ones(z_indices.shape, device=z_indices.device)) mask = mask.round().to(dtype=torch.int64) r_indices = torch.randint_like(z_indices, self.transformer.config.vocab_size) a_indices = mask*z_indices+(1-mask)*r_indices else: a_indices = z_indices cz_indices = torch.cat((c_indices, a_indices), dim=1) # target includes all sequence elements (no need to handle first one # differently because we are conditioning) target = z_indices # make the prediction logits, _ = self.transformer(cz_indices[:, :-1]) # cut off conditioning outputs - output i corresponds to p(z_i | z_{<i}, c) logits = logits[:, c_indices.shape[1]-1:] return logits, target def top_k_logits(self, logits, k): v, ix = torch.topk(logits, k) out = logits.clone() out[out < v[..., [-1]]] = -float('Inf') return out @torch.no_grad() def sample(self, x, c, steps, temperature=1.0, sample=False, top_k=None, callback=lambda k: None): x = torch.cat((c,x),dim=1) block_size = self.transformer.get_block_size() assert not self.transformer.training if self.pkeep <= 0.0: # one pass suffices since input is pure noise anyway assert len(x.shape)==2 noise_shape = (x.shape[0], steps-1) #noise = torch.randint(self.transformer.config.vocab_size, noise_shape).to(x) noise = c.clone()[:,x.shape[1]-c.shape[1]:-1] x = torch.cat((x,noise),dim=1) logits, _ = self.transformer(x) # take all logits for now and scale by temp logits = logits / temperature # optionally crop probabilities to only the top k options if top_k is not None: logits = self.top_k_logits(logits, top_k) # apply softmax to convert to probabilities probs = F.softmax(logits, dim=-1) # sample from the distribution or take the most likely if sample: shape = probs.shape probs = probs.reshape(shape[0]*shape[1],shape[2]) ix = torch.multinomial(probs, num_samples=1) probs = probs.reshape(shape[0],shape[1],shape[2]) ix = ix.reshape(shape[0],shape[1]) else: _, ix = torch.topk(probs, k=1, dim=-1) # cut off conditioning x = ix[:, c.shape[1]-1:] else: for k in range(steps): callback(k) assert x.size(1) <= block_size # make sure model can see conditioning x_cond = x if x.size(1) <= block_size else x[:, -block_size:] # crop context if needed logits, _ = self.transformer(x_cond) # pluck the logits at the final step and scale by temperature logits = logits[:, -1, :] / temperature # optionally crop probabilities to only the top k options if top_k is not None: logits = self.top_k_logits(logits, top_k) # apply softmax to convert to probabilities probs = F.softmax(logits, dim=-1) # sample from the distribution or take the most likely if sample: ix = torch.multinomial(probs, num_samples=1) else: _, ix = torch.topk(probs, k=1, dim=-1) # append to the sequence and continue x = torch.cat((x, ix), dim=1) # cut off conditioning x = x[:, c.shape[1]:] return x @torch.no_grad() def encode_to_z(self, x): quant_z, _, info = self.first_stage_model.encode(x) indices = info[2].view(quant_z.shape[0], -1) indices = self.permuter(indices) return quant_z, indices @torch.no_grad() def encode_to_c(self, c): if self.downsample_cond_size > -1: c = F.interpolate(c, size=(self.downsample_cond_size, self.downsample_cond_size)) quant_c, _, [_,_,indices] = self.cond_stage_model.encode(c) if len(indices.shape) > 2: indices = indices.view(c.shape[0], -1) return quant_c, indices @torch.no_grad() def decode_to_img(self, index, zshape): index = self.permuter(index, reverse=True) bhwc = (zshape[0],zshape[2],zshape[3],zshape[1]) quant_z = self.first_stage_model.quantize.get_codebook_entry( index.reshape(-1), shape=bhwc) x = self.first_stage_model.decode(quant_z) return x @torch.no_grad() def log_images(self, batch, temperature=None, top_k=None, callback=None, lr_interface=False, **kwargs): log = dict() N = 4 if lr_interface: x, c = self.get_xc(batch, N, diffuse=False, upsample_factor=8) else: x, c = self.get_xc(batch, N) x = x.to(device=self.device) c = c.to(device=self.device) quant_z, z_indices = self.encode_to_z(x) quant_c, c_indices = self.encode_to_c(c) # create a "half"" sample z_start_indices = z_indices[:,:z_indices.shape[1]//2] index_sample = self.sample(z_start_indices, c_indices, steps=z_indices.shape[1]-z_start_indices.shape[1], temperature=temperature if temperature is not None else 1.0, sample=True, top_k=top_k if top_k is not None else 100, callback=callback if callback is not None else lambda k: None) x_sample = self.decode_to_img(index_sample, quant_z.shape) # sample z_start_indices = z_indices[:, :0] index_sample = self.sample(z_start_indices, c_indices, steps=z_indices.shape[1], temperature=temperature if temperature is not None else 1.0, sample=True, top_k=top_k if top_k is not None else 100, callback=callback if callback is not None else lambda k: None) x_sample_nopix = self.decode_to_img(index_sample, quant_z.shape) # det sample z_start_indices = z_indices[:, :0] index_sample = self.sample(z_start_indices, c_indices, steps=z_indices.shape[1], sample=False, callback=callback if callback is not None else lambda k: None) x_sample_det = self.decode_to_img(index_sample, quant_z.shape) # reconstruction x_rec = self.decode_to_img(z_indices, quant_z.shape) log["inputs"] = x log["reconstructions"] = x_rec if self.cond_stage_key != "image": cond_rec = self.cond_stage_model.decode(quant_c) if self.cond_stage_key == "segmentation": # get image from segmentation mask num_classes = cond_rec.shape[1] c = torch.argmax(c, dim=1, keepdim=True) c = F.one_hot(c, num_classes=num_classes) c = c.squeeze(1).permute(0, 3, 1, 2).float() c = self.cond_stage_model.to_rgb(c) cond_rec = torch.argmax(cond_rec, dim=1, keepdim=True) cond_rec = F.one_hot(cond_rec, num_classes=num_classes) cond_rec = cond_rec.squeeze(1).permute(0, 3, 1, 2).float() cond_rec = self.cond_stage_model.to_rgb(cond_rec) log["conditioning_rec"] = cond_rec log["conditioning"] = c log["samples_half"] = x_sample log["samples_nopix"] = x_sample_nopix log["samples_det"] = x_sample_det return log def get_input(self, key, batch): x = batch[key] if len(x.shape) == 3: x = x[..., None] if len(x.shape) == 4: x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format) if x.dtype == torch.double: x = x.float() return x def get_xc(self, batch, N=None): x = self.get_input(self.first_stage_key, batch) c = self.get_input(self.cond_stage_key, batch) if N is not None: x = x[:N] c = c[:N] return x, c def shared_step(self, batch, batch_idx): x, c = self.get_xc(batch) logits, target = self(x, c) loss = F.cross_entropy(logits.reshape(-1, logits.size(-1)), target.reshape(-1)) return loss def training_step(self, batch, batch_idx): loss = self.shared_step(batch, batch_idx) self.log("train/loss", loss, prog_bar=True, logger=True, on_step=True, on_epoch=True) return loss def validation_step(self, batch, batch_idx): loss = self.shared_step(batch, batch_idx) self.log("val/loss", loss, prog_bar=True, logger=True, on_step=True, on_epoch=True) return loss def configure_optimizers(self): """ Following minGPT: This long function is unfortunately doing something very simple and is being very defensive: We are separating out all parameters of the model into two buckets: those that will experience weight decay for regularization and those that won't (biases, and layernorm/embedding weights). We are then returning the PyTorch optimizer object. """ # separate out all parameters to those that will and won't experience regularizing weight decay decay = set() no_decay = set() whitelist_weight_modules = (torch.nn.Linear, ) blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding) for mn, m in self.transformer.named_modules(): for pn, p in m.named_parameters(): fpn = '%s.%s' % (mn, pn) if mn else pn # full param name if pn.endswith('bias'): # all biases will not be decayed no_decay.add(fpn) elif pn.endswith('weight') and isinstance(m, whitelist_weight_modules): # weights of whitelist modules will be weight decayed decay.add(fpn) elif pn.endswith('weight') and isinstance(m, blacklist_weight_modules): # weights of blacklist modules will NOT be weight decayed no_decay.add(fpn) # special case the position embedding parameter in the root GPT module as not decayed no_decay.add('pos_emb') # validate that we considered every parameter param_dict = {pn: p for pn, p in self.transformer.named_parameters()} inter_params = decay & no_decay union_params = decay | no_decay assert len(inter_params) == 0, "parameters %s made it into both decay/no_decay sets!" % (str(inter_params), ) assert len(param_dict.keys() - union_params) == 0, "parameters %s were not separated into either decay/no_decay set!" \ % (str(param_dict.keys() - union_params), ) # create the pytorch optimizer object optim_groups = [ {"params": [param_dict[pn] for pn in sorted(list(decay))], "weight_decay": 0.01}, {"params": [param_dict[pn] for pn in sorted(list(no_decay))], "weight_decay": 0.0}, ] optimizer = torch.optim.AdamW(optim_groups, lr=self.learning_rate, betas=(0.9, 0.95)) return optimizer
15,002
42.613372
127
py
HeadlineCause
HeadlineCause-main/headline_cause/labse.py
import gc import numpy as np import tensorflow as tf import tensorflow_hub as hub import tensorflow_text as text from tqdm import tqdm from util import gen_batch DEFAULT_ENCODER_PATH = "https://tfhub.dev/google/LaBSE/2" DEFAULT_PREPROCESSOR_PATH = "https://tfhub.dev/google/universal-sentence-encoder-cmlm/multilingual-preprocess/2" def normalization(embeds): norms = np.linalg.norm(embeds, 2, axis=1, keepdims=True) return embeds / norms class LaBSE: def __init__(self, encoder_path=DEFAULT_ENCODER_PATH, preprocessor_path=DEFAULT_PREPROCESSOR_PATH): self.preprocessor = hub.KerasLayer(preprocessor_path) self.model = hub.KerasLayer(encoder_path) def embed_batch(self, texts): return normalization(self.model(self.preprocessor(texts))["default"]) def labse_get_embeddings(self, sentences, batch_size): embeddings = np.zeros((len(sentences), 768)) current_index = 0 for batch in gen_batch(sentences, batch_size): batch_embeddings = self.embed_batch(batch) embeddings[current_index:current_index+batch_size, :] = batch_embeddings current_index += batch_size tf.keras.backend.clear_session() gc.collect() return embeddings def __call__(self, sentences, batch_size=16): return self.labse_get_embeddings(sentences, batch_size)
1,380
33.525
112
py
HeadlineCause
HeadlineCause-main/headline_cause/util.py
import os import json import csv import random from urllib.parse import urlparse import torch import numpy as np def write_tsv(records, header, path): with open(path, "w") as w: writer = csv.writer(w, delimiter="\t", quotechar='"') writer.writerow(header) for r in records: row = [r[key] for key in header] writer.writerow(row) def read_tsv(file_name): records = [] with open(file_name, "r") as r: reader = csv.reader(r, delimiter="\t") header = next(reader) for row in reader: record = dict(zip(header, row)) records.append(record) return records def write_jsonl(records, path): with open(path, "w") as w: for r in records: w.write(json.dumps(r, ensure_ascii=False) + "\n") def read_jsonl(file_name): records = [] with open(file_name, "r") as r: for line in r: record = json.loads(line) records.append(record) return records def get_host(url): return '{uri.scheme}://{uri.netloc}/'.format(uri=urlparse(url)) def set_random_seed(seed): random.seed(seed) np.random.seed(seed) os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:2" os.environ["PL_GLOBAL_SEED"] = str(seed) os.environ["PYTHONHASHSEED"] = str(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True def gen_batch(records, batch_size): batch_start = 0 while batch_start < len(records): batch_end = batch_start + batch_size batch = records[batch_start: batch_end] batch_start = batch_end yield batch
1,720
23.585714
67
py
HeadlineCause
HeadlineCause-main/headline_cause/predict.py
import argparse import torch from tqdm import tqdm from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline import numpy as np from util import read_jsonl, write_jsonl def get_batch(data, batch_size): start_index = 0 while start_index < len(data): end_index = start_index + batch_size batch = data[start_index:end_index] yield batch start_index = end_index def pipe_predict(data, pipe, batch_size=64): raw_preds = [] for batch in tqdm(get_batch(data, batch_size), desc="predict"): raw_preds += pipe(batch) preds = np.array([int(max(labels, key=lambda x: x["score"])["label"][-1]) for labels in raw_preds]) pp = np.array([[label["score"] for label in labels] for labels in raw_preds]) return preds, pp def predict(model_path, input_path, output_path): model = AutoModelForSequenceClassification.from_pretrained(model_path) tokenizer = AutoTokenizer.from_pretrained(model_path) model.eval() pipe = pipeline( "text-classification", model=model, tokenizer=tokenizer, framework="pt", device=(0 if torch.cuda.is_available() else -1), return_all_scores=True ) data = read_jsonl(input_path) test_pairs = [(r["left_title"], r["right_title"]) for r in data] labels, probs = pipe_predict(test_pairs, pipe) for r, l, p in zip(data, labels, probs): r["pred_label"] = int(l) r["pred_prob"] = float(p[l]) write_jsonl(data, output_path) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--model-path", type=str, required=True) parser.add_argument("--input-path", type=str, required=True) parser.add_argument("--output-path", type=str, required=True) args = parser.parse_args() predict(**vars(args))
1,850
31.473684
103
py
HeadlineCause
HeadlineCause-main/headline_cause/train_clf.py
import argparse import json import random from statistics import mean import torch from torch.utils.data import Dataset from transformers import AutoTokenizer, AutoModelForSequenceClassification from transformers import Trainer, TrainingArguments, EarlyStoppingCallback from scipy.stats import entropy from sklearn.metrics import classification_report from augment import augment from util import read_jsonl class PairsDataset(Dataset): def __init__(self, records, max_tokens, tokenizer, result_key): self.tokenizer = tokenizer self.max_tokens = max_tokens self.records = list() for r in records: inputs = self.embed_pair(r) inputs["labels"] = r["label"] self.records.append(inputs) def __len__(self): return len(self.records) def embed_pair(self, pair): inputs = self.tokenizer( text=pair["left_title"], text_pair=pair["right_title"], add_special_tokens=True, max_length=self.max_tokens, padding="max_length", truncation="longest_first", return_tensors="pt" ) return {key: value.squeeze(0) for key, value in inputs.items()} def __getitem__(self, index): return self.records[index] def calc_labels(records, result_key): full_result_mapping = { "bad": 0, "same": 1, "rel": 2, "left_right_cause": 3, "right_left_cause": 4, "left_right_refute": 5, "right_left_refute": 6 } simple_result_mapping = { "not_cause": 0, "left_right": 1, "right_left": 2 } result_mapping = simple_result_mapping if "simple" in result_key else full_result_mapping for r in records: r["label"] = result_mapping[r[result_key]] return records def main( input_path, min_agreement, max_tokens, model_name, epochs, eval_steps, warmup_steps, lr, task, batch_size, grad_accum_steps, seed, out_dir ): res_key = "{}_result".format(task) agreement_key = "{}_agreement".format(task) records = read_jsonl(input_path) records = [r for r in records if r[agreement_key] > min_agreement] records = calc_labels(records, res_key) random.seed(seed) random.shuffle(records) val_border = int(len(records) * 0.8) test_border = int(len(records) * 0.9) train_records = records[:val_border] val_records = records[val_border:test_border] test_records = records[test_border:] train_records, val_records = augment(train_records, task), augment(val_records, task) labels_count = len({r[res_key] for r in train_records}) assert labels_count >= 3 print("Train records: ", len(train_records)) print("Val records: ", len(val_records)) print("Test records: ", len(test_records)) print("Labels: ", labels_count) tokenizer = AutoTokenizer.from_pretrained(model_name, do_lower_case=False) train_dataset = PairsDataset(train_records, max_tokens, tokenizer, res_key) val_dataset = PairsDataset(val_records, max_tokens, tokenizer, res_key) test_dataset = PairsDataset(test_records, max_tokens, tokenizer, res_key) device = "cuda" if torch.cuda.is_available() else "cpu" model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=labels_count) model = model.to(device) training_args = TrainingArguments( output_dir="checkpoints", evaluation_strategy="steps", save_strategy="steps", per_device_train_batch_size=batch_size, per_device_eval_batch_size=batch_size, logging_steps=eval_steps, save_steps=eval_steps, warmup_steps=warmup_steps, learning_rate=lr, num_train_epochs=epochs, gradient_accumulation_steps=grad_accum_steps, report_to="none", load_best_model_at_end=True, save_total_limit=1 ) trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=val_dataset ) trainer.train() model.save_pretrained(out_dir) train_dataset.tokenizer.save_pretrained(out_dir) y_true, y_pred = [], [] true_entropies = [] false_entropies = [] with torch.no_grad(): for item in test_dataset: input_ids = item["input_ids"].unsqueeze(0).to(device) mask = item["attention_mask"].unsqueeze(0).to(device) label = item["labels"] outputs = model(input_ids, mask, return_dict=True) logits = outputs.logits[0] pred = torch.argmax(logits).item() y_pred.append(pred) y_true.append(label) scores = torch.softmax(logits, dim=0).cpu().numpy() ent = entropy(scores) if pred == label: true_entropies.append(ent) else: false_entropies.append(ent) print("Avg true entropy: {}".format(mean(true_entropies))) if false_entropies: print("Avg false entropy: {}".format(mean(false_entropies))) print(classification_report(y_true, y_pred)) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--input-path", type=str, required=True) parser.add_argument("--out-dir", type=str, required=True) parser.add_argument("--max-tokens", type=int, default=60) parser.add_argument("--min-agreement", type=float, default=0.69) parser.add_argument("--epochs", type=int, default=4) parser.add_argument("--eval-steps", type=int, default=32) parser.add_argument("--warmup-steps", type=int, default=16) parser.add_argument("--lr", type=float, default=2e-05) parser.add_argument("--batch-size", type=int, default=8) parser.add_argument("--seed", type=int, default=42) parser.add_argument("--task", type=str, choices=("full", "simple"), default="simple") parser.add_argument("--grad-accum-steps", type=int, default=16) parser.add_argument("--model-name", type=str, default="xlm-roberta-large") args = parser.parse_args() main(**vars(args))
6,161
31.603175
99
py
HeadlineCause
HeadlineCause-main/headline_cause/active_learning/infer_clf.py
import argparse import json import torch import numpy as np from scipy.stats import entropy from scipy.special import softmax from tqdm import tqdm from transformers import AutoTokenizer, AutoModelForSequenceClassification from util import write_jsonl, read_jsonl, gen_batch def main( input_path, model_path, output_path, max_tokens, models_count, batch_size ): device = "cuda" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained(model_path) model = AutoModelForSequenceClassification.from_pretrained(model_path) model = model.to(device) model.train() # Monte-Carlo Dropout output_records = [] records = list(read_jsonl(input_path)) for batch in tqdm(gen_batch(records, batch_size)): s1 = [r["left_title"] for r in batch] s2 = [r["right_title"] for r in batch] inputs = tokenizer( text=s1, text_pair=s2, add_special_tokens=True, max_length=max_tokens, padding="max_length", truncation="longest_first", return_tensors="pt" ) inputs = inputs.to(device) num_labels = model.num_labels all_scores = torch.zeros((len(batch), models_count, num_labels)) with torch.no_grad(): for model_num in range(models_count): output = model(**inputs) logits = output.logits scores = torch.softmax(logits, dim=1).cpu() all_scores[:, model_num, :] = scores for sample_num in range(len(batch)): sample = batch[sample_num] sample_scores = all_scores[sample_num] avg_scores = torch.mean(sample_scores, dim=0).tolist() entropy_over_avg = float(entropy(avg_scores)) entropies = [float(entropy(scores)) for scores in sample_scores] avg_entropy = float(np.mean(entropies)) bald_score = entropy_over_avg - avg_entropy sample["entropy"] = entropy_over_avg sample["avg_entropy"] = avg_entropy sample["bald"] = bald_score sample["scores"] = avg_scores output_records.append(sample) write_jsonl(output_records, output_path) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--input-path", type=str, required=True) parser.add_argument("--model-path", type=str, required=True) parser.add_argument("--output-path", type=str, required=True) parser.add_argument("--batch-size", type=int, default=128) parser.add_argument("--models-count", type=int, default=3) parser.add_argument("--max-tokens", type=int, default=60) args = parser.parse_args() main(**vars(args))
2,774
32.035714
76
py
opioid-repurposing
opioid-repurposing-main/generate_bt_fps_mean.py
from fairseq.models.roberta import RobertaModel import argparse import sys import numpy as np import torch def load_pretrain_model(model_name_or_path, checkpoint_file, data_name_or_path, bpe='smi'): '''Currently only load to cpu()''' # load model pretrain_model = RobertaModel.from_pretrained( model_name_or_path, checkpoint_file, data_name_or_path, # dict_dir, bpe=bpe, ) pretrain_model.eval() return pretrain_model def extract_hidden(pretrain_model, target_file): sample_num = 0 for i, line in enumerate(open(target_file)): if len(line.strip()) == 0: continue sample_num += 1 hidden_features = {i: None for i in range(sample_num)} for i, line in enumerate(open(target_file)): if len(line.strip()) == 0: continue tokens = pretrain_model.encode(line.strip()) if len(tokens) > pretrain_model.args.max_positions: tokens = torch.cat( (tokens[:pretrain_model.args.max_positions - 1], tokens[-1].unsqueeze(0))) _, all_layer_hiddens = pretrain_model.model( tokens.unsqueeze(0), features_only=True, return_all_hiddens=True) #print(len(tokens)) hidden_info = all_layer_hiddens['inner_states'][-1] # last_hidden shape [tokens_num, sample_num(default=1), hidden_dim] # hidden_features.append(hidden_info.squeeze(1).cpu().detach().numpy()) hidden_features[i] = hidden_info.squeeze(1).cpu().detach().numpy() #print(hidden_features[i].shape) # hidden_features type: dict, length: samples_num return hidden_features def extract_features_from_hidden(hidden_info): samples_num = len(hidden_info) hidden_dim = np.shape(hidden_info[0])[-1] #print(np.shape(hidden_info[0])) samples_features = np.zeros([samples_num, hidden_dim]) for n_sample, hidden in hidden_info.items(): # hidden shape [tokens, embed_dim] samples_features[n_sample, :] =np.mean(hidden,axis=0) return samples_features def main(args): pretrain_model = load_pretrain_model( args.model_name_or_path, args.checkpoint_file, args.data_name_or_path, args.bpe) hidden_info = extract_hidden(pretrain_model, args.target_file) print('Generate features from hidden information') samples_features = extract_features_from_hidden(hidden_info) print(f'Features shape: {np.shape(samples_features)}') np.save(args.save_feature_path, samples_features) def parse_args(args): parser = argparse.ArgumentParser(description="Tools kit for downstream jobs") parser.add_argument('--model_name_or_path', default=None, type=str, help='Pretrained model folder') parser.add_argument('--checkpoint_file', default='checkpoint_best.pt', type=str, help='Pretrained model name') parser.add_argument('--data_name_or_path', default=None, type=str, help="Pre-training dataset folder") parser.add_argument('--dict_file', default='dict.txt', type=str, help="Pre-training dict filename(full path)") parser.add_argument('--bpe', default='smi', type=str) parser.add_argument('--target_file', default=None, type=str, help="Target file for feature extraction, default format is .smi") parser.add_argument('--save_feature_path', default='extract_f1.npy', type=str, help="Saving feature filename(path)") args = parser.parse_args() return args def cli_main(): args = parse_args(sys.argv[1:]) print(args) main(args) if __name__ == '__main__': cli_main() print('End!')
3,724
33.490741
91
py
LAMOL
LAMOL-master/test.py
import torch import csv import os import json import logging from fp16 import FP16_Module import GPUtil from collections import OrderedDict from settings import args, MODEL_CLASS, TOKENIZER, SPECIAL_TOKEN_IDS, init_logging from settings import MEMORY_FACTOR, LEN_FACTOR, TASK_DICT, MODEL_CONFIG, DATA_ATTRS, SPECIAL_TOKENS, CONFIG_CLASS, CONFIG_NAME from utils import QADataset, top_k_top_p_filtering, create_dataloader, logits_to_tokens, get_model_dir from utils import sample_sequence, remove_id, get_gen_token, lll_unbound_setting from metrics import compute_metrics logger = logging.getLogger(__name__) def test_one_to_one(task_load, task_eval, model, score_dict): logger.info("start to test { task: %s (load) %s (eval), seq train type: %s }" % (task_load, task_eval, args.seq_train_type)) test_qadata = QADataset(TASK_DICT[task_eval]["test"] , "test", SPECIAL_TOKEN_IDS[task_load]).sort() max_a_len = test_qadata.max_a_len test_dataloader = create_dataloader(test_qadata, "test") n_examples = len(test_qadata) logger.info("len of test dataset: {}".format(n_examples)) need_process = OrderedDict() qa_results = [0 for _ in range(n_examples)] all_pasts = [[0 for _ in range(n_examples)] for __ in range(MODEL_CONFIG.n_layer)] max_tot_lens = [0 for _ in range(n_examples)] cnt = 0 for n_steps, (cqs, len_cqs, _, _, _, _, _) in enumerate(test_dataloader): # assume n_gpus == 1 cqs = cqs[0] len_cqs = len_cqs[0] n_inputs = cqs.shape[0] all_outputs = model(input_ids=cqs.cuda()) outputs = all_outputs[0] if args.model_name == "gpt2": pasts = all_outputs[1] next_logits = outputs[range(n_inputs), len_cqs-1, :] / args.temperature_qa next_tokens = logits_to_tokens(next_logits).cpu() for i in range(n_inputs): max_tot_lens[cnt] = max_a_len + test_qadata[cnt][1] qa_results[cnt] = cqs[i][:len_cqs[i]] if next_tokens[i] != SPECIAL_TOKEN_IDS["eos_token"]: qa_results[cnt] = torch.cat((cqs[i][:len_cqs[i]], next_tokens[i])) if len(qa_results[cnt]) not in [max_tot_lens[cnt], args.max_len]: need_process.update([[cnt, None]]) if args.model_name == "gpt2": for layer_id in range(MODEL_CONFIG.n_layer): all_pasts[layer_id][cnt] = pasts[layer_id][:, i, ..., :len_cqs[i], :].type(torch.float32 if args.fp32 else torch.half) cnt += 1 if len(need_process) > int(12 * args.memory_sizes[0] / cqs.shape[1]): # dynamic threshold to avoid out of memory sample_sequence(model, need_process, qa_results, all_pasts, max_tot_lens) sample_sequence(model, need_process, qa_results, all_pasts, max_tot_lens) if task_eval in ['wikisql','woz.en','multinli.in.out']: ids = test_qadata.get_indices() test_qadata.sort_by_index() qa_results = [x[1] for x in sorted([(i, g) for i, g in zip(ids, qa_results)])] for i in range(len(test_qadata)): _, len_cq, _, _, Y, _, _, _ = test_qadata[i] if task_eval in ['wikisql','woz.en']: Y = test_qadata.answers[i] else: Y = list(filter(lambda x: x != -1, Y))[:-1] # remove eos Y = ' '.join([str(y) for y in Y]).split(str(SPECIAL_TOKEN_IDS["pad_token"])) Y = [TOKENIZER.decode(list(map(int, y.split()))) for y in Y] qa_results[i] = [TOKENIZER.decode(qa_results[i].tolist()[len_cq:]), Y] get_test_score(task_eval, qa_results, score_dict) model_dir = model.model_dir ep = model.ep results_path = os.path.join(model_dir,"qa_{}_{}.csv".format(task_eval,ep+1)) if not args.debug: with open(results_path, "w",encoding="utf-8") as f: qa_writer = csv.writer(f,delimiter=',') qa_writer.writerow(["y","pred"]) for pred, y in qa_results: if task_eval == 'wikisql': y = y["answer"] elif task_eval == 'woz.en': y = y[1] qa_writer.writerow([y,pred]) return model, score_dict def get_test_score(task_eval,qa_results,score_dict): score = compute_metrics( qa_results, bleu='iwslt.en.de' in task_eval or 'multinli.in.out' in task_eval, dialogue='woz.en' in task_eval, rouge='cnn_dailymail' in task_eval, logical_form='wikisql' in task_eval, corpus_f1='zre' in task_eval ) score_dict[task_eval] = score def test_one_to_many(task_load): score_dicts = [] for ep in range(args.n_train_epochs[task_load]): model_dir = get_model_dir([task_load]) model_path = os.path.join(model_dir, 'model-{}'.format(ep+1)) config_path = os.path.join(model_dir,CONFIG_NAME) gen_token = get_gen_token(task_load) TOKENIZER.add_tokens([gen_token]) SPECIAL_TOKENS[task_load] = gen_token SPECIAL_TOKEN_IDS[task_load] = TOKENIZER.convert_tokens_to_ids(gen_token) model_config = CONFIG_CLASS.from_json_file(config_path) model = MODEL_CLASS(model_config).cuda().eval() state_dict = torch.load(model_path, map_location='cuda:0') model.load_state_dict(state_dict) if not args.fp32: model = FP16_Module(model) model.ep = ep model.model_dir = model_dir logger.info("task: {}, epoch: {}".format(task_load, ep+1)) score_dict = {k:None for k in args.tasks} with torch.no_grad(): for task_eval in args.tasks: test_one_to_one(task_load, task_eval, model, score_dict) logger.info("score: {}".format(score_dict)) score_dicts.append(score_dict) with open(os.path.join(model_dir, "metrics.json"),"w") as f: json.dump(score_dicts, f) if __name__ == '__main__': if args.n_gpus > 1: raise NotImplementedError("test can be run with only one gpu currently!") if args.model_name == "gpt2": args.fp32 = False # always use fp16 in testing if not args.debug: logging.getLogger("pytorch_transformers").setLevel(logging.WARNING) logging.getLogger("pytorch_transformers.tokenization_utils").setLevel(logging.CRITICAL) init_logging(os.path.join(args.model_dir_root, 'log_test.txt')) logger.info('args = {}'.format(args)) if args.seq_train_type == "multitask": test_one_to_many('_'.join(args.tasks)) else: if args.unbound: TASK_DICT = lll_unbound_setting(split_size=args.unbound, data_type="test",test_target="origin") for task_load in args.splitted_tasks: test_one_to_many(task_load) else: for task_load in args.tasks: test_one_to_many(task_load)
6,865
41.645963
146
py
LAMOL
LAMOL-master/settings.py
import os import json import argparse import logging import datetime logger = logging.getLogger(__name__) import GPUtil from pytorch_transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, OpenAIGPTConfig from pytorch_transformers import GPT2LMHeadModel, GPT2Tokenizer, GPT2Config, CONFIG_NAME import torch BASE_DIR = os.path.dirname(os.path.abspath(__file__)) FILL_VAL = -1 LEN_FACTOR = 1.163 MEMORY_FACTOR = { "finetune": 0.58, "multitask": 0.58, "lll": 0.35, "ewc": 0.30, "mas": 0.18, "gem": 0.50, } TURING_ARCHS = {'Tesla V100', '2080 Ti'} MODEL_CLASSES = { 'gpt2': (GPT2LMHeadModel, GPT2Tokenizer, GPT2Config), 'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, OpenAIGPTConfig), } SAVE_NAME = 'model-' FINAL_SAVE_NAME = 'model-finish' def parse_args(): parser = argparse.ArgumentParser() parser.add_argument("--adam_epsilon", default=1e-4, type=float) parser.add_argument("--add_task_tokens", action="store_true") parser.add_argument("--data_dir", type=str, required=True) parser.add_argument("--debug", action="store_true") parser.add_argument("--decay_style", type=str, default="linear") parser.add_argument("--fp32", action="store_true") parser.add_argument("--real_sample", action="store_true") parser.add_argument("--unbound", type=int, default=0) parser.add_argument("--gen_lm_sample_percentage", type=float, default=0.05) parser.add_argument("--learning_rate", type=float, default=6.25e-5) parser.add_argument("--logging_steps", type=int, default=1000) parser.add_argument("--lm_lambda", type=float, default=0.25) parser.add_argument("--lr_schedule", type=str, default="warmup_linear") parser.add_argument("--max_grad_norm", type=int, default=1) parser.add_argument("--max_n_epochs", type=int, default=9) parser.add_argument("--min_batch_size", type=int, default=4) parser.add_argument("--min_n_steps", type=int, default=1500) parser.add_argument("--model_dir_root", type=str, required=True) parser.add_argument("--model_name", type=str, default="gpt2", choices=["gpt2", "openai-gpt"]) parser.add_argument("--n_gpus", type=int, default=1) parser.add_argument("--n_train_epochs", type=int, default=3) parser.add_argument("--dynamic_epochs", action="store_true") parser.add_argument("--n_warmup_ratio", type=float, default=0.005) parser.add_argument("--n_workers", type=int, default=4) parser.add_argument("--use_sep", action="store_true") parser.add_argument("--reg_lambda", type=float, default=1.) parser.add_argument("--seed", type=int, default=42) parser.add_argument("--seq_train_type", type=str, default="lll", choices=["lll","finetune","multitask","mas","ewc","gem"]) parser.add_argument("--tasks", nargs='+', default=["squad2"]) parser.add_argument("--skip_tasks", nargs='+') parser.add_argument("--temperature_lm", type=float, default=1.0) parser.add_argument("--temperature_qa", type=float, default=1.0) parser.add_argument("--test_batch_size", type=int, default=0) parser.add_argument("--tokens_weight", type=float, default=5) parser.add_argument("--top_k_lm", type=int, default=20) parser.add_argument("--top_k_qa", type=int, default=20) parser.add_argument("--top_p_lm", type=float, default=0.) parser.add_argument("--top_p_qa", type=float, default=0.) parser.add_argument("--train_batch_size", type=int, default=0) parser.add_argument("--weight_decay", type=float, default=0.01) parser.add_argument("--qp_margin", type=float, default=0.5) args = parser.parse_args() if args.debug: args.logging_steps = 1 torch.manual_seed(0) torch.backends.cudnn.deterministric = True args.model_dir_root = os.path.join(args.model_dir_root, args.model_name, args.seq_train_type, "{}_{}".format("_".join(args.tasks), args.gen_lm_sample_percentage) if "lll" in args.seq_train_type else "_".join(args.tasks)) args.device_ids = GPUtil.getAvailable(maxLoad=0.05, maxMemory=0.05, limit=args.n_gpus) if len(args.device_ids) == 0: logger.error('No available GPUs!') raise NotImplementedError("No CPU mode available!") if len(args.device_ids) < args.n_gpus: logger.warning('Available number of GPU = {} < n_gpus = {}'.format(len(args.device_ids), args.n_gpus)) args.n_gpus = len(args.device_ids) logger.warning('Continue training with {} GPUs'.format(args.n_gpus)) torch.cuda.set_device(args.device_ids[0]) gpus = GPUtil.getGPUs() gpu_names = [gpus[device_id].name for device_id in args.device_ids] if not all(any(turing_arch in gpu_name for turing_arch in TURING_ARCHS) for gpu_name in gpu_names): logger.warning('Not all gpus support fp16 training! Will use fp32 instead.') args.fp32 = True if args.model_name == "openai-gpt": args.fp32 = True # openai-gpt currently doesn't support fp16 if not args.fp32: global MEMORY_FACTOR MEMORY_FACTOR = dict([k, v*1.4] for k, v in MEMORY_FACTOR.items()) args.memory_sizes = [gpus[device_id].memoryTotal for device_id in args.device_ids] args.memory_sizes[0] = args.memory_sizes[0] * (1 - 0.04 * (args.n_gpus-1)) for i in range(1, args.n_gpus): args.memory_sizes[i] = args.memory_sizes[i] * 1.04 if args.train_batch_size <= 0: args.train_batch_size = [int(memory_size * MEMORY_FACTOR[args.seq_train_type]) for memory_size in args.memory_sizes] if args.test_batch_size <= 0: args.test_batch_size = [int(memory_size * MEMORY_FACTOR[args.seq_train_type]) for memory_size in args.memory_sizes] special_tokens = {"ans_token":'__ans__', "pad_token":'__pad__', "unk_token":'__unk__', "eos_token": '<|endoftext|>'} if args.use_sep: special_tokens["sep_token"] = '__sep__' model_class, tokenizer_class, config_class = MODEL_CLASSES[args.model_name] tokenizer = tokenizer_class.from_pretrained(args.model_name) tokenizer.add_tokens(list(special_tokens.values())) special_token_ids = {k:tokenizer.convert_tokens_to_ids(v) for k,v in special_tokens.items()} model_config = config_class.from_pretrained(args.model_name) model_config.vocab_size = len(tokenizer) tokens_weight = torch.ones([model_config.vocab_size], dtype=torch.float).cuda() tokens_weight[special_token_ids["ans_token"]] = args.tokens_weight if args.use_sep: tokens_weight[special_token_ids["sep_token"]] = args.tokens_weight args.max_len = model_config.n_positions data_attrs_path = os.path.join(BASE_DIR,"data_attrs.json") assert os.path.exists(data_attrs_path) with open(data_attrs_path, "r") as f: data_attrs = json.load(f) if args.seq_train_type == "multitask": args.n_train_epochs = {'_'.join(args.tasks): args.n_train_epochs} elif args.unbound: pass else: if "gem" in args.seq_train_type: args.memory_data = [] if args.dynamic_epochs: data_sizes = {task: data_attrs[task]["train"]["data_size"] for task in args.tasks} max_total_data_size = max(data_sizes.values()) * args.n_train_epochs args.n_train_epochs = {d[0]: min(args.max_n_epochs, max_total_data_size//d[1]) for d in data_sizes.items()} else: args.n_train_epochs = {task: args.n_train_epochs for task in args.tasks} return args, model_config, model_class, tokenizer, config_class, special_token_ids, special_tokens, data_attrs, tokens_weight class TimeFilter(logging.Filter): def filter(self, record): try: last = self.last except AttributeError: last = record.relativeCreated delta = record.relativeCreated/1000 - last/1000 record.relative = "{:.1f}".format(delta) record.uptime = str(datetime.timedelta(seconds=record.relativeCreated//1000)) self.last = record.relativeCreated return True def init_logging(filename): logging_format = "%(asctime)s - %(uptime)s - %(relative)ss - %(levelname)s - %(name)s - %(message)s" logging.basicConfig(format=logging_format, filename=filename, filemode='a', level=logging.INFO) console_handler = logging.StreamHandler() console_handler.setFormatter(logging.Formatter(logging_format)) root_logger = logging.getLogger() root_logger.addHandler(console_handler) for handler in root_logger.handlers: handler.addFilter(TimeFilter()) args, MODEL_CONFIG, MODEL_CLASS, TOKENIZER, CONFIG_CLASS, SPECIAL_TOKEN_IDS, SPECIAL_TOKENS, DATA_ATTRS, TOKENS_WEIGHT = parse_args() TASK_DICT = { "squad1": { "train":os.path.join(args.data_dir,"squad-train-v1.1.json"), "eval":os.path.join(args.data_dir,"squad-dev-v1.1.json"), "test":os.path.join(args.data_dir,"squad-dev-v1.1.json"), "n_train_epochs": args.n_train_epochs }, "squad2": { "train":os.path.join(args.data_dir,"squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"squad-dev-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "iwslt.en.de": { "train":os.path.join(args.data_dir,"iwslt.en.de_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"iwslt.en.de_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"iwslt.en.de_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "cnn_dailymail": { "train":os.path.join(args.data_dir,"cnn_dailymail_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"cnn_dailymail_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"cnn_dailymail_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "multinli.in.out": { "train":os.path.join(args.data_dir,"multinli.in.out_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"multinli.in.out_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"multinli.in.out_to_squad-dev-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "sst": { "train":os.path.join(args.data_dir,"sst_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"sst_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"sst_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "srl": { "train":os.path.join(args.data_dir,"srl_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"srl_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"srl_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "zre": { "train":os.path.join(args.data_dir,"zre_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"zre_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"zre_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "woz.en": { "train":os.path.join(args.data_dir,"woz.en_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"woz.en_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"woz.en_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "wikisql": { "train":os.path.join(args.data_dir,"wikisql_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"wikisql_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"wikisql_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "schema": { "train":os.path.join(args.data_dir,"schema_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"schema_to_squad-dev-v2.0.json"), "test":os.path.join(args.data_dir,"schema_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "ag": { "train":os.path.join(args.data_dir,"ag_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"ag_to_squad-test-v2.0.json"), "test":os.path.join(args.data_dir,"ag_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "dbpedia": { "train":os.path.join(args.data_dir,"dbpedia_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"dbpedia_to_squad-test-v2.0.json"), "test":os.path.join(args.data_dir,"dbpedia_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "yahoo": { "train":os.path.join(args.data_dir,"yahoo_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"yahoo_to_squad-test-v2.0.json"), "test":os.path.join(args.data_dir,"yahoo_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "amazon": { "train":os.path.join(args.data_dir,"amazon_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"amazon_to_squad-test-v2.0.json"), "test":os.path.join(args.data_dir,"amazon_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, "yelp": { "train":os.path.join(args.data_dir,"yelp_to_squad-train-v2.0.json"), "eval":os.path.join(args.data_dir,"yelp_to_squad-test-v2.0.json"), "test":os.path.join(args.data_dir,"yelp_to_squad-test-v2.0.json"), "n_train_epochs": args.n_train_epochs }, }
13,903
47.957746
133
py
LAMOL
LAMOL-master/utils.py
import torch from torch.utils.data import Dataset, DataLoader, Sampler import torch.nn.functional as F import re import csv import json import uuid import pickle as pkl import numpy as np from copy import deepcopy import os from glob import glob import logging import pathlib from collections import OrderedDict from settings import args, TASK_DICT, SPECIAL_TOKENS, SPECIAL_TOKEN_IDS, FILL_VAL from settings import TOKENIZER, LEN_FACTOR, DATA_ATTRS, MEMORY_FACTOR, MODEL_CONFIG, MODEL_CLASS from multiprocessing import Pool import sys import time import quadprog import io sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="UTF-8") logger = logging.getLogger(__name__) def make_dir(d): pathlib.Path(d).mkdir(parents=True, exist_ok=True) def get_gen_token(task): if args.add_task_tokens: return '__' + task + '__' else: return '__gen__' def get_model_dir(tasks): return os.path.join(args.model_dir_root, tasks[0]) if args.seq_train_type != "multitask" else args.model_dir_root def get_losses(parallel_model, cqa, Y, gen_X, gen_Y, loss_fct): if "lll" in args.seq_train_type: qa_logits = parallel_model(cqa) lm_logits = parallel_model(gen_X) qa_loss = loss_fct([torch.transpose(l, 1, 2) for l in qa_logits], Y) lm_loss = loss_fct([torch.transpose(l, 1, 2) for l in lm_logits], gen_Y) return torch.mean(qa_loss), args.lm_lambda * torch.mean(lm_loss) else: qa_logits = parallel_model(cqa) qa_loss = loss_fct([torch.transpose(l, 1, 2) for l in qa_logits], Y) return torch.mean(qa_loss), torch.tensor(0.) def pad_to_max_len(l, pad_len, val): return l + [val] * pad_len def pad_all_to_max_len(ls, val): max_len = max(len(l) for l in ls) return [pad_to_max_len(l, max_len-len(l), val) for l in ls] def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')): """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering Args: logits: logits distribution shape (vocabulary size) top_k > 0: keep only top k tokens with highest probability (top-k filtering). top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering). Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751) From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317 """ # assert logits.dim() == 1 # batch size 1 for now - could be updated for more but the code would be less clear top_k = min(top_k, logits.size(-1)) # Safety check if top_k > 0: # Remove all tokens with a probability less than the last token of the top-k indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] logits[indices_to_remove] = filter_value # if top_p > 0.0: # sorted_logits, sorted_indices = torch.sort(logits, descending=True) # cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1) # # Remove tokens with cumulative probability above the threshold # sorted_indices_to_remove = cumulative_probs > top_p # # Shift the indices to the right to keep also the first token above the threshold # sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() # sorted_indices_to_remove[..., 0] = 0 # indices_to_remove = sorted_indices[sorted_indices_to_remove] # logits[indices_to_remove] = filter_value return logits def varlen_collate_fn(data): batch_size = (len(data) + args.n_gpus - 1) // args.n_gpus cqs = torch.tensor(pad_all_to_max_len([datum[0] for datum in data], SPECIAL_TOKEN_IDS["pad_token"])).split(batch_size) len_cqs = torch.tensor([datum[1] for datum in data]).split(batch_size) cqas = torch.tensor(pad_all_to_max_len([datum[2] for datum in data], SPECIAL_TOKEN_IDS["pad_token"])).split(batch_size) len_cqas = torch.tensor([datum[3] for datum in data]).split(batch_size) Ys = torch.tensor(pad_all_to_max_len([datum[4] for datum in data], FILL_VAL)).split(batch_size) gen_Xs = torch.tensor(pad_all_to_max_len([datum[5] for datum in data], SPECIAL_TOKEN_IDS["pad_token"])).split(batch_size) gen_Ys = torch.tensor(pad_all_to_max_len([datum[6] for datum in data], FILL_VAL)).split(batch_size) return list(cqs), list(len_cqs), list(cqas), list(len_cqas), list(Ys), list(gen_Xs), list(gen_Ys) def dynamic_collate_fn(data, batch_size): def local_collate(): null_counter = 0 _cqs, _len_cqs, _cqas, _len_cqas, _Ys, _gen_Xs, _gen_Ys = [], [], [], [], [], [], [] Y_max_len = max(len(data[j][4]) for j in range(st, ed)) cq_max_len = max(len(data[j][0]) for j in range(st, ed)) for j in range(st, ed): if None in data[j] or [] in data[j]: null_counter+=1 logger.warning('null example in collate_fn, count: {}'.format(null_counter)) continue pad_len = cqa_max_len - len(data[j][2]) _cqs.append(pad_to_max_len(data[j][0], cq_max_len-len(data[j][0]), SPECIAL_TOKEN_IDS["pad_token"])) _len_cqs.append(data[j][1]) _cqas.append(pad_to_max_len(data[j][2], pad_len, SPECIAL_TOKEN_IDS["pad_token"])) _len_cqas.append(data[j][3]) _Ys.append(pad_to_max_len(data[j][4], Y_max_len - len(data[j][4]), FILL_VAL)) _gen_Xs.append(pad_to_max_len(data[j][5], pad_len, SPECIAL_TOKEN_IDS["pad_token"])) _gen_Ys.append(pad_to_max_len(data[j][6], pad_len, FILL_VAL)) cqs.append(torch.tensor(_cqs)) len_cqs.append(torch.tensor(_len_cqs)) cqas.append(torch.tensor(_cqas)) len_cqas.append(torch.tensor(_len_cqas)) Ys.append(torch.tensor(_Ys)) gen_Xs.append(torch.tensor(_gen_Xs)) gen_Ys.append(torch.tensor(_gen_Ys)) cqs, len_cqs, cqas, len_cqas, Ys, gen_Xs, gen_Ys = [], [], [], [], [], [], [] cqa_max_len, cnt, st = 0, 0, 0 for ed, datum in enumerate(data): ln = len(datum[2]) # use cqas to calibrate if max(cqa_max_len, ln)**LEN_FACTOR * (ed - st + 1) > batch_size[cnt]: local_collate() cnt += 1 cqa_max_len = 0 st = ed cqa_max_len = max(cqa_max_len, ln) ed += 1 # otherwise ed will be len(data)-1 local_collate() return cqs, len_cqs, cqas, len_cqas, Ys, gen_Xs, gen_Ys class QADataset(Dataset): def __init__(self, data_paths, data_type, gen_token, extra_data=[]): self.data_type = data_type self.gen_token = gen_token if args.use_sep: self.sep_token = SPECIAL_TOKEN_IDS["sep_token"] self.ans_token = SPECIAL_TOKEN_IDS["ans_token"] self.eos_token = SPECIAL_TOKEN_IDS["eos_token"] self.pad_token = SPECIAL_TOKEN_IDS["pad_token"] if not isinstance(data_paths, list): data_paths = [data_paths] data = [] for data_path in data_paths: if not data_path: continue with open(data_path, "r") as f: raw_ds = json.load(f) raw_ds = map(lambda x: x["paragraphs"], raw_ds["data"]) d = [] for raw_d in raw_ds: d.extend(raw_d) data += d self.data = [] self.max_a_len = 0 if len(data_paths)==1 and data_paths[0] is not None and ('wiki' in data_paths[0] or 'woz' in data_paths[0]): #data = self._sort_by_index(data) #args.n_workers = 1 if 'wiki' in data_paths[0]: answers_file = "wikisql_answers.json" elif 'woz' in data_paths[0]: answers_file = "woz.en_answers.json" with open(os.path.join(args.data_dir,answers_file),"r") as f: self.answers = json.load(f) if len(data) > 0: self.data_tokenization(data) if len(extra_data) > 0: extra_data = map(lambda x: self.etl_single_extra_data(x), extra_data) extra_data = list(filter(lambda x:x, extra_data)) if args.gen_lm_sample_percentage > 0. and len(extra_data) == 0: logger.warning("No good extra data but sample percentage > 0!") self.data += extra_data def etl_single_extra_data(self, data): gen_token = data[0] data = ' '.join([str(datum) for datum in data[1:]]) try: if args.use_sep: context, qa = re.split(str(SPECIAL_TOKEN_IDS["sep_token"]), data) else: context = "" qa = data question, answer = re.split(str(SPECIAL_TOKEN_IDS["ans_token"]), qa) context = [int(c) for c in context.strip().split()] question = [int(q) for q in question.strip().split()] answer = [int(a) for a in re.sub(str(SPECIAL_TOKEN_IDS["eos_token"]), "", answer).strip().split()] uid = uuid.uuid1().hex data = self.parse_example(gen_token, context, question, answer, uid) except ValueError: return return data def concat_example(self, gen_token, c, sep_token, q, ans_token, a, eos_token): example = sep_token + q + ans_token + a if len(example) + 1 > args.max_len: logger.warning('an example with len {} is too long!'.format(len(example) + 1)) return example = gen_token + c[:args.max_len-len(example)-1] + example + eos_token return example def parse_example(self, gen_token, context, question, answer, idx): if args.use_sep: cq_example = self.concat_example([], context, [self.sep_token], question, [self.ans_token], [], []) cqa_example = self.concat_example([], context, [self.sep_token], question, [self.ans_token], answer, []) else: cq_example = self.concat_example([], context, [], question, [self.ans_token], [], []) cqa_example = self.concat_example([], context, [], question, [self.ans_token], answer, []) Y_example = self.concat_example([], [], [], [], [], answer, [self.eos_token]) Y_example = [FILL_VAL] * (len(cqa_example) - len(Y_example)) + Y_example if args.use_sep: gen_X_example = self.concat_example([gen_token], context, [self.sep_token], question, [self.ans_token], answer, []) gen_Y_example = self.concat_example([], context, [self.sep_token], question, [self.ans_token], answer, [self.eos_token]) else: gen_X_example = self.concat_example([gen_token], context, [], question, [self.ans_token], answer, []) gen_Y_example = self.concat_example([], context, [], question, [self.ans_token], answer, [self.eos_token]) return cq_example, len(cq_example), cqa_example, len(cqa_example), Y_example, gen_X_example, gen_Y_example, idx def parallel_tokenization(self, d): examples = [] context = TOKENIZER.encode(d["context"]) max_a_len = 0 for qa in d["qas"]: question = TOKENIZER.encode(qa["question"]) raw_answers = qa["answers"] if len(raw_answers) == 0: assert qa["is_impossible"] raw_answers.append({"text": ""}) answer = [] for i, raw_answer in enumerate(raw_answers): answer.extend(TOKENIZER.encode(raw_answer["text"])) if i != len(raw_answers) - 1: answer.append(self.pad_token) max_a_len = max(max_a_len, len(answer)) examples.append(self.parse_example(self.gen_token, context, question, answer, qa.get("id", 0))) return examples, max_a_len def data_tokenization(self, data): if args.debug: data = data[:10] new_data = [] for datum in data: new_data.append(self.parallel_tokenization(datum)) data = new_data else: with Pool(args.n_workers) as pool: data = pool.map(self.parallel_tokenization, data) for datum, max_a_len in data: self.data.extend(datum) self.max_a_len = max(self.max_a_len, max_a_len) def sort(self): self.data.sort(key=lambda x: len(x[0])) return self def sort_by_index(self): self.data.sort(key=lambda x: x[-1]) def get_indices(self): return [d[-1] for d in self.data] #def _sort_by_index(self,data): # datum = [] # for d in data: # for qa in d["qas"]: # datum.append({"context":d["context"], "qas":[qa]}) # datum.sort(key=lambda x:x["qas"][0]["id"]) # return datum def __len__(self): return len(self.data) def __getitem__(self, index): return self.data[index] class EarlyStopping: def __init__(self, logger, patience=7, verbose=False): self.patience = patience self.verbose = verbose self.counter = 0 self.best_score = None self.early_stop = False self.val_loss_min = np.Inf self.logger = logger def __call__(self, val_loss, model, model_dir): score = -val_loss if self.best_score is None: self.best_score = score self.save_checkpoint(val_loss, model, model_dir) elif score < self.best_score: self.counter += 1 self.logger.info(f'EarlyStopping counter: {self.counter} out of {self.patience}') if self.counter >= self.patience: self.early_stop = True else: self.best_score = score self.save_checkpoint(val_loss, model, model_dir) self.counter = 0 def save_checkpoint(self, val_loss, model, model_dir): '''Saves model when validation loss decrease.''' if self.verbose: self.logger.info(f'Validation loss decreased ({self.val_loss_min:.6f} --> {val_loss:.6f}). Saving model ...') model.save_pretrained(model_dir) TOKENIZER.save_pretrained(model_dir) self.val_loss_min = val_loss class TrainStep: def __init__(self, model, optimizer, scheduler): self.model = model self.optimizer = optimizer self.scheduler = scheduler def __call__(self, loss, scheduler_steps): if not args.fp32: self.optimizer.backward(loss, update_master_grads=False) else: loss.backward() if not args.fp32: self.optimizer.update_master_grads() self.optimizer.clip_master_grads(args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(self.model.parameters(), args.max_grad_norm) if "gem" in args.seq_train_type and self.model.task_id >0: store_grad(self.model.parameters, self.model.grads, self.model.grad_dims,self.model.task_id) indx = torch.cuda.LongTensor([i for i in range(self.model.task_id)]) dotp = torch.mm(self.model.grads[:, self.model.task_id].unsqueeze(0), self.model.grads.index_select(1, indx)) if (dotp < 0).sum() != 0: project2cone2(self.model.grads[:, self.model.task_id].unsqueeze(1), self.model.grads.index_select(1, indx), args.qp_margin) # copy gradients back overwrite_grad(self.model.parameters, self.model.grads[:, self.model.task_id], self.model.grad_dims) if args.seq_train_type in args.REG_TYPE_KEYS: self.optimizer.step(self.model.reg_params) else: self.optimizer.step() if args.fp32 or (not self.optimizer.overflow): for i in range(scheduler_steps): self.scheduler.step() self.optimizer.zero_grad() class GEMStep: def __init__(self, model, parallel_model, train_loss_fct, optimizer): self.model = model self.parallel_model = parallel_model self.train_loss_fct = train_loss_fct self.optimizer = optimizer def __call__(self,current_task_id): for past_task_id, md in enumerate(args.memory_data): # Not saving current task's grads. if past_task_id >= current_task_id: return qadata = QADataset(None, "test", "gen", md) dataloader = create_dataloader(qadata, "test") grads_tmp = torch.zeros(sum(self.model.grad_dims),).cuda() if not args.fp32: grads_tmp = grads_tmp.half() for _, _, cqa, _, Y, gen_X, gen_Y in dataloader: #CHECK n_inputs = sum(_cqa.shape[0] for _cqa in cqa) self.optimizer.zero_grad() for i in range(len(cqa)): cqa[i] = (cqa[i].to(args.device_ids[i]),) Y[i] = Y[i].to(args.device_ids[i]) gen_X[i] = (gen_X[i].to(args.device_ids[i]),) gen_Y[i] = gen_Y[i].to(args.device_ids[i]) losses = get_losses(self.parallel_model, cqa, Y, gen_X, gen_Y, self.train_loss_fct) loss = sum(losses) if not args.fp32: self.optimizer.backward(loss, update_master_grads=False) else: loss.backward() if not args.fp32: #copy fp16 grads to fp32 grads self.optimizer.update_master_grads() self.optimizer.clip_master_grads(args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(self.model.parameters(), args.max_grad_norm) i = 0 for param in self.model.parameters(): if param.grad is not None: beg = 0 if i == 0 else sum(self.model.grad_dims[:i]) end = sum(self.model.grad_dims[:i+1]) grads_tmp[beg: end] += param.grad.data.view(-1)*n_inputs i += 1 grads_tmp /= len(qadata) self.model.grads[:, past_task_id].copy_(grads_tmp) self.optimizer.zero_grad() class DynamicBatchSampler(Sampler): def __init__(self, dataset, data_type, max_batch_size): self.dataset = dataset self.data_type = data_type if data_type == "train": self.batch_size = args.train_batch_size else: self.batch_size = args.test_batch_size self.n_samples = len(dataset) self.max_batch_size = max_batch_size def __iter__(self): if args.debug or self.data_type == "test": indices = range(self.n_samples) else: indices = np.random.permutation(self.n_samples) max_len, cnt, st = 0, 0, 0 batch = [] for ed, idx in enumerate(indices): ln = len(self.dataset[idx][2]) if max(max_len, ln)**LEN_FACTOR * (ed - st + 1) > self.batch_size[cnt]: st = ed cnt += 1 max_len = 0 if cnt == args.n_gpus: yield batch cnt = 0 batch = [] max_len = max(max_len, ln) batch.append(idx) if len(batch) == self.max_batch_size and self.data_type == "train": yield batch cnt, max_len, st = 0, 0, ed batch = [] if len(batch) > 0: yield batch def __len__(self): raise NotImplementedError def create_dataloader(dataset, data_type, max_batch_size=1000000000): if data_type == "train": batch_size = args.train_batch_size else: batch_size = args.test_batch_size if isinstance(batch_size, list): collate_fn=lambda x,bs=batch_size: dynamic_collate_fn(x, bs) shuffle = False batch_size = 1 batch_sampler = DynamicBatchSampler(dataset, data_type, max_batch_size) else: collate_fn=lambda x: varlen_collate_fn(x) shuffle = not (data_type != "train" or args.debug) batch_sampler = None dataloader = DataLoader(dataset, num_workers=args.n_workers, collate_fn=collate_fn, shuffle=shuffle, batch_size=batch_size, batch_sampler=batch_sampler) return dataloader class WrapModel(torch.nn.Module): def __init__(self, model): super(WrapModel, self).__init__() self.model = model def forward(self, input_ids): outputs = self.model(input_ids) return outputs[0] def remove_id(idx, need_process, all_pasts): assert idx in need_process del need_process[idx] for layer_id in range(MODEL_CONFIG.n_layer): all_pasts[layer_id][idx] = 0 def sample_sequence(model, need_process, qa_results, all_pasts, max_tot_lens): while len(need_process) > 0: first_id = next(iter(need_process)) shortest_len = len(qa_results[first_id]) decode_batch_size = int(args.memory_sizes[0] * MEMORY_FACTOR[args.seq_train_type] // (shortest_len+1)**LEN_FACTOR) it = iter(need_process) stop = False remove_ids = [] while not stop: batch_ids, input_ids, past = [], [], [[] for _ in range(MODEL_CONFIG.n_layer)] while True: try: cur_id = next(it) if len(qa_results[cur_id]) > shortest_len: stop = True break batch_ids.append(cur_id) if args.model_name == "gpt2": input_ids.append(qa_results[cur_id][-1:]) for layer_id in range(MODEL_CONFIG.n_layer): past[layer_id].append(all_pasts[layer_id][cur_id]) else: input_ids.append(qa_results[cur_id]) if len(input_ids) == decode_batch_size: break except StopIteration: stop = True break n_inputs = len(input_ids) if n_inputs == 0: break input_ids = torch.stack(input_ids) if args.model_name == "gpt2": for layer_id in range(MODEL_CONFIG.n_layer): past[layer_id] = torch.stack(past[layer_id], dim=1) all_outputs = model(input_ids=input_ids.cuda(), past=past) else: all_outputs = model(input_ids=input_ids.cuda()) outputs = all_outputs[0] if args.model_name == "gpt2": pasts = all_outputs[1] next_logits = outputs[..., -1, :] / args.temperature_qa next_tokens = logits_to_tokens(next_logits).cpu() for i, cur_id in enumerate(batch_ids): if next_tokens[i] == SPECIAL_TOKEN_IDS["eos_token"]: remove_ids.append(cur_id) else: qa_results[cur_id] = torch.cat((qa_results[cur_id], next_tokens[i])) if len(qa_results[cur_id]) in [max_tot_lens[cur_id], args.max_len]: remove_ids.append(cur_id) elif args.model_name == "gpt2": for layer_id in range(MODEL_CONFIG.n_layer): all_pasts[layer_id][cur_id] = pasts[layer_id][:, i].type(torch.float if args.fp32 else torch.half) for idx in remove_ids: remove_id(idx, need_process, all_pasts) def write_extra_data(dump_path, qa_results): logger.info(f"writing extra data in {dump_path} ...") with open(dump_path,"w",newline="",encoding="utf-8") as f: lm_writer = csv.writer(f,delimiter=',') lm_writer.writerow(["gen"]) for l in qa_results: lm_writer.writerow([l]) def parse_single_real_data(data,task): c = data["paragraphs"][0]["context"] q = data["paragraphs"][0]["qas"][0]["question"] a = data["paragraphs"][0]["qas"][0]["answers"][0]["text"] if args.use_sep: data = "{}{}{}{}{}{}{}".format(SPECIAL_TOKENS[task],c,SPECIAL_TOKENS["sep_token"],q,SPECIAL_TOKENS["ans_token"],a,SPECIAL_TOKENS["eos_token"]) else: data = "{}{} {}{}{}{}".format(SPECIAL_TOKENS[task],c,q,SPECIAL_TOKENS["ans_token"],a,SPECIAL_TOKENS["eos_token"]) return data def get_real_data(task, train_extra_data, accum=True, encode=True): task_idx = args.tasks.index(task) gen_size = DATA_ATTRS[task]["train"]["data_size"] if accum: prev_tasks = args.tasks[:task_idx] gen_size = int(np.ceil(gen_size * args.gen_lm_sample_percentage))//len(prev_tasks) else: prev_tasks = [args.tasks[task_idx-1]] gen_size = int(gen_size * args.gen_lm_sample_percentage) datum = [] for prev_task in prev_tasks: with open(TASK_DICT[prev_task]["train"],"r") as f: data = data_expand(json.load(f)["data"]) indices = np.random.choice(range(len(data)), gen_size) for i in indices: d = parse_single_real_data(data[i],prev_task) datum.append(d) if encode: train_extra_data.append(TOKENIZER.encode(d)) model_dir = get_model_dir([prev_task]) dump_path = os.path.join(model_dir,"real.csv") write_extra_data(dump_path, datum) return dump_path def read_extra_data(gen_path, train_extra_data): with open(gen_path,"r") as lm_file: reader = csv.reader(lm_file,delimiter=',') next(reader) for row in reader: row = TOKENIZER.encode(row[0].strip()) train_extra_data.append(row) def create_extra_data(task, prev_task, model, train_extra_data): if args.real_sample: logger.info(f"using real data as extra data") return get_real_data(task, train_extra_data) task_cnt = args.tasks.index(task) model_dir = get_model_dir([prev_task]) gen_path = os.path.join(model_dir,"lm.csv") if os.path.exists(gen_path): logger.info(f"extra data exists in {gen_path}, read it!") return read_extra_data(gen_path, train_extra_data) gen_size = DATA_ATTRS[task]["train"]["data_size"] gen_size = int(np.ceil(gen_size * args.gen_lm_sample_percentage)) gen_size -= (gen_size % task_cnt) if args.debug: gen_size = task_cnt model.eval() need_process = OrderedDict() qa_results = [] for task_name in args.tasks[:task_cnt]: qa_results.extend([torch.tensor([SPECIAL_TOKEN_IDS[task_name]]) for _ in range(gen_size//task_cnt)]) all_pasts = [[ torch.empty(2, MODEL_CONFIG.n_head, 0, MODEL_CONFIG.n_embd//MODEL_CONFIG.n_head, dtype=torch.float if args.fp32 else torch.half).cuda() for _ in range(gen_size) ] for __ in range(MODEL_CONFIG.n_layer)] max_tot_lens = [args.max_len for _ in range(gen_size)] for i in range(gen_size): need_process.update([[i, None]]) if len(need_process) > int(args.memory_sizes[0] * 0.12): sample_sequence(model, need_process, qa_results, all_pasts, max_tot_lens) sample_sequence(model, need_process, qa_results, all_pasts, max_tot_lens) model.train() qa_results = [res.tolist() for res in qa_results] train_extra_data.extend(qa_results) qa_results = [TOKENIZER.decode(res) for res in qa_results] write_extra_data(gen_path, qa_results) def logits_to_tokens(next_logits): filtered_logits = top_k_top_p_filtering(next_logits, top_k=args.top_k_qa, top_p=args.top_p_qa) log_probs = F.softmax(filtered_logits, dim=-1) next_tokens = torch.multinomial(log_probs, num_samples=1) return next_tokens def lll_unbound_setting(split_size=10,data_type="train",test_target="self"): data_dir = os.path.join(args.data_dir,"{}_{}".format("_".join(args.tasks),args.gen_lm_sample_percentage)) if data_type == "test": args.splitted_tasks = [f"task_{i}" for i in range(split_size)] args.n_train_epochs = {task: args.n_train_epochs for task in args.splitted_tasks} if test_target in ["self","all"]: for no in range(split_size): task = f"task_{no}" test_data_path = os.path.join(data_dir,f"{task}-test.json") TASK_DICT[task] = {} TASK_DICT[task]["test"] = test_data_path if test_target == "all": args.tasks += args.splitted_tasks else: args.tasks = args.splitted_tasks elif data_type == "train": create_lll_unbound_data(split_size) args.n_train_epochs = {task: args.n_train_epochs for task in args.tasks} return TASK_DICT def create_lll_unbound_data(split_size=10): data_dir = os.path.join(args.data_dir,"{}_{}".format("_".join(args.tasks),args.gen_lm_sample_percentage)) pathlib.Path(data_dir).mkdir(parents=True, exist_ok=True) datum = [] test_datum = [] data_sizes = [] chunk_sizes = [] for task in args.tasks: train_data_path = TASK_DICT[task]["train"] with open(train_data_path, "r") as f: data = json.load(f)["data"] data = data_expand(data) data_sizes.append(len(data)) datum += data test_data_path = TASK_DICT[task]["test"] with open(test_data_path, "r") as f: data = json.load(f)["data"] data = data_expand(data) test_datum.append(data) chunk_size = int(np.ceil(len(datum)/split_size)) tasks = [] for no, i in enumerate(range(0, len(datum), chunk_size)): task = f"task_{no}" tasks.append(task) chunk = datum[i:i + chunk_size] if i < len(datum)-chunk_size else datum[i:] chunk_sizes.append(len(chunk)) DATA_ATTRS[task] = {"train":{"data_size":None}} DATA_ATTRS[task]["train"]["data_size"] = len(chunk) train_data_path = os.path.join(data_dir,f"{task}-train.json") with open(train_data_path,"w") as f: json.dump({"data":chunk},f) TASK_DICT[task] = {} TASK_DICT[task]["train"] = train_data_path args.tasks = tasks sis = get_split_indices(data_sizes,chunk_sizes) test_split = [] for dic in sis.values(): merged_data = [] for k,v in dic.items(): from_index = int(len(test_datum[k])*v[0]) to_index = int(len(test_datum[k])*v[1]) merged_data+= test_datum[k][from_index:to_index] test_split.append(merged_data) for no, chunk in enumerate(test_split): task = f"task_{no}" test_data_path = os.path.join(data_dir,f"{task}-test.json") with open(test_data_path,"w") as f: json.dump({"data":chunk},f) TASK_DICT[task]["test"] = test_data_path def data_expand(data): datum = [] for d in data: para = d["paragraphs"] for p in para: for qa in p["qas"]: d = {"context": p["context"], "qas": [qa]} datum.append({"paragraphs":[d]}) return datum def get_split_indices(data_sizes,chunk_sizes): ds = deepcopy(data_sizes) records = {} tmp = {} order = 0 # data_sizes index i = 0 # chunk_sizes index while len(data_sizes)>0: d0 = data_sizes[0] c0 = chunk_sizes[0] if d0>c0: val = c0/ds[order] else: val = d0/ds[order] if order not in tmp: rec = (0,val) tmp[order] = val else: rec = (tmp[order],tmp[order]+val) tmp[order] += val if i in records: records[i][order] = rec else: records[i] = {order: rec} if d0>c0: data_sizes[0]-=c0 chunk_sizes.pop(0) i+=1 else: chunk_sizes[0]-=d0 data_sizes.pop(0) order+=1 if d0==c0: chunk_sizes.pop(0) i+=1 return records def store_grad(get_ps, grads, grad_dims, task_id): i = 0 for param in get_ps(): if param.grad is not None: beg = 0 if i == 0 else sum(grad_dims[:i]) end = sum(grad_dims[:i+1]) grads[beg: end, task_id].copy_(param.grad.data.view(-1)) i += 1 def overwrite_grad(pp, newgrad, grad_dims): cnt = 0 for param in pp(): if param.grad is not None: beg = 0 if cnt == 0 else sum(grad_dims[:cnt]) en = sum(grad_dims[:cnt + 1]) this_grad = newgrad[beg: en].contiguous().view( param.grad.data.size()) param.grad.data.copy_(this_grad) cnt += 1 def project2cone2(gradient, memories, margin=0.5, eps=1e-3): memories_np = memories.cpu().t().double().numpy() gradient_np = gradient.cpu().contiguous().view(-1).double().numpy() t = memories_np.shape[0] P = np.dot(memories_np, memories_np.transpose()) P = 0.5 * (P + P.transpose()) + np.eye(t) * eps q = np.dot(memories_np, gradient_np) * -1 G = np.eye(t) h = np.zeros(t) + margin v = quadprog.solve_qp(P, q, G, h)[0] x = np.dot(v, memories_np) + gradient_np gradient.copy_(torch.Tensor(x).view(-1, 1))
33,442
38.577515
150
py
LAMOL
LAMOL-master/scheduler.py
# coding=utf-8 # Copyright (c) 2019, NVIDIA CORPORATION. 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. """PyTorch DataLoader for TFRecords""" import torch from torch.optim.lr_scheduler import _LRScheduler import math class AnnealingLR(_LRScheduler): """Anneals the learning rate from start to zero along a cosine curve.""" DECAY_STYLES = ['linear', 'cos', 'exp', 'const', 'None'] def __init__(self, optimizer, start_lr, warmup_iter, num_iters, decay_style=None, last_iter=-1): self.optimizer = optimizer self.start_lr = start_lr self.warmup_iter = warmup_iter self.num_iters = last_iter + 1 self.end_iter = num_iters self.decay_style = decay_style.lower() if isinstance(decay_style, str) else None self.step(self.num_iters) def get_lr(self): # https://openreview.net/pdf?id=BJYwwY9ll pg. 4 if self.warmup_iter > 0 and self.num_iters <= self.warmup_iter: return float(self.start_lr) * self.num_iters / self.warmup_iter else: if self.decay_style == self.DECAY_STYLES[0]: return self.start_lr*((self.end_iter-(self.num_iters-self.warmup_iter))/self.end_iter) elif self.decay_style == self.DECAY_STYLES[1]: return self.start_lr / 2.0 * (math.cos(math.pi * (self.num_iters - self.warmup_iter) / self.end_iter) + 1) elif self.decay_style == self.DECAY_STYLES[2]: raise NotImplementedError("Exponential decay not yet implemented") else: return self.start_lr def step(self, step_num=None): if step_num is None: step_num = self.num_iters + 1 self.num_iters = step_num new_lr = self.get_lr() for group in self.optimizer.param_groups: group['lr'] = new_lr def state_dict(self): sd = { 'start_lr': self.start_lr, 'warmup_iter': self.warmup_iter, 'num_iters': self.num_iters, 'decay_style': self.decay_style, 'end_iter': self.end_iter } return sd def load_state_dict(self, sd): self.start_lr = sd['start_lr'] self.warmup_iter = sd['warmup_iter'] self.num_iters = sd['num_iters'] self.end_iter = sd['end_iter'] self.decay_style = sd['decay_style'] self.step(self.num_iters)
2,930
38.608108
122
py
LAMOL
LAMOL-master/fp16util.py
# coding=utf-8 # Copyright (c) 2019, NVIDIA CORPORATION. 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. import torch import torch.nn as nn from torch.autograd import Variable from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors # import mpu class tofp16(nn.Module): """ Utility module that implements:: def forward(self, input): return input.half() """ def __init__(self): super(tofp16, self).__init__() def forward(self, input): return input.half() def BN_convert_float(module): """ Utility function for network_to_half(). Retained for legacy purposes. """ if isinstance(module, torch.nn.modules.batchnorm._BatchNorm) and module.affine is True: module.float() for child in module.children(): BN_convert_float(child) return module def network_to_half(network): """ Convert model to half precision in a batchnorm-safe way. Retained for legacy purposes. It is recommended to use FP16Model. """ return nn.Sequential(tofp16(), BN_convert_float(network.half())) def convert_module(module, dtype): """ Converts a module's immediate parameters and buffers to dtype. """ for param in module.parameters(recurse=False): if param is not None: if param.data.dtype.is_floating_point: param.data = param.data.to(dtype=dtype) if param._grad is not None and param._grad.data.dtype.is_floating_point: param._grad.data = param._grad.data.to(dtype=dtype) for buf in module.buffers(recurse=False): if buf is not None and buf.data.dtype.is_floating_point: buf.data = buf.data.to(dtype=dtype) def convert_network(network, dtype): """ Converts a network's parameters and buffers to dtype. """ for module in network.modules(): if isinstance(module, torch.nn.modules.batchnorm._BatchNorm) and module.affine is True: continue convert_module(module, dtype) return network class FP16Model(nn.Module): """ Convert model to half precision in a batchnorm-safe way. """ def __init__(self, network): super(FP16Model, self).__init__() self.network = convert_network(network, dtype=torch.half) def forward(self, *inputs): inputs = tuple(t.half() for t in inputs) return self.network(*inputs) def backwards_debug_hook(grad): raise RuntimeError("master_params recieved a gradient in the backward pass!") def prep_param_lists(model, flat_master=False): """ Creates a list of FP32 master parameters for a given model, as in `Training Neural Networks with Mixed Precision: Real Examples`_. Args: model (torch.nn.Module): Existing Pytorch model flat_master (bool, optional, default=False): Flatten the master parameters into a single tensor, as a performance optimization. Returns: A tuple (``model_params``, ``master_params``). ``model_params`` is a list of the model's parameters for later use with :func:`model_grads_to_master_grads` and :func:`master_params_to_model_params`. ``master_params`` is a list of FP32 master gradients. If ``flat_master=True``, ``master_params`` will be a list with one element. Example:: model_params, master_params = prep_param_lists(model) .. warning:: Currently, if ``flat_master=True``, all the model's parameters must be the same type. If the model has parameters of different types, use ``flat_master=False``, or use :class:`FP16_Optimizer`. .. _`Training Neural Networks with Mixed Precision: Real Examples`: http://on-demand.gputechconf.com/gtc/2018/video/S81012/ """ model_params = [param for param in model.parameters() if param.requires_grad] if flat_master: # Give the user some more useful error messages try: # flatten_dense_tensors returns a contiguous flat array. # http://pytorch.org/docs/master/_modules/torch/_utils.html master_params = _flatten_dense_tensors([param.data for param in model_params]).float() except: print("Error in prep_param_lists: model may contain a mixture of parameters " "of different types. Use flat_master=False, or use F16_Optimizer.") raise master_params = torch.nn.Parameter(master_params) master_params.requires_grad = True # master_params.register_hook(backwards_debug_hook) if master_params.grad is None: master_params.grad = master_params.new(*master_params.size()) return model_params, [master_params] else: master_params = [param.clone().float().detach() for param in model_params] for param in master_params: param.requires_grad = True return model_params, master_params def model_grads_to_master_grads(model_params, master_params, flat_master=False): """ Copy model gradients to master gradients. Args: model_params: List of model parameters created by :func:`prep_param_lists`. master_params: List of FP32 master parameters created by :func:`prep_param_lists`. If ``master_params`` was created with ``flat_master=True``, ``flat_master=True`` should also be supplied to :func:`model_grads_to_master_grads`. """ if flat_master: # The flattening may incur one more deep copy than is necessary. master_params[0].grad.data.copy_( _flatten_dense_tensors([p.grad.data for p in model_params])) else: for model, master in zip(model_params, master_params): if model.grad is not None: if master.grad is None: master.grad = Variable(master.data.new(*master.data.size())) master.grad.data.copy_(model.grad.data) else: master.grad = None def master_params_to_model_params(model_params, master_params, flat_master=False): """ Copy master parameters to model parameters. Args: model_params: List of model parameters created by :func:`prep_param_lists`. master_params: List of FP32 master parameters created by :func:`prep_param_lists`. If ``master_params`` was created with ``flat_master=True``, ``flat_master=True`` should also be supplied to :func:`master_params_to_model_params`. """ if flat_master: for model, master in zip(model_params, _unflatten_dense_tensors(master_params[0].data, model_params)): model.data.copy_(master) else: for model, master in zip(model_params, master_params): model.data.copy_(master.data) # Backward compatibility fixes def to_python_float(t): if hasattr(t, 'item'): return t.item() else: return t[0] TORCH_MAJOR = int(torch.__version__.split('.')[0]) TORCH_MINOR = int(torch.__version__.split('.')[1]) # clip_grad_norm = mpu.clip_grad_norm #elif TORCH_MAJOR == 0 and TORCH_MINOR <= 4: # clip_grad_norm = torch.nn.utils.clip_grad_norm #else: clip_grad_norm = torch.nn.utils.clip_grad_norm_
7,680
36.468293
337
py
LAMOL
LAMOL-master/regularizers.py
import abc import math import torch from torch.optim import Optimizer, SGD from settings import args, FILL_VAL, TOKENS_WEIGHT from utils import get_losses, get_model_dir from parallel import DataParallelCriterion from torch.nn import CrossEntropyLoss, MSELoss import pickle as pkl import os from torch.nn.functional import softmax class Regularizer(abc.ABC): def __init__(self, model, parallel_model, dataloaders, task, prev_task=None): self.model = model self.parallel_model = parallel_model self.dataloaders = dataloaders self.task = task self.prev_task = prev_task @abc.abstractmethod def task_start_do(self): return NotImplemented @abc.abstractmethod def task_end_do(self): return NotImplemented def save_reg_params(self): model_dir = get_model_dir([self.task]) reg_params_path = os.path.join(model_dir, "reg_params.pkl") with open(reg_params_path, 'wb') as f: pkl.dump(self.model.reg_params,f) def load_reg_params(self): if self.prev_task: model_dir = get_model_dir([self.prev_task]) reg_params_path = os.path.join(model_dir, "reg_params.pkl") with open(reg_params_path, 'rb') as f: self.model.reg_params = pkl.load(f) input() class MAS(Regularizer): def task_start_do(self,freeze_layers=[]): #self.load_reg_params() task_start_do(self.model, freeze_layers) def task_end_do(self): updater = Omega_update(self.model.parameters(), lr=0.0001, momentum=0.9) compute_importance(self.model, self.parallel_model, updater, self.dataloaders) accumulate_reg_params(self.model) self.save_reg_params() class EWC(Regularizer): def task_start_do(self,freeze_layers=[]): #self.load_reg_params() task_start_do(self.model, freeze_layers) def task_end_do(self): updater = Omega_update(self.model.parameters(), lr=0.0001, momentum=0.9) compute_importance(self.model, self.parallel_model, updater, self.dataloaders, loss_type="ewc") accumulate_reg_params(self.model) self.save_reg_params() REG_TYPES = { "mas": MAS, "ewc": EWC, } args.REG_TYPE_KEYS = REG_TYPE_KEYS = list(REG_TYPES.keys()) def task_start_do(model, freeze_layers=[]): if not hasattr(model,"reg_params"): initialize_reg_params(model,freeze_layers) else: clean_omega_sum(model,freeze_layers) def initialize_reg_params(model,freeze_layers=[]): """initialize an omega for each parameter to zero""" reg_params={} for name, param in model.named_parameters(): if not name in freeze_layers: # print('initializing param',name) omega=torch.FloatTensor(param.size()).zero_() omega=omega.cuda() init_val=param.data.clone() init_val=init_val.cuda() reg_param={} reg_param['omega'] = omega reg_param['omega_sum'] = omega #initialize the initial value to that before starting training reg_param['init_val'] = init_val reg_params[param]=reg_param if 'data_count' not in reg_params: reg_params['data_count'] = 0 reg_params['lambda'] = args.reg_lambda model.reg_params = reg_params def clean_omega_sum(model,freeze_layers=[]): for name, param in model.named_parameters(): if not name in freeze_layers: omega=torch.FloatTensor(param.size()).zero_() omega=omega.cuda() reg_param = model.reg_params.get(param) reg_param['omega_sum'] = omega model.reg_params[param]=reg_param model.reg_params['data_count'] = 0 class Weight_Regularized_AdamW(Optimizer): """ Implements Adam algorithm with weight decay fix. Parameters: lr (float): learning rate. Default 1e-3. betas (tuple of 2 floats): Adams beta parameters (b1, b2). Default: (0.9, 0.999) eps (float): Adams epsilon. Default: 1e-6 weight_decay (float): Weight decay. Default: 0.0 correct_bias (bool): can be set to False to avoid correcting bias in Adam (e.g. like in Bert TF repository). Default True. """ def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-6, weight_decay=0.0, correct_bias=True): if lr < 0.0: raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr)) if not 0.0 <= betas[0] < 1.0: raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[0])) if not 0.0 <= betas[1] < 1.0: raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[1])) if not 0.0 <= eps: raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(eps)) defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, correct_bias=correct_bias) super(Weight_Regularized_AdamW, self).__init__(params, defaults) def step(self, reg_params, closure=None): """Performs a single optimization step. Arguments: closure (callable, optional): A closure that reevaluates the model and returns the loss. """ loss = None if closure is not None: loss = closure() reg_lambda=reg_params.get('lambda') for group in self.param_groups: for p in group['params']: if p.grad is None: continue grad = p.grad.data if grad.is_sparse: raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead') state = self.state[p] # State initialization if len(state) == 0: state['step'] = 0 # Exponential moving average of gradient values state['exp_avg'] = torch.zeros_like(p.data) # Exponential moving average of squared gradient values state['exp_avg_sq'] = torch.zeros_like(p.data) exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] beta1, beta2 = group['betas'] state['step'] += 1 # Decay the first and second moment running average coefficient # In-place operations to update the averages at the same time exp_avg.mul_(beta1).add_(1.0 - beta1, grad) exp_avg_sq.mul_(beta2).addcmul_(1.0 - beta2, grad, grad) denom = exp_avg_sq.sqrt().add_(group['eps']) step_size = group['lr'] if group['correct_bias']: # No bias correction for Bert bias_correction1 = 1.0 - beta1 ** state['step'] bias_correction2 = 1.0 - beta2 ** state['step'] step_size = step_size * math.sqrt(bias_correction2) / bias_correction1 p.data.addcdiv_(-step_size, exp_avg, denom) # 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 to 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. # Add weight decay at the end (fixed version) #Regularize PART CODE GOES HERE if p in reg_params: reg_param=reg_params.get(p) #get omega for this parameter omega=reg_param.get('omega') #initial value when the training start init_val=reg_param.get('init_val') curr_weight_val=p.data #get the difference weight_dif=curr_weight_val.add(-1,init_val) #compute the MAS penalty regulizer=weight_dif.mul(2*reg_lambda*omega) del weight_dif del curr_weight_val del omega del init_val #add the MAS regulizer to the gradient # grad.add_(regulizer) p.data.add_(-group['lr'], regulizer) del regulizer #Regularize PART CODE ENDS if group['weight_decay'] > 0.0: p.data.add_(-group['lr'] * group['weight_decay'], p.data) return loss # update omega for one task; use in compute_importance class Omega_update(SGD): """ Update the paramerter importance using the gradient of the function output norm. To be used at deployment time. reg_params:parameters omega to be updated batch_index,batch_size:used to keep a running average over the seen samples """ def __init__(self, params, lr=0.001, momentum=0, dampening=0, weight_decay=0, nesterov=False): super(Omega_update, self).__init__(params,lr,momentum,dampening,weight_decay,nesterov) def __setstate__(self, state): super(Omega_update, self).__setstate__(state) def step(self, reg_params, batch_size, closure=None): """ Performs a single parameters importance update setp """ #print('************************DOING A STEP************************') reg_params['data_count'] += batch_size loss = None if closure is not None: loss = closure() for group in self.param_groups: #if the parameter has an omega to be updated for p in group['params']: #print('************************ONE PARAM************************') if p.grad is None: continue if p in reg_params: #HERE MAS IMPOERANCE UPDATE GOES #get the gradient unreg_dp = p.grad.data.clone() reg_param = reg_params.get(p) #get parameter omega omega = reg_param.get('omega_sum') if args.seq_train_type == "ewc": omega = omega.add((unreg_dp)**2) else: omega = omega.add(unreg_dp.abs_()) reg_param['omega_sum'] = omega reg_params[p] = reg_param #HERE MAS IMPOERANCE UPDATE ENDS return loss#HAS NOTHING TO DO # update omega for one task def compute_importance(model, parallel_model, updater, dataloaders, loss_type="l2"): """Mimic the depoloyment setup where the model is applied on some samples and those are used to update the importance params Uses the L2norm of the function output. This is what we MAS uses as default """ # model.eval() # Set model to training mode so we get the gradient # train_loss_fct = DataParallelCriterion(CrossEntropyLoss(ignore_index=FILL_VAL), args.device_ids) softmax = torch.nn.Softmax(dim=-1) if loss_type == "l2": loss_fct = DataParallelCriterion(torch.nn.MSELoss(reduction='mean'), args.device_ids) elif loss_type == "l1": loss_fct = DataParallelCriterion(torch.nn.L1Loss(reduction='mean'), args.device_ids) elif loss_type == "ewc": CELoss = CrossEntropyLoss(ignore_index=FILL_VAL, reduction='mean', weight=TOKEN_WEIGHT) loss_fct = DataParallelCriterion(CELoss, args.device_ids) # Iterate over data. for dataloader in dataloaders: for cq, len_cq, cqa, len_cqa, Y, _, _ in dataloader: # get the inputs n_inputs = sum(len(_cq) for _cq in cq) for i in range(len(cqa)): cq[i] = (cq[i].to(args.device_ids[i]),) len_cq[i] = len_cq[i].to(args.device_ids[i]) cqa[i] = (cqa[i].to(args.device_ids[i]),) len_cqa[i] = len_cqa[i].to(args.device_ids[i]) Y[i] = Y[i].to(args.device_ids[i]) # zero the parameter gradients updater.zero_grad() # forward if loss_type != "ewc": logits = parallel_model(cq) logits = [logit[range(len(logit)), len_cq[i]-1, :] for i, logit in enumerate(logits)] #logits = [softmax(logit, dim=-1) for logit in logits] target_zeros = [torch.zeros(logit.size()).to(args.device_ids[i]) for i, logit in enumerate(logits)] logits = [softmax(logit) for logit in logits] if loss_type == "l2": targets = loss_fct(logits, target_zeros) elif loss_type == "l1": targets = loss_fct(logits, target_zeros) else: targets, _ = get_losses(parallel_model, cqa, Y, None, None, loss_fct) targets /= n_inputs #compute the gradients targets.backward() #update the parameters importance updater.step(model.reg_params, n_inputs) # omega of task1 + omega of task2 ... # new_omega=omega_sum/data_count; omega=new_omega+prev_omega def accumulate_reg_params(model, freeze_layers=[]): """accumelate the newly computed omega with the previously stroed one from the old previous tasks""" for name, param in model.named_parameters(): if not name in freeze_layers: if param in model.reg_params: reg_param=model.reg_params.get(param) # print('restoring previous omega',name) prev_omega=reg_param.get('omega') new_omega=reg_param.get('omega_sum') / model.reg_params["data_count"] acc_omega=torch.add(prev_omega,new_omega) del reg_param['omega_sum'] reg_param['omega'] = acc_omega model.reg_params[param]=reg_param del prev_omega del new_omega del acc_omega else: if param in model.reg_params: reg_param=model.reg_params.get(param) # print('removing unused omega',name) del reg_param['omega'] del model.reg_params[param] class Weight_Regularized_SGD(SGD): r"""Implements SGD training with importance params regulization. IT inherents stochastic gradient descent (optionally with momentum). Nesterov momentum is based on the formula from """ def __init__(self, params, lr=0.001, momentum=0, dampening=0, weight_decay=0, nesterov=False): super(Weight_Regularized_SGD, self).__init__(params, lr,momentum,dampening,weight_decay,nesterov) def __setstate__(self, state): super(Weight_Regularized_SGD, self).__setstate__(state) def step(self, reg_params,closure=None): loss = None if closure is not None: loss = closure() reg_lambda=reg_params.get('lambda') for group in self.param_groups: weight_decay = group['weight_decay'] momentum = group['momentum'] dampening = group['dampening'] nesterov = group['nesterov'] for p in group['params']: if p.grad is None: continue d_p = p.grad.data #MAS PART CODE GOES HERE #if this param has an omega to use for regulization if p in reg_params: reg_param=reg_params.get(p) #get omega for this parameter omega=reg_param.get('omega') #initial value when the training start init_val=reg_param.get('init_val') curr_wegiht_val=p.data #move the tensors to cuda init_val=init_val.cuda() omega=omega.cuda() #get the difference weight_dif=curr_wegiht_val.add(-1,init_val) #compute the MAS penalty regulizer=weight_dif.mul(2*reg_lambda*omega) del weight_dif del curr_wegiht_val del omega del init_val #add the MAS regulizer to the gradient d_p.add_(regulizer) del regulizer #MAS PARAT CODE ENDS if weight_decay != 0: d_p.add_(weight_decay,p.data.sign()) if momentum != 0: param_state = self.state[p] if 'momentum_buffer' not in param_state: buf = param_state['momentum_buffer'] = d_p.clone() else: buf = param_state['momentum_buffer'] buf.mul_(momentum).add_(1 - dampening, d_p) if nesterov: d_p = d_p.add(momentum, buf) else: d_p = buf p.data.add_(-group['lr'], d_p) return loss
17,354
39.549065
137
py
LAMOL
LAMOL-master/loss_scaler.py
# coding=utf-8 # Copyright (c) 2019, NVIDIA CORPORATION. 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. import torch # import mpu # item() is a recent addition, so this helps with backward compatibility. def to_python_float(t): if hasattr(t, 'item'): return t.item() else: return t[0] class LossScaler: """ Class that manages a static loss scale. This class is intended to interact with :class:`FP16_Optimizer`, and should not be directly manipulated by the user. Use of :class:`LossScaler` is enabled via the ``static_loss_scale`` argument to :class:`FP16_Optimizer`'s constructor. Args: scale (float, optional, default=1.0): The loss scale. """ def __init__(self, scale=1): self.cur_scale = scale # `params` is a list / generator of torch.Variable def has_overflow(self, params): return False # `x` is a torch.Tensor def _has_inf_or_nan(x): return False def update_scale(self, overflow): pass @property def loss_scale(self): return self.cur_scale def scale_gradient(self, module, grad_in, grad_out): return tuple(self.loss_scale * g for g in grad_in) def backward(self, loss, retain_graph=False): scaled_loss = loss*self.loss_scale scaled_loss.backward(retain_graph=retain_graph) class DynamicLossScaler: """ Class that manages dynamic loss scaling. It is recommended to use :class:`DynamicLossScaler` indirectly, by supplying ``dynamic_loss_scale=True`` to the constructor of :class:`FP16_Optimizer`. However, it's important to understand how :class:`DynamicLossScaler` operates, because the default options can be changed using the the ``dynamic_loss_args`` argument to :class:`FP16_Optimizer`'s constructor. Loss scaling is designed to combat the problem of underflowing gradients encountered at long times when training fp16 networks. Dynamic loss scaling begins by attempting a very high loss scale. Ironically, this may result in OVERflowing gradients. If overflowing gradients are encountered, :class:`DynamicLossScaler` informs :class:`FP16_Optimizer` that an overflow has occurred. :class:`FP16_Optimizer` then skips the update step for this particular iteration/minibatch, and :class:`DynamicLossScaler` adjusts the loss scale to a lower value. If a certain number of iterations occur without overflowing gradients detected, :class:`DynamicLossScaler` increases the loss scale once more. In this way :class:`DynamicLossScaler` attempts to "ride the edge" of always using the highest loss scale possible without incurring overflow. Args: init_scale (float, optional, default=2**32): Initial loss scale attempted by :class:`DynamicLossScaler.` scale_factor (float, optional, default=2.0): Factor used when adjusting the loss scale. If an overflow is encountered, the loss scale is readjusted to loss scale/``scale_factor``. If ``scale_window`` consecutive iterations take place without an overflow, the loss scale is readjusted to loss_scale*``scale_factor``. scale_window (int, optional, default=1000): Number of consecutive iterations without an overflow to wait before increasing the loss scale. """ def __init__(self, init_scale=2**32, scale_factor=2., scale_window=1000, min_scale=1, delayed_shift=1, consecutive_hysteresis=False): self.cur_scale = init_scale self.cur_iter = 0 self.last_overflow_iter = -1 self.scale_factor = scale_factor self.scale_window = scale_window self.min_scale = min_scale self.delayed_shift = delayed_shift self.cur_hysteresis = delayed_shift self.consecutive_hysteresis = consecutive_hysteresis # `params` is a list / generator of torch.Variable def has_overflow_serial(self, params): for p in params: if p.grad is not None and DynamicLossScaler._has_inf_or_nan(p.grad.data): return True return False def has_overflow(self, params): overflow = self.has_overflow_serial(params) # Since each model parallel GPU carries only part of the model, # make sure overflow flag is synced across all the model parallel GPUs # overflow_gpu = torch.cuda.ByteTensor([overflow]) # torch.distributed.all_reduce(overflow_gpu, # op=torch.distributed.ReduceOp.MAX, # group=mpu.get_model_parallel_group()) # overflow = overflow_gpu[0].item() return bool(overflow) # `x` is a torch.Tensor def _has_inf_or_nan(x): try: # if x is half, the .float() incurs an additional deep copy, but it's necessary if # Pytorch's .sum() creates a one-element tensor of the same type as x # (which is true for some recent version of pytorch). cpu_sum = float(x.float().sum()) # More efficient version that can be used if .sum() returns a Python scalar # cpu_sum = float(x.sum()) except RuntimeError as instance: # We want to check if inst is actually an overflow exception. # RuntimeError could come from a different error. # If so, we still want the exception to propagate. if "value cannot be converted" not in instance.args[0]: raise return True else: if cpu_sum == float('inf') or cpu_sum == -float('inf') or cpu_sum != cpu_sum: return True return False # `overflow` is boolean indicating whether the gradient overflowed def update_scale(self, overflow): if not hasattr(self, 'min_scale'): self.min_scale = 1 if not hasattr(self, 'delayed_shift'): self.delayed_shift = 1 if not hasattr(self, 'cur_hysteresis'): self.cur_hysteresis = 1 if not hasattr(self, 'consecutive_hysteresis'): self.consecutive_hysteresis = True if overflow: # self.cur_scale /= self.scale_factor if self.delayed_shift == 1 or self.cur_hysteresis == 1: self.cur_scale = max(self.cur_scale/self.scale_factor, self.min_scale) else: self.cur_hysteresis -= 1 self.last_overflow_iter = self.cur_iter else: if self.consecutive_hysteresis: self.cur_hysteresis = self.delayed_shift if (self.cur_iter - self.last_overflow_iter) % self.scale_window == 0: if not self.consecutive_hysteresis: self.cur_hysteresis = self.delayed_shift self.cur_scale *= self.scale_factor self.cur_iter += 1 @property def loss_scale(self): return self.cur_scale def scale_gradient(self, module, grad_in, grad_out): return tuple(self.loss_scale * g for g in grad_in) def backward(self, loss, retain_graph=False): scaled_loss = loss*self.loss_scale scaled_loss.backward(retain_graph=retain_graph) ############################################################## # Example usage below here -- assuming it's in a separate file ############################################################## """ TO-DO separate out into an example. if __name__ == "__main__": import torch from torch.autograd import Variable from dynamic_loss_scaler import DynamicLossScaler # N is batch size; D_in is input dimension; # H is hidden dimension; D_out is output dimension. N, D_in, H, D_out = 64, 1000, 100, 10 # Create random Tensors to hold inputs and outputs, and wrap them in Variables. x = Variable(torch.randn(N, D_in), requires_grad=False) y = Variable(torch.randn(N, D_out), requires_grad=False) w1 = Variable(torch.randn(D_in, H), requires_grad=True) w2 = Variable(torch.randn(H, D_out), requires_grad=True) parameters = [w1, w2] learning_rate = 1e-6 optimizer = torch.optim.SGD(parameters, lr=learning_rate) loss_scaler = DynamicLossScaler() for t in range(500): y_pred = x.mm(w1).clamp(min=0).mm(w2) loss = (y_pred - y).pow(2).sum() * loss_scaler.loss_scale print('Iter {} loss scale: {}'.format(t, loss_scaler.loss_scale)) print('Iter {} scaled loss: {}'.format(t, loss.data[0])) print('Iter {} unscaled loss: {}'.format(t, loss.data[0] / loss_scaler.loss_scale)) # Run backprop optimizer.zero_grad() loss.backward() # Check for overflow has_overflow = DynamicLossScaler.has_overflow(parameters) # If no overflow, unscale grad and update as usual if not has_overflow: for param in parameters: param.grad.data.mul_(1. / loss_scaler.loss_scale) optimizer.step() # Otherwise, don't do anything -- ie, skip iteration else: print('OVERFLOW!') # Update loss scale for next iteration loss_scaler.update_scale(has_overflow) """
9,788
40.130252
326
py
LAMOL
LAMOL-master/train.py
import torch from torch.utils.data import DataLoader from torch import nn from pytorch_transformers import AdamW, WEIGHTS_NAME, WarmupLinearSchedule import csv import numpy as np import os import logging from fp16 import FP16_Module, FP16_Optimizer from parallel import DataParallelModel, DataParallelCriterion from collections import OrderedDict from utils import * from settings import args, TASK_DICT, init_logging, MODEL_CONFIG, MODEL_CLASS, SPECIAL_TOKENS, CONFIG_CLASS from settings import TOKENIZER, SPECIAL_TOKEN_IDS, FILL_VAL, SAVE_NAME, FINAL_SAVE_NAME, TOKENS_WEIGHT, CONFIG_NAME from scheduler import AnnealingLR from regularizers import REG_TYPES, REG_TYPE_KEYS, Weight_Regularized_AdamW, Weight_Regularized_SGD from torch.nn import CrossEntropyLoss logger = logging.getLogger(__name__) def train(task_ids, model): tasks = [args.tasks[task_id] for task_id in task_ids] logger.info("start to train { task: %s, seq train type: %s }" % (tasks, args.seq_train_type)) model_dir = get_model_dir(tasks) make_dir(model_dir) train_dataset = [TASK_DICT[t]["train"] for t in tasks] train_extra_data = [] if "lll" in args.seq_train_type and task_ids[0] > 0 and not args.skip_tasks: prev_task = args.tasks[task_ids[0]-1] with torch.no_grad(): create_extra_data(tasks[0], prev_task, model, train_extra_data) elif "gem" in args.seq_train_type and task_ids[0] > 0: get_real_data(tasks[0], train_extra_data, accum=False, encode=True) args.memory_data.append(train_extra_data) train_extra_data = [] logger.info('extra training data size: {}'.format(len(train_extra_data))) if not model: # which_model_to_load = model_dir if os.path.isfile(os.path.join(model_dir, FINAL_SAVE_NAME)) else args.model_name model = MODEL_CLASS.from_pretrained(args.model_name).cuda() model.resize_token_embeddings(len(TOKENIZER)) if not args.fp32: model = FP16_Module(model) gen_token = get_gen_token(tasks[0]) TOKENIZER.add_tokens([gen_token]) TOKENIZER.save_pretrained(model_dir) SPECIAL_TOKENS[tasks[0]] = gen_token SPECIAL_TOKEN_IDS[tasks[0]] = TOKENIZER.convert_tokens_to_ids(gen_token) logger.info('gen token = {} , gen token id = {}'.format(gen_token, SPECIAL_TOKEN_IDS[tasks[0]])) MODEL_CONFIG.vocab_size = len(TOKENIZER) MODEL_CONFIG.to_json_file(os.path.join(model_dir,CONFIG_NAME)) global TOKENS_WEIGHT if len(TOKENIZER) != TOKENS_WEIGHT.shape[0]: TOKENS_WEIGHT = torch.cat((TOKENS_WEIGHT, torch.ones([1]).cuda())) if args.skip_tasks and len(tasks) == 1: logger.info("*********** skip task: {} ***********".format(tasks[0])) if tasks[0] in args.skip_tasks: if len(args.skip_tasks) == 1: model_dir = get_model_dir(tasks) model_path = os.path.join(model_dir, FINAL_SAVE_NAME) config_path = os.path.join(model_dir,CONFIG_NAME) model_config = CONFIG_CLASS.from_json_file(config_path) model = MODEL_CLASS(model_config).cuda() state_dict = torch.load(model_path) model.load_state_dict(state_dict) if not args.fp32: model = FP16_Module(model) if args.seq_train_type in REG_TYPE_KEYS: logger.info("calulating reg_params ...") train_qadata = QADataset(train_dataset, "train", SPECIAL_TOKEN_IDS[tasks[0]], train_extra_data) max_train_batch_size = max(len(train_qadata) // args.min_n_steps, args.min_batch_size) train_dataloader = create_dataloader(train_qadata, "train", max_train_batch_size) parallel_model = DataParallelModel(WrapModel(model), args.device_ids) regularizer = REG_TYPES[args.seq_train_type](model, parallel_model, [train_dataloader], tasks[0]) regularizer.task_start_do() regularizer.task_end_do() torch.save(model.state_dict(), os.path.join(model_dir, FINAL_SAVE_NAME)) logger.info("done reg_params!") args.skip_tasks.remove(tasks[0]) return model model.resize_token_embeddings(len(TOKENIZER)) if not args.fp32: # again because resize_token_embeddings makes embedding layer fp32 model = FP16_Module(model) parallel_model = DataParallelModel(WrapModel(model), args.device_ids) train_qadata = QADataset(train_dataset, "train", SPECIAL_TOKEN_IDS[tasks[0]], train_extra_data) max_train_batch_size = max(len(train_qadata) // args.min_n_steps, args.min_batch_size) train_dataloader = create_dataloader(train_qadata, "train", max_train_batch_size) if not args.unbound and args.seq_train_type != "multitask": #n_train_epochs = TASK_DICT[tasks[0]]["n_train_epochs"] n_train_epochs = args.n_train_epochs[tasks[0]] else: n_train_epochs = args.n_train_epochs['_'.join(tasks)] n_train_optimization_steps = len(train_qadata) * n_train_epochs logger.info('len of train dataset: {} , max train batch size {} , num of opt steps: {}'.format( len(train_qadata), max_train_batch_size, n_train_optimization_steps)) param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} ] if "gem" in args.seq_train_type: model.task_id = task_ids[0] if not hasattr(model, "grad_dims"): model.grad_dims = [] for param in model.parameters(): model.grad_dims.append(param.data.numel()) if not hasattr(model, "grads"): model.grads = torch.zeros(sum(model.grad_dims),len(args.tasks)) model.grads = model.grads.cuda() if args.seq_train_type in REG_TYPE_KEYS: optimizer = Weight_Regularized_AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) else: optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) if not args.fp32: optimizer = FP16_Optimizer(optimizer, static_loss_scale=None, dynamic_loss_scale=True, dynamic_loss_args={'scale_window': 100, 'min_scale': 1, 'delayed_shift': 2}) scheduler = AnnealingLR(optimizer, start_lr=args.learning_rate, warmup_iter=int(args.n_warmup_ratio*len(train_qadata)), num_iters=int(n_train_optimization_steps), decay_style=args.decay_style) train_loss_fct = DataParallelCriterion(CrossEntropyLoss(ignore_index=FILL_VAL, weight=TOKENS_WEIGHT), args.device_ids) if args.seq_train_type in REG_TYPE_KEYS: copy_train_dataloader = create_dataloader(train_qadata, "train", max_train_batch_size) prev_task = args.tasks[task_ids[0]-1] regularizer = REG_TYPES[args.seq_train_type](model, parallel_model, [copy_train_dataloader], tasks[0], prev_task) regularizer.task_start_do() tot_n_steps = 0 train_once = TrainStep(model, optimizer, scheduler) if "gem" in args.seq_train_type and task_ids[0] != 0: gem_step = GEMStep(model, parallel_model, train_loss_fct, optimizer) model.train() for ep in range(n_train_epochs): cum_loss, cum_qa_loss, cum_lm_loss, cur_n_inputs = 0, 0, 0, 0 for n_steps, (_, _, cqa, _, Y, gen_X, gen_Y) in enumerate(train_dataloader): n_inputs = sum(_cqa.shape[0] for _cqa in cqa) for i in range(len(cqa)): cqa[i] = (cqa[i].to(args.device_ids[i]),) Y[i] = Y[i].to(args.device_ids[i]) gen_X[i] = (gen_X[i].to(args.device_ids[i]),) gen_Y[i] = gen_Y[i].to(args.device_ids[i]) losses = get_losses(parallel_model, cqa, Y, gen_X, gen_Y, train_loss_fct) loss = sum(losses) if "gem" in args.seq_train_type and task_ids[0] != 0: gem_step(task_ids[0]) train_once(loss, n_inputs) qa_loss = losses[0].item() * n_inputs lm_loss = losses[1].item() * n_inputs cum_loss += (qa_loss + lm_loss) cum_qa_loss += qa_loss cum_lm_loss += lm_loss cur_n_inputs += n_inputs if (n_steps + 1 ) % args.logging_steps == 0: logger.info('progress {:.3f} , lr {:.1E} , loss {:.3f} , qa loss {:.3f} , lm loss {:.3f} , avg batch size {:.1f}'.format( ep + cur_n_inputs/len(train_qadata), scheduler.get_lr(), cum_loss/cur_n_inputs, cum_qa_loss/cur_n_inputs, cum_lm_loss/cur_n_inputs, cur_n_inputs/(n_steps + 1) )) torch.save(model.state_dict(), os.path.join(model_dir, SAVE_NAME+str(ep+1))) tot_n_steps += (n_steps + 1) logger.info('epoch {}/{} done , tot steps {} , lr {:.1E} , loss {:.2f} , qa loss {:.2f} , lm loss {:.2f} , avg batch size {:.1f}'.format( ep+1, n_train_epochs, tot_n_steps, scheduler.get_lr(), cum_loss/cur_n_inputs, cum_qa_loss/cur_n_inputs, cum_lm_loss/cur_n_inputs, cur_n_inputs/(n_steps+1) )) # task end do for reg if args.seq_train_type in REG_TYPE_KEYS: regularizer.task_end_do() torch.save(model.state_dict(), os.path.join(model_dir, FINAL_SAVE_NAME)) return model if __name__ == '__main__': if not args.debug: logging.getLogger("pytorch_transformers").setLevel(logging.WARNING) logging.getLogger("pytorch_transformers.tokenization_utils").setLevel(logging.CRITICAL) make_dir(args.model_dir_root) init_logging(os.path.join(args.model_dir_root, 'log_train.txt')) logger.info('args = {}'.format(str(args))) model = None if args.seq_train_type == "multitask": model = train(list(range(len(args.tasks))), model) else: if args.unbound: TASK_DICT = lll_unbound_setting(split_size=args.unbound) for task_id in range(len(args.tasks)): model = train([task_id], model)
10,296
48.267943
166
py
LAMOL
LAMOL-master/parallel.py
import threading import torch from torch.nn.parallel import DataParallel from torch.nn.parallel.parallel_apply import get_a_var from torch.nn.parallel.scatter_gather import scatter torch_ver = torch.__version__[:3] __all__ = ['DataParallelModel', 'DataParallelCriterion'] class DataParallelModel(DataParallel): def forward(self, inputs, **kwargs): kwargs = scatter(kwargs, self.device_ids[:len(inputs)], self.dim) if len(self.device_ids) == 1: return (self.module(*inputs[0], **kwargs[0]),) replicas = self.replicate(self.module, self.device_ids[:len(inputs)]) outputs = self.parallel_apply(replicas, inputs, kwargs) return outputs def replicate(self, module, device_ids): modules = super(DataParallelModel, self).replicate(module, device_ids) execute_replication_callbacks(modules) return modules class DataParallelCriterion(DataParallel): def forward(self, inputs, targets, **kwargs): # input should be already scatterd # scattering the targets instead if not self.device_ids: return self.module(inputs, *targets, **kwargs) kwargs = scatter(kwargs, self.device_ids[:len(inputs)], self.dim) if len(self.device_ids) == 1: return self.module(inputs[0], targets[0], **kwargs[0]) replicas = self.replicate(self.module, self.device_ids[:len(inputs)]) outputs = _criterion_parallel_apply(replicas, inputs, targets, kwargs) return self.gather(outputs, self.output_device) def _criterion_parallel_apply(modules, inputs, targets, kwargs_tup=None, devices=None): assert len(modules) == len(inputs) assert len(targets) == len(inputs) if kwargs_tup: assert len(modules) == len(kwargs_tup) else: kwargs_tup = ({},) * len(modules) if devices is not None: assert len(modules) == len(devices) else: devices = [None] * len(modules) lock = threading.Lock() results = {} if torch_ver != "0.3": grad_enabled = torch.is_grad_enabled() def _worker(i, module, input, target, kwargs, device=None): if torch_ver != "0.3": torch.set_grad_enabled(grad_enabled) if device is None: device = get_a_var(input).get_device() try: with torch.cuda.device(device): # this also avoids accidental slicing of `input` if it is a Tensor if not isinstance(input, (list, tuple)): input = (input,) if not isinstance(target, (list, tuple)): target = (target,) output = module(*(input + target), **kwargs) with lock: results[i] = output except Exception as e: with lock: results[i] = e if len(modules) > 1: threads = [threading.Thread(target=_worker, args=(i, module, input, target, kwargs, device),) for i, (module, input, target, kwargs, device) in enumerate(zip(modules, inputs, targets, kwargs_tup, devices))] for thread in threads: thread.start() for thread in threads: thread.join() else: _worker(0, modules[0], inputs[0], targets[0], kwargs_tup[0], devices[0]) outputs = [] for i in range(len(inputs)): output = results[i] if isinstance(output, Exception): raise output outputs.append(output) return outputs ########################################################################### # Adapted from Synchronized-BatchNorm-PyTorch. # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch # class CallbackContext(object): pass def execute_replication_callbacks(modules): """ Execute an replication callback `__data_parallel_replicate__` on each module created by original replication. The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)` Note that, as all modules are isomorphism, we assign each sub-module with a context (shared among multiple copies of this module on different devices). Through this context, different copies can share some information. We guarantee that the callback on the master copy (the first copy) will be called ahead of calling the callback of any slave copies. """ master_copy = modules[0] nr_modules = len(list(master_copy.modules())) ctxs = [CallbackContext() for _ in range(nr_modules)] for i, module in enumerate(modules): for j, m in enumerate(module.modules()): if hasattr(m, '__data_parallel_replicate__'): m.__data_parallel_replicate__(ctxs[j], i)
4,814
36.038462
91
py
LAMOL
LAMOL-master/fp16.py
# coding=utf-8 # Copyright (c) 2019, NVIDIA CORPORATION. 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. """Stable version of apex FP16 Optimizer""" import torch from torch import nn from torch.autograd import Variable from torch.nn.parameter import Parameter from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors from loss_scaler import DynamicLossScaler, LossScaler from fp16util import model_grads_to_master_grads, master_params_to_model_params, clip_grad_norm FLOAT_TYPES = (torch.FloatTensor, torch.cuda.FloatTensor) HALF_TYPES = (torch.HalfTensor, torch.cuda.HalfTensor) def conversion_helper(val, conversion): """Apply conversion to val. Recursively apply conversion if `val` is a nested tuple/list structure.""" if not isinstance(val, (tuple, list)): return conversion(val) rtn = [conversion_helper(v, conversion) for v in val] if isinstance(val, tuple): rtn = tuple(rtn) return rtn def fp32_to_fp16(val): """Convert fp32 `val` to fp16""" def half_conversion(val): val_typecheck = val if isinstance(val_typecheck, (Parameter, Variable)): val_typecheck = val.data if isinstance(val_typecheck, FLOAT_TYPES): val = val.half() return val return conversion_helper(val, half_conversion) def fp16_to_fp32(val): """Convert fp16 `val` to fp32""" def float_conversion(val): val_typecheck = val if isinstance(val_typecheck, (Parameter, Variable)): val_typecheck = val.data if isinstance(val_typecheck, HALF_TYPES): val = val.float() return val return conversion_helper(val, float_conversion) class FP16_Module(nn.Module): def __init__(self, module): super(FP16_Module, self).__init__() self.add_module('module', module.half()) def forward(self, *inputs, **kwargs): return fp16_to_fp32(self.module(*(fp32_to_fp16(inputs)), **kwargs)) def state_dict(self, destination=None, prefix='', keep_vars=False): return self.module.state_dict(destination, prefix, keep_vars) def load_state_dict(self, state_dict, strict=True): self.module.load_state_dict(state_dict, strict=strict) def resize_token_embeddings(self, len_tokenizer): self.module.resize_token_embeddings(len_tokenizer) # TODO: Update overflow check + downscale to use Carl's fused kernel. class FP16_Optimizer(object): """ :class:`FP16_Optimizer` is designed to wrap an existing PyTorch optimizer, and manage static or dynamic loss scaling and master weights in a manner transparent to the user. For standard use, only two lines must be changed: creating the :class:`FP16_Optimizer` instance, and changing the call to ``backward``. Example:: model = torch.nn.Linear(D_in, D_out).cuda().half() optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) # Name the FP16_Optimizer instance to replace the existing optimizer # (recommended but not required): optimizer = FP16_Optimizer(optimizer, static_loss_scale = 128.0) ... # loss.backward() becomes: optimizer.backward(loss) ... Example with dynamic loss scaling:: ... optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) # optional arg to control dynamic loss scaling behavior # dynamic_loss_args={'scale_window' : 500}) # Usually, dynamic_loss_args is not necessary. Args: init_optimizer (torch.optim.optimizer): Existing optimizer created with the parameters to optimize. Internally, :class:`FP16_Optimizer` replaces the passed optimizer's fp16 parameters, if any, with fp32 master parameters copied from the original ones. :class:`FP16_Optimizer` also stores references to the original fp16 parameters, and updates these fp16 parameters from the master fp32 copy at the end of each :attr:`step`. static_loss_scale (float, optional, default=1.0): Loss scale used internally to scale gradients computed by the model. Any fp16 gradients will be copied to fp32, then downscaled before being applied to the fp32 master params, so ``static_loss_scale`` should not affect learning rate. dynamic_loss_scale (bool, optional, default=False): Use dynamic loss scaling. If True, this will override any ``static_loss_scale`` option. dynamic_loss_args (dict, optional, default=None): Dict of kwargs that will be forwarded to the internal :class:`DynamicLossScaler` instance's constructor. Keys of this dict must match kwargs accepted by :class:`DynamicLossScaler`'s constructor. If ``dynamic_loss_args`` is unspecified, :class:`DynamicLossScaler`'s defaults will be used. verbose (bool, optional, default=True): By default, FP16_Optimizer's constructor prints out the parameters and parameter groups it is ingesting, as a sanity check. If this becomes annoying (e.g. for large models), it can be disabled by passing ``verbose=False``. ``verbose=False`` will not disable printing when the loss scale is readjusted during dynamic loss scaling. ``init_optimizer`` is expected to have been constructed in the ordinary way. It is recommended (although not required) that the newly constructed :class:`FP16_Optimizer` instance be named to replace ``init_optimizer``, for two reasons: First, it means that references to the same name later in the file will not have to change. Second, :class:`FP16_Optimizer` reserves the right (as an implementation detail) to modify ``init_optimizer``. If you do choose a unique name for the new :class:`FP16_Optimizer` instance, you should only work with this new instance, because the preexisting optimizer might no longer behave as expected. ``init_optimizer`` may be any Pytorch optimizer. It may contain a mixture of fp16 and fp32 parameters organized into any number of ``param_groups`` with different hyperparameters. The :class:`FP16_Optimizer` constructor will ingest these ``param_groups`` and remember them. Calls to :: loss.backward() must be replaced with :: optimizer.backward(loss) because :class:`FP16_Optimizer` requires ownership of the backward pass to implement loss scaling and copies to master gradients. .. note:: Loss scaling, either static or dynamic, is orthogonal to learning rate, because gradients are downscaled before being applied. This means that adjusting the loss scale, or using dynamic loss scaling, should not require retuning the learning rate or any other hyperparameters. **Advanced options** **Closures**: :class:`FP16_Optimizer` can wrap a Pytorch optimizer that receives a closure. See docstring for :attr:`step`. **Gradient clipping**: Use :attr:`clip_master_grads`. **Multiple losses**: If your model accumulates gradients from multiple losses, this can be made more efficient by supplying ``update_master_grads=False`` to :attr:`backward`. See docstring for :attr:`backward`. **Manually adjusting loss scale**: The current loss scale can be retrieved or set via :: print(optimizer.loss_scale) optimizer.loss_scale = new_loss_scale For static loss scaling, manually adjusting the loss scale over time is a reasonable thing to do. During later epochs, gradients may become smaller, and a higher loss scale may be required, analogous to scheduling the learning rate. Dynamic loss scaling is more subtle (see :class:`DynamicLossScaler`) and in this case, manually adjusting the loss scale is not recommended. **Multi_GPU training**: If the wrapped ``init_optimizer`` was created from a model wrapped in Pytorch DistributedDataParallel or Apex DistributedDataParallel, :class:`FP16_Optimizer` should still work as intended. """ def __init__(self, init_optimizer, static_loss_scale=1.0, dynamic_loss_scale=False, dynamic_loss_args=None, verbose=False): if not torch.cuda.is_available: raise SystemError("Cannot use fp16 without CUDA.") self.verbose = verbose self.optimizer = init_optimizer # init_state_dict sets up an alternative way to cast per-param state tensors. # Stashing here in case https://github.com/pytorch/pytorch/issues/7733 makes it necessary. # init_state_dict = init_optimizer.state_dict() self.fp16_groups = [] self.fp32_from_fp16_groups = [] self.fp32_from_fp32_groups = [] for i, param_group in enumerate(self.optimizer.param_groups): self.maybe_print("FP16_Optimizer processing param group {}:".format(i)) fp16_params_this_group = [] fp32_params_this_group = [] fp32_from_fp16_params_this_group = [] for i, param in enumerate(param_group['params']): if param.requires_grad: if param.type() == 'torch.cuda.HalfTensor': self.maybe_print("FP16_Optimizer received torch.cuda.HalfTensor with {}" .format(param.size())) fp16_params_this_group.append(param) master_param = param.detach().clone().float() master_param.requires_grad = True # Copythe model parallel flag. # master_param.model_parallel = param.model_parallel param_group['params'][i] = master_param fp32_from_fp16_params_this_group.append(master_param) # Reset existing state dict key to the new master param. # We still need to recast per-param state tensors, if any, to FP32. if param in self.optimizer.state: self.optimizer.state[master_param] = self.optimizer.state.pop(param) elif param.type() == 'torch.cuda.FloatTensor': self.maybe_print("FP16_Optimizer received torch.cuda.FloatTensor with {}" .format(param.size())) fp32_params_this_group.append(param) param_group['params'][i] = param else: raise TypeError("Wrapped parameters must be either " "torch.cuda.FloatTensor or torch.cuda.HalfTensor. " "Received {}".format(param.type())) self.fp16_groups.append(fp16_params_this_group) self.fp32_from_fp16_groups.append(fp32_from_fp16_params_this_group) self.fp32_from_fp32_groups.append(fp32_params_this_group) # Leverage state_dict() and load_state_dict() to recast preexisting per-param state tensors self.optimizer.load_state_dict(self.optimizer.state_dict()) # alternative way to cast per-param state tensors: # self.optimizer.load_state_dict(init_state_dict) if dynamic_loss_scale: self.dynamic_loss_scale = True if dynamic_loss_args is not None: self.loss_scaler = DynamicLossScaler(**dynamic_loss_args) else: self.loss_scaler = DynamicLossScaler() else: self.dynamic_loss_scale = False self.loss_scaler = LossScaler(static_loss_scale) self.overflow = False self.first_closure_call_this_step = True self.clip_grad_norm = clip_grad_norm def maybe_print(self, msg): if self.verbose: print(msg) def __getstate__(self): raise RuntimeError("FP16_Optimizer should be serialized using state_dict().") def __setstate__(self, state): raise RuntimeError("FP16_Optimizer should be deserialized using load_state_dict().") def zero_grad(self, set_grads_to_None=False): """ Zero fp32 and fp16 parameter grads. """ # In principle, only the .grad attributes of the model params need to be zeroed, # because gradients are copied into the FP32 master params. However, we zero # all gradients owned by the optimizer, just to be safe: for group in self.optimizer.param_groups: for p in group['params']: if set_grads_to_None: p.grad = None else: if p.grad is not None: p.grad.detach_() p.grad.zero_() # Zero fp16 gradients owned by the model: for fp16_group in self.fp16_groups: for param in fp16_group: if set_grads_to_None: param.grad = None else: if param.grad is not None: param.grad.detach_() # as in torch.optim.optimizer.zero_grad() param.grad.zero_() def _check_overflow(self): params = [] for group in self.fp16_groups: for param in group: params.append(param) for group in self.fp32_from_fp32_groups: for param in group: params.append(param) self.overflow = self.loss_scaler.has_overflow(params) def _update_scale(self, has_overflow=False): self.loss_scaler.update_scale(has_overflow) def _master_params_to_model_params(self): for fp16_group, fp32_from_fp16_group in zip(self.fp16_groups, self.fp32_from_fp16_groups): master_params_to_model_params(fp16_group, fp32_from_fp16_group) def _model_params_to_master_params(self): for fp16_group, fp32_from_fp16_group in zip(self.fp16_groups, self.fp32_from_fp16_groups): master_params_to_model_params(fp32_from_fp16_group, fp16_group) # To consider: Integrate distributed with this wrapper by registering a hook on each variable # that does the overflow check, gradient copy + downscale, and fp32 allreduce in a different stream. def _model_grads_to_master_grads(self): for fp16_group, fp32_from_fp16_group in zip(self.fp16_groups, self.fp32_from_fp16_groups): model_grads_to_master_grads(fp16_group, fp32_from_fp16_group) def _downscale_master(self): if self.loss_scale != 1.0: for group in self.optimizer.param_groups: for param in group['params']: if param.grad is not None: param.grad.data.mul_(1./self.loss_scale) def clip_master_grads(self, max_norm, norm_type=2): """ Clips fp32 master gradients via ``torch.nn.utils.clip_grad_norm``. Args: max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the current fp32 gradients (viewed as a single vector). .. warning:: Returns -1 if the most recently computed fp16 gradients overflowed (that is, if ``self.overflow`` is ``True``). """ if not self.overflow: fp32_params = [] for param_group in self.optimizer.param_groups: for param in param_group['params']: fp32_params.append(param) return self.clip_grad_norm(fp32_params, max_norm, norm_type) else: return -1 def state_dict(self): """ Returns a dict containing the current state of this :class:`FP16_Optimizer` instance. This dict contains attributes of :class:`FP16_Optimizer`, as well as the state_dict of the contained Pytorch optimizer. Example:: checkpoint = {} checkpoint['model'] = model.state_dict() checkpoint['optimizer'] = optimizer.state_dict() torch.save(checkpoint, "saved.pth") """ state_dict = {} state_dict['loss_scaler'] = self.loss_scaler state_dict['dynamic_loss_scale'] = self.dynamic_loss_scale state_dict['overflow'] = self.overflow state_dict['first_closure_call_this_step'] = self.first_closure_call_this_step state_dict['optimizer_state_dict'] = self.optimizer.state_dict() state_dict['fp32_from_fp16'] = self.fp32_from_fp16_groups return state_dict def load_state_dict(self, state_dict): """ Loads a state_dict created by an earlier call to state_dict(). If ``fp16_optimizer_instance`` was constructed from some ``init_optimizer``, whose parameters in turn came from ``model``, it is expected that the user will call ``model.load_state_dict()`` before ``fp16_optimizer_instance.load_state_dict()`` is called. Example:: model = torch.nn.Linear(D_in, D_out).cuda().half() optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) optimizer = FP16_Optimizer(optimizer, static_loss_scale = 128.0) ... checkpoint = torch.load("saved.pth") model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer']) """ # I think it should actually be ok to reload the optimizer before the model. self.loss_scaler = state_dict['loss_scaler'] self.dynamic_loss_scale = state_dict['dynamic_loss_scale'] self.overflow = state_dict['overflow'] self.first_closure_call_this_step = state_dict['first_closure_call_this_step'] self.optimizer.load_state_dict(state_dict['optimizer_state_dict']) # At this point, the optimizer's references to the model's fp32 parameters are up to date. # The optimizer's hyperparameters and internal buffers are also up to date. # However, the fp32 master copies of the model's fp16 params stored by the optimizer are still # out of date. There are two options. # 1: Refresh the master params from the model's fp16 params. # This requires less storage but incurs precision loss. # 2: Save and restore the fp32 master copies separately. # We choose option 2. # # Pytorch Optimizer.load_state_dict casts saved buffers (e.g. momentum) to the type and device # of their associated parameters, because it's possible those buffers might not exist yet in # the current optimizer instance. In our case, as long as the current FP16_Optimizer has been # constructed in the same way as the one whose state_dict we are loading, the same master params # are guaranteed to exist, so we can just copy_() from the saved master params. for current_group, saved_group in zip(self.fp32_from_fp16_groups, state_dict['fp32_from_fp16']): for current, saved in zip(current_group, saved_group): current.data.copy_(saved.data) def step(self, closure=None): # could add clip option. """ If no closure is supplied, :attr:`step` should be called after ``fp16_optimizer_obj.backward(loss)``. :attr:`step` updates the fp32 master copy of parameters using the optimizer supplied to :class:`FP16_Optimizer`'s constructor, then copies the updated fp32 params into the fp16 params originally referenced by :class:`FP16_Optimizer`'s constructor, so the user may immediately run another forward pass using their model. If a closure is supplied, :attr:`step` may be called without a prior call to :attr:`backward(loss)`. This control flow is identical to `ordinary Pytorch optimizer use`_ with closures. However, the user should take care that any ``loss.backward()`` call within the closure has been replaced by ``fp16_optimizer_obj.backward(loss)``. Args: closure (optional): Closure that will be supplied to the underlying optimizer originally passed to :class:`FP16_Optimizer`'s constructor. closure should call :attr:`zero_grad()` on the :class:`FP16_Optimizer` object, compute the loss, call :attr:`backward(loss)`, and return the loss. Example with closure:: # optimizer is assumed to be an FP16_Optimizer object, previously constructed from an # existing pytorch optimizer. for input, target in dataset: def closure(): optimizer.zero_grad() output = model(input) loss = loss_fn(output, target) # loss.backward() becomes: optimizer.backward(loss) return loss optimizer.step(closure) .. warning:: Currently, calling :attr:`step` with a closure is not compatible with dynamic loss scaling. .. _`ordinary Pytorch optimizer use`: http://pytorch.org/docs/master/optim.html#optimizer-step-closure """ scale = self.loss_scaler.loss_scale self._update_scale(self.overflow) if self.overflow: self.maybe_print("OVERFLOW! Skipping step. Attempted loss scale: {}, reducing to {}" .format(scale, self.loss_scale)) return if closure is not None: retval = self._step_with_closure(closure) else: retval = self.optimizer.step() self._master_params_to_model_params() return retval def _step_with_closure(self, closure): def wrapped_closure(): # helpful for debugging # print("Calling wrapped_closure, first_closure_call_this_step = {}" # .format(self.first_closure_call_this_step)) if self.first_closure_call_this_step: # We expect that the fp16 params are initially fresh on entering self.step(), # so _master_params_to_model_params() is unnecessary the first time wrapped_closure() # is called within self.optimizer.step(). self.first_closure_call_this_step = False else: # If self.optimizer.step() internally calls wrapped_closure more than once, # it may update the fp32 params after each call. However, self.optimizer # doesn't know about the fp16 params at all. If the fp32 params get updated, # we can't rely on self.optimizer to refresh the fp16 params. We need # to handle that manually: self._master_params_to_model_params() # Our API expects the user to give us ownership of the backward() call by # replacing all calls to loss.backward() with optimizer.backward(loss). # This requirement holds whether or not the call to backward() is made within a closure. # If the user is properly calling optimizer.backward(loss) within "closure," # calling closure() here will give the fp32 master params fresh gradients # for the optimizer to play with, so all wrapped_closure needs to do is call # closure() and return the loss. temp_loss = closure() while(self.overflow): scale = self.loss_scaler.loss_scale self._update_scale(self.overflow) self.maybe_print("OVERFLOW within closure! Skipping step. Attempted loss scale: {}, " "reducing to {}".format(scale, self.loss_scale)) temp_loss = closure() return temp_loss retval = self.optimizer.step(wrapped_closure) self.first_closure_call_this_step = True return retval def backward(self, loss, update_master_grads=True, retain_graph=False): """ :attr:`backward` performs the following conceptual steps: 1. fp32_loss = loss.float() (see first Note below) 2. scaled_loss = fp32_loss*loss_scale 3. scaled_loss.backward(), which accumulates scaled gradients into the ``.grad`` attributes of the model's leaves (which may be fp16, fp32, or a mixture, depending how your model was defined). 4. fp16 grads are then copied to the master params' ``.grad`` attributes (see second Note), which are guaranteed to be fp32. 5. Finally, master grads are divided by loss_scale. In this way, after :attr:`backward`, the master params have fresh gradients, and :attr:`step` may be called. .. note:: :attr:`backward` internally converts the loss to fp32 before applying the loss scale. This provides some additional safety against overflow if the user has supplied an fp16 loss value. However, for maximum overflow safety, the user should compute the loss criterion (MSE, cross entropy, etc) in fp32 before supplying it to :attr:`backward`. .. warning:: The gradients found in a model's leaves after the call to :attr:`backward` should not be regarded as valid in general, because it's possible they have been scaled (and in the case of dynamic loss scaling, the scale factor may change over time). If the user wants to inspect gradients after a call to :attr:`backward`, only the master gradients should be regarded as valid. These can be retrieved via :attr:`inspect_master_grad_data()`. Args: loss: The loss output by the user's model. loss may be either float or half (but see first Note above). update_master_grads (bool, optional, default=True): Option to copy fp16 grads to fp32 grads on this call. By setting this to False, the user can delay the copy, which is useful to eliminate redundant fp16->fp32 grad copies if :attr:`backward` is being called on multiple losses in one iteration. If set to False, the user becomes responsible for calling :attr:`update_master_grads` before calling :attr:`step`. retain_graph (bool, optional, default=False): Forwards the usual ``retain_graph=True`` option to the internal call to ``loss.backward``. If ``retain_graph`` is being used to accumulate gradient values from multiple backward passes before calling ``optimizer.step``, passing ``update_master_grads=False`` is also recommended (see Example below). Example:: # Ordinary operation: optimizer.backward(loss) # Naive operation with multiple losses (technically valid, but less efficient): # fp32 grads will be correct after the second call, but # the first call incurs an unnecessary fp16->fp32 grad copy. optimizer.backward(loss1) optimizer.backward(loss2) # More efficient way to handle multiple losses: # The fp16->fp32 grad copy is delayed until fp16 grads from all # losses have been accumulated. optimizer.backward(loss1, update_master_grads=False) optimizer.backward(loss2, update_master_grads=False) optimizer.update_master_grads() """ # To consider: try multiple backward passes using retain_grad=True to find # a loss scale that works. After you find a loss scale that works, do a final dummy # backward pass with retain_graph=False to tear down the graph. Doing this would avoid # discarding the iteration, but probably wouldn't improve overall efficiency. self.loss_scaler.backward(loss.float(), retain_graph=retain_graph) if update_master_grads: self.update_master_grads() def update_master_grads(self): """ Copy the ``.grad`` attribute from stored references to fp16 parameters to the ``.grad`` attribute of the fp32 master parameters that are directly updated by the optimizer. :attr:`update_master_grads` only needs to be called if ``fp16_optimizer_obj.backward`` was called with ``update_master_grads=False``. """ if self.dynamic_loss_scale: self._check_overflow() if self.overflow: return self._model_grads_to_master_grads() self._downscale_master() def inspect_master_grad_data(self): """ When running with :class:`FP16_Optimizer`, ``.grad`` attributes of a model's fp16 leaves should not be regarded as truthful, because they might be scaled. After a call to :attr:`fp16_optimizer_obj.backward(loss)`, if no overflow was encountered, the fp32 master params' ``.grad`` attributes will contain valid gradients properly divided by the loss scale. However, because :class:`FP16_Optimizer` flattens some parameters, accessing them may be nonintuitive. :attr:`inspect_master_grad_data` allows those gradients to be viewed with shapes corresponding to their associated model leaves. Returns: List of lists (one list for each parameter group). The list for each parameter group is a list of the ``.grad.data`` attributes of the fp32 master params belonging to that group. """ if self.overflow: print("Warning: calling FP16_Optimizer.inspect_master_grad_data while in an overflow state. " "Gradients are currently invalid (may be inf, nan, or stale). Returning None.") return None else: # The optimizer owns only references to master params. master_grads_data = [] for param_group in self.optimizer.param_groups: master_grads_this_group = [] for param in param_group['params']: if param.grad is not None: master_grads_this_group.append(param.grad.data) else: master_grads_this_group.append(None) master_grads_data.append(master_grads_this_group) return master_grads_data # Promote loss scale so it can be retrieved or set via "fp16_optimizer_instance.loss_scale" def _get_loss_scale(self): return self.loss_scaler.loss_scale def _set_loss_scale(self, value): self.loss_scaler.cur_scale = value loss_scale = property(_get_loss_scale, _set_loss_scale) # Promote state so it can be retrieved or set via "fp16_optimizer_instance.state" def _get_state(self): return self.optimizer.state def _set_state(self, value): self.optimizer.state = value state = property(_get_state, _set_state) # Promote param_groups so it can be retrieved or set via "fp16_optimizer_instance.param_groups" # (for example, to adjust the learning rate) def _get_param_groups(self): return self.optimizer.param_groups def _set_param_groups(self, value): self.optimizer.param_groups = value param_groups = property(_get_param_groups, _set_param_groups)
31,831
49.28752
437
py
capacity-approaching-autoencoders
capacity-approaching-autoencoders-master/gammaDIME.py
from keras import backend as K # gamma-DIME loss def gamma_dime_loss(args): # define the parameter gamma gamma = 1 t_xy = args[0] t_xy_bar = args[1] loss = -(gamma*K.mean(K.log(t_xy)) - K.mean(K.pow(t_xy_bar, gamma))+1) return loss
256
24.7
74
py
capacity-approaching-autoencoders
capacity-approaching-autoencoders-master/Capacity-Approaching_AE.py
from __future__ import absolute_import, division, print_function, unicode_literals from keras.layers import Input, Dense, GaussianNoise, Concatenate, Lambda, Reshape, Flatten, Dropout from keras.layers import BatchNormalization, Activation, ZeroPadding2D from keras.models import Sequential, Model, load_model from keras.optimizers import Adam from keras import backend as K from uniform_noise import UniformNoise from gammaDIME import gamma_dime_loss import os import argparse K.clear_session() import scipy.io as sio import numpy as np import tensorflow as tf from numpy.random import seed seed(1) from tensorflow import set_random_seed set_random_seed(2) # custom cross-entropy to allow gradient separation and label smoothing regularization def categorical_crossentropy(y_true, y_pred, from_logits=False, label_smoothing=0.2): y_pred = K.constant(y_pred) if not K.is_tensor(y_pred) else y_pred y_true = K.cast(y_true, y_pred.dtype) if label_smoothing is not 0: smoothing = K.cast_to_floatx(label_smoothing) def _smooth_labels(): num_classes = K.cast(K.shape(y_true)[1], y_pred.dtype) return y_true * (1.0 - smoothing) + (smoothing / num_classes) y_true = K.switch(K.greater(smoothing, 0), _smooth_labels, lambda: y_true) return K.categorical_crossentropy(y_true, y_pred, from_logits=from_logits) # MINE loss def mine_loss(args): t_xy = args[0] t_xy_bar = args[1] loss = -(K.mean(t_xy) - K.logsumexp(t_xy_bar) + K.log(tf.cast(K.shape(t_xy)[0], tf.float32))) return loss # mixed loss def customLoss(MI): def dice(yTrue, yPred): beta = 0.2 # BETA PARAMETER IN THE PAPER, to choose at the beginning return categorical_crossentropy(yTrue, yPred) - beta*MI return dice # shuffling for MINE input def data_generation_mi(data_x, data_y): data_xy = np.hstack((data_x, data_y)) data_y_shuffle = np.take(data_y, np.random.permutation(data_y.shape[0]), axis=0, out=data_y) data_x_y = np.hstack((data_x, data_y_shuffle)) return data_xy, data_x_y # BLER computation def compute_BER(x,y): sh = x.shape p = sh[0] ber = 0 for i in range(p): if np.argmax(y[i,:]) != np.argmax(x[i,:]): ber = ber +1 return ber/p # get a "l2 norm of gradients" tensor def get_gradient_norm(model): with K.name_scope('gradient_norm'): grads = K.gradients(model.total_loss, model.trainable_weights) norm = K.sqrt(sum([K.sum(K.square(g)) for g in grads])) return norm # shuffling tensor def shuffleColumns(x): return tf.gather(x, tf.random.shuffle(tf.range(tf.shape(x)[0]))) # for parser def str2bool(v): if isinstance(v, bool): return v if v.lower() in ('yes', 'true', 't', 'y', '1'): return True elif v.lower() in ('no', 'false', 'f', 'n', '0'): return False else: raise argparse.ArgumentTypeError('Boolean value expected.') class CAAE(): def __init__(self, encoder_path, decoder_path, output_directory, MI_type='MINE'): block_size = 5 # k parameter num_enc_inputs = pow(2, block_size) # M parameter = 2^k num_hidden = 2 # 2*N parameter R = block_size / num_hidden # rate EbN0dB = 7 # training parameter N = pow(10, -0.1 * EbN0dB) / (2 * R) # training variance K.set_learning_phase(1) # set flag to allow noise layer during training self.encoder_path = encoder_path self.decoder_path = decoder_path self.output_directory = output_directory # if model is already available, load it if os.path.exists(self.encoder_path) and os.path.exists(self.decoder_path): self.encoder = load_model(self.encoder_path) self.decoder = load_model(self.decoder_path) print("Loaded models...") else: # input shape self.num_enc_inputs = num_enc_inputs self.num_hidden = num_hidden self.num_joint = 2*num_hidden # set optimizers optimizer = Adam(0.01, 0.5) optimizer_MI = Adam(0.01, 0.5) optimizer_MI_est = Adam(0.01, 0.5) # build the transmitter self.transmitter = self.build_transmitter() # build the receiver self.receiver = self.build_receiver() # build the mutual information block estimator self.discriminator = self.build_discriminator(MI_type) # the transmitter encodes the bits in s s_in = Input(shape=(self.num_enc_inputs,)) x = self.transmitter(s_in) # build the channel #x_n = Lambda(lambda x: np.sqrt(num_hidden)*K.l2_normalize(x,axis=1))(x) # if CONSTANT POWER x_n = BatchNormalization(axis=-1, center=False, scale=False)(x) # if AVERAGE POWER, use batch normalization ch = Lambda(lambda x: x)(x_n) # delta channel, you can set up different channels in a different function, e.g., Rayleigh y = GaussianNoise(np.sqrt(N))(ch) # gaussian layer #y = UniformNoise(minval=-np.sqrt(3*N), maxval=np.sqrt(3*N)(x_n) # uniform noise layer # receiver connection s_out = self.receiver(y) # build the keras model self.encoder = Model(s_in,x_n) # set up the tensors for the mutual information block T1 = Concatenate(name='network/concatenate_layer_1')([x_n, y]) y_bar_input = Lambda(lambda x:shuffleColumns(x))(y) # shuffle y input as y_bar T2 = Concatenate(name='network/concatenate_layer_2')([x_n, y_bar_input]) # estimation of joint and marginals t_xy = self.discriminator(T1) t_x_y = self.discriminator(T2) if MI_type == 'gammaDIME': loss = Lambda(gamma_dime_loss, name='gamma_dime_loss')([t_xy, t_x_y]) else: # use MINE is default loss = Lambda(mine_loss, name='mine_loss')([t_xy, t_x_y]) output_MI = Lambda(lambda x: -x)(loss) self.loss_model = Model(s_in, output_MI) self.loss_model.add_loss(loss) self.loss_model.compile(optimizer=optimizer_MI) # decoder model y_in = Input(shape=(self.num_hidden,)) s_dec = self.receiver(y_in) self.decoder = Model(y_in,s_dec) # combined model self.combined = Model(s_in, s_out) loss_model_2 = customLoss(output_MI) self.combined.compile(loss=loss_model_2, optimizer=optimizer) # model for the estimation of MI at different SNRs T1_est = Input(shape=(self.num_joint,)) T2_est = Input(shape=(self.num_joint,)) t_xy = self.discriminator(T1_est) t_x_y = self.discriminator(T2_est) if MI_type == 'gammaDIME': loss_est = Lambda(gamma_dime_loss, name='gamma_dime_loss')([t_xy, t_x_y]) else: loss_est = Lambda(mine_loss, name='mine_loss')([t_xy, t_x_y]) output_MI_est = Lambda(lambda x: -x)(loss_est) self.loss_model_est = Model([T1_est, T2_est], output_MI_est) self.loss_model_est.add_loss(loss_est) self.loss_model_est.compile(optimizer=optimizer_MI_est) def build_transmitter(self): model = Sequential() model.add(Dense(self.num_enc_inputs, activation="relu", input_dim=self.num_enc_inputs)) # model.add(Dense(100)) if high rate model.add(Dense(self.num_hidden)) model.summary() s_in = Input(shape=(self.num_enc_inputs,)) x = model(s_in) return Model(s_in, x) def build_receiver(self): model = Sequential() model.add(Dense(self.num_enc_inputs, activation="relu", input_dim=self.num_hidden)) model.add(Dense(self.num_enc_inputs, activation='softmax')) model.summary() y = Input(shape=(self.num_hidden,)) s_out = model(y) return Model(y, s_out) def build_discriminator(self, MI_type): if MI_type == 'gammaDIME': activation_type = 'softplus' else: activation_type = 'linear' model = Sequential() model.add(Dense(200, activation="relu", input_dim=self.num_joint)) model.add(GaussianNoise(0.3)) # It works only during the training model.add(Dense(200, activation="sigmoid")) model.add(Dense(1, activation = activation_type)) model.summary() T = Input(shape=(self.num_joint,)) MI = model(T) return Model(T,MI) def train(self, epochs, batch_size_AE, batch_size_MI, k , n): block_size = k # k parameter num_enc_inputs = pow(2, block_size) # M parameter = 2^k p = 1000 # realizations alphabet = np.eye(num_enc_inputs, dtype='float32') # one-hot encoded values s_in = np.transpose(np.tile(alphabet, p)) m = n # to avoid confusion with noise R = block_size / m # (half) rate EbN0dB = 7 N = pow(10, -0.1 * EbN0dB) / (2 * R) # noise power MI = np.zeros((1,epochs)) epochs_AE = [10,100,1000] # training epochs for the autoencoder block epochs_MI = [10,100,100] # training epochs for the MINE block for e in range(3): for ij in range(epochs): # training the MI estimator for epoch in range(epochs_MI[e]): idx = np.random.randint(0, s_in.shape[0], batch_size_MI) s_in_batch = s_in[idx] loss_MI = self.loss_model.train_on_batch(s_in_batch,[]) # Plot the progress. if needed # print("%d [MI loss: %f" % (epoch, loss_MI)) mutual_information = self.loss_model.predict(s_in) MI[0,ij] = np.mean(mutual_information) # traning auto encoder ON BATCH loss_AE = np.zeros((epochs_AE[e],)) for epoch in range(epochs_AE[e]): idx = np.random.randint(0, s_in.shape[0], batch_size_AE) s_in_batch = s_in[idx] loss_AE[epoch] = self.combined.train_on_batch(s_in_batch, s_in_batch) print("%d [CAAE_loss: %f, MI_loss: %f]" % (ij, np.mean(loss_AE), MI[0,ij])) # median for a stable output # if you want to visualize the constellation and the BLER during the training progress on Matlab EbN0_dB = range(-14, 29) ber = np.zeros((43,)) j = 0 p_test = 1000 s_in_t = np.transpose(np.tile(alphabet, p_test)) for EbN0 in EbN0_dB: N = pow(10, -0.1 * EbN0) / (2 * R) mean_t = np.zeros((m,)) cov_t = np.dot(N, np.eye(m)) features_r = self.encoder.predict(s_in_t) # get the code noise = np.random.multivariate_normal(mean_t, cov_t, p_test * num_enc_inputs) # add channel influence features_n = np.add(features_r, noise) # received code s_out_t = self.decoder.predict(features_n) # decoded message ber[j] = compute_BER(s_in_t,s_out_t) j = j + 1 sio.savemat('BER_%d.mat'%e,{'ber': ber}) features_r = self.encoder.predict(s_in_t) sio.savemat('code_%d.mat'%e, {'code': features_r}) # save the model save_path = self.output_directory + "/Models_AE" if not os.path.exists(save_path): os.makedirs(save_path) self.encoder.save(save_path + "/encoder.h5") self.decoder.save(save_path + "/decoder.h5") def train_MI(self, epochs_MI, batch_size_MI, k, n): block_size = k # k parameter num_enc_inputs = pow(2, block_size) # M parameter = 2^k m = n # N parameter R = block_size/m p = 10000 alphabet = np.eye(num_enc_inputs, dtype='float32') # One-hot encoded values s_in = np.transpose(np.tile(alphabet, p)) EbN0_dB = range(-14, 19) j = 0 MI = np.zeros((1,33)) for EbN0 in EbN0_dB: N = pow(10, -0.1 * EbN0)/ (2 * R) mean_t = np.zeros((m,)) cov_t = np.dot(N, np.eye(m)) features_r = self.encoder.predict(s_in) noise = np.random.multivariate_normal(mean_t,cov_t,p*num_enc_inputs) features_n = np.add(features_r,noise) # training the MI estimator data_xy, data_x_y = data_generation_mi(features_r, features_n) for epoch in range(epochs_MI): idx = np.random.randint(0, s_in.shape[0], batch_size_MI) data_xy_batch = data_xy[idx] data_x_y_batch = data_x_y[idx] self.loss_model_est.train_on_batch([data_xy_batch, data_x_y_batch],[]) mutual_information = self.loss_model_est.predict([data_xy, data_x_y]) MI[0,j] = np.mean(mutual_information) print(MI[0,j]) j = j+1 # save the estimated mutual information for Matlab sio.savemat('MI_estimation.mat', {'Eb':EbN0_dB, 'MI': MI}) def test(self, batch_size_AE, k, n): block_size = k # k parameter num_enc_inputs = pow(2, block_size) # M parameter = 2^k alphabet = np.eye(num_enc_inputs, dtype='float32') # One-hot encoded values m = n R = block_size / m p_test = batch_size_AE s_in_t = np.transpose(np.tile(alphabet, p_test)) EbN0_dB = range(-14, 19) ber = np.zeros((33,)) j = 0 for EbN0 in EbN0_dB: N = pow(10, -0.1 * EbN0) / (2 * R) mean_t = np.zeros((m,)) cov_t = np.dot(N, np.eye(m)) features_r = self.encoder.predict(s_in_t) noise = np.random.multivariate_normal(mean_t, cov_t, p_test * num_enc_inputs) features_n = np.add(features_r, noise) s_out_t = self.decoder.predict(features_n) ber[j] = compute_BER(s_in_t, s_out_t) j = j + 1 sio.savemat('BLER_test.mat', {'ber': ber}) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--load_encoder', help='Path to existing generator1 weights file', default="Models_AE/encoder.h5") parser.add_argument('--load_decoder', help='Path to existing discriminator1 weights file', default="Models_AE/decoder.h5") parser.add_argument('--output_directory', help="Directoy to save weights and images to.", default="Output") parser.add_argument('--train', type=str2bool, help="Start the training process.", default=False) parser.add_argument('--MI_type', help="Select the type of estimator to use.", default="MINE") args = parser.parse_args() # Load the model CAAE = CAAE(args.load_encoder, args.load_decoder, args.output_directory, args.MI_type) # Check if the folder with the encoder is empty and the flag training is on if not(os.path.exists(args.load_encoder)) and args.train: print('Training the model') CAAE.train(epochs=10, batch_size_AE = 1000, batch_size_MI = 1000, k = 5, n = 2) CAAE.train_MI(epochs_MI=100, batch_size_MI=1000, k = 5, n = 2) CAAE.test(batch_size_AE = 4000, k = 5, n = 2) # Just testing elif os.path.exists(args.load_encoder): print('Testing using the loaded models in your folder. If you want to train the model, delete the models') CAAE.test(batch_size_AE = 4000, k = 5, n = 2) else: print('Error in the models loading, please check the correct path')
15,913
37.626214
132
py
capacity-approaching-autoencoders
capacity-approaching-autoencoders-master/uniform_noise.py
from keras.engine import Layer from keras import backend as K class UniformNoise(Layer): """Apply additive uniform noise Only active at training time since it is a regularization layer. # Arguments minval: Minimum value of the uniform distribution maxval: Maximum value of the uniform distribution # Input shape Arbitrary. # Output shape Same as the input shape. """ def __init__(self, minval=-1.0, maxval=1.0, **kwargs): super(UniformNoise, self).__init__(**kwargs) self.supports_masking = True self.minval = minval self.maxval = maxval def call(self, inputs, training=None): def noised(): return inputs + K.random_uniform(shape=K.shape(inputs), minval=self.minval, maxval=self.maxval) return K.in_train_phase(noised, inputs, training=training) def get_config(self): config = {'minval': self.minval, 'maxval': self.maxval} base_config = super(UniformNoise, self).get_config() return dict(list(base_config.items()) + list(config.items()))
1,190
30.342105
69
py
vermouth-martinize
vermouth-martinize-master/doc/source/conf.py
# -*- coding: utf-8 -*- # # Configuration file for the Sphinx documentation builder. # # This file does only contain a selection of the most common options. For a # full list see the documentation: # http://www.sphinx-doc.org/en/stable/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. # from pkg_resources import get_distribution import os # import sys # sys.path.insert(0, os.path.abspath('.')) # Do not generate APIdocs for members missing docstrings (undoc-members) os.environ['APIDOC_OPTIONS'] = 'members,show-inheritence,inherited-members' # Set APIDOC options #os.environ['SPHINX_APIDOC_OPTIONS'] = 'members,undoc-members,show-inheritance,special-members' os.environ['SPHINX_APIDOC_OPTIONS'] = 'members' # -- Project information ----------------------------------------------------- project = 'VerMoUTH' copyright = '2018, University of Groningen' author = 'Peter C Kroon, Jonathan Barnoud, Tsjerk A Wassenaar, Siewert-Jan Marrink' # The full version, including alpha/beta/rc tags release = get_distribution('vermouth').version # The short X.Y version version = '.'.join(release.split('.')[:2]) # -- General configuration --------------------------------------------------- # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # 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.autodoc', 'sphinx.ext.intersphinx', 'sphinx.ext.mathjax', 'sphinx.ext.viewcode', 'sphinx.ext.napoleon', 'sphinx.ext.autosectionlabel', 'sphinxcontrib.apidoc', ] # Add any paths that contain templates here, relative to this directory. templates_path = ['_templates'] # The suffix(es) of source filenames. # You can specify multiple suffix as a list of string: # # source_suffix = ['.rst', '.md'] source_suffix = '.rst' # The master toctree document. master_doc = 'index' # 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 = 'en' # 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 = 'sphinx' nitpick_ignore = [ ('py:class', 'networkx.algorithms.isomorphism.isomorphvf2.GraphMatcher'), ('py:class', 'optional'), ] # -- 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 = 'alabaster' # 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 = {} # 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'] # Custom sidebar templates, must be a dictionary that maps document names # to template names. # # The default sidebars (for documents that don't match any pattern) are # defined by theme itself. Builtin themes are using these templates by # default: ``['localtoc.html', 'relations.html', 'sourcelink.html', # 'searchbox.html']``. # # html_sidebars = {} # -- Options for HTMLHelp output --------------------------------------------- # Output file base name for HTML help builder. htmlhelp_basename = 'VerMoUTHdoc' # -- Options for LaTeX output ------------------------------------------------ latex_elements = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # # 'pointsize': '10pt', # Additional stuff for the LaTeX preamble. # # 'preamble': '', # Latex figure (float) alignment # # 'figure_align': 'htbp', } # Grouping the document tree into LaTeX files. List of tuples # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ (master_doc, 'VerMoUTH.tex', 'VerMoUTH Documentation', author, 'manual'), ] # -- Options for manual page output ------------------------------------------ # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ (master_doc, 'vermouth', 'VerMoUTH Documentation', [author], 1) ] # -- Options for Texinfo output ---------------------------------------------- # Grouping the document tree into Texinfo files. List of tuples # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ (master_doc, 'VerMoUTH', 'VerMoUTH Documentation', author, 'VerMoUTH', 'One line description of project.', 'Miscellaneous'), ] # -- Extension configuration ------------------------------------------------- apidoc_module_dir = '../../vermouth' apidoc_output_dir = 'api' apidoc_separate_modules = True apidoc_excluded_paths = ['tests', 'redistributed'] autodoc_inherit_docstrings = False autoclass_content = 'both' autodoc_default_options = {'members': None, 'undoc-members': None, 'show-inheritance': None} napoleon_google_docstring = False napoleon_numpy_docstring = True napoleon_preprocess_types = False napoleon_type_aliases = { 'Molecule': 'vermouth.molecule.Molecule', } # -- Options for intersphinx extension --------------------------------------- # Example configuration for intersphinx: refer to the Python standard library. intersphinx_mapping = { 'python': ('https://docs.python.org', None), 'networkx': ('https://networkx.github.io/documentation/latest', None), 'numpy': ('http://docs.scipy.org/doc/numpy', None), 'scipy': ('http://docs.scipy.org/doc/scipy/reference', None), } # Borrowed from https://github.com/sphinx-doc/sphinx/issues/5603 # On top of that, networkx.isomorphism.GraphMatcher is not documented, so link # to the VF2 isomorphism module instead. # See https://github.com/networkx/networkx/issues/3239 intersphinx_aliases = { #('py:class', 'networkx.classes.graph.Graph'): ('py:class', 'networkx.Graph'), #('py:class', 'networkx.algorithms.isomorphism.vf2userfunc.GraphMatcher'): ('py:class', 'networkx.isomorphism.GraphMatcher'), #('py:class', 'networkx.algorithms.isomorphism.vf2userfunc.GraphMatcher'): ('py:module','networkx.algorithms.isomorphism.isomorphvf2'), ('py:class', 'networkx.isomorphism.GraphMatcher'): ('py:module', 'networkx.algorithms.isomorphism.isomorphvf2') } autosectionlabel_prefix_document = True def add_intersphinx_aliases_to_inv(app): from sphinx.ext.intersphinx import InventoryAdapter inventories = InventoryAdapter(app.builder.env) for alias, target in app.config.intersphinx_aliases.items(): alias_domain, alias_name = alias target_domain, target_name = target try: found = inventories.main_inventory[target_domain][target_name] try: inventories.main_inventory[alias_domain][alias_name] = found except KeyError: continue except KeyError: continue def setup(app): app.add_config_value('intersphinx_aliases', {}, 'env') app.connect('builder-inited', add_intersphinx_aliases_to_inv)
8,080
32.953782
139
py
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks-master/network_training/SWEC_ETHZ.py
import os import numpy as np import torch from torch import nn from torch.utils.data import DataLoader, Dataset from torchvision import datasets, transforms from torchsummary import summary import torch.nn.functional as F from sklearn.model_selection import KFold from sklearn import preprocessing import matplotlib.pyplot as plt import math from sklearn.decomposition import PCA import sklearn import brevitas.nn as qnn import pandas as pd from Network import ParallelConvolution from utils import foldretrieve # ----------------------------------------------------------- rootPath = '/scratch/jcu/cl/TBioCAS/processed_data/' device = 'cuda' if torch.cuda.is_available() else 'cpu' seed = 8 patients = ['01', '02', '03', '05', '06'] dataset = 'SWEC_ETHZ' dataType = 'features' numFold = 5 bS = 32 num_epoch = 100 network_size = 32 # ----------------------------------------------------------- print('Using {} device'.format(device),'\n') ### Data Loading (preictal 0, interictal 1) torch.manual_seed(seed) np.random.seed(seed) torch.backends.cudnn.benchmark = False torch.use_deterministic_algorithms(True) torch.backends.cudnn.deterministic = True os.environ['PYTHONHASHSEED'] = str(seed) df = pd.DataFrame(columns=['Patient', 'Fold', 'Epoch', 'Loss', 'Train Accuracy', 'Test Accuracy']) for patient in patients: allData = np.empty((0, 176)) rawLabel = np.empty((0,), dtype=int) newData = np.load(rootPath+dataType+'/'+dataset+'_'+'patient'+'_'+patient+'_'+'synthetic_preictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.zeros((newData.shape[0],),dtype=int),axis=0) newData = np.load(rootPath+dataType+'/'+dataset+'_'+'patient'+'_'+patient+'_'+'preictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.zeros((newData.shape[0],),dtype=int),axis=0) newData = np.load(rootPath+dataType+'/'+dataset+'_'+'patient'+'_'+patient+'_'+'interictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.ones((newData.shape[0],),dtype=int),axis=0) allLabel = np.zeros((rawLabel.size, rawLabel.max()+1)) allLabel[np.arange(rawLabel.size),rawLabel] = 1 pca = PCA(n_components=64) allData = pca.fit_transform(allData) inputbits=6 inputstep=(np.amax(allData) - np.amin(allData))/(2**inputbits-1) allData = np.round(allData/inputstep) allData *= inputstep randInd = np.arange(0,len(allData)) np.random.shuffle(randInd) allData = allData[randInd] allLabel = allLabel[randInd] allData = allData[:math.floor(allData.shape[0]/5)*5] allLabel = allLabel[:math.floor(allLabel.shape[0]/5)*5] foldsData = np.split(allData,numFold) foldsLabel = np.split(allLabel,numFold) loss_function = nn.BCEWithLogitsLoss() print('--------------------------------') print(f'Patient {patient}') for fold in range(0,numFold): trainData, testData, trainLabel, testLabel = foldretrieve(fold, foldsData, foldsLabel) print(f'Fold Number {fold}') network = ParallelConvolution(size=network_size).to(device) optimizer = torch.optim.Adam(network.parameters()) lossHist = [] testAccHist = [] trainAccHist = [] bestAcc = 0 bestSen = 0 bestSpe = 0 bestFPC = 1e10 bestAUR = 0 bestEpoch = [0] * 5 for epoch in range(num_epoch): trainCorrect = 0 trainTotal = 0 testCorrect = 0 testTotal = 0 trainOutputs = np.empty((0,1)) trainLabels = np.empty((0,1)) trainCM = np.zeros((2, 2)) testOutputs = np.empty((0,1)) testLabels = np.empty((0,1)) testCM = np.zeros((2, 2)) for i in range(0,len(trainData),bS): data = trainData[i:i+bS].to(device,dtype=torch.float) label = trainLabel[i:i+bS].to(device,dtype=torch.float) # Zero the gradients optimizer.zero_grad() network.eval() # Perform forward pass outputs = network(data) network.train() # Compute loss loss = loss_function(outputs, label) # Perform backward pass loss.backward() # Perform optimization optimizer.step() network.eval() outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) trainCM = np.add(trainCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) trainCorrect += np.sum(outputs==label) trainTotal += len(outputs) trainOutputs = np.append(trainOutputs,outputs.reshape((-1,1)),axis=0) trainLabels = np.append(trainLabels,label.reshape((-1,1)),axis=0) for i in range(0,len(testData),bS): data = testData[i:i+bS].to(device,dtype=torch.float) label = testLabel[i:i+bS].to(device,dtype=torch.float) network.eval() # Perform forward pass outputs = network(data) # Caclulate statistics testCorrect += np.sum(np.argmax(outputs.cpu().detach().numpy(),axis=1)==np.argmax(label.cpu().detach().numpy(),axis=1)) testTotal += len(outputs) outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) testCM = np.add(testCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) testCorrect += np.sum(outputs==label) testTotal += len(outputs) testOutputs = np.append(testOutputs,outputs.reshape((-1,1)),axis=0) testLabels = np.append(testLabels,label.reshape((-1,1)),axis=0) if testCorrect/testTotal > bestAcc: torch.save(network.state_dict(),'SWEC_ETHZ_P{}F{}.pt'.format(patient,fold)) trainF1 = sklearn.metrics.f1_score(trainLabels,trainOutputs) trainAccuracy = trainCorrect/trainTotal trainTN, trainFP, trainFN, trainTP = trainCM.ravel() trainSensitivity = trainTP / (trainTP+trainFN) trainSpecitivity = trainTN / (trainTN+trainFP) try: trainAUROC = sklearn.metrics.roc_auc_score(trainLabels, trainOutputs) except: trainAUROC = 0. testF1 = sklearn.metrics.f1_score(testLabels,testOutputs) testAccuracy = testCorrect/testTotal testTN, testFP, testFN, testTP = testCM.ravel() testSensitivity = testTP / (testTP+testFN) testSpecitivity = testTN / (testTN+testFP) try: testAUROC = sklearn.metrics.roc_auc_score(testLabels, testOutputs) except: testAUROC = 0. if testAccuracy > bestAcc: bestAcc = testAccuracy bestEpoch[0] = epoch if testSensitivity > bestSen: bestSen = testSensitivity bestEpoch[1] = epoch if testSpecitivity > bestSpe: bestSpe = testSpecitivity bestEpoch[2] = epoch if testFP < bestFPC: bestFPC = testFP bestEpoch[3] = epoch if testAUROC > bestAUR: bestAUR = testAUROC bestEpoch[4] = epoch print('Epoch:', epoch, 'Train Loss:', loss.item()) print('Train Accuracy:', trainAccuracy, 'Train Sensitivity:', trainSensitivity, 'Train Specitivity:', trainSpecitivity, 'Train FP Count:', trainFP, 'Train AUROC:', trainAUROC) print('Test Accuracy:', testAccuracy, 'Test Sensitivity:', testSensitivity, 'Test Specitivity:', testSpecitivity, 'Test FP Count:', testFP, 'Test AUROC:', testAUROC,) print('\n') lossHist.append(loss) trainAccHist.append(trainCorrect/trainTotal) testAccHist.append(testCorrect/testTotal) df = df.append({'Patient': patient, 'Fold': fold, 'Epoch': epoch, 'Loss': loss.cpu().item(), 'Train Accuracy': trainAccuracy * 100., 'Test Accuracy': testAccuracy * 100.}, ignore_index=True) df.to_csv('SWEC_ETHZ.csv', index=False)
8,593
41.756219
202
py
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks-master/network_training/utils.py
import torch import numpy as np def foldretrieve(fold,foldsData,foldsLabel): testData = foldsData[fold] testLabel = foldsLabel[fold] allData = foldsData[0:fold]+foldsData[fold:-1] allLabel = foldsLabel[0:fold]+foldsLabel[fold:-1] try: trainData = np.concatenate([*allData]) except: trainData = allData try: trainLabel = np.concatenate([*allLabel]) except: trainLabel = allLabel trainData = np.expand_dims(trainData,1) testData = np.expand_dims(testData,1) trainData = torch.tensor(trainData) testData = torch.tensor(testData) trainLabel = torch.tensor(trainLabel) testLabel = torch.tensor(testLabel) return trainData,testData,trainLabel,testLabel
743
26.555556
53
py
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks-master/network_training/Network.py
import torch from torch import nn import torch.nn.functional as F import brevitas.nn as qnn ### Network Definition class ParallelConvolution(nn.Module): def __init__(self, size=32): super(ParallelConvolution, self).__init__() self.conv1 = qnn.QuantConv1d(1,32,size,weight_bit_width=6) self.conv2 = qnn.QuantConv1d(1,32,62-size,weight_bit_width=6) self.avgpool = nn.AvgPool1d(2) self.flatten = nn.Flatten() self.dropout = nn.Dropout(0.5) self.linear = qnn.QuantLinear(1088,8,bias=True,weight_bit_width=4) self.classification = qnn.QuantLinear(8,2,bias=True,weight_bit_width=4) def forward(self,x): out1 = F.relu(self.conv1(x)) out2 = F.relu(self.conv2(x)) x = torch.cat((out1, out2),dim=2) x = self.avgpool(x) x = self.flatten(x) x = self.dropout(x) x = self.linear(x) out = self.classification(x) return out
954
33.107143
79
py
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks-master/network_training/CHBMIT.py
import os import numpy as np import torch from torch import nn from torch.utils.data import DataLoader, Dataset from torchvision import datasets, transforms from torchsummary import summary import torch.nn.functional as F from sklearn.model_selection import KFold from sklearn import preprocessing import matplotlib.pyplot as plt import math from sklearn.decomposition import PCA import sklearn import brevitas.nn as qnn import pandas as pd from Network import ParallelConvolution from utils import foldretrieve # ----------------------------------------------------------- rootPath = '/scratch/jcu/cl/TBioCAS/processed_data/' device = 'cuda' if torch.cuda.is_available() else 'cpu' seed = 8 patients = ['01', '02', '03', '05', '08'] dataset = 'CHBMIT' dataType = 'features' numFold = 5 bS = 32 num_epoch = 100 network_size = 32 # ----------------------------------------------------------- print('Using {} device'.format(device),'\n') ### Data Loading (preictal 0, interictal 1) torch.manual_seed(seed) np.random.seed(seed) torch.backends.cudnn.benchmark = False torch.use_deterministic_algorithms(True) torch.backends.cudnn.deterministic = True os.environ['PYTHONHASHSEED'] = str(seed) df = pd.DataFrame(columns=['Patient', 'Fold', 'Epoch', 'Loss', 'Train Accuracy', 'Test Accuracy']) for patient in patients: allData = np.empty((0, 176)) rawLabel = np.empty((0,), dtype=int) newData = np.load(rootPath+dataType+'/'+dataset+'_'+'patient'+'_'+patient+'_'+'synthetic_preictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.zeros((newData.shape[0],),dtype=int),axis=0) newData = np.load(rootPath+dataType+'/'+dataset+'_'+'patient'+'_'+patient+'_'+'preictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.zeros((newData.shape[0],),dtype=int),axis=0) newData = np.load(rootPath+dataType+'/'+dataset+'_'+'patient'+'_'+patient+'_'+'interictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.ones((newData.shape[0],),dtype=int),axis=0) allLabel = np.zeros((rawLabel.size, rawLabel.max()+1)) allLabel[np.arange(rawLabel.size),rawLabel] = 1 pca = PCA(n_components=64) allData = pca.fit_transform(allData) inputbits=6 inputstep=(np.amax(allData) - np.amin(allData))/(2**inputbits-1) allData = np.round(allData/inputstep) allData *= inputstep randInd = np.arange(0,len(allData)) np.random.shuffle(randInd) allData = allData[randInd] allLabel = allLabel[randInd] allData = allData[:math.floor(allData.shape[0]/5)*5] allLabel = allLabel[:math.floor(allLabel.shape[0]/5)*5] foldsData = np.split(allData,numFold) foldsLabel = np.split(allLabel,numFold) loss_function = nn.BCEWithLogitsLoss() print('--------------------------------') print(f'Patient {patient}') for fold in range(0,numFold): trainData, testData, trainLabel, testLabel = foldretrieve(fold, foldsData, foldsLabel) print(f'Fold Number {fold}') network = ParallelConvolution(size=network_size).to(device, dtype=torch.float) optimizer = torch.optim.Adam(network.parameters()) lossHist = [] testAccHist = [] trainAccHist = [] bestAcc = 0 bestSen = 0 bestSpe = 0 bestFPC = 1e10 bestAUR = 0 bestEpoch = [0, 0, 0, 0, 0] for epoch in range(num_epoch): trainCorrect = 0 trainTotal = 0 testCorrect = 0 testTotal = 0 trainOutputs = np.empty((0,1)) trainLabels = np.empty((0,1)) trainCM = np.zeros((2, 2)) testOutputs = np.empty((0,1)) testLabels = np.empty((0,1)) testCM = np.zeros((2, 2)) for i in range(0,len(trainData),bS): data = trainData[i:i+bS].to(device,dtype=torch.float) label = trainLabel[i:i+bS].to(device,dtype=torch.float) # Zero the gradients optimizer.zero_grad() network.eval() # Perform forward pass outputs = network(data) network.train() # Compute loss loss = loss_function(outputs, label) # Perform backward pass loss.backward() # Perform optimization optimizer.step() network.eval() outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) trainCM = np.add(trainCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) trainCorrect += np.sum(outputs==label) trainTotal += len(outputs) trainOutputs = np.append(trainOutputs,outputs.reshape((-1,1)),axis=0) trainLabels = np.append(trainLabels,label.reshape((-1,1)),axis=0) for i in range(0,len(testData),bS): data = testData[i:i+bS].to(device,dtype=torch.float) label = testLabel[i:i+bS].to(device,dtype=torch.float) network.eval() # Perform forward pass outputs = network(data) # Caclulate statistics testCorrect += np.sum(np.argmax(outputs.cpu().detach().numpy(),axis=1)==np.argmax(label.cpu().detach().numpy(),axis=1)) testTotal += len(outputs) outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) testCM = np.add(testCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) testCorrect += np.sum(outputs==label) testTotal += len(outputs) testOutputs = np.append(testOutputs,outputs.reshape((-1,1)),axis=0) testLabels = np.append(testLabels,label.reshape((-1,1)),axis=0) if testCorrect/testTotal > bestAcc: torch.save(network.state_dict(),'CHBMIT_P{}F{}.pt'.format(patient,fold)) trainF1 = sklearn.metrics.f1_score(trainLabels,trainOutputs) trainAccuracy = trainCorrect/trainTotal trainTN, trainFP, trainFN, trainTP = trainCM.ravel() trainSensitivity = trainTP / (trainTP+trainFN) trainSpecitivity = trainTN / (trainTN+trainFP) try: trainAUROC = sklearn.metrics.roc_auc_score(trainLabels, trainOutputs) except: trainAUROC = 0. testF1 = sklearn.metrics.f1_score(testLabels,testOutputs) testAccuracy = testCorrect/testTotal testTN, testFP, testFN, testTP = testCM.ravel() testSensitivity = testTP / (testTP+testFN) testSpecitivity = testTN / (testTN+testFP) try: testAUROC = sklearn.metrics.roc_auc_score(testLabels, testOutputs) except: testAUROC = 0. if testAccuracy > bestAcc: bestAcc = testAccuracy bestEpoch[0] = epoch if testSensitivity > bestSen: bestSen = testSensitivity bestEpoch[1] = epoch if testSpecitivity > bestSpe: bestSpe = testSpecitivity bestEpoch[2] = epoch if testFP < bestFPC: bestFPC = testFP bestEpoch[3] = epoch if testAUROC > bestAUR: bestAUR = testAUROC bestEpoch[4] = epoch print('Epoch:', epoch, 'Train Loss:', loss.item()) print('Train Accuracy:', trainAccuracy, 'Train Sensitivity:', trainSensitivity, 'Train Specitivity:', trainSpecitivity, 'Train FP Count:', trainFP, 'Train AUROC:', trainAUROC) print('Test Accuracy:', testAccuracy, 'Test Sensitivity:', testSensitivity, 'Test Specitivity:', testSpecitivity, 'Test FP Count:', testFP, 'Test AUROC:', testAUROC,) print('\n') lossHist.append(loss) trainAccHist.append(trainCorrect/trainTotal) testAccHist.append(testCorrect/testTotal) df = df.append({'Patient': patient, 'Fold': fold, 'Epoch': epoch, 'Loss': loss.cpu().item(), 'Train Accuracy': trainAccuracy, 'Test Accuracy': testAccuracy}, ignore_index=True) df.to_csv('CHBMIT.csv', index=False)
8,597
41.776119
188
py
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks
Memristive-Seizure-Detection-and-Prediction-by-Parallel-Convolutional-Neural-Networks-master/network_training/Transfer_CHBMIT_SWEC_ETHZ.py
import os import numpy as np import torch from torch import nn from torch.utils.data import DataLoader, Dataset from torchvision import datasets, transforms from torchsummary import summary import torch.nn.functional as F from sklearn.model_selection import KFold from sklearn import preprocessing import matplotlib.pyplot as plt import math from sklearn.decomposition import PCA import sklearn import brevitas.nn as qnn import pandas as pd from itertools import permutations from Network import ParallelConvolution from utils import foldretrieve # ----------------------------------------------------------- rootPath = '/scratch/jcu/cl/TBioCAS/processed_data/' device = 'cuda' if torch.cuda.is_available() else 'cpu' seed = 8 patients = {'CHBMIT': ['01', '02', '03', '05', '08'], 'SWEC_ETHZ': ['01', '02', '03', '05', '06']} datasets = ['CHBMIT', 'SWEC_ETHZ'] dataType = 'features' numFold = 5 bS = 32 num_epoch_retrain = 10 network_size = 32 # ----------------------------------------------------------- print('Using {} device'.format(device),'\n') ### Data Loading (preictal 0, interictal 1) torch.manual_seed(seed) np.random.seed(seed) torch.backends.cudnn.benchmark = False torch.use_deterministic_algorithms(True) torch.backends.cudnn.deterministic = True os.environ['PYTHONHASHSEED'] = str(seed) dataset_permutations = permutations(datasets) for dataset_permutation in dataset_permutations: print('--------------------------------') print(dataset_permutation) pretrained_dataset = dataset_permutation[0] evaluation_dataset = dataset_permutation[1] df = pd.DataFrame(columns=['Pretrained Patient', 'Evaluation Patient', 'Fold', 'Epoch', 'Loss', 'Train Accuracy', 'Test Accuracy']) for patient_idx, patient in enumerate(patients[evaluation_dataset]): allData = np.empty((0, 176)) rawLabel = np.empty((0,), dtype=int) newData = np.load(rootPath+dataType+'/'+evaluation_dataset+'_'+'patient'+'_'+patient+'_'+'synthetic_preictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.zeros((newData.shape[0],),dtype=int),axis=0) newData = np.load(rootPath+dataType+'/'+evaluation_dataset+'_'+'patient'+'_'+patient+'_'+'preictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.zeros((newData.shape[0],),dtype=int),axis=0) newData = np.load(rootPath+dataType+'/'+evaluation_dataset+'_'+'patient'+'_'+patient+'_'+'interictal.npy') allData = np.append(allData,newData,axis=0) rawLabel = np.append(rawLabel,np.ones((newData.shape[0],),dtype=int),axis=0) allLabel = np.zeros((rawLabel.size, rawLabel.max()+1)) allLabel[np.arange(rawLabel.size),rawLabel] = 1 pca = PCA(n_components=64) allData = pca.fit_transform(allData) inputbits=6 inputstep=(np.amax(allData) - np.amin(allData))/(2**inputbits-1) allData = np.round(allData/inputstep) allData *= inputstep randInd = np.arange(0,len(allData)) np.random.shuffle(randInd) allData = allData[randInd] allLabel = allLabel[randInd] allData = allData[:math.floor(allData.shape[0]/5)*5] allLabel = allLabel[:math.floor(allLabel.shape[0]/5)*5] foldsData = np.split(allData,numFold) foldsLabel = np.split(allLabel,numFold) loss_function = nn.BCEWithLogitsLoss() print('--------------------------------') print(f'Patient {patient}') for fold in range(0,numFold): trainData, testData, trainLabel, testLabel = foldretrieve(fold, foldsData, foldsLabel) print(f'Fold Number {fold}') # Load the pretrained model network = ParallelConvolution(size=network_size).to(device) network.load_state_dict(torch.load('{}_P{}F{}.pt'.format(pretrained_dataset,patients[pretrained_dataset][patient_idx],fold))) # Freeze the weights of all layers for name, param in network.named_parameters(): param.requires_grad = False # Unfreeze the weights of the last two (linear) layers network.linear.weight.requires_grad = True network.linear.bias.requires_grad = True network.classification.weight.requires_grad = True network.classification.bias.requires_grad = True optimizer = torch.optim.Adam(network.parameters()) lossHist = [] testAccHist = [] trainAccHist = [] bestAcc = 0 bestSen = 0 bestSpe = 0 bestFPC = 1e10 bestAUR = 0 bestEpoch = [0] * 5 for epoch in range(num_epoch_retrain): if epoch == 0: with torch.no_grad(): testCorrect = 0 testTotal = 0 testOutputs = np.empty((0,1)) testLabels = np.empty((0,1)) testCM = np.zeros((2, 2)) for i in range(0,len(testData),bS): data = testData[i:i+bS].to(device,dtype=torch.float) label = testLabel[i:i+bS].to(device,dtype=torch.float) network.eval() # Perform forward pass outputs = network(data) # Caclulate statistics testCorrect += np.sum(np.argmax(outputs.cpu().detach().numpy(),axis=1)==np.argmax(label.cpu().detach().numpy(),axis=1)) testTotal += len(outputs) outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) testCM = np.add(testCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) testCorrect += np.sum(outputs==label) testTotal += len(outputs) testOutputs = np.append(testOutputs,outputs.reshape((-1,1)),axis=0) testLabels = np.append(testLabels,label.reshape((-1,1)),axis=0) testF1 = sklearn.metrics.f1_score(testLabels,testOutputs) testAccuracy = testCorrect/testTotal testTN, testFP, testFN, testTP = testCM.ravel() testSensitivity = testTP / (testTP+testFN) testSpecitivity = testTN / (testTN+testFP) try: testAUROC = sklearn.metrics.roc_auc_score(testLabels, testOutputs) except: testAUROC = 0. df = df.append({'Pretrained Patient': patients[pretrained_dataset][patient_idx], 'Evaluation Patient': patient, 'Fold': fold, 'Epoch': -1, 'Loss': -1, 'Train Accuracy': -1, 'Test Accuracy': testAccuracy * 100.}, ignore_index=True) trainCorrect = 0 trainTotal = 0 testCorrect = 0 testTotal = 0 trainOutputs = np.empty((0,1)) trainLabels = np.empty((0,1)) trainCM = np.zeros((2, 2)) testOutputs = np.empty((0,1)) testLabels = np.empty((0,1)) testCM = np.zeros((2, 2)) for i in range(0,len(trainData),bS): data = trainData[i:i+bS].to(device,dtype=torch.float) label = trainLabel[i:i+bS].to(device,dtype=torch.float) # Zero the gradients optimizer.zero_grad() network.eval() # Perform forward pass outputs = network(data) network.train() # Compute loss loss = loss_function(outputs, label) # Perform backward pass loss.backward() # Perform optimization optimizer.step() network.eval() outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) trainCM = np.add(trainCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) trainCorrect += np.sum(outputs==label) trainTotal += len(outputs) trainOutputs = np.append(trainOutputs,outputs.reshape((-1,1)),axis=0) trainLabels = np.append(trainLabels,label.reshape((-1,1)),axis=0) for i in range(0,len(testData),bS): data = testData[i:i+bS].to(device,dtype=torch.float) label = testLabel[i:i+bS].to(device,dtype=torch.float) network.eval() # Perform forward pass outputs = network(data) # Caclulate statistics testCorrect += np.sum(np.argmax(outputs.cpu().detach().numpy(),axis=1)==np.argmax(label.cpu().detach().numpy(),axis=1)) testTotal += len(outputs) outputs = outputs.cpu().detach().numpy() label = label.cpu().detach().numpy() outputs = np.argmax(outputs,axis=1) label = np.argmax(label,axis=1) testCM = np.add(testCM, sklearn.metrics.confusion_matrix(outputs, label, labels=[0, 1])) testCorrect += np.sum(outputs==label) testTotal += len(outputs) testOutputs = np.append(testOutputs,outputs.reshape((-1,1)),axis=0) testLabels = np.append(testLabels,label.reshape((-1,1)),axis=0) if testCorrect/testTotal > bestAcc: torch.save(network.state_dict(),'{}_{}_P{}F{}.pt'.format(pretrained_dataset, evaluation_dataset, patient, fold)) trainF1 = sklearn.metrics.f1_score(trainLabels,trainOutputs) trainAccuracy = trainCorrect/trainTotal trainTN, trainFP, trainFN, trainTP = trainCM.ravel() trainSensitivity = trainTP / (trainTP+trainFN) trainSpecitivity = trainTN / (trainTN+trainFP) try: trainAUROC = sklearn.metrics.roc_auc_score(trainLabels, trainOutputs) except: trainAUROC = 0. testF1 = sklearn.metrics.f1_score(testLabels,testOutputs) testAccuracy = testCorrect/testTotal testTN, testFP, testFN, testTP = testCM.ravel() testSensitivity = testTP / (testTP+testFN) testSpecitivity = testTN / (testTN+testFP) try: testAUROC = sklearn.metrics.roc_auc_score(testLabels, testOutputs) except: testAUROC = 0. if testAccuracy > bestAcc: bestAcc = testAccuracy bestEpoch[0] = epoch if testSensitivity > bestSen: bestSen = testSensitivity bestEpoch[1] = epoch if testSpecitivity > bestSpe: bestSpe = testSpecitivity bestEpoch[2] = epoch if testFP < bestFPC: bestFPC = testFP bestEpoch[3] = epoch if testAUROC > bestAUR: bestAUR = testAUROC bestEpoch[4] = epoch print('Epoch:', epoch, 'Train Loss:', loss.item()) print('Train Accuracy:', trainAccuracy, 'Train Sensitivity:', trainSensitivity, 'Train Specitivity:', trainSpecitivity, 'Train FP Count:', trainFP, 'Train AUROC:', trainAUROC) print('Test Accuracy:', testAccuracy, 'Test Sensitivity:', testSensitivity, 'Test Specitivity:', testSpecitivity, 'Test FP Count:', testFP, 'Test AUROC:', testAUROC,) print('\n') lossHist.append(loss) trainAccHist.append(trainCorrect/trainTotal) testAccHist.append(testCorrect/testTotal) df = df.append({'Pretrained Patient': patients[pretrained_dataset][patient_idx], 'Evaluation Patient': patient, 'Fold': fold, 'Epoch': epoch, 'Loss': loss.cpu().item(), 'Train Accuracy': trainAccuracy * 100., 'Test Accuracy': testAccuracy * 100.}, ignore_index=True) df.to_csv('Transfer_%s_%s.csv' % (pretrained_dataset, evaluation_dataset), index=False)
12,878
48.918605
282
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Apply_ML_Models_v1_3.py
import warnings # warnings.filterwarnings("ignore") from sklearn.exceptions import ConvergenceWarning warnings.simplefilter("ignore", category=ConvergenceWarning) import os warnings.simplefilter("ignore", category=FutureWarning) import sys import numpy as np from matplotlib import pyplot as plt import seaborn as sns import tensorflow as tf # import gpflow import pandas as pd # import torch # import torch.nn as nn import sklearn from sklearn.preprocessing import StandardScaler, FunctionTransformer from sklearn.model_selection import RepeatedKFold, ShuffleSplit, LeaveOneGroupOut, GroupKFold, GroupShuffleSplit import sklearn.linear_model from sklearn.ensemble import GradientBoostingRegressor from sklearn.ensemble import RandomForestRegressor from sklearn.neural_network import MLPRegressor from sklearn.gaussian_process import GaussianProcessRegressor from sklearn.gaussian_process.kernels import RBF, ExpSineSquared, WhiteKernel, ConstantKernel from sklearn.neighbors import KNeighborsRegressor from xgboost import XGBRegressor from sklearn.decomposition import PCA import random import yaml import datetime import argparse import superconductors_3D.machine_learning.Custom_Machine_Learning_v1_3 as ML # from superconductors_3D.machine_learning.RGM_own import RGM_sklearn import csv from collections import namedtuple import json from copy import deepcopy import time # import mlflow from superconductors_3D.machine_learning.own_libraries.analysis.Experiments.Run import MLRun, get_hparams # from superconductors_3D.machine_learning.own_libraries.models.GNN.MEGNet_tf import MEGNet_tf, read_json_file # from contextlib import redirect_stdout # from superconductors_3D.machine_learning.own_libraries.utils import Refactoring # from superconductors_3D.machine_learning.own_libraries.models.GPflow_GP import GPflow_GP # from superconductors_3D.machine_learning.own_libraries.own_functions import movecol from superconductors_3D.utils.projectpaths import projectpath # from superconductors_3D.machine_learning.own_libraries.models.NN import MLP_Lightning from superconductors_3D.machine_learning.own_libraries.utils.Scalers import restricted_arcsinh, restricted_sinh, restricted_exp, restricted_log def none_or_str(value): if value == 'None': return None return value def parse_arguments(args_from_fn, use_models, experiment, output_note, outdirname, calcdir, save_models, CV, n_folds, n_repeats, CV_keep_groups, n_reps, train_frac, domain_colname, is_optimizing_NN, dataset, hparams_file, n_jobs, sample_weights, metrics_sample_weights, use_data_frac, add_params): """Parse arguments from command line and if not given from command line insert argument from function.""" parser = argparse.ArgumentParser(description='Apply different ML models to some data and save output.') parser.add_argument('--dataset', '-d', type=str, help='The input dataset.') parser.add_argument('--calcdir', '-c', type=str, nargs='?', help='Change directory to this path.') parser.add_argument('--outdir', '-u', dest='outdirname', type=str, nargs='?', help='Directory of the output.') # parser.add_argument('--trainfracs', '-t', dest='all_trainfracs', type=float, nargs='*', help='List of train fractions.') parser.add_argument('--experiment', '-e', type=str, help='Identifier for a specific experiment.') parser.add_argument('--note', dest='output_note', type=str, help='Output text to add to main output file.') parser.add_argument('--n_reps', dest='n_reps', type=str, help='Number of repetitions of train test split if CV is Random.') parser.add_argument('--train-frac', dest='train_frac', type=float, help='Train fraction when random CV splitting is used.') parser.add_argument('--use-models', '-a', dest='use_models', type=str, nargs='*', help='Names of models that shall be trained.') parser.add_argument('--cv', dest='CV', type=str, help='Cross Validation mode.') parser.add_argument('--domain-colname', dest='domain_colname', type=none_or_str, help='Name of column in dataset that indicates groups.') parser.add_argument('--n_folds', dest='n_folds', type=str, help='Number of folds if CV is KFold.') parser.add_argument('--n_repeats', dest='n_repeats', type=str, help='Number of folds if CV is KFold.') parser.add_argument('--save-models', dest='save_models', type=bool, help='If models should be saved.') parser.add_argument('--optimizing', '-o', dest='is_optimizing_NN', help='If optimizing Neural Network.') parser.add_argument('--hparams-file', dest='hparams_file', help='Path to hyperparameter file.') parser.add_argument('--n-jobs', dest='n_jobs', help='Number of jobs to run in parallel when possible.') parser.add_argument('--sample-weights', dest='sample_weights', type=none_or_str, help='String indicating the weights column for each data point for training the models.') parser.add_argument('--metrics-sample-weights', dest='metrics_sample_weights', type=none_or_str, help='String indicating the weights column for each data point only for the metrics.') parser.add_argument('--CV-keep-groups', dest='CV_keep_groups', help='Group column name for a KFold with data points of each group in either test or train.') parser.add_argument('--use-data-frac', dest='use_data_frac', help='For debugging, use only this fraction of data.') parser.add_argument('--add-params', dest='add_params', type=json.loads, help='Add these parameters to the experiment parameters for easier recognition of results.') args = argparse.Namespace() # Add manually defined arguments to args. args.dataset = dataset args.experiment = experiment args.output_note = output_note args.outdirname = outdirname args.calcdir = calcdir args.n_reps = n_reps args.train_frac = train_frac args.is_optimizing_NN = is_optimizing_NN args.dataset = dataset args.use_models = use_models args.CV = CV args.domain_colname = domain_colname args.n_folds = n_folds args.n_repeats = n_repeats args.save_models = save_models args.hparams_file = hparams_file args.n_jobs = n_jobs args.sample_weights = sample_weights args.metrics_sample_weights = metrics_sample_weights args.CV_keep_groups = CV_keep_groups args.use_data_frac = use_data_frac args.add_params = add_params # args given as cmd line arguments have the second highest priority. if __name__ == '__main__': cmd_args = parser.parse_args() if len(sys.argv) > 1: print('Parsed arguments:', cmd_args) args_dict = vars(args) cmd_args_dict = {key: val for key, val in vars(cmd_args).items() if not val is None} args_dict.update(cmd_args_dict) args = argparse.Namespace(**args_dict) # args given as function arguments of main() have the highest priority. if args_from_fn: print(f'Updating args with arguments from function call: {args_from_fn}') args_dict = vars(args) args_dict.update(args_from_fn) args = argparse.Namespace(**args_dict) print(f'New arguments: {args}') return(args) def print_title(string): print(""" ============================ {} ============================""".format(string)) return() def is_round(num, prec=6): """Checks if a given number is a round value with given precision.""" is_round = round(num) == round(num, prec) return(is_round) def make_output_directory(outdirname, label): """Get numbered output directory `outdir` in parent directory `outdirname`.""" os.makedirs(outdirname, exist_ok=True) # Get number that's not already used for a directory. num = 0 dir_list = os.listdir(outdirname) while 'results_{}'.format(num) in '\t'.join(dir_list): num += 1 outdir = os.path.join(outdirname, f'results_{num}_{label}') if not os.path.exists(outdir): os.mkdir(outdir) else: raise Warning('You should not end up here.') return(outdir) def get_train_test_data(df_data, CV, n_folds, domain_colname, trainfrac=None, random_n_reps=1, n_repeats=1, group=None): """Gets train and test data doing the specified cross validation. """ data_array = df_data.to_numpy() if CV == 'KFold': # Convert train fraction and number of total repetititions to KFold input parameters. if group is None: split = RepeatedKFold(n_splits=n_folds, n_repeats=n_repeats).split(data_array) else: assert n_repeats == 1, 'This is not yet implemented.' groups = df_data[group].to_numpy() split = GroupKFold(n_splits=n_folds).split(data_array, groups=groups) elif CV == 'Random': assert trainfrac not in (0,1), "ShuffleSplit won't understand that this is a fraction." assert trainfrac != None if group is None: split = ShuffleSplit(train_size=trainfrac, n_splits=random_n_reps).split(data_array) else: groups = df_data[group].to_numpy() split = GroupShuffleSplit(train_size=trainfrac, n_splits=random_n_reps).split(data_array, groups=groups) elif CV == 'LOGO': # Work around: LeaveOneGroupOut.split() doesn't seem to work without specified y, probably a bug. y = np.zeros(len(df_data)) domains = df_data[domain_colname] split = LeaveOneGroupOut().split(data_array, y, domains) # Concatenate train and test indices of each repetition. # train_indices, test_indices = [], [] # for train_index, test_index in split: # train_indices.append(list(train_index)) # test_indices.append(list(test_index)) for i, (train_indices, test_indices) in enumerate(split): n_samples = len(df_data) assert n_samples == len(train_indices) + len(test_indices) empty = '' test_or_train = pd.Series(np.full(n_samples, empty)) test_or_train[train_indices] = 'train' test_or_train[test_indices] = 'test' # So that I can be sure that in case the indices of df and series don't align it still just adds everything in the right order. test_or_train = list(test_or_train) colname = ML.All_Data.All_Data.name_CV_col(i) df_data[colname] = test_or_train assert all([df_data[colname].iloc[idx] == 'train' for idx in train_indices]) assert all([df_data[colname].iloc[idx] == 'test' for idx in test_indices]) assert not (df_data[colname] == empty).any(), 'Some of the columns are neither test nor train.' return(df_data) # class Sin(nn.Module): # """Sin activation function.""" # def __init__(self): # super().__init__() # # def forward(self, input): # result = torch.Sin(input) # return(result) # class Multiple_Acts(nn.Module): # """Layer with multiple different activation functions. # """ # def __init__(self, acts): # """`acts` must have as many activation functions (torch functions) as the length of the input will be. # """ # super().__init__() # self.acts = acts # def forward(self, input): # n_input = len(input) # assert n_input == len(self.acts) # result = torch.empty(n_input) # for i in range(n_input): # result[i] = self.acts[i](input[i]) # result = torch.Sin(input) # return(result) # # def get_activation_fn(activation: str): # """Returns torch activation function based on string activation. # """ # if activation == 'relu': # activation_fn = nn.ReLU() # elif activation == 'logistic': # activation_fn = nn.Sigmoid() # elif activation == 'tanh': # activation_fn = nn.Tanh() # elif activation == 'sin': # activation_fn = Sin() # else: # raise ValueError(f'Activation function {activation} not recognized. Activation functions are lowercase always.') # return(activation_fn) # # # def get_sequential_NN(input_layer_size: int, hidden_layer_sizes: list, activation: str, output_layers: str): # """Returns a sequential (Feed Forward) NN. `last_linear` means if the last layer should be linear or with activation function.""" # activation_fn = get_activation_fn(activation) # out_act_fn = output_layers # layers = [] # num_layers = len(hidden_layer_sizes) # for i in range(num_layers): # if i == 0: # in_size = input_layer_size # else: # in_size = hidden_layer_sizes[i-1] # out_size = hidden_layer_sizes[i] # layers.append(nn.Linear(in_size, out_size)) # last_layer = i == num_layers - 1 # if not last_layer: # layers.append(activation_fn) # elif out_act_fn != None: # out_activation_fn = get_activation_fn(out_act_fn) # layers.append(out_activation_fn) # # layers = tuple(layers) # network = nn.Sequential(*layers) # return(network) # # # def get_featurizer(input_layer_size, hparams, mode): # """Returns the first part of the RGM, the featurizer or representation NN. # """ # hidden_layer_sizes = ML.net_pattern( # hparams['nn_layers'], # hparams['nn_base_dim'], # hparams['nn_end_dim'] # ) # activation = hparams['nn_act'] # # last_linear = False # if last layer linear or with activation fn # output_layers = hparams['nn_act'] # if mode == 'FeedForward': # featurizer = get_sequential_NN(input_layer_size, hidden_layer_sizes, activation, output_layers) # else: # raise ValueError('mode of Featurizer not recognized.') # return(featurizer) # # # def get_classifier(output_layer_size, hparams, output_layers=None): # """Returns a torch sequential NN with specific layers and output_layer_size. # """ # if hparams['RGM_classifier_layers'] < 1: # raise ValueError('Invalid "RGM_classifier_layers": {hparams["RGM_classifier_layers"]}') # # num_hidden_layers = hparams['RGM_classifier_layers'] - 1 # activation = hparams['nn_act'] # classifier_layers = [hparams['nn_end_dim'] for _ in range(num_hidden_layers)] # classifier_layers.append(output_layer_size) # # classifier = get_sequential_NN(hparams['nn_end_dim'], classifier_layers, activation, output_layers) # return(classifier) def get_validation_columns(df_data, args, domain_colname): """Adds validation columns to a df based on the current CV columns, so that only the train rows are split again in test and train. The old CV columns will be renamed to 'test_`CV_col`' and the new validation columns will be named from 0 to nfolds*nfolds-1. """ all_CV_cols = [col for col in df_data.columns if col.startswith('CV_')] # Keep old CV columns around for sanity checks. df_data = df_data.rename(columns={cv: 'test_'+cv for cv in all_CV_cols}) counter = 0 for cv in all_CV_cols: old_cv = 'test_' + cv df = df_data.loc[df_data[old_cv] == 'train'] # Use same type of CV for validation columns as for test columns. df = get_train_test_data( df_data=df, CV=args.CV, n_folds=args.n_folds, n_repeats=args.n_repeats, domain_colname=domain_colname, trainfrac=args.train_frac, random_n_reps=args.n_reps, group=args.CV_keep_groups ) cv_df = df[all_CV_cols] # Name validation columns from 0 to nfold*nfold-1. n_cvs = len(all_CV_cols) rename_cols = {f'CV_{i}': f'CV_{n_cvs*counter+i}' for i in range(n_cvs)} cv_df = cv_df.rename(columns=rename_cols) counter +=1 df_data = df_data.join(cv_df) # Sanity check. new_cv_cols = list(rename_cols.values()) assert df_data.loc[df_data[old_cv] == 'test', new_cv_cols].isna().all().all(), 'Some of the old test columns do not have NaN columns in the validation columns!' return df_data def get_all_models(hparams, n_features, n_targets, use_models, n_domains=1, domain_col=None, output_layers=None, outdir=None, scaler=None, args=None): """Definitions of models that I regularly use. """ all_models = {} #################### # 1 NEAREST NEIGHBOR #################### if '1NN' in use_models: Nearest_Neighbors = KNeighborsRegressor(n_neighbors=1) all_models['1NN'] = Nearest_Neighbors ################### # LINEAR REGRESSION ################### if 'LR' in use_models: Linear_Regression = sklearn.linear_model.LinearRegression() all_models['LR'] = Linear_Regression ################ # Neural Network ################ # Set some hyperparameter variables for the NN. net_dims = ML.net_pattern( hparams['nn_layers'], hparams['nn_base_dim'], hparams['nn_end_dim'] ) net_dims2 = ML.net_pattern( hparams['RGM_classifier_layers'], hparams['nn_end_dim'], hparams['nn_end_dim'] ) net_dims = net_dims + net_dims2 if 'NNsk' in use_models: NNsk = MLPRegressor( hidden_layer_sizes=net_dims, activation=hparams['nn_act'], solver='adam', max_iter=hparams["n_epochs"], early_stopping=True, validation_fraction=0.2, alpha=hparams["nn_l2"], batch_size=hparams['nn_batch_size'], learning_rate_init=hparams["learning_rate"], n_iter_no_change=hparams["nn_patience"] ) all_models['NNsk'] = NNsk # ############### # # Lightning MLP # ############### # if 'NNL' in use_models: # NNL = MLP_Lightning.MLP( # hidden_layer_sizes=net_dims, # activation=hparams['nn_act'], # solver='adam', # n_epochs=hparams["n_epochs"], # validation_fraction=0.2, # alpha=hparams["nn_l2"], # batch_size=hparams['nn_batch_size'], # learning_rate=hparams["learning_rate"], # patience=hparams["nn_patience"] # ) # all_models['NNL'] = NNL ############### # Random Forest ############### if 'RF' in use_models: n_trees = hparams["RF_n_estimators"] Random_Forest = RandomForestRegressor(n_estimators=n_trees) all_models['RF'] = Random_Forest ############################ # Gradient Boosting ############################ if 'GB' in use_models: n_trees = hparams["GB_n_estimators"] Gradient_Boosting = GradientBoostingRegressor(n_estimators=n_trees) all_models['GB'] = Gradient_Boosting ######### # XGBoost ######### if 'XGB' in use_models: XGBoost = XGBRegressor() all_models['XGB'] = XGBoost ############################ # Gaussian Process ############################ batch_size = 100 epochs = 1000 learning_rate = 0.1 n_inducing_points = 100 lengthscales = np.full(n_features, hparams['GP_lengthscale']) noise = hparams["GP_alpha"] if 'GPsk' in use_models: kernel = ConstantKernel() * RBF(length_scale=lengthscales) Gaussian_Process = GaussianProcessRegressor(kernel=kernel, alpha=noise**2, normalize_y=True) all_models['GPsk'] = Gaussian_Process # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'GPR' in use_models: # model = gpflow.models.GPR # GPR = GPflow_GP(model, kernel, alpha=noise) # all_models['GPR'] = GPR # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SGPR' in use_models: # model = gpflow.models.SGPR # SGPR = GPflow_GP(model, kernel, alpha=noise, n_inducing_points=n_inducing_points, standard_scale=True) # all_models['SGPR'] = SGPR # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'VGP' in use_models: # model = gpflow.models.VGP # VGP = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate) # all_models['VGP'] = VGP # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SVGP' in use_models: # model = gpflow.models.SVGP # SVGP = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, train_noise=True, diff_std_for_sc_and_non_sc=False, natgrad=False, train_noise_scale=False, predict_y=False) # all_models['SVGP'] = SVGP # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SVGP_single' in use_models: # model = gpflow.models.SVGP # SVGP = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, train_noise=True, diff_std_for_sc_and_non_sc=False, natgrad=False, train_noise_scale=False, predict_y=False) # all_models['SVGP_single'] = SVGP # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SVGP_sc_non-sc' in use_models: # model = gpflow.models.SVGP # SVGP = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, train_noise=True, diff_std_for_sc_and_non_sc=True, natgrad=False, train_noise_scale=False, predict_y=False) # all_models['SVGP_sc_non-sc'] = SVGP # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SVGP_features' in use_models: # model = gpflow.models.SVGP # SVGP = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, train_noise=False, diff_std_for_sc_and_non_sc=False, natgrad=False, train_noise_scale=True, predict_y=True) # all_models['SVGP_features'] = SVGP # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) + gpflow.kernels.WhiteKernel(variance=noise**2) # if 'SVGP_white' in use_models: # model = gpflow.models.SVGP # SVGP = GPflow_GP(model, kernel, alpha=0.0011, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, train_noise=True, diff_std_for_sc_and_non_sc=False, natgrad=False, train_noise_scale=False, predict_y=False) # all_models['SVGP_white'] = SVGP # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SVGP_RGM' in use_models: # model = gpflow.models.SVGP # NN_path = 'RGM' # SVGP_RGM = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, NN_path=NN_path) # all_models['SVGP_RGM'] = SVGP_RGM # # kernel = gpflow.kernels.Constant() * gpflow.kernels.RBF(lengthscales=lengthscales) # if 'SVGP_NN' in use_models: # model = gpflow.models.SVGP # NN_path = 'NN' # SVGP_NN = GPflow_GP(model, kernel, alpha=noise, batch_size=batch_size, epochs=epochs, learning_rate=learning_rate, n_inducing_points=n_inducing_points, standard_scale=True, NN_path=NN_path) # all_models['SVGP_NN'] = SVGP_NN # # # # For all pytorch models: # input_layer_size = n_features # featurizer = get_featurizer(input_layer_size, hparams, mode='FeedForward') # classifier = get_classifier(output_layer_size=n_targets, hparams=hparams, output_layers=output_layers) # # # ############################ # # Regret Minimization Network # ############################ # if 'RGM' in use_models: # RGM = RGM_sklearn( # solver=hparams['nn_solver'], # max_iter=hparams["n_epochs"], # batch_size=hparams['nn_batch_size'], # learning_rate_init=hparams["learning_rate"], # featurizer=deepcopy(featurizer), # classifier=deepcopy(classifier), # batch_mode=hparams['RGM_batch_mode'], # weight_decay=hparams["nn_l2"], # rgm_e=hparams['RGM_rgm_e'], # erm_e=hparams['RGM_erm_e'], # holdout_e=hparams['RGM_holdout_e'], # detach_classifier=hparams['RGM_detach_classifier'], # oracle=hparams['RGM_oracle'], # ensemble_pred=hparams['RGM_ensemble_pred'], # validation_fraction=0.2, # early_stopping=hparams['RGM_early_stopping'], # 'False', 'valid', 'extrapol', ... # n_iter_no_change=hparams["nn_patience"], # clip_grad=hparams['NN_clip_grad'], # num_train_domains=hparams['RGM_num_train_domains'], # max_n_classifiers=10, # if_log_metrics=False, # coeff_lr_classifier=hparams['coeff_lr_classifier'], # reduce_lr_factor=hparams['NN_reduce_lr_factor'], # use_tensorboard=False # ) # RGM.domain_col = domain_col # all_models['RGM'] = RGM # # # ############################ # # Regret Minimization Network without domains # ############################ # if 'NN' in use_models: # NN = RGM_sklearn( # solver=hparams['nn_solver'], # max_iter=hparams["n_epochs"], # batch_size=hparams['nn_batch_size'], # learning_rate_init=hparams["learning_rate"], # featurizer=deepcopy(featurizer), # classifier=deepcopy(classifier), # batch_mode='Conserve_ratio', # weight_decay=hparams["nn_l2"], # rgm_e=1, # erm_e=1, # holdout_e=1, # detach_classifier=False, # oracle=False, # ensemble_pred=False, # validation_fraction=0.2, # early_stopping='valid', # n_iter_no_change=hparams["nn_patience"], # clip_grad=hparams['NN_clip_grad'], # num_train_domains=1, # if_log_metrics=False, # reduce_lr_factor=hparams['NN_reduce_lr_factor'], # use_tensorboard=False # ) # all_models['NN'] = NN # # ################# # # Original MEGNet # ################# # if 'MEGNet' in use_models: # validation_frac = 0.2 # # transfer_model = None if args.add_params['prev_model'] is None else projectpath(args.add_params['prev_model']) # MEGNet0 = MEGNet_tf( # use_learnt_elemental_embedding=False, # epochs=hparams['n_epochs'], # lr=args.add_params['lr'],#hparams['learning_rate'], # batch_size=args.add_params['batch_size'],#hparams['nn_batch_size'], # patience=hparams['nn_patience'], # l2_coef=args.add_params['l2'],#None, # dropout=args.add_params['dropout'], # r_cutoff=4, # early_stopping=args.add_params['early_stopping'], # validation_frac=validation_frac, # loss='mse', # domain_col=domain_col, # optimizer_kwargs={'clipnorm': args.add_params['clipnorm']}, # tensorboard=False, # nblocks=args.add_params['nblocks'], # n1=args.add_params['n1'], # n2=args.add_params['n2'], # n3=args.add_params['n3'], # lr_exp_decay=args.add_params['lr_exp_decay'],#0.997, # prev_model=None, # act=args.add_params['act'], # npass=args.add_params['npass'], # n_feat_bond=args.add_params['n_feat_bond'] # ) # all_models['MEGNet'] = MEGNet0 return(all_models) def train_with_args(args): os.chdir(args.calcdir) print("JOB STARTED") print_title("Experiment: {}".format(args.experiment)) print("Current working directory: %s"%(args.calcdir)) # Get hyperparameters ) args.hparams = get_hparams(args.hparams_file) args.random_seed = args.hparams['random_seed'] if args.random_seed != None: print(f"RANDOM SEED FIXED!!!: {args.random_seed}") else: args.random_seed = random.randint(0, 1e3) np.random.seed(args.random_seed) random.seed(args.random_seed) # torch.manual_seed(args.random_seed) tf.random.set_seed(args.random_seed) domain_colname = args.domain_colname print(f'Domain: {domain_colname}') # Add CV mode to name of output directory outdir = make_output_directory(args.outdirname, label=args.experiment) print(f'Output directory: {outdir}') # TODO # assert not 'n_repetitions' in args.add_params # wished_CV_cols = [] # # wished_CV_cols = [f'CV_{i}' for i in range(args.add_params['n_repetitions'])] # if args.add_params['features'] == 'graph': # usecols = lambda x: (not (x.startswith('SOAP') or x.startswith('PCA') or x.startswith('MAGPIE'))) and ((not x.startswith('CV_')) or (x in wished_CV_cols)) # else: # usecols = lambda x: (not x.startswith('CV_')) or (x in wished_CV_cols) # print(f'n_repetitions = {len(wished_CV_cols)}') # Get dataset. df_data = pd.read_csv(args.dataset, header=1) if args.use_data_frac != None: warnings.warn(f'Using only a fraction of {args.use_data_frac} of data for debugging purposes.') df_data = df_data.sample(frac=args.use_data_frac).reset_index() # Add absolute path to graphs. if args.add_params['features'] == 'graph': df_data['graph'] = df_data['graph'].apply(projectpath) magpie_features = [magpie for magpie in df_data.columns if magpie.startswith('MAGPIE')] soap_features = [soap for soap in df_data.columns if soap.startswith('SOAP')] pca_soap_features = [f'PCA_SOAP_{i}' for i in range(100)] electronic_features = ['band_gap_2', 'energy_2', 'energy_per_atom_2', 'formation_energy_per_atom_2', 'total_magnetization_2', 'num_unique_magnetic_sites_2', 'true_total_magnetization_2'] lattice_features = ['lata_2', 'latb_2', 'latc_2'] sym_features = ['cubic', 'hexagonal', 'monoclinic', 'orthorhombic', 'tetragonal', 'triclinic', 'trigonal', 'primitive', 'base-centered', 'body-centered', 'face-centered'] features = [] if 'MAGPIE' in args.add_params['features']: features += magpie_features if 'PCA' in args.add_params['features']: features += pca_soap_features elif 'SOAP' in args.add_params['features']: features += soap_features if 'electro' in args.add_params['features']: features += electronic_features if not (args.add_params['features'] == 'MAGPIE' or args.add_params['features'] == 'MAGPIE (all)' or args.add_params['features'] == 'graph'): features += ['crystal_temp_2'] features += sym_features + lattice_features if args.add_params['features'] == 'graph': features += ['graph'] targets = ['tc'] # Define scaling/ transformation of features and targets. log_transform = FunctionTransformer( func=restricted_arcsinh, inverse_func=restricted_sinh, check_inverse=False ) Column_Transformer = {} for feat in features: Column_Transformer[feat] = StandardScaler() if features != ['graph'] else 'passthrough' for target in targets: Column_Transformer[target] = log_transform#StandardScaler() assert len(Column_Transformer) == len(features + targets) output_layers = None # None or any from get_activation_fn() # # Train by physical group. # group = args.add_params['group'] # print(f'Training only group {group}') # df_data = df_data.loc[df_data['sc_class'] == group] # # TODO: remove. For CV experiment: Exclude data points with unclear physical group # df_data = df_data[df_data['sc_class_unique_sc']] # Get indices for order in which groups shall be displayed. if domain_colname != None: groupcol = df_data[domain_colname] sorted_grouplabels = ML.unique_sorted(groupcol) index_group_map = {index: grouplabel for index, grouplabel in enumerate(sorted_grouplabels)} print('Group indices:') print(index_group_map) # Get train and test columns with specified CV algorithm. if args.CV != None: df_data = get_train_test_data( df_data=df_data, CV=args.CV, n_folds=args.n_folds, n_repeats=args.n_repeats, domain_colname=domain_colname, trainfrac=args.train_frac, random_n_reps=args.n_reps, group=args.CV_keep_groups ) # # Make some tests which parts of the matching & adaptation algorithm have which effects. # # if args.add_params['n_exclude_if_too_many_structures']: # # n_max = 1 # # n_structures = df_data.groupby('formula_sc')['formula_sc'].transform(len) # # too_many_structures = n_structures > n_max # # n_exclude_sc = len(df_data[too_many_structures].drop_duplicates("formula_sc")) # # df_data = df_data[~too_many_structures] # # print(f'Excluding {n_exclude_sc} superconductors because they have more than {n_max} structures.') # if args.add_params['drop_duplicate_superconductors']: # n_before = len(df_data) # n_sc_before = df_data['formula_sc'].nunique() # df_data = df_data.drop_duplicates('formula_sc') # assert n_sc_before == df_data['formula_sc'].nunique() # print(f'Lost {n_before - len(df_data)} duplicate structures but kept all superconductors.') # # if args.add_params['only_totreldiff=0']: # # n_sc_before = df_data['formula_sc'].nunique() # # df_data = df_data[df_data['totreldiff'] == 0] # # n_sc_lost = n_sc_before - df_data['formula_sc'].nunique() # # print(f'Lost {n_sc_lost} superconductors by filtering by totreldiff=0.') # # if args.add_params['only_abs_matches']: # # n_sc_before = df_data['formula_sc'].nunique() # # df_data = df_data[df_data['correct_formula_frac']] # # n_sc_lost = n_sc_before - df_data['formula_sc'].nunique() # # print(f'Lost {n_sc_lost} superconductors by keeping only absolute matches.') # # if args.add_params['without_lattice_feats']: # # remove_features = sym_features + lattice_features # # features = [feat for feat in features if not feat in remove_features] # df_data['weight'] = 1 / df_data.groupby('formula_sc')['formula_sc'].transform(len) # Make a train data experiment. # if not args.add_params['train_frac'] is None: # fraction = args.add_params['train_frac'] # CV_cols = [col for col in df_data.columns if col.startswith('CV_')] # for cv in CV_cols: # is_test = df_data[cv] == 'test' # is_train = df_data[cv] == 'train' # all_train_groups = df_data.loc[is_train, args.CV_keep_groups].unique() # n_train_groups = int(len(all_train_groups) * fraction) # train_groups = np.random.choice(all_train_groups, size=n_train_groups, replace=False) # df_data.loc[is_train, cv] = np.where(df_data.loc[is_train, args.CV_keep_groups].isin(train_groups), 'train', np.nan) # assert all(df_data[is_test] == 'test') # assert all(df_data[is_train].isna() | (df_data[is_train] == 'train')) # mean_n_sc = (df_data.drop_duplicates('formula_sc')[CV_cols] == 'train').sum().mean() # args.add_params['mean_n_train_sc'] = float(mean_n_sc) # Rename test columns and make validation columns instead. # if args.add_params['HPO'] and CV != None: # df_data = get_validation_columns(df_data, args, domain_colname) # Sanity check. if args.CV_keep_groups == 'chemical_composition_sc': CV_cols = [col for col in df_data.columns if col.startswith('CV_')] for CV_col in CV_cols: is_test = df_data[CV_col] == 'test' is_train = df_data[CV_col] == 'train' test_comps = df_data.loc[is_test, 'chemical_composition_sc'].unique() train_comps = df_data.loc[is_train, 'chemical_composition_sc'].unique() assert len(np.intersect1d(test_comps, train_comps)) == 0 # Select data by criteria. # if CV != None: # from dataset_preparation import _6_select_best_matches_and_prepare_df # df_data = _6_select_best_matches_and_prepare_df.keep_only_best_matches(df=df_data, criteria=args.add_params['criteria'], n_exclude_if_more_structures=args.add_params['n_exclude_if_more_structures'], output_graph_dir=None) args.add_params['n_data_points'] = len(df_data) print('Number of data points:', len(df_data)) # Save values of some variables in the final output file for later convenience. save_value_of_variables = {key: val for key, val in vars(args).items()} save_value_of_variables['Domain_colname'] = domain_colname n_domains = len(df_data[domain_colname].unique()) if domain_colname != None else 1 all_models = get_all_models(args.hparams, len(features), len(targets), args.use_models, n_domains, domain_colname, output_layers, outdir=outdir, args=args) use_models = {modelname: all_models[modelname] for modelname in args.use_models} ml = ML.Machine_Learning( data=df_data, features=features, targets=targets, domain=domain_colname, sample_weights=args.sample_weights, metrics_sample_weights=args.metrics_sample_weights, Column_Transformer=Column_Transformer, save_models=args.save_models, save_torch_models=True, is_optimizing_NN=args.is_optimizing_NN, save_value_of_variables=save_value_of_variables, print_features=False, print_targets=True, print_domain_score=True, random_seed=args.random_seed, save_all_values_and_predictions=True, n_jobs=args.n_jobs, copy_files=[args.hparams_file] ) print('Start training.') ml.train(use_models, outdir) # # Make lots of plots. # plt.ioff() # plot_dir = os.path.join(outdir, 'plots') # models_without_loss_curve = ['LR', '1NN', 'RF', 'XGB', 'GPsk', 'GPR', 'SGPR'] # run = MLRun(outdir) # plot_models = [modelname for modelname in args.use_models if not modelname in models_without_loss_curve] # run.final_plots(plot_dir, plot_models, df_data, domain_colname, features, targets, use_models, outdir) # Check if refactoring was successful and everything has stayed the same as in comparison_dir. # comparison_dir = '/home/timo/superconductors_3D/analysis/results/testing/results_202_' # refactoring = Refactoring.Refactoring(comparison_dir) # refactoring.check(outdir) return(ml, outdir) def main(args_from_fn): # ============================================================================= # Define options. # ============================================================================= # use_models = ['1NN', 'LR', 'XGB', 'SVGP', 'NNsk', 'NN', 'RGM'] use_models = ['XGB'] experiment = '' add_params = { # 'features': 'graph', # 'database': 'MP', # "act": "relu", # "dropout": 0.8762999083316979, # "lr": 0.000008567057310854599, # "lr_exp_decay": 0.983512961357719, # "n1": 46, # "n2": 52, # "n3": 73, # "n_feat_bond": 18, # "nblocks": 2, # "npass": 8, # 'batch_size': 56, # 'clipnorm': 1.063448501785796, # 'l2': 2.1555727094418956e-7, # 'n_exclude_if_too_many_structures': False, # 'drop_duplicate_superconductors': False, # 'only_totreldiff=0': False, # 'only_abs_matches': False, # 'same_n_sc': True, # 'without_lattice_feats': True, # 'criteria': ['no_crystal_temp_given_2'] # 'group': 'Oxide', # 'train_frac': 0.3, # 'early_stopping': True, 'features': 'MAGPIE+DSOAP', # 'CV': 'LOGO', 'CV_keep_groups': 'chemical_composition_sc', # 'domain_colname': None,#'num_elements_sc', # 'CV_name': 'LOCO_phys', } output_note = '' outdirname = '/home/timo/superconductors_3D/analysis/results/testing'#'/media/timo/ext42/academic_projects/superconductors_3D/analysis/results/testing' calcdir = os.getcwd() # Cross validation CV = 'Random' # `KFold`, `LOGO`, `Random` or None n_folds = 5 # for KFold n_repeats = 1 # for KFold CV_keep_groups = 'chemical_composition_sc' # for KFold, Random n_reps = 3 # for Random train_frac = 0.8 # for Random domain_colname = None # for LOGO # Weights sample_weights = 'weight' metrics_sample_weights = 'weight' # Dataset dataset = '/home/timo/superconductors_3D/data_before/final/MP/SC_MP_matches.csv'#projectpath('data', 'final', 'ICSD', 'SC_ICSD_matches.csv')#projectpath('data', 'intermediate', 'MP', '5_features_SC_MP.csv') # Hyperparameters hparams_file = 'hparams.yml' n_jobs = 1 is_optimizing_NN = False # Only run NN and don't plot anything when optimizing NN save_models = True # Debugging use_data_frac = None # None for using all data. # ============================================================================= # End define options. # ============================================================================= print('Input:\n', sys.argv) args = parse_arguments(args_from_fn, use_models, experiment, output_note, outdirname, calcdir, save_models, CV, n_folds, n_repeats, CV_keep_groups, n_reps, train_frac, domain_colname, is_optimizing_NN, dataset, hparams_file, n_jobs, sample_weights, metrics_sample_weights, use_data_frac, add_params) print('args.add_params:\n', args.add_params) # Run ML and measure time of run. starttime = datetime.datetime.now() ml, _ = train_with_args(args) duration = datetime.datetime.now() - starttime print(f'Duration: {duration}') return ml if __name__ == "__main__": ml = main(args_from_fn={})
45,766
46.328852
303
py
3DSC
3DSC-main/superconductors_3D/machine_learning/RGM_own.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Tue Apr 20 17:24:04 2021 @author: timo This script contains my wrapper of the RGM of 2020 Jin to go with the standard sklearn API. """ import os import numpy as np from superconductors_3D.machine_learning.Algorithms.RGM_Jin import RGM as RGM_Jin import torch import torch.nn as nn from torch.utils.tensorboard import SummaryWriter import random import warnings from torch.utils.data import Dataset, DataLoader from chemprop.nn_utils import initialize_weights from collections import namedtuple from sklearn.model_selection import StratifiedShuffleSplit from copy import deepcopy import mlflow import time # import torchviz from superconductors_3D.machine_learning.own_libraries.own_functions import enforce_types def log_metrics(metrics: dict, step: float = None, max_tries: float = 10, if_log_metrics: bool = True) -> None: """Logs metrics with mlflow but also checks for errors and tries again up to max_tries times if there is a connection error. Logs only if if_log_metrics = True. """ if if_log_metrics: i = 0 while i < max_tries: try: mlflow.log_metrics(metrics, step=step) i = max_tries except ConnectionError: warnings.warn('Connection Error, try again.') time.sleep(0.1) i += 1 return() def get_domain_batch(x_data, y_data, w_data, d_data, domain): """Returns the parts of x_data, y_data, w_data that are from domain as defined in d_data. """ is_domain = d_data == domain x = x_data[is_domain] y = y_data[is_domain] d = d_data[is_domain] w = w_data[is_domain] return(x, y, d, w) def get_all_domain_batches(x_data, y_data, w_data, d_data, all_domains): """Returns a list of batches where each entry in the list is a batch of a certain domain. """ batches = [] for i, domain in enumerate(all_domains): batch = get_domain_batch(x_data, y_data, w_data, d_data, domain) batches.append(batch) return(batches) class DomainDataset(Dataset): def __init__(self, x, y, d, w, device): self.x = x self.y = y self.d = d self.w = w self.device = device # Assert same lengths length = len(self.x) assert all([length == len(self.y), length == len(self.d), length == len(self.w)]), 'At least one of features, targets, domains and weights has different length.' def __len__(self): return len(self.x) def __getitem__(self, index): x = self.x[index] y = self.y[index] d = self.d[index] w = self.w[index] return x, y, d, w def collate_fn(self, batch): x, y, d, w = zip(*batch) x = torch.stack(x) y = torch.stack(y) d = torch.stack(d) w = torch.stack(w) return x, y, d, w class EarlyStopping: """Stops early if validation loss doesn't improve after a given patience.""" def __init__(self, n_iter_no_change, delta=0): """ Args: n_iter_no_change (int): How long to wait after last time validation loss improved. delta (float): Minimum change in the monitored quantity to qualify as an improvement. Default: 0 """ self.n_iter_no_change = n_iter_no_change self.counter = 0 self.best_score = None self.early_stop = False self.valid_loss_min = np.Inf self.delta = delta def check_loss(self, valid_loss, model, epoch): score = -valid_loss if self.best_score is None: self.best_score = score self.save_checkpoint(valid_loss, model, epoch) elif score < self.best_score + self.delta: self.counter += 1 if self.counter >= self.n_iter_no_change: self.early_stop = True else: self.best_score = score self.save_checkpoint(valid_loss, model, epoch) self.counter = 0 def save_checkpoint(self, valid_loss, model, epoch): '''Saves model when validation loss decrease.''' # Deepcopy is important, otherwise only reference is passed. self.best_model_state_dict = deepcopy(model.state_dict()) self.valid_loss_min = valid_loss self.best_epoch = epoch class RGM_sklearn(nn.Module): """A wrapper around the RGM implementation of 2020 Jin to be able to be called like a standard sklearn module. """ @enforce_types def __init__(self, solver:str, max_iter: int, batch_size: int, learning_rate_init: (float, int), featurizer, classifier, batch_mode: str, weight_decay: (float, int), rgm_e: (float, int), erm_e: (float, int), holdout_e: (float, int), detach_classifier: bool, oracle: bool, ensemble_pred: bool, validation_fraction: float, early_stopping: str, n_iter_no_change: int, num_train_domains: int, max_n_classifiers: int=999999999, clip_grad: (float, int)=np.inf, if_log_metrics: bool=True, random_seed: (int, type(None))=None, coeff_lr_classifier: (float, int)=1., reduce_lr_factor: float=1., use_tensorboard: bool=True): super(RGM_sklearn, self).__init__() self.max_epochs = max_iter self.batch_size = batch_size self.learning_rate_init = learning_rate_init self.coeff_lr_classifier = coeff_lr_classifier self.batch_mode = batch_mode # Conserve_ratio' self.rgm_e = rgm_e self.erm_e = erm_e self.holdout_e = holdout_e self.detach_classifier = detach_classifier self.oracle = oracle self.seed = random_seed self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") print(f'device: {self.device}') self.loss_func = nn.MSELoss(reduction='none') self.featurizer = deepcopy(featurizer) self.classifier = deepcopy(classifier) self.solver = solver self.weight_decay = weight_decay self.ensemble_pred = ensemble_pred self.validation_fraction = validation_fraction self.early_stopping = early_stopping self.clip_grad = clip_grad self.if_log_metrics = if_log_metrics self.num_train_domains = num_train_domains self.max_n_classifiers = max_n_classifiers self.use_tensorboard = use_tensorboard self.reduce_lr_factor = reduce_lr_factor if early_stopping != 'False': assert n_iter_no_change != None self.n_iter_no_change = n_iter_no_change self.input_args = deepcopy(locals()) self.input_args['estimator_name'] = type(self).__name__ del self.input_args['self'] del self.input_args['featurizer'] del self.input_args['classifier'] self.backward_graphs = [] def get_optimizer(self, solver, parameters, lr, weight_decay): """Returns correct optimizer based on `solver`.""" if solver == 'sgd': optimizer = torch.optim.SGD(parameters, lr=lr, weight_decay=weight_decay) elif solver == 'adam': optimizer = torch.optim.Adam(parameters, lr=lr, weight_decay=weight_decay) elif solver == 'adamw': optimizer = torch.optim.AdamW(parameters, lr=lr, weight_decay=weight_decay) elif solver == 'rmsprop': optimizer = torch.optim.RMSprop(parameters, lr=lr, weight_decay=weight_decay) else: raise Warning('Solver "f{solver} not known. Solvers are lowercase always.') return(optimizer) def calc_params_norm(self, parameters, norm_type=2): """Calculates parameter norm as in torch.nn.utils.clip_grad. Returns torch.tensor(). """ parameters = [p for p in parameters if p.grad is not None] if len(parameters) == 0: return torch.tensor(0.) device = parameters[0].grad.device total_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]), norm_type) return(total_norm) # def make_backwards_graph(self, loss, batch_idx): # """"Make graph of backward pass.""" # if self.epoch == 0 and batch_idx == 0: # backward_graph = \ # torchviz.make_dot(loss, dict(self.trainer.named_parameters())) # self.backward_graphs.append(backward_graph) # return() def initialize_model_weights(self, model): """Initializes all weights of the model.""" model.apply(self.weight_initialization) return() def weight_initialization(self, m): """Initializes the layer m with a uniform xavier distribution if it is linear.""" if isinstance(m, nn.Linear): nn.init.xavier_uniform_(m.weight.data) return() def make_consistent_torch_tensors(self, x_train, y_train, d_train, w_train): """Make data to torch tensors for consistency. If weights are not given give it the neutral value 1 and make it a consistent tensor. """ if w_train is None: shape = y_train.shape w_train = torch.ones(shape, dtype=torch.float, device=self.device) else: w_train = torch.tensor(w_train, dtype=torch.float, device=self.device) x_train = torch.tensor(x_train, dtype=torch.float, device=self.device) if y_train.ndim == 1: y_train = y_train.reshape(-1, 1) # fix sklearn inconsistency. y_train = torch.tensor(y_train, dtype=torch.float, device=self.device) try: if len(d_train) == 0: num_datapoints = len(x_train) d_train = torch.zeros(num_datapoints, dtype=torch.float, device=self.device) else: d_train = torch.tensor(d_train, dtype=torch.float, device=self.device) except TypeError: print('Internally converting string domain labels to integers.') labels, counts = np.unique(d_train, return_counts=True) sort_idc = np.argsort(counts) sorted_labels = labels[sort_idc] domain_dict = {label: idx for idx, label in enumerate(sorted_labels)} d_train = np.array([domain_dict[d] for d in d_train]) d_train = torch.tensor(d_train, dtype=torch.float, device=self.device) assert all([length == len(x_train) for length in (len(y_train), len(d_train), len(w_train))]) return(x_train, y_train, d_train, w_train) def get_validation_data(self, x_train, y_train, d_train, w_train): """Splits data in train and validation data.""" assert isinstance(self.validation_fraction, float), 'validation_fraction must be float.' train_indices, valid_indices = next(StratifiedShuffleSplit(n_splits=1, test_size=self.validation_fraction).split(x_train.cpu(), d_train.cpu())) x_train, x_valid = x_train[train_indices], x_train[valid_indices] y_train, y_valid = y_train[train_indices], y_train[valid_indices] d_train, d_valid = d_train[train_indices], d_train[valid_indices] w_train, w_valid = w_train[train_indices], w_train[valid_indices] x_valid = x_valid.to(self.device) y_valid = y_valid.to(self.device) d_valid = d_valid.to(self.device) w_valid = w_valid.to(self.device) return(x_train, x_valid, y_train, y_valid, d_train, d_valid, w_train, w_valid) def clip_grad_norm(self): """Clips the gradient norm, returns gradient norm before and after and logs them. """ grad_norm_before = torch.nn.utils.clip_grad_norm_(self.trainer.parameters(), self.clip_grad).item() grad_norm_clipped = self.calc_params_norm(self.trainer.parameters()).item() # self.writer.add_scalar('grad_norm_before', grad_norm_before) # self.writer.add_scalar('grad_norm_clipped', grad_norm_clipped) return(grad_norm_before, grad_norm_clipped) def weighted_average(self, y, weights): y = y * weights norm = weights.sum() return y.sum() / norm def loss_forward(self, y_pred, y_true, weights): # TODO assert (y_pred >= 0).all() pred_losses = self.loss_func(y_pred, y_true) pred_loss = self.weighted_average(pred_losses, weights) return pred_loss def get_domain_data(self, x_data, y_data, w_data, d_data, domain): """Returns the dataset for a single domain. x_data, y_data, w_data, d_data are the datasets with all domains together. domain is the domain label. """ x, y, d, w = get_domain_batch(x_data, y_data, w_data, d_data, domain) data = DomainDataset(x, y, d, w, self.device) return(data) def determine_batch_ratios(self, mode): """This function determines the batch sizes per domain ('batch_ratios') based on `mode`. """ if mode == 'Conserve_ratio': # Each domain will have a batch size according to its ratio in the total dataset. ratios = self.all_domain_counts / sum(self.all_domain_counts) batch_ratios = self.batch_size * ratios batch_ratios = torch.round(batch_ratios.float()).int() drop_last = True else: raise ValueError(f'Unknown batch_mode "{mode}".') if any(batch_ratios == 0): warnings.warn('One of the domains has very few examples and should have zero examples per batch acccording to batch_size. Setting minimum number of examples to 1.') min_batch_ratios = torch.ones_like(batch_ratios) batch_ratios = torch.maximum(batch_ratios, min_batch_ratios) elif any(batch_ratios > self.all_domain_counts): warnings.warn('For at least one domain the batchsize is greater than the number of samples, unpredictable behaviour.') return(batch_ratios, drop_last) def get_batches(self, x_data, y_data, d_data, w_data, mode): """Generates a list of batches for each domain. """ batch_ratios, drop_last = self.determine_batch_ratios(mode) # Get data loaders for each domain that return minibatches scaled like the ratio in the total dataset. data_loaders = {} for domain, batch_ratio in zip(self.all_domains, batch_ratios): batch_ratio = batch_ratio.item() domain = domain.item() data = self.get_domain_data(x_data, y_data, w_data, d_data, domain) data_loaders[domain] = iter(DataLoader(data, batch_size=batch_ratio, drop_last=drop_last, shuffle=True, collate_fn=data.collate_fn)) # Go through batches of each domain. If a domain has no examples left in the dataset, resample it and continue until all domains are through at least once. num_batches = [len(loader) for loader in data_loaders.values()] # print('The number of batches per domain is', num_batches) num_iters = max(num_batches) for batch_idx in range(num_iters): batches = [] for domain, batch_ratio in zip(self.all_domains, batch_ratios): batch_ratio = batch_ratio.item() domain = domain.item() d_loader = data_loaders[domain] d_batch = next(d_loader, None) # If one domain has no data points left reload data points. if d_batch == None: assert self.num_domains > 1 data = self.get_domain_data(x_data, y_data, w_data, d_data, domain) d_loader = iter(DataLoader(data, batch_size=batch_ratio, drop_last=drop_last, shuffle=True, collate_fn=data.collate_fn)) d_batch = next(d_loader, None) data_loaders[domain] = d_loader assert not d_batch == None batches.append(d_batch) yield(batches) def prepare_trainer(self): """Builds up the Neural Network trainer with all it's layers, it's optimizer etc. """ trainer = RGM_Jin(featurizer=self.featurizer, classifier=self.classifier, num_domains=self.num_domains, rgm_e=self.rgm_e, erm_e=self.erm_e, holdout_e=self.holdout_e, detach_classifier=self.detach_classifier, oracle=self.oracle, loss_forward=self.loss_forward, num_train_domains=self.num_train_domains, max_n_classifiers=self.max_n_classifiers ) # Initialize weights new. Very important because otherwise all classifiers have the same initialization! initialize_weights(trainer) # self.initialize_model_weights(trainer) trainer = trainer.to(self.device) # Get parameter lists for classifier and representation separately for different learning rates. classifier_params = [] for classifier in (trainer.classifier, trainer.f_k, trainer.g_k): classifier_params.extend(list(classifier.parameters())) assert len(list(trainer.parameters())) == len(classifier_params) + len(list(trainer.copy_f_k.parameters())) + len(list(trainer.featurizer.parameters())) learning_rate_classifier = self.coeff_lr_classifier * self.learning_rate_init parameters = [ {'params': trainer.featurizer.parameters()}, {'params': classifier_params, 'lr': learning_rate_classifier} ] trainer.optimizer = self.get_optimizer(solver=self.solver, parameters=parameters, lr=self.learning_rate_init, weight_decay=self.weight_decay ) trainer.scheduler = torch.optim.lr_scheduler.ExponentialLR(trainer.optimizer, gamma=self.reduce_lr_factor) return(trainer) def validate(self, x_train, x_valid, y_train, y_valid, w_train, w_valid): """Calculates train and validation loss.""" y_train_pred = self.evaluate(x_train) y_valid_pred = self.evaluate(x_valid) train_loss = self.loss_forward(y_pred=y_train_pred, y_true=y_train, weights=w_train).item() valid_loss = self.loss_forward(y_pred=y_valid_pred, y_true=y_valid, weights=w_valid).item() return(train_loss, valid_loss) def train_epoch(self, x_train, y_train, d_train, w_train): # d_train: domain labels "Train one epoch." self.trainer.train() all_batches = self.get_batches(x_train, y_train, d_train, w_train, mode=self.batch_mode) # Record all losses and the gradient. losses = {sub_loss: 0 for sub_loss in self.trainer.loss_curve_iter.keys()} losses['grad_norm_before'] = 0 losses['grad_norm_clipped'] = 0 for batch_idx, d_batches in enumerate(all_batches): self.trainer.optimizer.zero_grad() loss = self.trainer(d_batches) # self.make_backwards_graph(loss, batch_idx) loss.backward() grad_norm_before, grad_norm_clipped = self.clip_grad_norm() self.trainer.optimizer.step() for sub_loss in self.trainer.loss_curve_iter.keys(): losses[sub_loss] += self.trainer.loss_curve_iter[sub_loss][-1] losses['grad_norm_before'] += grad_norm_before losses['grad_norm_clipped'] += grad_norm_clipped for sub_loss, loss_value in losses.items(): self.loss_curve_[sub_loss].append(loss_value) def train_until_stop(self, x_train, y_train, d_train, w_train, x_valid, y_valid, w_valid): """Trains the model until either all epochs are trained or early_stopping criterion is reached. Label is a string to differentiate the run to train the representation and the run to train the classifier if these are seperated. """ if self.early_stopping != 'False': stop = EarlyStopping(self.n_iter_no_change) self.loss_curve_['train'] = [] self.loss_curve_['valid'] = [] for epoch in range(self.max_epochs): self.epoch = epoch self.train_epoch(x_train, y_train, d_train, w_train) # Validation on validation set. train_loss, valid_loss = self.validate(x_train, x_valid, y_train, y_valid, w_train, w_valid) self.loss_curve_['train'].append(train_loss) self.loss_curve_['valid'].append(valid_loss) # metrics = {'train_loss': train_loss, 'valid_loss': valid_loss} # log_metrics(metrics, step=epoch, if_log_metrics=self.if_log_metrics) # self.writer.add_scalars('loss', {'train': train_loss, 'valid': valid_loss}, global_step=epoch) self.trainer.scheduler.step() # Early stopping, but only after min 20% of epochs. if self.early_stopping != 'False' and epoch > self.max_epochs / 5: valid_score = self.loss_curve_[self.early_stopping][-1] stop.check_loss(valid_score, self.trainer, epoch) if stop.early_stop == True: print(f'Stopped early in epoch {epoch}.') break if self.early_stopping != 'False' and epoch > self.max_epochs / 5: self.trainer.load_state_dict(stop.best_model_state_dict) self.best_epoch = stop.best_epoch print(f'Best model loaded from epoch {self.best_epoch}.') return() def fit(self, x_train, y_train, d_train=[], sample_weight=[]): """Train the model for all epochs. """ if self.seed != None: warnings.warn('torch.manual_seed and random.seed behave a bit dangerous, different than np.random_seed. If the torch seed is set it not only is set for afterwards in this run, but it is set even if the variables are cleaned! Only a kernel restart also unsets the torch.manual_seed. I don\'t know if random.seed also persists after cleaning variables but I think so and it definitely persists across modules.') torch.manual_seed(self.seed) np.random.seed(self.seed) random.seed(self.seed) x_train, y_train, d_train, w_train = self.make_consistent_torch_tensors(x_train, y_train, d_train, sample_weight) # Get validation data. if not self.validation_fraction == None: x_train, x_valid, y_train, y_valid, d_train, d_valid, w_train, w_valid = \ self.get_validation_data(x_train, y_train, d_train, w_train) else: raise Warning('Code currently asserts validation set.') # Initialize variables. self.num_domains = len(torch.unique(d_train)) self.output_size = y_train.shape[1] if len(y_train.shape) == 2 else 1 self.n_train_samples = len(y_train) if self.batch_size > len(x_train): warnings.warn('Got batch size larger than sample size. It is going to be clipped.') self.batch_size = len(x_train) self.all_domains, self.all_domain_counts = torch.unique(d_train, return_counts=True) self.use_classifier = False if self.ensemble_pred else True self.trainer = self.prepare_trainer() self.loss_curve_ = {sub_loss: [] for sub_loss in self.trainer.loss_curve_iter.keys()} self.loss_curve_['grad_norm_before'] = [] self.loss_curve_['grad_norm_clipped'] = [] self.train_until_stop(x_train, y_train, d_train, w_train, x_valid, y_valid, w_valid) # self.trainer.log_losses = False # batches_data = get_all_domain_batches(x_train, y_train, w_train, d_train, self.all_domains) # self.writer.add_graph(self.trainer, [batches_data]) # self.writer.add_embedding(x_train, y_train, tag='data') # self.writer.flush() # torch.onnx.export(self.trainer, (batches_data,), self.outpath, export_params=True, opset_version=11, input_names=['input'], output_names=['output'], example_outputs=torch.tensor(1)) # torch.save(self.trainer, self.outpath) # for loss_name, loss_curve in self.trainer.loss_curve_iter.items(): # self.loss_curve_[loss_name] = loss_curve # Push model to cpu after training. Could be useful sometimes but not really needed. self = self.to('cpu') self.device = 'cpu' def evaluate(self, x_test): """Forward evaluates the network. Can either use the classifier or an ensemble of all extrapolator classifiers. """ if self.use_classifier: model = self.trainer.network model.eval() with torch.no_grad(): y_pred = model(x_test) else: shape = (self.trainer.n_classifiers, len(x_test), self.output_size) all_y_pred = torch.zeros(shape, device=self.device) for k, model in enumerate(self.trainer.ensemble): model.eval() with torch.no_grad(): all_y_pred[k, :, :] = model(x_test) y_pred = torch.quantile(all_y_pred, 0.5, dim=0) # median return(y_pred) def predict(self, x_test): """Infers data points and outputs a numpy array with the same dtype as the input array. """ input_dtype = x_test.dtype x_test = torch.from_numpy(x_test).float().to(self.device) self.use_classifier = True if not self.ensemble_pred else False y_pred = self.evaluate(x_test) y_pred = np.array(y_pred.cpu(), dtype=input_dtype) return(y_pred)
26,342
45.054196
617
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Custom_Machine_Learning_v1_3.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Fri Sep 11 10:14:31 2020 @author: timo This module is for the class MachineLearning that automatically executes a lot of different models and prints and saves all the output. """ import warnings # warnings.filterwarnings("ignore") import os # os.environ['TF_CPP_MIN_LOG_LEVEL']='2' import numpy as np import tensorflow as tf from matplotlib import pyplot as plt # import torch import pandas as pd from sklearn.preprocessing import StandardScaler from sklearn.utils.validation import check_is_fitted from sklearn.pipeline import Pipeline from sklearn.compose import TransformedTargetRegressor, ColumnTransformer import sklearn.metrics import sklearn.linear_model import random import copy import yaml import datetime from shutil import copyfile from superconductors_3D.machine_learning.own_libraries.own_functions import movecol, isfloat # import torch.optim import itertools import superconductors_3D.machine_learning.own_libraries.data.All_Data as All_Data import superconductors_3D.machine_learning.own_libraries.data.All_scores as All_Scores from superconductors_3D.machine_learning.own_libraries.data.Domain_statistics import save_Domain_statistics from superconductors_3D.machine_learning.own_libraries.data import Feature_Importances as FI from contextlib import redirect_stdout import json from superconductors_3D.machine_learning.own_libraries.utils.Models import Models, get_modelpath def load_df_and_metadata(path): """Loads the dataframe and if possible the metadata. """ try: with open(path, 'r') as f: first_line = f.readline() metadata = json.loads(first_line) df = pd.read_csv(path, header=1) # print(f'Loaded df and metadata from {path}.') except json.JSONDecodeError: df = pd.read_csv(path) metadata = {} print(f'Metadata not found. Loaded df from {path}.') return(df, metadata) def save_df_and_metadata(df, metadata, outpath): """Save a df as csv with all the important metadata saved as json comment in the first line of the file. """ if os.path.exists(outpath): os.remove(outpath) with open(outpath, 'a') as f: json.dump(metadata, f) f.write('\n') df.to_csv(f, index=False) return def regressor_from_pipeline(pipe): """Returns the ML model from a given sklearn Pipeline or TransformedTargetRegressor. """ return pipe.regressor_['model'] def get_saved_model(modelname: str, repetition: int, run_dir: str, regressor=False): """Finds a single saved model with given name and repetition in run_dir/models and returns it. """ model = Models().load(modelname=modelname, repetition=repetition, rundir=run_dir, regressor=regressor) return(model) def assert_allclose(x1, x2, atol=1e-5, rtol=1e-3): """Asserts that arrays x1 and x2 are either equal or at least close. """ try: assert np.allclose(x1, x2, rtol, atol) except TypeError: # If x1 and x2 are e.g. string arrays. assert all(x1 == x2) return def net_pattern(n_layers, base_size, end_size): """Calculates layer sizes for each layer from first and last layer so that the layer size continously increases/decreases. """ if n_layers != 1: factor = (end_size / base_size)**(1/(n_layers - 1)) else: factor = 1 layer_sizes = [int(round(base_size*factor**n)) for n in range(n_layers)] return(layer_sizes) def unique_sorted(groupcol): """Returns a list with the unique group label sorted by frequency. groupcol must be a pandas series.""" group_occ = groupcol.value_counts().reset_index() groups = group_occ['index'].tolist() return(groups) def print_row(values, name, uncertainties=None, width=7, num_width=3, dec=2, delim=' | '): """Prints a nicely formatted row. Name is the left-most entry. Values and uncertainties need to be iterables. """ values = np.asanyarray(values) if uncertainties != None: uncertainties = np.asanyarray(uncertainties) assert isinstance(values[0], float) or isinstance(values[0], int) num_width = 2 print_vals = [f'{val:^.{num_width}g}±{unc:^.{num_width}g}' for val, unc in zip(values, uncertainties)] print_vals = [f'{string:^{width}.{width}}' for string in print_vals] else: if isinstance(values[0], str): print_vals = [f'{val:^{width}.{width}}' for val in values] elif isfloat(values[0]): print_vals = [f'{val:^{width}.{num_width}g}' for val in values] cells = [f'{name:5.5}'] cells = cells + print_vals print(delim.join(cells)) print('-'*75) return() def Sc_classification(*Tc_arrays): """Returns the classification in 0 (non-sc) or 1 (sc) based on the continous value of Tc. """ all_sc_class_arrays = [] for Tc in Tc_arrays: Tc = np.asarray(Tc) assert all(Tc >= 0), 'We found a negative Tc!' # Reduce prediction of Tc to sc classes, i.e 0 or 1. Sc_class = np.where(Tc > 0, 1, 0) all_sc_class_arrays.append(Sc_class) return all_sc_class_arrays warnings.simplefilter("ignore", category=sklearn.exceptions.UndefinedMetricWarning) def logr2(y_true, y_pred, sample_weight): """Calculates the r2 score after taking the arcsinh (like the log).""" y_true = np.arcsinh(y_true) y_pred = np.arcsinh(y_pred) logr2 = sklearn.metrics.r2_score(y_true=y_true, y_pred=y_pred, sample_weight=sample_weight) return logr2 def Sc_F1(Tc_true, Tc_pred, sample_weight): """Calculates the F1 score for a superconductor (with regression of Tc), i.e. before calculating the usual F1 score it sets class=0 if Tc==0 and class=1 if Tc>0. """ sc_true_class, sc_pred_class = Sc_classification(Tc_true, Tc_pred) score = sklearn.metrics.f1_score(sc_true_class, sc_pred_class, sample_weight=sample_weight) return score def Sc_accuracy(Tc_true, Tc_pred, sample_weight): """Calculates the accuracy score for a superconductor (with regression of Tc), i.e. before calculating the usual accuracy score it sets class=0 if Tc==0 and class=1 if Tc>0. """ sc_true_class, sc_pred_class = Sc_classification(Tc_true, Tc_pred) score = sklearn.metrics.accuracy_score(sc_true_class, sc_pred_class, sample_weight=sample_weight) return score def Sc_precision(Tc_true, Tc_pred, sample_weight): """Calculates the precision score for a superconductor (with regression of Tc), i.e. before calculating the usual precision score it sets class=0 if Tc==0 and class=1 if Tc>0. """ sc_true_class, sc_pred_class = Sc_classification(Tc_true, Tc_pred) score = sklearn.metrics.precision_score(sc_true_class, sc_pred_class, sample_weight=sample_weight) return score def Sc_recall(Tc_true, Tc_pred, sample_weight): """Calculates the recall score for a superconductor (with regression of Tc), i.e. before calculating the usual recall score it sets class=0 if Tc==0 and class=1 if Tc>0. """ sc_true_class, sc_pred_class = Sc_classification(Tc_true, Tc_pred) score = sklearn.metrics.recall_score(sc_true_class, sc_pred_class, sample_weight=sample_weight) return score def specificity(y_true, y_pred, sample_weight): """Calculate the specificity (the recall of the negative class). """ # With sklearn. score_with_sklearn = sklearn.metrics.recall_score(y_true, y_pred, pos_label=0, sample_weight=sample_weight) # Own implementation. # Invert classes. y_true = np.abs(y_true - 1) y_pred = np.abs(y_pred - 1) # Calculate recall of inverted classes. score = sklearn.metrics.recall_score(y_true, y_pred, sample_weight=sample_weight) # Double checking both implementations. assert score == score_with_sklearn return score def Sc_specificity(Tc_true, Tc_pred, sample_weight): """Calculates the specificity score for a superconductor (with regression of Tc), i.e. before calculating the usual specificity score it sets class=0 if Tc==0 and class=1 if Tc>0. """ sc_true_class, sc_pred_class = Sc_classification(Tc_true, Tc_pred) score = specificity(sc_true_class, sc_pred_class, sample_weight=sample_weight) return score def Sc_G_means(Tc_true, Tc_pred, sample_weight): """Calculates the G-means score (geometric mean of Recall and specificity) for a superconductor (with regression of Tc), i.e. before calculating the usual specificity score it sets class=0 if Tc==0 and class=1 if Tc>0. """ recall = Sc_recall(Tc_true, Tc_pred, sample_weight=sample_weight) spec = Sc_specificity(Tc_true, Tc_pred, sample_weight=sample_weight) score = np.sqrt(recall*spec) return score def Sc_OoB(Tc_true, Tc_pred, sample_weight, bound_max=200): """Calculates what fraction of data points is Out of Bounds (OoB), i.e. how many are outliers. Doesn't need Tc_true, this is just for the sake of consistency. """ is_OoB = np.isclose(Tc_pred, bound_max, atol=1).astype(int) if not (sample_weight is None): is_OoB = sample_weight * is_OoB score = sum(is_OoB) / len(is_OoB) return score def Sc_MARE(Tc_true, Tc_pred, sample_weight, min_Tc=0): """Calculates the mean absolute relative error. Note, this definition is slightly different from the usual MAPE to accomodate for superconductors, e.g. it only looks at data points with a Tc higher than min_Tc. """ Tc_true, Tc_pred = np.asarray(Tc_true), np.asarray(Tc_pred) is_sc = (Tc_true > min_Tc) & (Tc_pred > min_Tc) Tc_true = Tc_true[is_sc] Tc_pred = Tc_pred[is_sc] if not (sample_weight is None): sample_weight = sample_weight[is_sc] norm = np.maximum(Tc_true, Tc_pred) diff = np.abs(Tc_true - Tc_pred) score = diff / norm if len(score) > 0: score = np.average(score, weights=sample_weight) else: warnings.warn('The MARE can not be calculated, setting to 0.') score = 0 return score def Sc_SMAPE(Tc_true, Tc_pred, sample_weight, min_Tc=0): """Calculates the mean absolute relative error. Note, this definition is slightly different from the usual MAPE to accomodate for superconductors, e.g. it only looks at data points with a Tc higher than min_Tc. """ Tc_true, Tc_pred = np.asarray(Tc_true), np.asarray(Tc_pred) is_sc = (Tc_true > min_Tc) & (Tc_pred > min_Tc) Tc_true = Tc_true[is_sc] Tc_pred = Tc_pred[is_sc] if not (sample_weight is None): sample_weight = sample_weight[is_sc] norm = (Tc_true + Tc_pred) diff = np.abs(Tc_true - Tc_pred) score = diff / norm if len(score) > 0: score = np.average(score, weights=sample_weight) else: warnings.warn('The SMAPE can not be calculated, setting to 0.') score = 0 return score def out_of_sigma(y_true, y_pred, sigma_lower_bound, sigma_upper_bound): """Calculates which fraction of data points is out of the lower and upper uncertainty bounds. """ y_true = np.asarray(y_true) y_pred = np.asarray(y_pred) sigma_lower_bound = np.asarray(sigma_lower_bound) sigma_upper_bound = np.asarray(sigma_upper_bound) assert all(y_pred >= sigma_lower_bound) and all(y_pred <= sigma_upper_bound), 'Predicions don\'t lie between sigma bounds.' # Calculate uncertainty for each data point. Use upper bound when overestimating and lower bound when underestimating. error = y_pred - y_true overestimated = error > 0 uncertainty_bounds = np.where(overestimated, sigma_lower_bound, sigma_upper_bound) uncertainty = np.abs(y_pred - uncertainty_bounds) error = np.abs(error) out_of_sigma = error > uncertainty # Fraction. result = sum(out_of_sigma) / len(out_of_sigma) return result def name_score_column(target, scorename, CV): """Returns the name of the columns of `All_scores.csv`. """ return f'{target}_{scorename}_{CV}' def get_scores(score, model, target, CV, all_scores_path): """Returns scores of specified model and target from the file `all_scores_path` which is an instance of `All_scores.csv`. """ df, _ = load_df_and_metadata(all_scores_path) # df = pd.read_csv(all_scores_path) df = df[df['Model'] == model] colname = name_score_column(target, score, CV) scores = df[colname] return(scores) def tolist(*variables): results = [] for var in variables: if isinstance(var, str) or isinstance(var, int) or isinstance(var, float): var = [var] results.append(var) return(results) def inverse_transform_std(mu, std, scaler): """Makes the inverse transform of the std by transforming upper and lower bound. Returns upper and lower bound after the inverse transform. """ lower_conf = mu - std upper_conf = mu + std lower_conf_trans = scaler.inverse_transform(lower_conf) upper_conf_trans = scaler.inverse_transform(upper_conf) return lower_conf_trans, upper_conf_trans # For file Domain_statistics.csv N_REPETITION = 'rand_instance' MODEL = 'Model' TEST_OR_TRAIN = 'test_or_train' SIZE = 'Size' # How many sigma uncertainty should be used for the uncertainty of the targets of data points returned from models (e.g. a Gaussian Process). This does NOT apply to the displayed uncertainty of metrics of models. Throughout all scripts one should refer to this variable. SIGMA = 2 class Machine_Learning(): """ This class contains functions to train with data on several models and automatically write the output to several files. """ def __init__(self, data, features, targets, domain=None, sample_weights=None, metrics_sample_weights=None, Column_Transformer=None, save_models=False, is_optimizing_NN=False, save_value_of_variables=[], NN_valfrac=0.2, print_features=True, print_targets=True, print_domain_score=False, random_seed=None, average_func=np.mean, save_all_values_and_predictions=True, save_torch_models=False, n_jobs=1, copy_files=[]): # For debugging, to be able to always get the same output. self.random_seed = random_seed # "None" for randomization. np.random.seed(self.random_seed) random.seed(self.random_seed) # torch.manual_seed(self.random_seed) tf.random.set_seed(self.random_seed) df_data = copy.deepcopy(data.reset_index(drop=True)) # Important for make_result_df() self.targets = targets self.df_data = df_data self.n_datapoints = len(self.df_data) self.features = features self.n_features = len(features) self.domain = domain self.sample_weights = sample_weights self.metrics_sample_weights = metrics_sample_weights assert all([feat in self.df_data.columns for feat in self.features]), "One or more of the given features does not exists in the data." assert all([target in self.df_data.columns for target in self.targets]), "One or more of the targets does not exist in the data." self.n_targets = len(targets) self.Column_Transformer = Column_Transformer if Column_Transformer != None else {col: StandardScaler() for col in features + targets} self.CV_cols = [col for col in self.df_data.columns if col.startswith('CV_')] self.n_repetitions = len(self.CV_cols) self.all_n_repetitions = list(range(self.n_repetitions)) assert self.CV_cols == [All_Data.All_Data.name_CV_col(i) for i in self.all_n_repetitions], 'Could not recognize CV columns with train and test indices in df_data.' self.train_lengths = [sum(self.df_data[col] == 'train') for col in self.CV_cols] if len(set(self.train_lengths)) == 1: self.train_lengths = self.train_lengths[0] self.test_lengths = [sum(self.df_data[col] == 'test') for col in self.CV_cols] if len(set(self.test_lengths)) == 1: self.test_lengths = self.test_lengths[0] # Check for NaN features nan_feature_cols = [col for col in self.features if self.df_data[col].isna().any()] if nan_feature_cols: raise ValueError(f'Error: Some of the values of the following features are NaN: {",".join(nan_feature_cols)}') self.save_models = save_models self.average_func = average_func self.NN_n_reps = 2 self.NN_valfrac = NN_valfrac self.is_optimizing_NN = is_optimizing_NN self.save_all_values_and_predictions = save_all_values_and_predictions self.save_torch_models = save_torch_models self.experiment = save_value_of_variables['experiment'] self.n_jobs = n_jobs self.copy_files = copy_files self.all_scores = {"r2": sklearn.metrics.r2_score, "logr2": logr2, "MAE": sklearn.metrics.mean_absolute_error, # "MdAE": sklearn.metrics.median_absolute_error, "MSLE": sklearn.metrics.mean_squared_log_error, # "MARE": Sc_MARE, "SMAPE": Sc_SMAPE, # "OoB": Sc_OoB, # "ScSpec": Sc_specificity, # "ScRecall": Sc_recall, # "ScG": Sc_G_means, # "ScAcc": Sc_accuracy, # "ScPrec": Sc_precision } self.all_scorenames = list(self.all_scores.keys()) if self.is_optimizing_NN == True: self.save_models = False # Write these variables to the final output file for later convenience. self.store_variables = {} self.store_variables["Length of train/ test set"] = (self.train_lengths, self.test_lengths) self.store_variables["Features"] = self.features self.store_variables["Targets"] = self.targets self.store_variables["Save models"] = self.save_models self.store_variables["Random seed"] = self.random_seed self.store_variables["Average function"] = self.average_func self.store_variables["Repetitions per model"] = self.n_repetitions self.store_variables["Repetitions for NN"] = self.NN_n_reps self.store_variables["Sigma (Uncertainty)"] = SIGMA for varname, value in save_value_of_variables.items(): self.store_variables[varname] = value # Initiate variables for later data storage. self.df_all_score_results = pd.DataFrame() self.all_values_and_predictions = self.df_data self.all_domain_stats = pd.DataFrame() self.all_loss_curves = {} # Set internal variables. self.num_features = len(self.features) self.print_features = print_features self.print_targets = print_targets self.print_domain_score = print_domain_score def get_train_and_test_data(self, CV_col, domain_col=None, sample_weights_col=None, metrics_sample_weights_col=None): """Return all train and test data.""" df = self.df_data x = df[self.features].to_numpy() y = df[self.targets].to_numpy() test = df[CV_col] == 'test' train = df[CV_col] == 'train' x_train, x_test = x[train], x[test] y_train, y_test = y[train], y[test] if domain_col != None: d_train = df[train][domain_col].to_numpy() d_test = df[test][domain_col].to_numpy() else: d_train, d_test = np.zeros(sum(train)), np.zeros(sum(test)) if sample_weights_col != None: w_train = df[train][sample_weights_col].to_numpy() w_test = df[test][sample_weights_col].to_numpy() else: w_train, w_test = None, None if metrics_sample_weights_col != None: mw_train = df[train][metrics_sample_weights_col].to_numpy() mw_test = df[test][metrics_sample_weights_col].to_numpy() else: mw_train, mw_test = None, None return(x, y, x_train, x_test, y_train, y_test, d_train, d_test, w_train, w_test, mw_train, mw_test) def all_score_results(self, all_train_scores, all_test_scores, modelname): """ Write all scores to dataframe. """ df_scores = pd.DataFrame() df_scores['Repetition'] = list(range(self.n_repetitions)) for target_idx, target_name in enumerate(self.targets): for score_idx, score_name in enumerate(self.all_scores.keys()): train_scorename = name_score_column(target_name, score_name, 'train') test_scorename = name_score_column(target_name, score_name, 'test') df_scores[train_scorename] = all_train_scores[:,score_idx, target_idx] df_scores[test_scorename] = all_test_scores[:,score_idx, target_idx] df_scores["Model"] = modelname return(df_scores) def unique_transformer(self): """Returns a unique transformer for all targets because sklearn doesn\'t suppourt multiple target scalers. """ all_target_scalers = [self.Column_Transformer[target] for target in self.targets] # Check for uniqueness for t1, t2 in itertools.product(all_target_scalers, all_target_scalers): if not type(t1) == type(t2): raise Warning('Sklearn doesn\'t support different scaling of differet targets.') # Get unique target scaler. target_scaler = all_target_scalers[0] return target_scaler def train_regressor(self, model, x_train, y_train, d_train, w_train): """Train a regressor of a model with data and return the regressor. """ if self.sample_weights != None: x_train, y_train, d_train, w_train = sklearn.utils.shuffle(x_train, y_train, d_train, w_train) else: x_train, y_train, d_train = sklearn.utils.shuffle(x_train, y_train, d_train) # Transform and scale features and targets in a pipeline. transform_columns = [] for idx, colname in enumerate(self.features): transformer = self.Column_Transformer[colname] entry = (colname, transformer, [idx]) transform_columns.append(entry) feature_transformer = ColumnTransformer(transform_columns) pipe = Pipeline([ ('ColumnTransformer', feature_transformer), ('model', model) ]) target_transformer = self.unique_transformer() regr = CustomTransformedTargetRegressor(regressor=pipe, transformer=target_transformer) # Fit with domain and sample weights if possible. train_kwargs = {} if hasattr(model, 'domain_col'): print('Fitting model with specified domains.') train_kwargs['model__d_train'] = d_train try: regr.fit(x_train, y_train, model__sample_weight=w_train, **train_kwargs) if self.sample_weights != None: print('Fitted model with sample weights.') except TypeError: if self.sample_weights != None: print('Model doesn\'t support sample weights.') regr.fit(x_train, y_train, **train_kwargs) return(regr) def apply_model(self, modelname, init_model, outdir): """Applies the model to the data to get the r² and MAE score and writes them into the numerical output file. """ # For debugging, to be able to always get the same output. # TODO: Remove this shit. # np.random.seed(self.random_seed) # random.seed(self.random_seed) # torch.manual_seed(self.random_seed) print(f"\n ### {modelname}:") n_scores = len(self.all_scores) all_train_scores = np.zeros((self.n_repetitions, n_scores, self.n_targets)) all_test_scores = np.zeros((self.n_repetitions, n_scores, self.n_targets)) self.all_loss_curves[modelname] = [] for i in range(self.n_repetitions): # Very important to have a new model for each run! Otherwise they might influence each other. # try: model = copy.deepcopy(init_model) # except TypeError: # print('CAN\'T COPY MODEL!!!') # model = init_model # Setup SummaryWriter for Tensorboard. # TODO # ============================================================================= # if hasattr(model, 'use_tensorboard') and model.use_tensorboard: # tb_dir = os.path.join(outdir, '../0tensorboard', os.path.basename(outdir) + f'_{modelname}_{i}') # model.writer = SummaryWriter(tb_dir) # model.outpath = get_modelpath(outdir, modelname, i) + '.pt' # ============================================================================= # Add NN_path to GP to train on transformed features of model `NN_path`. if hasattr(model, 'NN_path') and model.NN_path is not None: NN_path = model.NN_path if not os.path.exists(NN_path) and NN_path in self.all_models: NN_path = get_modelpath(outdir, NN_path, i) + '.pkl' model.NN_path = NN_path # Get train and test data. try: domain_col = model.domain_col assert domain_col == self.domain except AttributeError: domain_col = None CV_col = All_Data.All_Data.name_CV_col(i) x, y, x_train, x_test, y_train, y_test, d_train, d_test, w_train, w_test, mw_train, mw_test = self.get_train_and_test_data(CV_col=CV_col, domain_col=domain_col, sample_weights_col=self.sample_weights, metrics_sample_weights_col=self.metrics_sample_weights) # Train regressor in pipeline. regr = self.train_regressor(model, x_train, y_train, d_train, w_train) # Get predictions for scoring. y_pred_train = regr.predict(x_train) y_pred_test = regr.predict(x_test) assert y_train.shape == y_pred_train.shape assert y_test.shape == y_pred_test.shape # train/test_scores have shape (n_scores, n_targets) where n_scores is the number of score functions. train_scores = self.scores(y_train, y_pred_train, weights=mw_train) test_scores = self.scores(y_test, y_pred_test, weights=mw_test) all_train_scores[i,:,:] = train_scores all_test_scores[i,:,:] = test_scores # TODO: Make logging with tensorboard or something else. # ============================================================================= # if hasattr(model, 'use_tensorboard'): # r2_idx = list(self.all_scores.keys()).index('r2') # # add_hparams only takes these data types. # hparams = {key: val if type(val) in [int, float, str, bool, torch.tensor] else str(val) for key, val in model.input_args.items()} # model.writer.add_hparams( # hparams, # {'r2_train': train_scores[r2_idx], # 'r2_test': test_scores[r2_idx] # } # ) # model.writer.flush() # # Pickle can't deal with tensorboard logger. # delattr(model, 'writer') # ============================================================================= if self.save_models == True: Models().save(regr=regr, rundir=outdir, modelname=modelname, repetition=i) # Construct dataframe to save all true and predicted data together. if not self.is_optimizing_NN: # Get predictions and uncertainty if possible. # The std is a tuple of upper and lower bound because it will usually not be symmetrical due to the scaling. # The here called output `y_pred_std` is actually SIGMA * std (SIGMA is a global variable). This is so that one can dynamically change which degreee of uncertainty one wants. This is implemented in CustomTransformedTargetRegressor. try: y_pred, y_pred_std = regr.predict(x, return_std=True) scaled_unc, y_pred_std_lower, y_pred_std_upper = y_pred_std assert (y_pred >= y_pred_std_lower).all() and (y_pred <= y_pred_std_upper).all(), 'Prediction not between uncertainty bounds.' except TypeError: y_pred = regr.predict(x) for idx in range(self.n_targets): target = self.targets[idx] preds = y_pred[:,idx] colname = All_Data.All_Data.name_preds_col(modelname, i, target) self.all_values_and_predictions[colname] = preds # Add uncertainty to df if it exists. try: scaled_unc = scaled_unc[:,idx] std_lower = y_pred_std_lower[:,idx] std_upper = y_pred_std_upper[:,idx] std_lower_colname = All_Data.All_Data.name_unc_col(modelname, i, target, kind='lower') std_upper_colname = All_Data.All_Data.name_unc_col(modelname, i, target, kind='upper') scaled_unc_colname = All_Data.All_Data.name_unc_col(modelname, i, target, kind='scaled_unc') self.all_values_and_predictions[std_lower_colname] = std_lower self.all_values_and_predictions[std_upper_colname] = std_upper self.all_values_and_predictions[scaled_unc_colname] = scaled_unc except UnboundLocalError: pass # Some sanity tests due to paranoia. assert_allclose(x, self.all_values_and_predictions[self.features].to_numpy()) assert_allclose(x_train, x[self.all_values_and_predictions[CV_col] == 'train']) assert_allclose(y_pred_test, y_pred[self.all_values_and_predictions[CV_col] == 'test']) if not self.is_optimizing_NN: df_all_scores = self.all_score_results(all_train_scores, all_test_scores, modelname) self.df_all_score_results = self.df_all_score_results.append(df_all_scores) # Calculate averages and standard deviations over all repetitions. train_score = self.average_func(all_train_scores, axis=0) test_score = self.average_func(all_test_scores, axis=0) train_score_std = np.std(all_train_scores, axis=0) test_score_std = np.std(all_test_scores, axis=0) self.print_overall_performance_scores(train_score, train_score_std, test_score, test_score_std) if self.domain != None and self.print_domain_score: # Get scores and statistics for each domain df_domain_stats = self.get_domain_stats(modelname, self.domain) self.all_domain_stats = self.all_domain_stats.append(df_domain_stats) self.print_domain_stats(df_domain_stats, self.domain) # Write score output to file self.save_numerical_output(outdir, modelname, train_score, test_score, train_score_std, test_score_std) return def train(self, all_models, outdir): print(" ### Start training") self.all_models = all_models # Make directories for the output. outdir = self.prepare_directories(outdir) # Copy specified files to outdir. for file in self.copy_files: file_in_new_dir = os.path.join(outdir, file) copyfile(file, file_in_new_dir) # Store all arguments. with open(outdir + "/arguments", 'w') as f: yaml.dump(self.store_variables, f) # Print input information. print(" --- Data:") if self.print_features: print(" Features: {}".format(self.features)) if self.print_targets: print(" Targets: {}".format(self.targets)) print(" Num features: {}".format(len(self.features))) print(" Train data size: {}".format(self.train_lengths)) print(" Test data size: {}".format(self.test_lengths)) print(" Num repetitions: {}".format(self.n_repetitions)) # ============================================================================= # Train and evaluate all given models. # ============================================================================= for modelname, model in all_models.items(): starttime = datetime.datetime.now() self.apply_model(modelname, model, outdir) duration = datetime.datetime.now() - starttime print(f'Training time of {self.n_repetitions} instances of {modelname}: {duration}') # Save data if not self.is_optimizing_NN: all_models_names = list(all_models.keys()) if self.save_all_values_and_predictions: All_Data_file = os.path.join(outdir, "All_values_and_predictions.csv") All_Data.save_All_Data(df=self.all_values_and_predictions, outpath=All_Data_file, targets=self.targets, n_repetitions=self.n_repetitions, features=self.features, domains=self.domain, models=all_models_names, sample_weights=self.sample_weights, SIGMA=SIGMA, ) All_scores_file = os.path.join(outdir, "All_scores.csv") All_Scores.save_All_scores(df=self.df_all_score_results, outpath=All_scores_file, targets=self.targets, scorenames=self.all_scorenames, CVs=['train', 'test'], models=all_models_names ) if len(self.all_domain_stats) > 0: Domain_statistics_file = os.path.join(outdir, "Domain_statistics.csv") save_Domain_statistics(df=self.all_domain_stats, outpath=Domain_statistics_file, domains=self.domain, targets=self.targets, scorenames=self.all_scorenames, models=all_models_names ) try: # Try to get all feature importances if self.save_models: importances = FI.feature_importances_from_models( rundir=outdir, features=self.features, modelnames=all_models_names, repetitions=self.all_n_repetitions ) # Save feature importances feat_imps_file = os.path.join(outdir, FI.data_name) FI.save_Feature_Importances( df=importances, outpath=feat_imps_file, features=self.features, modelnames=all_models_names, repetitions=self.all_n_repetitions ) else: print('Could not try to get feature importances because no model saved.') except AttributeError: pass # No models with feature importances found. print("Successfully saved all data.") return() def prepare_directories(self, outdir): """Make directories for the output.""" if not os.path.exists(outdir): raise ValueError('outdir doesn\'t exist!') if (not os.path.exists("%s/models"%(outdir))) and self.save_models: os.makedirs("%s/models"%(outdir)) return(outdir) def print_overall_performance_scores(self, train_scores, train_scores_std, test_scores, test_scores_std): """Prints overall_scores in a nice format. The uncertainty is noted so that it is ± half of the std. """ for target_idx, target_name in enumerate(self.targets): print(f' --- Target: {target_name}') for score_idx, score_name in enumerate(self.all_scores.keys()): train_score = train_scores[score_idx][target_idx] train_score_std = train_scores_std[score_idx][target_idx] test_score = test_scores[score_idx][target_idx] test_score_std = test_scores_std[score_idx][target_idx] print(f' --- {score_name}:\tTraining: {train_score:.3f} ± {train_score_std:.3f}\t\tTesting: {test_score:.3f} ± {test_score_std:.3f}') return def scores(self, all_y_true, all_y_pred, weights=None): """Returns an array with shape (num_scores, num_targets) with the scores calculated with the functions in all_scores for every target in the second dimension of all_y_true. """ assert all_y_true.ndim == 2 all_y_true = np.array(all_y_true) all_y_pred = np.array(all_y_pred) num_targets = all_y_true.shape[1] num_data = all_y_true.shape[0] if not (weights is None): assert len(weights) == num_data num_scores = len(self.all_scores) # Calculate scores scores = np.zeros((num_scores, num_targets)) for score_idx, name in enumerate(self.all_scores.keys()): func = self.all_scores[name] for target_idx, (y_true, y_pred) in enumerate(zip(all_y_true.T, all_y_pred.T)): score = func(y_true, y_pred, sample_weight=weights) scores[score_idx, target_idx] = score return np.array(scores) def result_statistics(self, y_true, y_pred, targetname, weights=None): """Calculates several statistical properties of y_true and y_pred. """ y_true, y_pred = np.asanyarray(y_true), np.asanyarray(y_pred) if not (weights is None): assert len(y_true) == len(y_pred) and len(y_pred) == len(weights) stats = {} for score_name, score_fn in self.all_scores.items(): score = score_fn(y_true, y_pred, sample_weight=weights) stats[f'{score_name}_{targetname}'] = score stats[f'mean_{targetname}_true'] = y_true.mean() stats[f'std_{targetname}_true'] = y_true.std() stats[f'mean_{targetname}_pred'] = y_pred.mean() stats[f'std_{targetname}_pred'] = y_pred.std() return(stats) def get_domain_stats(self, modelname, domain_col): """Returns a df with the scores for each domain. """ domains = pd.unique(self.all_values_and_predictions[domain_col]) stats = {} counter = 0 for n, domain in itertools.product(self.all_n_repetitions, domains): is_domain = self.all_values_and_predictions[domain_col] == domain df_domain = self.all_values_and_predictions[is_domain] CV_col = All_Data.All_Data.name_CV_col(n) CVs = sorted(pd.unique(df_domain[CV_col])) for CV in CVs: df = df_domain[df_domain[CV_col] == CV] # Append descriptive data for this row. stats[counter] = { N_REPETITION: n, MODEL: modelname, TEST_OR_TRAIN: CV, domain_col: domain } # Calculate statistics for each target. for target in self.targets: col = All_Data.All_Data.name_preds_col(modelname, n, target) y_pred = df[col] y_true = df[target] weights = None if self.metrics_sample_weights is None else df[self.metrics_sample_weights] interesting_stats = self.result_statistics(y_true, y_pred, target, weights=weights) stats[counter].update(interesting_stats) stats[counter].update({SIZE: len(df)}) counter += 1 df_domain_stats = pd.DataFrame.from_dict(data=stats, orient='index') return(df_domain_stats) def print_domain_stats(self, df_domain_stats, domain_col): """Prints test scores and other stats per domain. """ # Print only test results otherwise it would be too much. df_domain_stats = df_domain_stats[df_domain_stats['test_or_train'] == 'test'] # Average over multiple runs to have the final test score of each domain. df_domain_stats = df_domain_stats.groupby([MODEL, domain_col]).mean().reset_index() # Print domains with many datapoints first. df_domain_stats = df_domain_stats.sort_values(by='Size', ascending=False) domains = df_domain_stats[domain_col].tolist() print('\n') print_row(domains, 'Domain') # Print scores of targets. for score in self.all_scores.keys(): for target in self.targets: score_name = f'{score}_{target}' vals = df_domain_stats[score_name] print_row(vals, score_name) # Print statistics of targets. for target in self.targets: for end in ['true', 'pred']: name = f'{target}_{end}' vals = df_domain_stats[f'mean_{target}_{end}'].tolist() stds = df_domain_stats[f'std_{target}_{end}'].tolist() print_row(vals, name, stds) def save_numerical_output(self, outdir, modelname, train_scores, test_scores, train_scores_std, test_scores_std): """Write formatted output scores to file. """ with open(outdir + "/Numerical results", "a+") as f: f.write("\n ### {}:\n".format(modelname)) with redirect_stdout(f): self.print_overall_performance_scores(train_scores, train_scores_std, test_scores, test_scores_std) f.close() return() def plot_feature_correlations(self, outdir, x, y, x_scaler, y_scaler): """Plot some feature correlations.""" print("Plotting feature correlations...") if not os.path.exists("%s/feature_correlation"%(outdir)): os.makedirs("%s/feature_correlation"%(outdir)) for idx in range(len(self.features)): #print("do best feature %i"%(idx)) plt.figure() feature_name=self.features[idx] if "pH" not in feature_name and "cross" not in feature_name: plt.scatter(x_scaler.inverse_transform(x)[:,idx],y_scaler.inverse_transform(y)) else: xs_unscaled_here=x_scaler.inverse_transform(x)[:,idx] xs_set=[xs_unscaled_here[0]] for value in xs_unscaled_here: distances=[abs(value-v2) for v2 in xs_set] if max(distances)>0.01: xs_set.append(value) indeces=[] for v2 in xs_set: indeces.append([]) for v_index, value in enumerate(xs_unscaled_here): d=abs(value-v2) if d<0.01: indeces[-1].append(v_index) #x_min=np.min(x[:,idx]) #x_max=np.max(x[:,idx]) #x_mean=0.5*(x_min+x_max) #indeces_low=np.where(x[:,idx]<x_mean)[0] #indeces_high=np.where(x[:,idx]>x_mean)[0] for counter,indeces_here in enumerate(indeces): parts=plt.violinplot([y_scaler.inverse_transform(y)[indeces_here].T[0]], positions=[float(xs_set[counter])], vert=True, widths=0.18, showmeans=False, showextrema=True, showmedians=False) for pcidx, pc in enumerate(parts['bodies']): pc.set_facecolor("C%i"%(counter)) pc.set_edgecolor('k') pc.set_alpha(0.7) parts['cbars'].set_color("k") parts['cmaxes'].set_color("k") parts['cmins'].set_color("k") #parts=plt.violinplot([y_scaler.inverse_transform(y)[indeces_high].T[0]], positions=[1.0], vert=True, widths=0.18, showmeans=False, showextrema=True, showmedians=False) #for pcidx, pc in enumerate(parts['bodies']): # pc.set_facecolor("C2") # pc.set_edgecolor('k') # pc.set_alpha(0.7) #parts['cbars'].set_color("k") #parts['cmaxes'].set_color("k") #parts['cmins'].set_color("k") #if "pH" in feature_name: # plt.xticks(xs_set, ["pH=1","pH=7"]) # #plt.xlim([-0.7,1.7]) #num_off=len(indeces_low) #num_on=len(indeces_high) #plt.text(0.2,14.0,"%i"%(num_off)) #plt.text(1.2,14.0,"%i"%(num_on)) # r2 = sklearn.metrics.r2_score(y,x[:,idx]) #plt.ylim([7.0,13.0]) #plt.xlabel("%s (r$^2$ = %.3f)"%(feature_name, r2), fontname=fontname) plt.xlabel("%s"%(feature_name)) plt.ylabel("Radius difference [px]") plt.savefig("%s/feature_correlation/best_20_%s.png"%(outdir, feature_name), dpi=300) plt.close() return() def save_scatter_plots(self, model_short, y_train, y_test, y_pred_train, y_pred_test, r2_train, r2_test, outdir, colors, index=None): """Make scatter plots of measured vs predicted data.""" plt.figure() plt.scatter(y_train, y_pred_train, marker="o", c=colors[0], label="Training: r$^2$ = %.3f"%(r2_train)) plt.scatter(y_test, y_pred_test, marker="o", c=colors[1], label="Testing: r$^2$ = %.3f"%(r2_test)) plt.xlabel("Measured normalized radius (%)") plt.ylabel("Predicted normalized radius (%)") plt.legend(loc="upper left") if index != None: savepath = "%s/scatter_plots/full_data_%s_%s.png"%(outdir, model_short, index) else: savepath = "%s/scatter_plots/full_data_%s.png"%(outdir, model_short) plt.savefig(savepath,dpi=300) plt.close() return() def make_result_df(self, train, test, y_pred_train, y_pred_test, modelname, i, x_train, y_train, x_test, y_test): """Returns a dataframe with all the initial data points and the predictions on them. """ # Names of additional columns. test_train_colname = 'test_or_train' model_colname = 'Model' rand_instance_colname = 'rand_instance' # Get data and assert that it is in the correct order. df = self.df_data train_index = df.index[train] test_index = df.index[test] train_entries = df.index.isin(train_index) test_entries = df.index.isin(test_index) assert np.allclose(df[train_entries][self.features].to_numpy(), x_train) assert np.allclose(df[test_entries][self.features].to_numpy(), x_test) assert np.allclose(df[train_entries][self.targets].to_numpy(), y_train) assert np.allclose(df[test_entries][self.targets].to_numpy(), y_test) # Get true and predicted target names and rename df accordingly. pred_target_names = [] true_target_names = [] for target in self.targets: df = df.rename(columns={target: target+'_true'}) true_target_names.append(target+'_true') pred_target_names.append(target+'_pred') # Build dataframes of predictions of test and train data. train_pred = pd.DataFrame( data=y_pred_train, index=train_index, columns=pred_target_names ) train_pred[test_train_colname] = 'train' test_pred = pd.DataFrame( data=y_pred_test, index=test_index, columns=pred_target_names ) test_pred[test_train_colname] = 'test' # Merge predictions and initial data. assert df.equals(df.reset_index(drop=True)) predictions = train_pred.append(test_pred) df = df.join(predictions) df = movecol(df, cols=pred_target_names, to=true_target_names[-1]) # Add name of used model and random instance. df[model_colname] = modelname df[rand_instance_colname] = i new_cols = [rand_instance_colname, model_colname, test_train_colname] column_order = new_cols + [col for col in df.columns.tolist() if not col in new_cols] df = df[column_order] return(df) def custom_median(self, array): """Calculates the median of an array and make sure to return an element of the array even if it has even length. """ sorted_array = np.sort(array) length = len(array) if length%2 == 1: idx = int(length/2 - 0.5) else: average = array.mean() possible_indices = [int(length/2), int(length/2-1)] if abs(sorted_array[possible_indices[0]] - average) <= abs(sorted_array[possible_indices[1]] - average): idx = possible_indices[0] else: idx = possible_indices[1] median = sorted_array[idx] return(median) def standard_models(self, hparams, n_features, n_targets, n_domains): """Definitions of models that I regularly use. Rather used as a library to quickly look up and copy paste models than as a function. """ pass # ============================================================================= # ################### # # LINEAR REGRESSION # ################### # Linear_Regression = sklearn.linear_model.LinearRegression() # # ################ # # Neural Network # ################ # # Set some hyperparameter variables for the NN. # net_dims = net_pattern( # hparams['nn_layers'], # hparams['nn_base_dim'], # hparams['nn_end_size'] # ) # Neural_Network = MLPRegressor( # hidden_layer_sizes=net_dims, # activation=hparams['nn_act'], # solver="adam", # max_iter=hparams["n_epochs"], # early_stopping=False, # # validation_fraction=0.1, # alpha=0,#hparams["nn_l2"], # # momentum=0, # # learning_rate="invscaling", # learning_rate="constant", # batch_size=hparams['nn_batch_size'], # learning_rate_init=hparams["learning_rate"], # n_iter_no_change=9999#hparams["nn_patience"] # ) # # # ############### # # Random Forest # ############### # n_trees = hparams["RF_n_estimators"] # Random_Forest = RandomForestRegressor(n_estimators=n_trees) # # # ############################ # # Gradient Boosting # ############################ # n_trees = hparams["GB_n_estimators"] # Gradient_Boosting = GradientBoostingRegressor(n_estimators=n_trees) # # # ############################ # # Gaussian Process # ############################ # kernel_scale = np.ones(n_features) # noise = hparams["GP_alpha"] # Gaussian_Process = GaussianProcessRegressor(kernel=RBF_sk(length_scale=kernel_scale), alpha=noise) # # # ############################ # # Regret Minimization Network with sklearn # ############################ # input_layer_size = n_features # net_dims = net_pattern( # hparams['nn_layers'], # hparams['nn_base_dim'], # hparams['nn_end_size'] # ) # module = RGM_sk(input_layer_size, net_dims, n_targets) # RGM_sk_model = NeuralNetRegressor( # module, # optimizer=torch.optim.Adam, # lr=hparams["learning_rate"], # max_epochs=hparams["n_epochs"], # batch_size=hparams['nn_batch_size'], # train_split=None, # callbacks='disable' # ) # # # def get_featurizer(input_layer_size, hidden_layer_sizes, activation): # """Returns the first part of the RGM, the featurizer or representation NN. # """ # layers = [] # for i in range(len(hidden_layer_sizes)): # if i == 0: # in_size = input_layer_size # else: # in_size = hidden_layer_sizes[i-1] # out_size = hidden_layer_sizes[i] # layers.append(nn.Linear(in_size, out_size)) # if activation == 'relu': # activation_fn = nn.ReLU() # else: # raise ValueError('Activation function not known.') # layers.append(activation_fn) # layers = tuple(layers) # featurizer = nn.Sequential(*layers) # return(featurizer) # # ############################ # # Regret Minimization Network # ############################ # input_layer_size = n_features # net_dims = net_pattern( # hparams['nn_layers'], # hparams['nn_base_dim'], # hparams['nn_end_size'] # ) # featurizer = get_featurizer(input_layer_size, net_dims, 'relu') # # # Set options for RGM. # RGM_args = namedtuple('RGM_args', ['linear', 'hidden_size', 'output_size', 'num_domains', 'rgm_e', 'erm_e', 'holdout_e', 'detach_classifier', 'oracle']) # RGM_args.linear = True # RGM_args.hidden_size = net_dims[-1] # RGM_args.output_size = n_targets # RGM_args.num_domains = n_domains # RGM_args.rgm_e = 1 # RGM_args.erm_e = 1 # RGM_args.holdout_e = 1 # RGM_args.detach_classifier = True # RGM_args.oracle = False # # RGM = RGM_sklearn(hidden_layer_sizes=net_dims, # activation='relu', # solver='adam', # max_iter=hparams["n_epochs"], # batch_size=hparams['nn_batch_size'], # learning_rate_init=hparams["learning_rate"], # featurizer=featurizer, # batch_mode='Conserve_ratio', # RGM_args=RGM_args # ) # # all_models = { # "Linear_Regression": Linear_Regression, # "Neural_Network": Neural_Network, # "Random_Forest": Random_Forest, # "Gradient_Boosting": Gradient_Boosting, # "Gaussian_Process": Gaussian_Process, # "RGM_sk": RGM_sk_model, # "RGM": RGM # } # return(all_models) # ============================================================================= class CustomTransformedTargetRegressor(TransformedTargetRegressor): def predict(self, X, **predict_params): """Predict using the base regressor, applying inverse. The regressor is used to predict and the ``inverse_func`` or ``inverse_transform`` is applied before returning the prediction. Parameters """ check_is_fitted(self) pred = self.regressor_.predict(X, **predict_params) returns_std = 'return_std' in predict_params.keys() and predict_params['return_std'] == True if returns_std: pred, std = pred # Use the specified degree of uncertainty throughout all scripts. std = std * SIGMA assert pred.shape == std.shape, 'Maybe you are trying to fit a GP to several targets or so? This doesn\'t work. Anyway, there\'s a shape mismatch.' if pred.ndim == 1: pred = pred.reshape(-1, 1) pred_trans = self.transformer_.inverse_transform(pred) if returns_std: if std.ndim == 1: std = std.reshape(-1, 1) lower_conf_trans, upper_conf_trans = inverse_transform_std(pred, std, self.transformer_) # Squeeze if prediction will be squeezed as well. if ( self._training_dim == 1 and lower_conf_trans.ndim == 2 and lower_conf_trans.shape[1] == 1 ): lower_conf_trans = lower_conf_trans.squeeze(axis=1) upper_conf_trans = upper_conf_trans.squeeze(axis=1) if ( self._training_dim == 1 and pred_trans.ndim == 2 and pred_trans.shape[1] == 1 ): pred_trans = pred_trans.squeeze(axis=1) if returns_std: assert len(set([pred_trans.shape, std.shape, lower_conf_trans.shape, upper_conf_trans.shape])) == 1, 'Prediction and uncertainty have different shapes.' return pred_trans, (std, lower_conf_trans, upper_conf_trans) else: return pred_trans # Ideas for refactoring: # Instead of giving n_repetitions, directly give a list with the CV column so that for repeatibility we can have CV columns for several things in the dataset. Write the columns to metadata of All_Data. # Make if possible to have different features/ targets per model? # Make it possible to calculate scores more flexible and calculate only some scores for each target (and maybe model?). # Train loop over all given models: # Function: Train all repetitions, save models. # Next functions: Either take passed on models (in case model can't be saved like tensorflow model) or read in saved models (for later analysis or if memory is not big enough) # Function: Calculate df All_Data. # Function: Calculate all train and test scores from All_Data. # For plotting function also either pass on all models or read in saved models. # cd into the correct directory at the begginning of Apply_ML_models and then cd into previos directory again. # In Apply_ML_models add single flag to debug and then set settings accordingly.
61,052
46.697656
418
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Algorithms/RGM_Jin_original.py
import torch import torch.nn as nn import torch.nn.functional as F import math import numpy as np import random class GradientReversal(torch.autograd.Function): beta = 1. @staticmethod def forward(self, x): return x.view_as(x) @staticmethod def backward(self, grad_output): return -GradientReversal.beta * grad_output class RGM(nn.Module): def __init__(self, featurizer, linear, hidden_size, output_size, num_domains, rgm_e, erm_e, holdout_e, detach_classifier, oracle, loss_forward, num_train_domains): # Custom: Replaced args.(...) by the regarding variable. super(RGM, self).__init__() self.featurizer = featurizer if linear: self.classifier = nn.Linear(hidden_size, output_size) self.copy_f_k = nn.ModuleList( [nn.Linear(hidden_size, output_size).requires_grad_(False) for _ in range(num_domains)] ) self.f_k = nn.ModuleList( [nn.Linear(hidden_size, output_size) for _ in range(num_domains)] ) self.g_k = nn.ModuleList( [nn.Linear(hidden_size, output_size) for _ in range(num_domains)] ) else: self.classifier = nn.Sequential( nn.Linear(hidden_size, hidden_size), nn.ReLU(), nn.Linear(hidden_size, output_size), ) self.copy_f_k = nn.ModuleList([ nn.Sequential( nn.Linear(hidden_size, hidden_size), nn.ReLU(), nn.Linear(hidden_size, output_size), ).requires_grad_(False) for _ in range(num_domains) ]) self.f_k = nn.ModuleList([ nn.Sequential( nn.Linear(hidden_size, hidden_size), nn.ReLU(), nn.Linear(hidden_size, output_size), ) for _ in range(num_domains) ]) self.g_k = nn.ModuleList([ nn.Sequential( nn.Linear(hidden_size, hidden_size), nn.ReLU(), nn.Linear(hidden_size, output_size), ) for _ in range(num_domains) ]) self.network = nn.Sequential(self.featurizer, self.classifier) self.loss_forward = loss_forward self.rgm_e = rgm_e self.num_domains = num_domains self.register_buffer('update_count', torch.tensor([0])) self.loss_curve_iter = {} # Custom self.train_representation = True # Custom self.train_classifier = True # Custom # def loss_forward(self, preds, batch_y, mask): # Custom # pred_loss = self.loss_func(preds, batch_y) * mask # return pred_loss.sum() / mask.sum() def forward(self, batches): assert len(batches) == 2 for k in range(self.num_domains): self.copy_f_k[k].load_state_dict(self.f_k[k].state_dict()) erm_loss = 0 all_phis = [] for batch_x, batch_y, batch_d, mask in batches: phi_x = self.featurizer(batch_x) all_phis.append(phi_x) preds = self.classifier(phi_x) erm_loss = erm_loss + self.loss_forward(preds, batch_y, mask) regret = 0 for k in range(self.num_domains): _, batch_y, _, mask = batches[k] phi_k = all_phis[k] preds = self.copy_f_k[k](phi_k) oracle_preds = self.g_k[k](GradientReversal.apply(phi_k)) regret = regret + self.loss_forward(preds, batch_y, mask) + self.loss_forward(oracle_preds, batch_y, mask) holdout_loss = 0 for k in range(self.num_domains): _, batch_y, _, mask = batches[1 - k] # hardcode: 2 domains phi_x = all_phis[1 - k].detach() # phi does not help f_{-e} preds = self.f_k[k](phi_x) holdout_loss = holdout_loss + self.loss_forward(preds, batch_y, mask) loss = erm_loss + holdout_loss + self.rgm_e * regret return loss / self.num_domains
4,194
37.842593
167
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Algorithms/RGM_Jin_210519_with_train_seperately_and_all_train_separately.py
import torch import torch.nn as nn import torch.nn.functional as F from torch.jit import TracerWarning import math import numpy as np import random from copy import deepcopy import warnings class GradientReversal(torch.autograd.Function): beta = 1. @staticmethod def forward(self, x): return x.view_as(x) @staticmethod def backward(self, grad_output): return -GradientReversal.beta * grad_output class RGM(nn.Module): def __init__(self, featurizer, classifier, num_domains, rgm_e, erm_e, holdout_e, detach_classifier, oracle, loss_forward, num_train_domains): super(RGM, self).__init__() # Deepcopy to not influence the original featurizer and classifier. self.featurizer = deepcopy(featurizer) self.classifier = deepcopy(classifier) # extrapolator f_{-e} self.f_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) # oracle f_{e} self.g_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) self.copy_f_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) for copy_f_k in self.copy_f_k: copy_f_k.requires_grad_(False) self.copy_g_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) for copy_g_k in self.copy_g_k: copy_g_k.requires_grad_(False) self.copy_classifier = deepcopy(self.classifier) self.copy_classifier.requires_grad_(False) self.network = nn.Sequential(self.featurizer, self.classifier) self.init_network = deepcopy(self.network) self.ensemble = nn.ModuleList([nn.Sequential(self.featurizer, self.f_k[k]) for k in range(num_domains)]) self.rgm_e = rgm_e self.erm_e = erm_e self.holdout_e = holdout_e self.detach_classifier = detach_classifier self.oracle = oracle self.num_domains = num_domains if num_train_domains == None: self.num_train_domains = num_domains - 1 else: self.num_train_domains = num_train_domains if self.num_train_domains > num_domains: raise ValueError('num_train_domains must be <= num_domains.') self.train_representation = True self.train_classifier = True self.train_only_classifiers = True self.train_only_representation = True self.loss_forward = loss_forward self.register_buffer('update_count', torch.tensor([0])) self.log_losses = True self.loss_curve_iter = {} self.loss_curve_iter[f'extrapol'] = [] self.loss_curve_iter[f'oracle'] = [] self.loss_curve_iter['erm'] = [] self.loss_curve_iter['holdout'] = [] self.loss_curve_iter['regret'] = [] def get_train_and_extrapolation_domains(self): """Returns all train and extrapolation domain indices (second dimension) for all auxiliary classifiers (first dimension). """ if self.num_train_domains == self.num_domains - 1: all_train_domains = [[i for i in range(self.num_domains) if i != k] for k in range(self.num_domains)] all_extra_domains = [[k] for k in range(self.num_domains)] elif self.num_train_domains == 1: all_train_domains = [[k] for k in range(self.num_domains)] all_extra_domains = [[i for i in range(self.num_domains) if i != k] for k in range(self.num_domains)] else: raise Warning('num_train_domains other than 1 or num_domains not implemented yet.') return(all_train_domains, all_extra_domains) def forward(self, batches): warnings.filterwarnings("ignore", category=TracerWarning) # Check that each batch in batches is from one domain. for _, _, batch_d, _ in batches: assert len(torch.unique(batch_d)) == 1 assert len(np.unique([torch.unique(batch_d).item() for _, _, batch_d, _ in batches])) == self.num_domains for k in range(self.num_domains): self.copy_f_k[k].load_state_dict(self.f_k[k].state_dict()) self.copy_g_k[k].load_state_dict(self.g_k[k].state_dict()) self.copy_classifier.load_state_dict(self.classifier.state_dict()) # Evaluate representation all_phis = [] for batch_x, _, _, _ in batches: phi_x = self.featurizer(batch_x) phi_x = phi_x if self.train_only_representation else phi_x.detach() all_phis.append(phi_x) if self.train_only_representation: assert phi_x.grad_fn != None else: assert phi_x.grad_fn == None # Compute L(f◦phi) erm_loss = torch.tensor(0) if self.train_classifier: for k, (_, batch_y, batch_d, mask) in enumerate(batches): phi_x = all_phis[k] if self.detach_classifier: phi_x = phi_x.detach() if self.train_only_representation and not self.detach_classifier: assert phi_x.grad_fn != None else: assert phi_x.grad_fn == None if self.train_only_classifiers: preds = self.classifier(phi_x) assert preds.grad_fn != None else: preds = self.copy_classifier(phi_x) if self.train_only_representation and not self.detach_classifier: assert preds.grad_fn != None else: assert preds.grad_fn == None norm = self.num_domains erm_loss = erm_loss + self.loss_forward(preds, batch_y, mask) / norm # Set train and extrapolation domain indices for each of the auxiliary classifiers. all_train_domains, all_extra_domains = self.get_train_and_extrapolation_domains() # Compute regret R^e(phi). extra_loss = torch.tensor(0) oracle_loss = torch.tensor(0) if self.num_domains > 1 and self.train_representation: for k, extra_domains in enumerate(all_extra_domains): # Loop classifiers for j in extra_domains: # Loop domains _, batch_y, batch_d, mask = batches[j] phi_k = all_phis[j] preds = self.copy_f_k[k](phi_k) # f_{-e}◦phi (extrapolator) if self.train_only_representation: assert preds.grad_fn != None else: assert preds.grad_fn == None norm = len(extra_domains) extra_loss = extra_loss + self.loss_forward(preds, batch_y, mask) / norm # The minus in the loss function in the paper 2020 Jin here is introduced by the GradientReversal Layer. if self.oracle: if self.train_only_classifiers: oracle_preds = self.g_k[k](GradientReversal.apply(phi_k)) # f_{e}◦phi assert oracle_preds.grad_fn != None else: oracle_preds = self.copy_g_k[k](GradientReversal.apply(phi_k)) # f_{e}◦phi if self.train_only_representation: assert oracle_preds.grad_fn != None else: assert oracle_preds.grad_fn == None oracle_loss = \ oracle_loss + self.loss_forward(oracle_preds, batch_y, mask) / norm if self.train_only_representation: assert extra_loss.grad_fn != None else: assert extra_loss.grad_fn == None if self.oracle: assert oracle_loss.grad_fn != None else: assert oracle_loss.grad_fn == None regret = extra_loss + oracle_loss # Compute L^{-e}(f_{-e}◦phi). Only for training f_{-e}, gradient is not backpropagated on phi (detached). holdout_loss = torch.tensor(0) if self.num_domains > 1 and self.train_representation and self.train_only_classifiers: for k, train_domains in enumerate(all_train_domains): # Loop classifiers # Train the kth classifier on all train_domains. for j in train_domains: # Loop domains _, batch_y, batch_d, mask = batches[j] phi_x = all_phis[j].detach() # phi does not help f_{-e} preds = self.f_k[k](phi_x) assert preds.grad_fn != None norm = len(train_domains) holdout_loss = holdout_loss + self.loss_forward(preds, batch_y, mask) / norm assert holdout_loss.grad_fn != None loss = (self.erm_e * erm_loss + self.rgm_e * (self.holdout_e * holdout_loss + regret)) / self.num_domains self.regret = regret.item() if self.train_representation and self.log_losses: self.loss_curve_iter[f'extrapol'].append(extra_loss.item()) self.loss_curve_iter[f'oracle'].append(oracle_loss.item()) self.loss_curve_iter['regret'].append(regret.item()) self.loss_curve_iter['holdout'].append(holdout_loss.item()) if self.train_classifier and self.log_losses: self.loss_curve_iter['erm'].append(erm_loss.item()) return loss
9,738
44.723005
145
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Algorithms/RGM_Jin.py
import torch import torch.nn as nn import torch.nn.functional as F from torch.jit import TracerWarning import math import numpy as np import random from copy import deepcopy import warnings from itertools import combinations class GradientReversal(torch.autograd.Function): beta = 1. @staticmethod def forward(self, x): return x.view_as(x) @staticmethod def backward(self, grad_output): return -GradientReversal.beta * grad_output class RGM(nn.Module): def __init__(self, featurizer, classifier, num_domains, rgm_e, erm_e, holdout_e, detach_classifier, oracle, loss_forward, num_train_domains, max_n_classifiers): super(RGM, self).__init__() self.num_domains = num_domains if num_train_domains < 0: num_train_domains = self.num_domains + num_train_domains if self.num_domains != 1 and (num_train_domains > self.num_domains - 1 or num_train_domains < 0): raise ValueError(f'Invalid value for num_train_domains: {num_train_domains}') # Set train and extrapolation domain indices for each of the auxiliary classifiers. self.all_train_domains, self.all_extra_domains = self.get_train_and_extrapolation_domains(num_train_domains, max_n_classifiers) self.n_classifiers = len(self.all_train_domains) # Deepcopy to not influence the original featurizer and classifier. self.featurizer = deepcopy(featurizer) self.classifier = deepcopy(classifier) # extrapolator f_{-e} self.f_k = nn.ModuleList(deepcopy(classifier) for _ in range(self.n_classifiers)) # oracle f_{e} self.g_k = nn.ModuleList(deepcopy(classifier) for _ in range(self.n_classifiers)) self.copy_f_k = nn.ModuleList(deepcopy(classifier) for _ in range(self.n_classifiers)) for copy_f_k in self.copy_f_k: copy_f_k.requires_grad_(False) self.network = nn.Sequential(self.featurizer, self.classifier) self.ensemble = nn.ModuleList([nn.Sequential(self.featurizer, self.f_k[k]) for k in range(self.n_classifiers)]) self.rgm_e = rgm_e self.erm_e = erm_e self.holdout_e = holdout_e self.detach_classifier = detach_classifier self.oracle = oracle self.loss_forward = loss_forward self.register_buffer('update_count', torch.tensor([0])) self.log_losses = True self.loss_curve_iter = {} self.loss_curve_iter['extrapol'] = [] self.loss_curve_iter['oracle'] = [] self.loss_curve_iter['erm'] = [] self.loss_curve_iter['holdout'] = [] self.loss_curve_iter['regret'] = [] self.loss_curve_iter['eff_regret'] = [] self.loss_curve_iter['total'] = [] self.loss_curve_iter['eff_loss'] = [] self.loss_curve_iter['rep_loss'] = [] def get_train_and_extrapolation_domains(self, num_train_domains, max_n_classifiers): """Returns all train and extrapolation domain indices (second dimension) for all auxiliary classifiers (first dimension). """ if self.num_domains > 1: domain_idc = range(self.num_domains) all_train_domains = tuple(combinations(domain_idc, num_train_domains)) n_train_domains = len(all_train_domains) if n_train_domains > max_n_classifiers: all_train_d_idc = range(0, n_train_domains) choose_random_idc = np.random.choice(all_train_d_idc, size=max_n_classifiers, replace=False) all_train_domains = tuple([domains for i, domains in enumerate(all_train_domains) if i in choose_random_idc]) # Extrapolation domains are all domains that are not train domains. all_extra_domains = [] for train_domains in all_train_domains: extra_domains = [] for i in range(self.num_domains): if not i in train_domains: extra_domains.append(i) all_extra_domains.append(tuple(extra_domains)) all_extra_domains = tuple(all_extra_domains) else: all_train_domains = () all_extra_domains = () return(all_train_domains, all_extra_domains) def forward(self, batches): warnings.filterwarnings("ignore", category=TracerWarning) # Check that each batch in batches is from one domain. for _, _, batch_d, _ in batches: assert len(torch.unique(batch_d)) == 1 assert len(np.unique([torch.unique(batch_d).item() for _, _, batch_d, _ in batches])) == self.num_domains assert len(batches) == self.num_domains # Get copy of f_k without gradient backprop that emulates frozen weights of f_k. for k in range(self.n_classifiers): self.copy_f_k[k].load_state_dict(self.f_k[k].state_dict()) # Evaluate representation all_phis = [] for batch_x, _, _, _ in batches: phi_x = self.featurizer(batch_x) all_phis.append(phi_x) # Compute L(f◦phi) erm_loss = torch.tensor(0) for k, (_, batch_y, batch_d, mask) in enumerate(batches): phi_x = all_phis[k] if self.detach_classifier: phi_x = phi_x.detach() preds = self.classifier(phi_x) erm_loss = erm_loss + self.loss_forward(preds, batch_y, mask) norm = self.num_domains erm_loss = erm_loss / norm # Compute regret R^e(phi). extra_loss = torch.tensor(0) oracle_loss = torch.tensor(0) for k, extra_domains in enumerate(self.all_extra_domains): # Loop classifiers for j in extra_domains: # Loop domains _, batch_y, batch_d, mask = batches[j] phi_k = all_phis[j] preds = self.copy_f_k[k](phi_k) # f_{-e}◦phi (extrapolator) norm = len(extra_domains) * len(self.all_extra_domains) extra_loss = extra_loss + self.loss_forward(preds, batch_y, mask) / norm # The minus in the loss function in the paper 2020 Jin is introduced by the GradientReversal Layer. if self.oracle: oracle_preds = self.g_k[k](GradientReversal.apply(phi_k)) # f_{e}◦phi oracle_loss = \ oracle_loss + self.loss_forward(oracle_preds, batch_y, mask) / norm regret = extra_loss + oracle_loss eff_regret = extra_loss.item() - oracle_loss.item() # Compute L^{-e}(f_{-e}◦phi). Only for training f_{-e}, gradient is not backpropagated on phi (detached). holdout_loss = torch.tensor(0) for k, train_domains in enumerate(self.all_train_domains): # Loop classifiers # Train the kth classifier on all train_domains. for j in train_domains: # Loop domains _, batch_y, batch_d, mask = batches[j] phi_x = all_phis[j].detach() # phi does not help f_{-e} preds = self.f_k[k](phi_x) norm = len(train_domains) * len(self.all_train_domains) holdout_loss = holdout_loss + self.loss_forward(preds, batch_y, mask) / norm erm_loss = self.erm_e * erm_loss holdout_loss = self.rgm_e * self.holdout_e * holdout_loss regret = self.rgm_e * regret eff_regret = self.rgm_e * eff_regret oracle_loss = self.rgm_e * oracle_loss extra_loss = self.rgm_e * extra_loss eff_loss = (erm_loss + holdout_loss + extra_loss - oracle_loss).item() loss = erm_loss + holdout_loss + regret if self.log_losses: if self.detach_classifier: rep_loss = eff_regret else: rep_loss = erm_loss.item() + eff_regret self.loss_curve_iter['extrapol'].append(extra_loss.item()) self.loss_curve_iter['oracle'].append(oracle_loss.item()) self.loss_curve_iter['regret'].append(regret.item()) self.loss_curve_iter['eff_regret'].append(eff_regret) self.loss_curve_iter['holdout'].append(holdout_loss.item()) self.loss_curve_iter['erm'].append(erm_loss.item()) self.loss_curve_iter['total'].append(loss.item()) self.loss_curve_iter['rep_loss'].append(rep_loss) self.loss_curve_iter['eff_loss'].append(eff_loss) return loss
8,734
44.494792
164
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Algorithms/RGM_Jin_210424_only_one_class.py
import torch import torch.nn as nn import torch.nn.functional as F import math import numpy as np import random class GradientReversal(torch.autograd.Function): beta = 1. @staticmethod def forward(self, x): return x.view_as(x) @staticmethod def backward(self, grad_output): return -GradientReversal.beta * grad_output class RGM(nn.Module): def __init__(self, featurizer, loss_func, args): super(RGM, self).__init__() self.featurizer = featurizer if args.linear: # f,trained on all environments self.classifier = nn.Linear(args.hidden_size, args.output_size) # copy of f_{-e} self.copy_f_k = nn.ModuleList( [nn.Linear(args.hidden_size, args.output_size).requires_grad_(False) for _ in range(args.num_domains)] ) # extrapolator f_{-e} self.f_k = nn.ModuleList( [nn.Linear(args.hidden_size, args.output_size) for _ in range(args.num_domains)] ) # oracle predictor f_{e} self.g_k = nn.ModuleList( [nn.Linear(args.hidden_size, args.output_size) for _ in range(args.num_domains)] ) else: # f,trained on all environments self.classifier = nn.Sequential( nn.Linear(args.hidden_size, args.hidden_size), nn.ReLU(), nn.Linear(args.hidden_size, args.output_size), ) # copy of f_{-e} self.copy_f_k = nn.ModuleList([ nn.Sequential( nn.Linear(args.hidden_size, args.hidden_size), nn.ReLU(), nn.Linear(args.hidden_size, args.output_size), ).requires_grad_(False) for _ in range(args.num_domains) ]) # extrapolator f_{-e} self.f_k = nn.ModuleList([ nn.Sequential( nn.Linear(args.hidden_size, args.hidden_size), nn.ReLU(), nn.Linear(args.hidden_size, args.output_size), ) for _ in range(args.num_domains) ]) # oracle predictor f_{e} self.g_k = nn.ModuleList([ nn.Sequential( nn.Linear(args.hidden_size, args.hidden_size), nn.ReLU(), nn.Linear(args.hidden_size, args.output_size), ) for _ in range(args.num_domains) ]) self.network = nn.Sequential(self.featurizer, self.classifier) self.loss_func = loss_func self.rgm_e = args.rgm_e self.num_domains = args.num_domains self.register_buffer('update_count', torch.tensor([0])) def loss_forward(self, preds, batch_y, mask): pred_loss = self.loss_func(preds, batch_y) * mask return pred_loss.sum() / mask.sum() def forward(self, batches): # assert len(batches) == 2 # TODO for k in range(self.num_domains): self.copy_f_k[k].load_state_dict(self.f_k[k].state_dict()) # Compute L(f◦phi) erm_loss = 0 all_phis = [] for batch_x, batch_y, mask in batches: phi_x = self.featurizer(batch_x) all_phis.append(phi_x) preds = self.classifier(phi_x) erm_loss = erm_loss + self.loss_forward(preds, batch_y, mask) # Compute regret R^e(phi). regret = 0 # ============================================================================= # for k in range(self.num_domains): # _, batch_y, mask = batches[k] # phi_k = all_phis[k] # preds = self.copy_f_k[k](phi_k) # f_{-e}◦phi # oracle_preds = self.g_k[k](GradientReversal.apply(phi_k)) # f_{e}◦phi # # The minus in the loss function in the paper 2020 Jin here is introduced by the GradientReversal Layer. # regret = regret + self.loss_forward(preds, batch_y, mask) + self.loss_forward(oracle_preds, batch_y, mask) # ============================================================================= # TODO # Compute L^{-e}(f_{-e}◦phi). Only for training f_{-e}, gradient is not backpropagated on phi (detached). holdout_loss = 0 # ============================================================================= # for k in range(self.num_domains): # _, batch_y, mask = batches[1 - k] # hardcode: 2 domains # phi_x = all_phis[1 - k].detach() # phi does not help f_{-e} # preds = self.f_k[k](phi_x) # holdout_loss = holdout_loss + self.loss_forward(preds, batch_y, mask) # ============================================================================= loss = erm_loss + holdout_loss + self.rgm_e * regret return loss / self.num_domains
5,006
40.040984
122
py
3DSC
3DSC-main/superconductors_3D/machine_learning/Algorithms/RGM_Jin_210515_old_train_seperetely.py
import torch import torch.nn as nn import torch.nn.functional as F from torch.jit import TracerWarning import math import numpy as np import random from copy import deepcopy import warnings class GradientReversal(torch.autograd.Function): beta = 1. @staticmethod def forward(self, x): return x.view_as(x) @staticmethod def backward(self, grad_output): return -GradientReversal.beta * grad_output class RGM(nn.Module): def __init__(self, featurizer, classifier, num_domains, rgm_e, erm_e, holdout_e, detach_classifier, oracle, loss_forward, num_train_domains): super(RGM, self).__init__() # Deepcopy to not influence the original featurizer and classifier. self.featurizer = deepcopy(featurizer) self.classifier = deepcopy(classifier) self.copy_f_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) for copy_f_k in self.copy_f_k: copy_f_k.requires_grad_(False) # extrapolator f_{-e} self.f_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) # oracle f_{e} self.g_k = nn.ModuleList(deepcopy(classifier) for _ in range(num_domains)) self.network = nn.Sequential(self.featurizer, self.classifier) self.init_network = deepcopy(self.network) self.ensemble = nn.ModuleList([nn.Sequential(self.featurizer, self.f_k[k]) for k in range(num_domains)]) self.rgm_e = rgm_e self.erm_e = erm_e self.holdout_e = holdout_e self.detach_classifier = detach_classifier self.oracle = oracle self.num_domains = num_domains if num_train_domains == None: self.num_train_domains = num_domains - 1 else: self.num_train_domains = num_train_domains if self.num_train_domains > num_domains: raise ValueError('num_train_domains must be <= num_domains.') self.train_representation = True self.train_classifier = True self.loss_forward = loss_forward self.register_buffer('update_count', torch.tensor([0])) self.log_losses = True self.loss_curve_iter = {} self.loss_curve_iter[f'extrapol'] = [] self.loss_curve_iter[f'oracle'] = [] self.loss_curve_iter['erm'] = [] self.loss_curve_iter['holdout'] = [] self.loss_curve_iter['regret'] = [] def get_train_and_extrapolation_domains(self): """Returns all train and extrapolation domain indices (second dimension) for all auxiliary classifiers (first dimension). """ if self.num_train_domains == self.num_domains - 1: all_train_domains = [[i for i in range(self.num_domains) if i != k] for k in range(self.num_domains)] all_extra_domains = [[k] for k in range(self.num_domains)] elif self.num_train_domains == 1: all_train_domains = [[k] for k in range(self.num_domains)] all_extra_domains = [[i for i in range(self.num_domains) if i != k] for k in range(self.num_domains)] else: raise Warning('num_train_domains other than 1 or num_domains not implemented yet.') return(all_train_domains, all_extra_domains) def forward(self, batches): warnings.filterwarnings("ignore", category=TracerWarning) # Check that each batch in batches is from one domain. for _, _, batch_d, _ in batches: assert len(torch.unique(batch_d)) == 1 assert len(np.unique([torch.unique(batch_d).item() for _, _, batch_d, _ in batches])) == self.num_domains for k in range(self.num_domains): self.copy_f_k[k].load_state_dict(self.f_k[k].state_dict()) # Evaluate representation all_phis = [] for batch_x, _, _, _ in batches: phi_x = self.featurizer(batch_x) all_phis.append(phi_x) # Compute L(f◦phi) erm_loss = torch.tensor(0) if self.train_classifier: for k, (_, batch_y, batch_d, mask) in enumerate(batches): phi_x = all_phis[k] if self.detach_classifier: phi_x = phi_x.detach() preds = self.classifier(phi_x) erm_loss = erm_loss + self.loss_forward(preds, batch_y, mask) # Set train and extrapolation domain indices for each of the auxiliary classifiers. all_train_domains, all_extra_domains = self.get_train_and_extrapolation_domains() # Compute regret R^e(phi). extra_loss = torch.tensor(0) oracle_loss = torch.tensor(0) if self.num_domains > 1 and self.train_representation: for k, extra_domains in enumerate(all_extra_domains): # Loop classifiers for j in extra_domains: # Loop domains _, batch_y, batch_d, mask = batches[j] phi_k = all_phis[j] preds = self.copy_f_k[k](phi_k) # f_{-e}◦phi (extrapolator) extra_loss = extra_loss + self.loss_forward(preds, batch_y, mask) # The minus in the loss function in the paper 2020 Jin here is introduced by the GradientReversal Layer. if self.oracle: oracle_preds = self.g_k[k](GradientReversal.apply(phi_k)) # f_{e}◦phi oracle_loss = \ oracle_loss + self.loss_forward(oracle_preds, batch_y, mask) regret = extra_loss + oracle_loss # Compute L^{-e}(f_{-e}◦phi). Only for training f_{-e}, gradient is not backpropagated on phi (detached). holdout_loss = torch.tensor(0) if self.num_domains > 1 and self.train_representation: for k, train_domains in enumerate(all_train_domains): # Loop classifiers # Train the kth classifier on all train_domains. for j in train_domains: # Loop domains _, batch_y, batch_d, mask = batches[j] phi_x = all_phis[j].detach() # phi does not help f_{-e} preds = self.f_k[k](phi_x) norm = len(train_domains) holdout_loss = holdout_loss + self.loss_forward(preds, batch_y, mask) / norm loss = (self.erm_e * erm_loss + self.rgm_e * (self.holdout_e * holdout_loss + regret)) / self.num_domains if self.train_representation and self.log_losses: self.loss_curve_iter[f'extrapol'].append(extra_loss.item()) self.loss_curve_iter[f'oracle'].append(oracle_loss.item()) self.loss_curve_iter['regret'].append(regret.item()) self.loss_curve_iter['holdout'].append(holdout_loss.item()) if self.train_classifier and self.log_losses: self.loss_curve_iter['erm'].append(erm_loss.item()) return loss
7,049
43.620253
145
py
3DSC
3DSC-main/superconductors_3D/machine_learning/own_libraries/models/GPflow_GP.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Sun Oct 3 21:13:25 2021 @author: Timo Sommer This script contains an implementation of a Gaussian Process that works with the GPflow models. """ import torch from sklearn.base import RegressorMixin, BaseEstimator import tensorflow as tf import tensorflow_probability as tfp import numpy as np import gpflow import pickle from copy import deepcopy from sklearn.preprocessing import StandardScaler import os import warnings from gpflow.utilities import print_summary, set_trainable from gpflow.optimizers import NaturalGradient class GPflow_GP(RegressorMixin, BaseEstimator): def __init__(self, model, kernel, alpha, mean_function=None, batch_size=100, epochs=1000, learning_rate=0.1, n_inducing_points=100, NN_path=None, standard_scale=False, natgrad=False, diff_std_for_sc_and_non_sc=False, train_noise=True, train_noise_scale=False, predict_y=False): """alpha: Estimated std of data. float for constant noise, array of floats for these noises as absolute std. Pass function to compute noise std from input y. `natgrad`: If the variational parameters should be trained using Natural Gradients instead of Adam. """ self.model = model self.kernel = kernel self.alpha = alpha self.mean_function = mean_function self.batch_size = batch_size self.epochs = epochs self.learning_rate = learning_rate self.n_inducing_points = n_inducing_points self.NN_path = NN_path self.standard_scale = standard_scale self.natgrad = natgrad self.diff_std_for_sc_and_non_sc = diff_std_for_sc_and_non_sc self.train_noise = train_noise self.train_noise_scale = train_noise_scale self.predict_y = predict_y def get_array_noise_var(self, alpha, y): try: # If alpha is function. noise_std = alpha(y) print('Using specified function for noise.') except TypeError: try: # If alpha is float. alpha = float(alpha) noise_std = np.full(y.shape, alpha) print('Using single float for noise, but in likelihood.') except TypeError: # If alpha is array. noise_std = np.asarray(alpha) print('Using specified array for alpha') noise_std = noise_std.reshape(y.shape) noise_var = noise_std**2 return noise_var def get_noise_var(self, y): """Implement variance per data point if model allows it, otherwise set the same variance per data point. """ if self.model in self.homogeneous_noise_models: # GPflow model only takes a single float for noise as `noise_var`. try: noise_variance = {'noise_variance': float(self.alpha)**2} print('Using single float as noise.') except TypeError: raise ValueError('Parameter `alpha` must be float for this model.') else: if y.shape[1] == 1: if self.diff_std_for_sc_and_non_sc: # Allow different variances for all data points with Tc=0 and all other data points. print('Using different noise for Tc=0 and Tc>0.') likelihood = gpflow.likelihoods.SwitchedLikelihood( [gpflow.likelihoods.Gaussian(variance=self.alpha**2), gpflow.likelihoods.Gaussian(variance=self.alpha**2)] ) non_sc_tc = min(y) groups = np.where(y > non_sc_tc, 0, 1) y = np.hstack([y, groups]) elif self.train_noise_scale: # Train a separarate GP to learn the scale of the noise. likelihood = gpflow.likelihoods.HeteroskedasticTFPConditional( distribution_class=tfp.distributions.Normal, scale_transform=tfp.bijectors.Softplus() ) self.kernel = gpflow.kernels.SeparateIndependent([ self.kernel, # This is k1, the kernel of f1 self.kernel, # this is k2, the kernel of f2 ]) else: try: # If alpha is float. print('Using single float as noise.') likelihood = gpflow.likelihoods.Gaussian(variance=float(self.alpha)**2) except TypeError: # If alpha is iterable or function. print('Alpha is iterable or function.') noise_var = self.get_array_noise_var(self.alpha, y) likelihood = HeteroskedasticGaussian() y = np.hstack([y, noise_var]) noise_variance = {'likelihood': likelihood} else: raise NotImplementedError('Expression above not yet working with multiple y (see num_latent_GPs of HeteroskedasticGaussian()).') return noise_variance, y def get_inducing_variable(self, X): """Choose `M` inducing points for sparse model from data `X`. """ if self.model in self.sparse_models: assert not (self.n_inducing_points is None) M = self.n_inducing_points Z = X[:M, :].copy() if self.train_noise_scale: Z = gpflow.inducing_variables.SeparateIndependentInducingVariables( [ gpflow.inducing_variables.InducingPoints(Z), gpflow.inducing_variables.InducingPoints(Z), ] ) inducing_variable = {'inducing_variable': Z} else: inducing_variable = {} return inducing_variable def run_adam(self, X, y): """Utility function running the Adam optimizer. """ assert not (self.batch_size is None) and not (self.epochs is None) # Commented out because seems wrong though it was in tutorial. # try: # gpflow.set_trainable(self.gp.inducing_variable, False) # except AttributeError: # pass # Train the variational parameters using natural gradients. if self.natgrad: set_trainable(self.gp.q_mu, False) set_trainable(self.gp.q_sqrt, False) natgrad = NaturalGradient(gamma=1) variational_params = [(self.gp.q_mu, self.gp.q_sqrt)] train_dataset = tf.data.Dataset.from_tensor_slices((X, y)).repeat().shuffle(len(X)) train_iter = iter(train_dataset.batch(self.batch_size)) # Create an Adam Optimizer action training_loss = self.gp.training_loss_closure(train_iter, compile=True) adam = tf.optimizers.Adam(self.learning_rate) @tf.function def optimization_step(): adam.minimize(training_loss, self.gp.trainable_variables) if self.natgrad: natgrad.minimize(training_loss, var_list=variational_params) self.loss_curve_ = [] for step in range(self.epochs): optimization_step() self.loss_curve_.append(training_loss().numpy()) return def transform_X_with_NN(self, X): """Load pytorch NN and use it to transform X. """ if not hasattr(self, 'NN_featurizer'): if os.path.exists(self.NN_path): with open(self.NN_path, 'rb') as f: NN = pickle.load(f) self.NN_featurizer = NN.regressor['model'].featurizer else: warnings.warn('Model {self.NN_path} not found. Continue training on original input features.') return(X) print(f'Training GP on output of {self.NN_path}.') else: print(f'Evaluating GP on output of {self.NN_path}.') X = deepcopy(X) X = torch.tensor(X, dtype=torch.float) X = self.NN_featurizer(X) X = X.cpu().detach().numpy() X = np.float64(X) # As it was before return X def inverse_transform_std(self, mu, std, scaler): """Makes the inverse transform of the std by transforming upper and lower bound. Returns upper and lower bound after the inverse transform. """ lower_conf = mu - std upper_conf = mu + std lower_conf_trans = scaler.inverse_transform(lower_conf) upper_conf_trans = scaler.inverse_transform(upper_conf) mu_trans = scaler.inverse_transform(mu) assert np.allclose(np.mean([lower_conf_trans, upper_conf_trans], axis=0), mu_trans), 'Std bounds not symmetrical.' std_trans = (upper_conf_trans - lower_conf_trans) / 2 return std_trans def fit(self, X, y): self.homogeneous_noise_models = [gpflow.models.GPR, gpflow.models.SGPR] self.sparse_models = [gpflow.models.SGPR, gpflow.models.SVGP] self.variational_models = [gpflow.models.VGP, gpflow.models.SVGP] # GPflow needs 2D input for y, otherwise bugs happen. X = deepcopy(np.asarray(X)) y = deepcopy(np.asarray(y)) if y.ndim == 1: y = y.reshape(-1, 1) n_targets = y.shape[1] X, y = self._validate_data(X, y, multi_output=True, y_numeric=True) if self.NN_path != None: X = self.transform_X_with_NN(X) self.fitted_on_NN = True else: self.fitted_on_NN = False # Standard scale X and y to have mean 0 and std 1. if self.standard_scale: self.X_scaler, self.y_scaler = StandardScaler(), StandardScaler() X = self.X_scaler.fit_transform(X) y = self.y_scaler.fit_transform(y) noise_variance, y = self.get_noise_var(y) if self.model != gpflow.models.SVGP: data = {'data': (X, y)} else: data = {'num_data': len(X)} inducing_variable = self.get_inducing_variable(X) if self.train_noise_scale: num_latent_GPs = noise_variance['likelihood'].latent_dim else: num_latent_GPs = n_targets print(f'Number of latent GPs: {num_latent_GPs}') self.gp = self.model(**data, kernel=self.kernel, mean_function=self.mean_function, num_latent_gps=num_latent_GPs, **noise_variance, **inducing_variable ) if not self.train_noise: # Keep specified noise fixed. try: set_trainable(self.gp.likelihood.variance, False) # set_trainable(self.gp.kernel.kernels[1].variance, False) except AttributeError: print('Noise can not be trained.') # print_summary(self.gp) # Make fit. if not self.model in self.variational_models: opt = gpflow.optimizers.Scipy() opt.minimize(self.gp.training_loss, self.gp.trainable_variables) else: self.run_adam(X, y) # print_summary(self.gp) return(self) def predict(self, X, return_std=False): X = np.asarray(X) if self.fitted_on_NN: X = self.transform_X_with_NN(X) if self.standard_scale: X = self.X_scaler.transform(X) if self.predict_y: y, var = self.gp.predict_y(X) else: y, var = self.gp.predict_f(X) y, var = np.asarray(y), np.asarray(var) std = np.sqrt(var) # if self.diff_std_for_sc_and_non_sc: # y = y[:,:-1] # Remove group column # max_uncertain_tc = std[0,1] # std = np.where(y.squeeze() > max_uncertain_tc, std[:,0], std[:,1]) # std = std.reshape(y.shape) # Transform mean and std back. if self.standard_scale: std = self.inverse_transform_std(y, std, self.y_scaler) y = self.y_scaler.inverse_transform(y) if return_std: return(y, std) else: return(y) class HeteroskedasticGaussian(gpflow.likelihoods.Likelihood): def __init__(self, **kwargs): # this likelihood expects a single latent function F, and two columns in the data matrix Y: super().__init__(latent_dim=1, observation_dim=2, **kwargs) def _log_prob(self, F, Y): # log_prob is used by the quadrature fallback of variational_expectations and predict_log_density. # Because variational_expectations is implemented analytically below, this is not actually needed, # but is included for pedagogical purposes. # Note that currently relying on the quadrature would fail due to https://github.com/GPflow/GPflow/issues/966 Y, NoiseVar = Y[:, 0], Y[:, 1] return gpflow.logdensities.gaussian(Y, F, NoiseVar) def _variational_expectations(self, Fmu, Fvar, Y): Y, NoiseVar = Y[:, 0], Y[:, 1] Fmu, Fvar = Fmu[:, 0], Fvar[:, 0] return ( -0.5 * np.log(2 * np.pi) - 0.5 * tf.math.log(NoiseVar) - 0.5 * (tf.math.square(Y - Fmu) + Fvar) / NoiseVar ) # The following two methods are abstract in the base class. # They need to be implemented even if not used. def _predict_log_density(self, Fmu, Fvar, Y): raise NotImplementedError def _predict_mean_and_var(self, Fmu, Fvar): raise NotImplementedError
14,397
39.787535
281
py
3DSC
3DSC-main/superconductors_3D/machine_learning/own_libraries/models/NN/MLP_Lightning.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Thu Nov 4 10:36:50 2021 @author: Timo Sommer This script includes a standard Neural Network based on pytorch lightning. """ import torch from torch.nn import functional as F from torch import nn from torch.utils.data import DataLoader from pytorch_lightning.core.lightning import LightningModule from pytorch_lightning.core.datamodule import LightningDataModule import numpy as np import pytorch_lightning as pl from pytorch_lightning.callbacks.early_stopping import EarlyStopping from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint from sklearn.model_selection import train_test_split import warnings warnings.filterwarnings("ignore", ".*does not have many workers.*") def get_activation_fn(activation: str): """Returns torch activation function based on string activation. """ if activation == 'relu': activation_fn = nn.ReLU() elif activation == 'logistic': activation_fn = nn.Sigmoid() elif activation == 'tanh': activation_fn = nn.Tanh() else: raise ValueError(f'Activation function {activation} not recognized. Activation functions are lowercase always.') return(activation_fn) def get_sequential_NN(input_layer_size: int, hidden_layer_sizes: list, activation: str): """Returns a sequential (Feed Forward) NN. `last_linear` means if the last layer should be linear or with activation function. """ activation_fn = get_activation_fn(activation) layers = [] num_layers = len(hidden_layer_sizes) for i in range(num_layers): if i == 0: in_size = input_layer_size else: in_size = hidden_layer_sizes[i-1] out_size = hidden_layer_sizes[i] layers.append(nn.Linear(in_size, out_size)) last_layer = i == num_layers - 1 if not last_layer: layers.append(activation_fn) layers = tuple(layers) network = nn.Sequential(*layers) return(network) class DataModule(LightningDataModule): def __init__(self, X, y, batch_size, validation_fraction): super().__init__() self.X = X self.y = y self.batch_size = batch_size self.validation_fraction = validation_fraction def prepare_data(self): # called only on 1 GPU return def setup(self, stage): # called on every GPU X_train, X_val, y_train, y_val = train_test_split( self.X, self.y, test_size=self.validation_fraction ) self.train_data = TrainDataset(X_train, y_train) self.val_data = TrainDataset(X_val, y_val) def train_dataloader(self): train_dataloader = DataLoader( self.train_data, batch_size=self.batch_size, shuffle=True ) return train_dataloader def val_dataloader(self): val_dataloader = DataLoader( self.val_data, batch_size=self.batch_size ) return val_dataloader class TrainDataset(torch.utils.data.Dataset): def __init__(self, X, y): self.X = torch.from_numpy(X).float() self.y = torch.from_numpy(y).float() def __len__(self): return len(self.y) def __getitem__(self, idx): return self.X[idx], self.y[idx] class TestDataset(torch.utils.data.Dataset): def __init__(self, X): self.X = torch.from_numpy(X).float() def __len__(self): return len(self.X) def __getitem__(self, idx): return self.X[idx] class LightningMLP(LightningModule): def __init__(self, solver, learning_rate, alpha, n_features, n_targets, hidden_layer_sizes, activation ): super().__init__() self.solver = solver self.learning_rate = learning_rate self.alpha = alpha self.n_features = n_features self.n_targets = n_targets self.hidden_layer_sizes = hidden_layer_sizes self.activation = activation self.loss_func = self.get_loss_func() self.loss_curve_ = {'train': [], 'valid': []} self.make_architecture() return def configure_optimizers(self): if self.solver == 'adam': optimizer = torch.optim.Adam( self.parameters(), lr=self.learning_rate, weight_decay=self.alpha ) else: raise NotImplementedError(f'Optimizer {self.solver} not found.') return optimizer def make_architecture(self): """Generates `self.network`. """ layer_sizes = list(self.hidden_layer_sizes) + [self.n_targets] self.network = get_sequential_NN( input_layer_size=self.n_features, hidden_layer_sizes=layer_sizes, activation=self.activation ) return def get_loss_func(self): loss_func = nn.MSELoss() return loss_func def get_loss(self, y_pred, y_true): loss = self.loss_func(y_pred, y_true) return loss def forward(self, X): """Used for inference. """ y_pred = self.network(X) return y_pred def training_step(self, batch, batch_idx): """Used for training. """ X, y_true = batch y_pred = self(X) loss = self.get_loss(y_pred, y_true) # log self.log('train_loss', loss, on_epoch=True) self.loss_curve_['train'].append(loss.item()) return loss def validation_step(self, batch, batch_idx): """Used for validation. """ X, y_true = batch y_pred = self(X) loss = self.get_loss(y_pred, y_true) # log self.log("val_loss", loss) self.loss_curve_['valid'].append(loss.item()) return loss class MLP(): def __init__(self, hidden_layer_sizes: tuple=(100), activation: str='relu', solver: str='adam', n_epochs: int=200, batch_size: int=200, learning_rate: float=1e-3, alpha: float=1e-4, validation_fraction: (int, type(None))=None, patience: int=10, # clip_grad: float=np.inf, # log_metrics: bool=True, # random_seed: (int, type(None))=None, # use_tensorboard: bool=True **kwargs ): self.hidden_layer_sizes = hidden_layer_sizes self.activation = activation self.solver = solver self.n_epochs = n_epochs self.batch_size = batch_size self.learning_rate = learning_rate self.alpha = alpha self.validation_fraction = validation_fraction self.patience = patience self.kwargs = kwargs def standardize_input(self, X, y): """Standardizes input parameters. """ # Module parameters self.n_epochs = int(self.n_epochs) self.batch_size = int(self.batch_size) self.learning_rate = float(self.learning_rate) self.alpha = float(self.alpha) if self.validation_fraction is not None: self.early_stopping = True self.validation_fraction = float(self.validation_fraction) assert (0 < self.validation_fraction < 1) else: self.early_stopping = False self.validation_fraction = 0 self.patience = int(self.patience) # Input arrrays self.n_features = X.shape[1] if y.ndim == 1: y = y.reshape(-1, 1) self.targets_1D = True else: self.targets_1D = False self.n_targets = y.shape[1] return X, y def configure_callbacks(self): """Configure all callbacks here. """ callbacks = [] # Early stopping if self.early_stopping: early_stopping = EarlyStopping( monitor="val_loss", patience=self.patience ) callbacks.append(early_stopping) return callbacks def fit(self, X, y): # TODO # Add repeatibility (https://pytorch-lightning.readthedocs.io/en/latest/common/trainer.html) X, y = self.standardize_input(X, y) self.callbacks = self.configure_callbacks() data = DataModule(X, y, self.batch_size, self.validation_fraction) model = LightningMLP( solver=self.solver, learning_rate=self.learning_rate, alpha=self.alpha, n_features = self.n_features, n_targets=self.n_targets, hidden_layer_sizes=self.hidden_layer_sizes, activation=self.activation ) self.trainer = pl.Trainer( max_epochs=self.n_epochs, callbacks=self.callbacks, enable_progress_bar=False, logger=False, **self.kwargs ) self.trainer.fit(model=model, datamodule=data) self.loss_curve_ = self.trainer.model.loss_curve_ return def convert_predictions_to_numpy(self, y_pred): """Sanitizes the output of `self.trainer.predict` to make a numpy array with the same number of dimensions as the input targets. """ y_pred = np.array(y_pred[0].tolist()).reshape(-1, self.n_targets) if self.targets_1D: y_pred = y_pred.ravel() return y_pred def predict(self, X): dataset = TestDataset(X) dataloader = DataLoader(dataset, batch_size=len(X)) # model_callback y_pred = self.trainer.predict( model=self.trainer.model, dataloaders=dataloader, ckpt_path='best' ) y_pred = self.convert_predictions_to_numpy(y_pred) return y_pred
11,648
30.569106
136
py
3DSC
3DSC-main/superconductors_3D/machine_learning/own_libraries/models/GNN/MEGNet_tf.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Tue Sep 14 09:12:44 2021 @author: Timo Sommer This script contains an implementation of MEGNet from the original authors based on tensorflow with an sklearn API. """ from sklearn.base import BaseEstimator, RegressorMixin from megnet.models import MEGNetModel from megnet.data.crystal import CrystalGraph, CrystalGraphDisordered, _AtomEmbeddingMap import json import numpy as np import os import sklearn import megnet import tensorflow as tf import keras.backend as K import networkx as nx import matplotlib.pyplot as plt from superconductors_3D.machine_learning.own_libraries.utils.Models import Models, regressor_from_pipeline, Pickle_tf import warnings warnings.filterwarnings('ignore', '.*The `lr` argument is deprecated, use `learning_rate` instead.*') def read_json_file(path_to_file): try: with open(path_to_file) as f: return json.load(f) except Exception as e: print(f'File {path_to_file} throws an error:') raise Exception(e) # TODO # Early Stopping with grouped validation set. # Loss function with only positive values and with Huber Loss. def sc_huber(ytrue, ypred, min_value=0): """Computes the huber loss and also sets all negative predicted y values to 0 beforehand. Useful for superconductors. """ ypred_min = tf.maximum(ypred, min_value) loss = tf.keras.losses.huber(ytrue, ypred_min) return loss # def train_test_split(X, y, sample_weight, test_size, groups): # """Splits data into train and test. If groups is given, a GroupShuffleSplit is used instead of a usual ShuffleSplit. # """ # if groups is None: # CV = sklearn.model_selection.ShuffleSplit(test_size=test_size, random_state=1) # TODO # train_inds, test_inds = next(CV.split(X, y, sample_weight)) # else: # CV = sklearn.model_selection.GroupShuffleSplit(test_size=test_size) # train_inds, test_inds = next(CV.split(X, y, groups)) # X_train, X_test, y_train, y_test = list(np.array(X)[train_inds]), list(np.array(X)[test_inds]), y[train_inds], y[test_inds] # return X_train, X_test, y_train, y_test class MEGNet_tf(BaseEstimator, RegressorMixin): def __init__(self, use_learnt_elemental_embedding, epochs, batch_size, patience, lr, l2_coef, dropout, optimizer_kwargs, nblocks, n1, n2, n3, r_cutoff=4, n_feat_bond=100, early_stopping=False, validation_frac=None, save_checkpoint=False, prev_model=None, loss='mse', domain_col=None, tensorboard=False, lr_exp_decay=None, act='softplus2', npass=3): """ Args: graphs_path: (str) path to a json file of all graphs kwargs: nfeat_edge: (int) number of bond features nfeat_global: (int) number of state features nfeat_node: (int) number of atom features nblocks: (int) number of MEGNetLayer blocks lr: (float) learning rate n1: (int) number of hidden units in layer 1 in MEGNetLayer n2: (int) number of hidden units in layer 2 in MEGNetLayer n3: (int) number of hidden units in layer 3 in MEGNetLayer nvocal: (int) number of total element embedding_dim: (int) number of embedding dimension nbvocal: (int) number of bond types if bond attributes are types bond_embedding_dim: (int) number of bond embedding dimension ngvocal: (int) number of global types if global attributes are types global_embedding_dim: (int) number of global embedding dimension npass: (int) number of recurrent steps in Set2Set layer ntarget: (int) number of output targets act: (object) activation function l2_coef: (float or None) l2 regularization parameter is_classification: (bool) whether it is a classification task loss: (object or str) loss function metrics: (list or dict) List or dictionary of Keras metrics to be evaluated by the model during training and testing dropout: (float) dropout rate graph_converter: (object) object that exposes a "convert" method for structure to graph conversion target_scaler: (object) object that exposes a "transform" and "inverse_transform" methods for transforming the target values optimizer_kwargs (dict): extra keywords for optimizer, for example clipnorm and clipvalue sample_weight_mode (str): sample weight mode for compilation kwargs (dict): in the case where bond inputs are pure distances (not the expanded distances nor integers for embedding, i.e., nfeat_edge=None and bond_embedding_dim=None), kwargs can take additional inputs for expand the distance using Gaussian basis. centers (np.ndarray): array for defining the Gaussian expansion centers width (float): width for the Gaussian basis """ self.use_learnt_elemental_embedding = use_learnt_elemental_embedding self.epochs = epochs self.batch_size = batch_size self.patience = patience self.r_cutoff = r_cutoff self.n_feat_bond = n_feat_bond self.early_stopping = early_stopping self.validation_frac = validation_frac self.save_checkpoint = save_checkpoint self.prev_model = prev_model self.loss = loss self.domain_col = domain_col self.tensorboard = tensorboard self.lr = lr self.lr_exp_decay = lr_exp_decay self.l2_coef = l2_coef self.dropout = dropout self.optimizer_kwargs = optimizer_kwargs self.nblocks = nblocks self.n1 = n1 self.n2 = n2 self.n3 = n3 self.act = act self.npass = npass def get_loss_fn(self): """Returns the loss function. """ if self.loss == 'mse': loss_function = 'mse' elif self.loss == 'huber': loss_function = 'huber' elif self.loss == 'sc_huber': loss_function = sc_huber else: raise ValueError(f'Loss function {self.loss} of MEGNet model not recognized.') return loss_function def get_act_fn(self): """Returns the activation function. """ if self.act == 'softplus2' or self.act == 'softplus': act_function = megnet.activations.softplus2 elif self.act == 'swish': act_function = megnet.activations.swish elif self.act == 'relu': act_function = tf.keras.activations.relu else: raise ValueError(f'Activation function {self.act} of MEGNet model not recognized.') return act_function def get_paths(self, X): assert X.shape[1] == 1 paths = [os.path.expanduser(path) for path in X[:,0]] return paths def get_graphs(self, X): """Returns the graphs of each input. X must be a list of paths to these graphs. """ paths = self.get_paths(X) graphs = [read_json_file(path) for path in paths] # Get already learnt elemental embedding features. if self.use_learnt_elemental_embedding: print('Convert atom features to learnt embedding.') embedding = _AtomEmbeddingMap() for i in range(len(graphs)): graphs[i]['atom'] = embedding.convert(graphs[i]['atom']) return graphs def get_prev_model_path(self): """Saves only the MEGNet part of a previously trained model in a temporary file and returns this filename. """ # megnet_path = Pickle_tf.get_tf_filepath(self.prev_model) return self.prev_model def get_feature_dimensions(self, example_graph): """Returns the input dimensions of the features in the correct `encoding`. """ # No encoding here, just use the plain provided features. nfeat_state = len(example_graph['state'][0]) # Get `encoding` of correct input of atom. Z_with_embedding = all([isinstance(feat, int) for feat in example_graph['atom']]) embedding_from_disordered_dict = all([isinstance(feat, dict) for feat in example_graph['atom']]) if Z_with_embedding: print('Use atom number Z with embedding as atom features.') nfeat_node = None elif embedding_from_disordered_dict: print('Use already learnt embedding of 16 features per element.') nfeat_node = 16 else: print('Use plain features as provided for the atoms.') nfeat_node = len(example_graph['atom'][0]) # Expand distance in gaussian basis nfeat_edge = None return nfeat_node, nfeat_edge, nfeat_state def exp_decay_scheduler(self, epoch, lr): """This learning rate scheduler reduces the learning rate exponentially. """ return lr * self.lr_exp_decay def init_callbacks(self): """Add callbacks. """ callbacks = [] if self.early_stopping: assert not self.validation_frac is None, 'Early stopping needs validation data!' callbacks.append( tf.keras.callbacks.EarlyStopping( monitor='val_loss', min_delta=0, patience=self.patience, verbose=0, mode='min', restore_best_weights=True )) # Early stopping with monitoring anything else than val_mae only works if not saving checkpoints, MEGNet is a bit weird there. self.save_checkpoint = False if self.tensorboard: callbacks.append( tf.keras.callbacks.TensorBoard( log_dir='tensorboard', write_graph=False, )) if not self.lr_exp_decay is None: callbacks.append( tf.keras.callbacks.LearningRateScheduler( schedule=self.exp_decay_scheduler )) return callbacks def train_val_split(self, graphs, y, sample_weight, groups): """Splits data into train and validation data. If groups is given, this gives the groups that will be kept together in either train or validation set. This is useful e.g. for superconductors, where all datapoints with the same chemical composition should be kept together. """ if not groups is None: raise NotImplementedError('groups is not implemented because the train_test_split with the groups can not deal with sample_weight yet.') if self.validation_frac == None: train_graphs = graphs val_graphs = None train_y = y val_y = None train_w = sample_weight val_w = None else: assert 0 < self.validation_frac < 1 if sample_weight is None: train_graphs, val_graphs, train_y, val_y = sklearn.model_selection.train_test_split(graphs, y, test_size=self.validation_frac)#, groups=groups) train_w = None val_w = None else: train_graphs, val_graphs, train_y, val_y, train_w, val_w = sklearn.model_selection.train_test_split(graphs, y, sample_weight, test_size=self.validation_frac)#, groups=groups) return train_graphs, val_graphs, train_y, val_y, train_w, val_w def sanitize_input(self, d_train): """Sanitize input. """ # We want either None or several groups, therefore set to None if there is only one group. This is the encoding of the ML script which is a bit unfortunate. if len(np.unique(d_train)) == 1: d_train = None if self.validation_frac == 0: self.validation_frac = None if self.lr_exp_decay == 1: self.lr_exp_decay = None if not self.lr_exp_decay is None: if not 0 < self.lr_exp_decay < 1: raise ValueError('`lr_exp_decay` must be a float between 0 and 1.') if self.lr_exp_decay < 0.95: warnings.warn('`lr_exp_decay` is very small, should usually be very close to 1.') return d_train def fit(self, X, y, d_train=None, sample_weight=None): """The input X must be a list with paths to json files of graph structures. """ d_train = self.sanitize_input(d_train) self.metrics = ['mse'] # Load graphs graphs = self.get_graphs(X) # Define model # Dimensions example_graph = graphs[0] self.nfeat_node, self.nfeat_edge, self.nfeat_state = self.get_feature_dimensions(example_graph) self.n_target = y.shape[1] if y.ndim == 2 else 1 # Expansion of bond features for continous representation gaussian_centers = np.linspace(0, self.r_cutoff + 1, self.n_feat_bond) gaussian_width = 0.5 self.graph_converter = CrystalGraph(cutoff=self.r_cutoff) self.model = FixedMEGNetModel( nfeat_node=self.nfeat_node, nfeat_global=self.nfeat_state, nfeat_edge=self.nfeat_edge, graph_converter=self.graph_converter, centers=gaussian_centers, width=gaussian_width, ntarget=self.n_target, loss=self.get_loss_fn(), metrics=self.metrics, lr=self.lr, l2_coef=self.l2_coef, dropout=self.dropout, optimizer_kwargs=self.optimizer_kwargs, nblocks=self.nblocks, n1=self.n1, n2=self.n2, n3=self.n3, act=self.get_act_fn(), npass=self.npass, ) train_graphs, val_graphs, train_y, val_y, train_w, val_w = self.train_val_split(graphs, y, sample_weight, d_train) # if not self.prev_model is None: # prev_model = self.get_prev_model_path() self.callbacks = self.init_callbacks() model = self.model.train_from_graphs(train_graphs=train_graphs, train_targets=train_y, validation_graphs=val_graphs, validation_targets=val_y, epochs=self.epochs, batch_size=self.batch_size, patience=self.patience, callbacks=self.callbacks, save_checkpoint=self.save_checkpoint, prev_model=self.prev_model, sample_weights=train_w, val_sample_weights=val_w, ) # Get loss curve self.loss_curve_ = {} self.loss_curve_['train'] = model.history.history['loss'] if not self.validation_frac is None: self.loss_curve_['valid'] = model.history.history['val_loss'] for metric in self.metrics: self.loss_curve_[f'{metric} (train)'] = model.history.history[metric] if not self.validation_frac is None: self.loss_curve_[f'{metric} (valid)'] = model.history.history[f'val_{metric}'] return self def predict(self, X): """The input X must be a list with paths to json files of graph structures. """ graphs = self.get_graphs(X) y_pred = self.model.predict_graphs(graphs) return y_pred from typing import Dict, List, Union from tensorflow.keras.callbacks import Callback from megnet.callbacks import ModelCheckpointMAE, ManualStop, ReduceLRUponNan class FixedMEGNetModel(MEGNetModel): """Added validation sample weights. """ def train_from_graphs( self, train_graphs: List[Dict], train_targets: List[float], validation_graphs: List[Dict] = None, validation_targets: List[float] = None, sample_weights: List[float] = None, epochs: int = 1000, batch_size: int = 128, verbose: int = 1, callbacks: List[Callback] = None, prev_model: str = None, lr_scaling_factor: float = 0.5, patience: int = 500, save_checkpoint: bool = True, automatic_correction: bool = False, dirname: str = "callback", val_sample_weights: List[float] = None, # Added this line **kwargs, ): """ Args: train_graphs: (list) list of graph dictionaries train_targets: (list) list of target values validation_graphs: (list) list of graphs as validation validation_targets: (list) list of validation targets sample_weights: (list) list of sample weights epochs: (int) number of epochs batch_size: (int) training batch size verbose: (int) keras fit verbose, 0 no progress bar, 1 only at the epoch end and 2 every batch callbacks: (list) megnet or keras callback functions for training prev_model: (str) file name for previously saved model lr_scaling_factor: (float, less than 1) scale the learning rate down when nan loss encountered patience: (int) patience for early stopping save_checkpoint: (bool) whether to save checkpoint automatic_correction: (bool) correct nan errors dirname: (str) the directory in which to save checkpoints, if `save_checkpoint=True` **kwargs: """ # load from saved model if prev_model: self.load_weights(prev_model) is_classification = "entropy" in str(self.model.loss) monitor = "val_acc" if is_classification else "val_mae" mode = "max" if is_classification else "min" has_sample_weights = sample_weights is not None if not os.path.isdir(dirname): os.makedirs(dirname) if callbacks is None: # with this call back you can stop the model training by `touch STOP` callbacks = [ManualStop()] train_nb_atoms = [len(i["atom"]) for i in train_graphs] train_targets = [self.target_scaler.transform(i, j) for i, j in zip(train_targets, train_nb_atoms)] if (validation_graphs is not None) and (validation_targets is not None): filepath = os.path.join(dirname, "%s_{epoch:05d}_{%s:.6f}.hdf5" % (monitor, monitor)) val_nb_atoms = [len(i["atom"]) for i in validation_graphs] validation_targets = [self.target_scaler.transform(i, j) for i, j in zip(validation_targets, val_nb_atoms)] val_inputs = self.graph_converter.get_flat_data(validation_graphs, validation_targets) val_generator = self._create_generator(*val_inputs, sample_weights=val_sample_weights, batch_size=batch_size) # Fixed this line. steps_per_val = int(np.ceil(len(validation_graphs) / batch_size)) if save_checkpoint: callbacks.extend( [ ModelCheckpointMAE( filepath=filepath, monitor=monitor, mode=mode, save_best_only=True, save_weights_only=False, val_gen=val_generator, steps_per_val=steps_per_val, target_scaler=self.target_scaler, ) ] ) # avoid running validation twice in an epoch val_generator = None # type: ignore steps_per_val = None # type: ignore if automatic_correction: callbacks.extend( [ ReduceLRUponNan( filepath=filepath, monitor=monitor, mode=mode, factor=lr_scaling_factor, patience=patience, has_sample_weights=has_sample_weights, ) ] ) else: val_generator = None # type: ignore steps_per_val = None # type: ignore train_inputs = self.graph_converter.get_flat_data(train_graphs, train_targets) # check dimension match self.check_dimension(train_graphs[0]) train_generator = self._create_generator(*train_inputs, sample_weights=sample_weights, batch_size=batch_size) steps_per_train = int(np.ceil(len(train_graphs) / batch_size)) self.fit( train_generator, steps_per_epoch=steps_per_train, validation_data=val_generator, validation_steps=steps_per_val, epochs=epochs, verbose=verbose, callbacks=callbacks, **kwargs, ) return self
21,677
40.688462
352
py
3DSC
3DSC-main/superconductors_3D/machine_learning/own_libraries/analysis/Experiments/Run.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Wed Aug 18 14:23:49 2021 @author: Timo Sommer This script contains a class to plot stuff for ML runs. """ import os import pandas as pd from pandas.api.types import is_string_dtype from pandas.api.types import is_numeric_dtype import numpy as np import matplotlib.pyplot as plt from superconductors_3D.machine_learning.own_libraries.utils.Scalers import Arcsinh_Scaler import matplotlib.colors import sklearn from scipy import stats from matplotlib.lines import Line2D from matplotlib.ticker import ScalarFormatter, LogLocator, NullFormatter, FormatStrFormatter import seaborn as sns import superconductors_3D.machine_learning.Custom_Machine_Learning_v1_3 as ML from superconductors_3D.machine_learning.own_libraries.data import All_Data from itertools import product import warnings import yaml import copy from superconductors_3D.machine_learning.own_libraries.data.All_scores import All_scores from superconductors_3D.machine_learning.own_libraries.data.Domain_statistics import Domain_statistics from typing import Union from superconductors_3D.machine_learning.own_libraries.data import Feature_Importances from superconductors_3D.dataset_preparation.utils.check_dataset import get_chem_dict sns.set_theme() from superconductors_3D.machine_learning.own_libraries.utils.Scores import SMAPE def unroll_dictionaries(dictionary, sep='__'): """Unrolls all dictionaries in the given dictionary. """ unrolled_dictionary = copy.deepcopy(dictionary) for key, val in dictionary.items(): if isinstance(val, dict): for subkey, subval in val.items(): unrolled_key = key + sep + subkey assert not unrolled_key in dictionary unrolled_dictionary[unrolled_key] = subval del unrolled_dictionary[key] return unrolled_dictionary def disassemble_dict(value): if isinstance(value, dict): val = list(value.values())[0] key = list(value.keys())[0] else: val = value key = value return key, val def color_for_gaussian_datapoints(y_std_train, y_std_test): """Set colors for datapoints so that datapoints with higher std have a higher transparency. Only useful for the Gaussian Process. """ color_train=matplotlib.colors.to_rgba("C1") colors_train=np.outer(np.ones(len(y_std_train)),np.array(color_train)) alpha_train=1.0-y_std_train/(np.max(y_std_train)+0.1) if np.min(alpha_train)<0.01: alpha_train-=np.min(alpha_train) if np.max(alpha_train)>=1.0: alpha_train/=(np.max(alpha_train)+0.01) colors_train[:,-1]=alpha_train.reshape(-1) color_test=matplotlib.colors.to_rgba("C2") colors_test=np.outer(np.ones(len(y_std_test)),np.array(color_test)) alpha_test=1.0-y_std_test/(np.max(y_std_test)+0.1) if np.min(alpha_test)<0.01: alpha_test-=np.min(alpha_test) if np.max(alpha_test)>=1.0: alpha_test/=(np.max(alpha_test)+0.01) colors_test[:,-1]=alpha_test.reshape(-1) return(colors_train, colors_test) def is_run_dir(absdir, accept_deprecated=False): """Checks if a given directory is an ML run directory. """ has_scores = os.path.exists(os.path.join(absdir, 'All_scores.csv')) has_arguments = os.path.exists(os.path.join(absdir, 'arguments')) has_hparams = os.path.exists(os.path.join(absdir, 'hparams.yml')) if has_arguments and has_hparams: if has_scores: return True else: warnings.warn(f'Failed ML run in directory {absdir}!') # old versions if accept_deprecated: if has_scores and not (has_arguments or has_hparams): warnings.warn(f'Old run with deprecated structure found! Accepting old run.') return True return False def get_hparams(hparams_file): """Get hyperparameter from yaml file.""" if os.path.exists(hparams_file): hparams = yaml.load(open(hparams_file,"r"), Loader=yaml.FullLoader) else: raise ValueError(f'No hyperparameter file "{hparams_file}" found.') return(hparams) def hparams_file(run_dir): return os.path.join(run_dir, 'hparams.yml') def analysis_dir(run_dir): return os.path.join(run_dir, 'plots') def all_scores_file(run_dir): return os.path.join(run_dir, 'All_scores.csv') def params_file(run_dir): return os.path.join(run_dir, 'arguments') class MLRun(): """This class contains functions to plot ML runs. """ def __init__(self, run_dir, get_arguments=True): self.run_dir = run_dir self.analysis_dir = analysis_dir(self.run_dir) self.all_scores_file = all_scores_file(self.run_dir) self.hparams_file = hparams_file(self.run_dir) self.params_file = params_file(self.run_dir) # self.all_scores = All_scores(self.all_scores_file) self.all_Data_path = os.path.join(self.run_dir, 'All_values_and_predictions.csv') if get_arguments: self.arguments = self.get_params() def get_hparams(self): """Returns a dictionary of the hyperparameters of this run. """ return get_hparams(self.hparams_file) def get_params(self, unroll=False): """Returns a dictionary of the parameters of this run (saved under `arguments`). """ if os.path.exists(self.params_file): params = yaml.load(open(self.params_file,"r"), Loader=yaml.Loader) else: raise ValueError(f'No hyperparameter file "{self.params_file}" found.') if unroll: params = unroll_dictionaries(params) return params def get_all_params(self, unroll=False): """Returns all parameters and hyperparameters as a dictionary. """ hparams = self.get_hparams() params = self.get_params() all_params = params.update(hparams) assert len(all_params) == len(hparams) + len(params), 'There is overlap between the names of parameters and hyperparameters!' if unroll: all_params = unroll_dictionaries(all_params) return all_params def get_scores(self): """Returns the df of all scores. """ all_scores = All_scores(self.all_scores_file) return all_scores def savefig(self, filename, **kwargs): directory = os.path.dirname(filename) if not os.path.exists(directory): os.mkdir(directory) plt.savefig(filename, **kwargs) return def plot_score_of_models(self, models: list, score_name: Union[str,dict], target: str, CVs: list=['test'], save_file=None): """Makes bar plots of all given scores of all models. `score_name` can be either a string or a dictionary where the key is the score_name in the csv file and the value is the name in the plot. """ # Allow to override plotted name of score by passing a dict. score_name, plot_score_name = disassemble_dict(score_name) all_scores = All_scores(self.all_scores_file) data = [] for model, CV in product(models, CVs): scores = all_scores.get_scores(target, score_name, model, CV) for score in scores: data.append( {'model': model, score_name: score, 'CV': CV, 'target': target }) data = pd.DataFrame(data) plt.figure() hue = 'CV' if len(CVs) > 1 else None sns.catplot(data=data, x='model', y=score_name, hue=hue, kind='bar', legend=False) plt.ylabel(plot_score_name) plt.legend(loc='best') plt.tight_layout() if save_file == None: save_file = f'scores_{target}_{score_name}_{"+".join(models)}_{"+".join(CVs)}.png' filename = os.path.join(self.analysis_dir, 'scores', save_file) self.savefig(filename, dpi=300) plt.show() plt.clf() return def plot_domain_score_of_model(self, domain_col: str, models: list, score_name: Union[str,dict], target: str, CVs: list=['test'], save_file=None, domain_order=None, rotate_xlabels=None, yscale: str='linear', ylim: tuple=None): """Makes a barplot of the domain scores of some models. """ # Allow to override plotted name of score by passing a dict. score_name, plot_score_name = disassemble_dict(score_name) # Get domain data. domain_statistics_file = os.path.join(self.run_dir, 'Domain_statistics.csv') domain_stats = Domain_statistics(domain_statistics_file, domain_col) domains = domain_stats.all_domains # Build up df. data = [] for model, CV, domain in product(models, CVs, domains): model, plot_model_name = disassemble_dict(model) score = domain_stats.get_score(model, domain, score_name, target, CV) data.append({ 'model': plot_model_name, 'group': domain, 'target': target, 'CV': CV, plot_score_name: score }) data = pd.DataFrame(data) plot_models = list(data['model'].unique()) # Plot. plt.figure() hue = 'group' if len(domains) > 1 else None if domain_order == None: domain_order = domains sns.catplot(data=data, x='model', y=plot_score_name, hue=hue, hue_order=domain_order, kind='bar', legend=False) # Scale plt.yscale(yscale) ax = plt.gca() ax.yaxis.set_major_formatter(ScalarFormatter()) plt.ylim(ylim) plt.legend(loc='best') if rotate_xlabels != None: ax.set_xticklabels(rotation=rotate_xlabels, ha="right") plt.tight_layout() if save_file == None: save_file = f'domain_scores_{target}_{score_name}_{"+".join(plot_models)}_{"+".join(CVs)}.png' filename = os.path.join(self.analysis_dir, 'domain_scores', save_file) self.savefig(filename, dpi=300) plt.show() plt.clf() return def plot_2D_preds(self, model, feature_dict, outpath, x_true=[], y_true=[], target_true=[], scatter_kwargs={}, target_name='target', target_idx=0, res=30): """Plots the prediction surface of the target of a model dependent on two features in given ranges. `feature_dict` must have feature names as keys and as value an iterable of [feature_min, feature_max]. """ features = list(feature_dict.keys()) x = features[0] y = features[1] xmin = feature_dict[x][0] xmax = feature_dict[x][1] ymin = feature_dict[y][0] ymax = feature_dict[y][1] x_grid = np.outer(np.linspace(xmin, xmax, res), np.ones(res)) y_grid = np.outer(np.linspace(ymin, ymax, res), np.ones(res)).T x_flat = x_grid.flatten() y_flat = y_grid.flatten() feats = np.array([x_flat, y_flat]).T # feats = model.x_scaler.transform(feats) z_flat = model.predict(feats) # z_flat = model.y_scaler.inverse_transform(z_flat) if len(z_flat.shape) > 1: z_flat = z_flat[:, target_idx] z_grid = z_flat.reshape(res, res) fig = plt.figure() ax = plt.axes(projection='3d') ax.plot_surface(x_grid, y_grid, z_grid, cmap='plasma', edgecolor='none', alpha=0.3) if len(x_true) > 0 and len(y_true) > 0 and len(target_true) > 0: ax.scatter(x_true, y_true, target_true, **scatter_kwargs) ax.set_title('Prediction plot') ax.set_xlabel(x) ax.set_ylabel(y) ax.set_zlabel(target_name) plt.savefig(outpath, dpi=300) plt.close() def plot_1D_preds(self, model_dict, data_dict, outpath, scatter_kwargs={}, y_idx=0, res=300, xlabel='x', ylabel='target', x_limits=[], y_limits=[], add_fn={}): """Plots the prediction surface of the target of some models dependent on one feature in given range. With `add_fn` you can add other functions to the plot, it needs the name of the function as key and the function as value. """ plt.figure() sns.set_theme() # Get x for the prediction. if x_limits == []: xmax = -np.inf xmin = np.inf for _, (x, _) in data_dict.items(): if max(x) > xmax: xmax = max(x) if min(x) < xmin: xmin = min(x) else: xmin = x_limits[0] xmax = x_limits[1] x = np.linspace(xmin, xmax, res) x = x.reshape((-1, 1)) # Plot true data points. for label, (x_, y_) in data_dict.items(): plt.plot(x_, y_, label=label, marker='.', linestyle='None', **scatter_kwargs) # Plot prediction of model. for modelname, model in model_dict.items(): x_pred = x y_pred = model.predict(x_pred) if len(y_pred.shape) > 1: y_pred = y_pred[:, y_idx] plt.plot(x_pred, y_pred, label=modelname) # Plot additional functions. for fn_name, f in add_fn.items(): y = f(x) plt.plot(x, y, ':', label=fn_name) plt.legend() # plt.title('Predictions') plt.xlabel(xlabel) plt.ylabel(ylabel) if not y_limits == []: plt.ylim(y_limits) plt.savefig(outpath, dpi=300, transparent=False) plt.close() def reduce_duplicates(self, df, duplicate_col: str, mean_cols: list): """Reduces the df so that of duplicates in `duplicate_col` will be taken the mean if all the values are numeric and otherwise it is asserted that it is the same for all the duplicates. Useful for making a df with crystals to a df based on individual superconductors. """ all_columns = df.columns n_unique = sum(~ df.duplicated([duplicate_col, 'repetition'])) groupby_cols = [col for col in df.columns if not col in mean_cols] assert duplicate_col in groupby_cols df = df.groupby(groupby_cols).mean().reset_index() assert len(df) == n_unique assert sorted(df.columns) == sorted(all_columns) return df def parity_plot_all_test_data(self, target, model, repetitions, domain_col, duplicate_col=None, log_log=False): """This makes a parity plot with all test data from all CV runs of all repetitions. """ true_target = f'true {target}' pred_target = f'pred {target}' pred_lower_bound = f'${ML.SIGMA} \sigma$ lower bound' pred_upper_bound = f'${ML.SIGMA} \sigma$ upper bound' rel_pred_lower_bound = f'${ML.SIGMA} \sigma$ lower bound (relative)' pred_scaled_unc = f'scaled ${ML.SIGMA} \sigma$' uncertainty = f'relative uncertainty' # When All_Data is finished this should be written to use it. data_path = os.path.join(self.run_dir, 'All_values_and_predictions.csv') df, _ = ML.load_df_and_metadata(data_path) other_cols = [duplicate_col] if not duplicate_col == None else [] data = All_Data.All_Data.get_test_data(df, target, model, repetitions, domain_col, other_cols=other_cols, true_target=true_target, pred_target=pred_target, pred_lower_bound=pred_lower_bound, pred_upper_bound=pred_upper_bound, pred_scaled_unc=pred_scaled_unc ) # Reduce data so that we have only one entry per superconductor instead of one entry per crystal and take mean of true/ pred/ unc tc columns. if duplicate_col != None: data = self.reduce_duplicates(data, duplicate_col, mean_cols=[true_target, pred_target, pred_lower_bound, pred_upper_bound, pred_scaled_unc]) # Sort so that domains with few data points are plotted on top. if domain_col != None: data['count'] = data.groupby('group')['group'].transform(pd.Series.count) data.sort_values('count', inplace=True, ascending=False) data = data.drop(columns='count') # Start plotting. if domain_col != None: plt.figure(figsize=(8, 4.8)) else: plt.figure() # To remove title of legend. if domain_col != None: data = data.rename(columns={'group': ''}) hue = '' else: hue = None # If uncertainty given, use color for uncertainty, not for group anymore. has_unc = pred_lower_bound in data.columns if has_unc: # rel_unc = data[pred_lower_bound] / data[pred_target] # data[rel_pred_lower_bound] = rel_unc unc = data[pred_scaled_unc] - min(data[pred_scaled_unc]) unc /= max(unc) data[uncertainty] = unc norm = plt.Normalize(min(unc), max(unc)) cmap = sns.cubehelix_palette(as_cmap=True, reverse=True) sm = plt.cm.ScalarMappable(cmap=cmap, norm=norm) sm.set_array([]) hue = uncertainty else: cmap = None # Plot scatter plot. marker_size = 15 ax = sns.scatterplot(data=data, x=true_target, y=pred_target, hue=hue, alpha=1, palette=cmap, s=marker_size) # Plot dotted line of perfect fit. x_min = min(data[true_target]) x_max = 200#max(data[true_target]) line = np.linspace(x_min, x_max, 300) ax.plot(line, line, '--k', label='perfect fit') # Plot title. plt.title(model) # Add legend or colorbar. if has_unc: ax.get_legend().remove() cbar = ax.figure.colorbar(sm) cbar.set_label(hue, labelpad=10) elif domain_col != None: ax.legend(loc='center left', bbox_to_anchor=(1, 0.5)) # Ticks. if log_log: ax.set(xscale='symlog', yscale='symlog') ax.xaxis.set_major_formatter(FormatStrFormatter('%.12g')) ax.yaxis.set_major_formatter(FormatStrFormatter('%.12g')) # Make labels nicer for tc. if target == 'tc': plt.ylabel('pred $T_c$ (K)') plt.xlabel('true $T_c$ (K)') ticks = [0, 0.3, 1, 3, 10, 30, 100, 200] ax.set_xticks(ticks) ax.set_yticks(ticks) ticklabels = ['0', '0.3', '1', '3', '10', '30', '100', '200'] ax.set_xticklabels(ticklabels) ax.set_yticklabels(ticklabels) # Save plot plt.tight_layout() save_dir = os.path.join(self.run_dir, 'plots/parity_plots') if not os.path.exists(save_dir): os.mkdir(save_dir) save_name = os.path.join(save_dir, f'all_together_parity_plot_{target}_{model}.png') plt.savefig(save_name, dpi=300) plt.show() return() def score_over_elemental_prevalence(self, model, target, score, repetitions, domains, chem_formula, duplicate_col, log=False): """This makes a parity plot of the uncertainty and the error with all test data from all CV runs of all repetitions. """ true_target = f'true {target}' pred_target = f'pred {target}' pred_lower_bound = f'${ML.SIGMA} \sigma$ lower bound' pred_upper_bound = f'${ML.SIGMA} \sigma$ upper bound' pred_scaled_unc = f'${ML.SIGMA} \sigma$ (scaled)' # Rename x and y axis to make them nicer for Tc. if target == 'tc': error = f'{score} of $T_c$' else: error = f'{score} of {target}' # When All_Data is finished this should be written to use it. data_path = os.path.join(self.run_dir, 'All_values_and_predictions.csv') df, _ = ML.load_df_and_metadata(data_path) other_cols = [duplicate_col] if duplicate_col != None else [] data = All_Data.All_Data.get_test_data(df, target, model, repetitions=repetitions, domain=domains, other_cols=other_cols, true_target=true_target, pred_target=pred_target, pred_lower_bound=pred_lower_bound, pred_upper_bound=pred_upper_bound, pred_scaled_unc=pred_scaled_unc ) # Reduce data so that we have only one entry per superconductor instead of one entry per crystal and take mean of true/ pred/ unc tc columns. if duplicate_col != None: mean_cols=[true_target, pred_target, pred_lower_bound, pred_upper_bound, pred_scaled_unc] data = self.reduce_duplicates(data, duplicate_col, mean_cols=mean_cols) assert score == 'SMAPE' data[score] = SMAPE(data[true_target], data[pred_target]) elemental_data = {'element': [], score: []} for formula, value in zip(data[chem_formula], data[score]): elements = list(get_chem_dict(formula).keys()) for el in elements: elemental_data['element'].append(el) elemental_data[score].append(value) elemental_data = pd.DataFrame(elemental_data) # Take mean per element. els = elemental_data.groupby('element') df = els.size().reset_index().rename(columns={0: 'occurrences of element'}) df[score] = list(els[score].mean()) df['std'] = list(els[score].std()) # Start plotting. plt.figure() # Plot scatter plot. ax = plt.errorbar(x=df['occurrences of element'], y=df[score], yerr=df['std'], fmt='.') # Add axis labels. plt.xlabel('occurrences of element') plt.ylabel(score) # Add title. plt.title(model) if log: plt.xscale('log') # # Add legend for quantile bars # label = '25%/ 75% quantiles' # handles, labels = plt.gca().get_legend_handles_labels() # line = Line2D([0], [0], label=label, color='k') # handles.extend([line]) # plt.legend(handles=handles) # Save plot plt.tight_layout() save_dir = os.path.join(self.run_dir, 'plots') if not os.path.exists(save_dir): os.mkdir(save_dir) save_name = os.path.join(save_dir, f'score_elemental_prevalence_{target}_{model}_{score}.png') plt.savefig(save_name, dpi=300) plt.show() return() return() def hist_error_over_x(self, x, target, model, repetitions, domain_col, duplicate_col=None, ylim=None, errortype='MdAE', uncertainty='quantiles', log_bars=False): """This makes a parity plot of the uncertainty and the error with all test data from all CV runs of all repetitions. `uncertainty´: 'quantiles' or 'sem' """ true_target = f'true {target}' pred_target = f'pred {target}' pred_lower_bound = f'${ML.SIGMA} \sigma$ lower bound' pred_upper_bound = f'${ML.SIGMA} \sigma$ upper bound' pred_scaled_unc = f'${ML.SIGMA} \sigma$ (scaled)' # Rename x and y axis to make them nicer for Tc. if target == 'tc': error = f'{errortype} of $T_c$' else: error = f'{errortype} of {target}' # When All_Data is finished this should be written to use it. data_path = os.path.join(self.run_dir, 'All_values_and_predictions.csv') df, _ = ML.load_df_and_metadata(data_path) other_cols = [duplicate_col, x] if duplicate_col != None else [x] data = All_Data.All_Data.get_test_data(df, target, model, repetitions, domain_col, other_cols=other_cols, true_target=true_target, pred_target=pred_target, pred_lower_bound=pred_lower_bound, pred_upper_bound=pred_upper_bound, pred_scaled_unc=pred_scaled_unc ) data = data[(data[true_target] > 0) & (data[pred_target] > 0)] # Reduce data so that we have only one entry per superconductor instead of one entry per crystal and take mean of true/ pred/ unc tc columns. if duplicate_col != None: mean_cols=[true_target, pred_target, pred_lower_bound, pred_upper_bound, pred_scaled_unc, x] data = self.reduce_duplicates(data, duplicate_col, mean_cols=mean_cols) # Sort so that domains with few data points are plotted on top. if domain_col != None: data['count'] = data.groupby('group')['group'].transform(pd.Series.count) data.sort_values('count', inplace=True, ascending=False) data = data.drop(columns='count') # Start plotting. if domain_col != None: plt.figure(figsize=(8, 4.8)) else: plt.figure() if log_bars: scaler = np.arcsinh inv_scaler = np.sinh else: scaler = lambda x: x inv_scaler = lambda x: x # Get errors to plot. if errortype == 'MdAE': data[error] = np.abs(data[pred_target] - data[true_target]) reduce_fn = 'median' elif errortype == 'SMAPE': data[error] = SMAPE(data[true_target], data[pred_target], multioutput='raw_values') reduce_fn = 'mean' # Bin data because there are too many points for a scatterplot. n_bins = 17 bin_space = scaler(data[x]) bin_width = (bin_space.max() - bin_space.min()) / (n_bins - 1) bins = np.linspace(bin_space.min(), bin_space.max(), n_bins) bin_width = inv_scaler(bins + bin_width) - inv_scaler(bins) bins = inv_scaler(bins) # if target == 'tc': # # Bin to the left for tc==0. # bins = np.insert(bins, 0, - bins[1], axis=0) # bin_width = np.insert(bin_width, 0, bin_width[0], axis=0) data['binned'] = pd.cut(data[x], bins=bins, include_lowest=True) if reduce_fn == 'median': data_binned = data.groupby('binned').median() elif reduce_fn == 'mean': data_binned = data.groupby('binned').mean() bars = bins[:-1] bin_width = bin_width[:-1] # Get uncertainties. if uncertainty == 'quantiles': data_binned['25_quantile'] = data.groupby('binned')[error].quantile(0.25) data_binned['75_quantile'] = data.groupby('binned')[error].quantile(0.75) yerr = (data_binned[error] - data_binned['25_quantile'], data_binned['75_quantile'] - data_binned[error]) elif uncertainty == 'sem': yerr = data.groupby('binned')[error].apply(stats.sem) # Plot bar plot. plt.bar(bars, data_binned[error], yerr=yerr, width=bin_width, align='edge', capsize=5) # Add axis labels. convert_label = {'totreldiff': '$\Delta_{totrel}$', 'tc': '$T_c$ (K)'} xlabel = convert_label[x] if x in convert_label else x if log_bars: # plt.xscale('symlog') plt.xscale(value='function', functions=(scaler, inv_scaler)) # Change log-scaled x-axis back into interpretable numbers. if target == 'tc' and log_bars: ax = plt.gca() ticks = np.array([0, 1, 3, 10, 30, 100]) ax.set_xticks(ticks) ticklabels = ['0', '1', '3', '10', '30', '100'] ax.set_xticklabels(ticklabels) plt.xlabel(xlabel) plt.ylabel(error) # Add title. plt.title(model) # Add legend for quantile bars labeldict = {'quantiles': '25-75% quantiles', 'sem': 'SEM'} label = labeldict[uncertainty] handles, labels = plt.gca().get_legend_handles_labels() line = Line2D([0], [0], label=label, color='k') handles.extend([line]) plt.legend(handles=handles) # Add plot limits. plt.ylim(ylim) # Save plot plt.tight_layout() save_dir = os.path.join(self.run_dir, 'plots') if not os.path.exists(save_dir): os.mkdir(save_dir) save_name = os.path.join(save_dir, f'quality_plot_all_test_data_{target}_{model}_{x}.png') plt.savefig(save_name, dpi=300) plt.show() return() def parity_plot_uncertainty_all_test_data(self, target, model, repetitions, domain_col, duplicate_col=None, log_log=False): """This makes a parity plot of the uncertainty and the error with all test data from all CV runs of all repetitions. """ true_target = f'true {target}' pred_target = f'pred {target}' pred_lower_bound = f'${ML.SIGMA} \sigma$ lower bound' pred_upper_bound = f'${ML.SIGMA} \sigma$ upper bound' pred_scaled_unc = f'${ML.SIGMA} \sigma$ (scaled)' uncertainty = f'relative uncertainty' # Rename x and y axis to make them nicer for Tc. if target == 'tc': error = 'error of $T_c$' uncertainty = f'${ML.SIGMA} \sigma$ bound of $T_c$' else: error = f'error of {target}' uncertainty = f'${ML.SIGMA} \sigma$ bound of {target}' # When All_Data is finished this should be written to use it. data_path = os.path.join(self.run_dir, 'All_values_and_predictions.csv') df, _ = ML.load_df_and_metadata(data_path) other_cols = [duplicate_col] if duplicate_col != None else [] data = All_Data.All_Data.get_test_data(df, target, model, repetitions, domain_col, other_cols=other_cols, true_target=true_target, pred_target=pred_target, pred_lower_bound=pred_lower_bound, pred_upper_bound=pred_upper_bound, pred_scaled_unc=pred_scaled_unc ) # If this model has no uncertainty given leave. has_uncertainty = pred_lower_bound in data.columns and pred_upper_bound in data.columns if not has_uncertainty: return # Reduce data so that we have only one entry per superconductor instead of one entry per crystal and take mean of true/ pred/ unc tc columns. if duplicate_col != None: data = self.reduce_duplicates(data, duplicate_col, mean_cols=[true_target, pred_target, pred_lower_bound, pred_upper_bound, pred_scaled_unc]) # Sort so that domains with few data points are plotted on top. data['count'] = data.groupby('group')['group'].transform(pd.Series.count) data.sort_values('count', inplace=True, ascending=False) data = data.drop(columns='count') # Start plotting. _ = plt.figure(figsize=(8, 4.8)) # To remove title of legend. data = data.rename(columns={'group': ''}) hue = '' # Get errors to plot. data[error] = data[pred_target] - data[true_target] # Get uncertainties to plot. overestimating = data[error] > 0 underestimating = data[error] <= 0 upper_uncertainties = data.loc[underestimating, pred_upper_bound] - data.loc[underestimating, pred_target] lower_uncertainties = data.loc[overestimating, pred_target] - data.loc[overestimating, pred_lower_bound] data.loc[underestimating, uncertainty] = upper_uncertainties data.loc[overestimating, uncertainty] = lower_uncertainties # Get percentage of data points out of sigma bound. out_of_sigma = ML.out_of_sigma(data[true_target], data[pred_target], data[pred_lower_bound], data[pred_upper_bound]) own_oos = data[error].abs() > data[uncertainty] assert sum(own_oos) / len(own_oos) == out_of_sigma, 'Implementations of out_of_sigma don\'t match.' # For debugging, can be deleted. # unc = data[pred_scaled_unc] - min(data[pred_scaled_unc]) # unc /= max(unc) # data['uncertainty'] = unc # Plot scatter plot. ax = sns.scatterplot(data=data, x=uncertainty, y=error, hue=hue, alpha=1) # Plot dotted line where the error is exactly at the boundary of the uncertainty. x_min = 0 x_max = max(data[uncertainty]) line = np.linspace(x_min, x_max, 300) ax.plot(line, line, '--k', label=f'${ML.SIGMA} \sigma$ boundary') ax.plot(line, -line, '--k') # Plot title. out_of_sigma_str = f'Out of ${ML.SIGMA} \sigma$: {out_of_sigma:.2f}' title = f'{model} ({out_of_sigma_str})' plt.title(title) # Add legend. ax.legend(loc='best')#, bbox_to_anchor=(1, 0.5)) # Ticks. if log_log: ax.set(xscale='symlog', yscale='symlog') ax.xaxis.set_major_formatter(FormatStrFormatter('%.12g')) ax.yaxis.set_major_formatter(FormatStrFormatter('%.12g')) # Make labels nicer for tc. if target == 'tc' and log_log: yticks = [-100, -30, -10, -3, -1, 0, 1, 3, 10, 30, 100] xticks = [0, 1, 3, 10, 30, 100, 200] ax.set_xticks(xticks) ax.set_yticks(yticks) yticklabels = ['-100', '-30', '-10', '-3', '-1', '0', '1', '3', '10', '30', '100'] xticklabels = ['0', '1', '3', '10', '30', '100', '200'] ax.set_xticklabels(xticklabels) ax.set_yticklabels(yticklabels) # Save plot plt.tight_layout() save_dir = os.path.join(self.run_dir, 'plots/parity_plots') if not os.path.exists(save_dir): os.mkdir(save_dir) save_name = os.path.join(save_dir, f'all_test_data_uncertainty_parity_plot_{target}_{model}.png') plt.savefig(save_name, dpi=300) plt.show() return() def raw_parity_plot(self, df, true_target, pred_target, hue, style, ax=None, log_log=False): """Plots a parity plot of predicted vs true target. """ true_target = df[true_target] if not isinstance(true_target, pd.Series) else true_target pred_target = df[pred_target] if not isinstance(pred_target, pd.Series) else pred_target if hue != None: hue = df[hue] if not isinstance(hue, pd.Series) else hue # Setup hue colors to be categorical. hue = hue.convert_dtypes().astype(str) palette = sns.color_palette(n_colors=hue.nunique()) hue_order = sorted(hue.unique()) # Remove legend title because it's annoying. hue = hue.rename('') else: palette = None hue_order = None style = df[style] if not isinstance(style, pd.Series) else style style_order = ['train', 'test'] # Plot ax = sns.scatterplot(data=df, x=true_target, y=pred_target, hue=hue, style=style, style_order=style_order, palette=palette, hue_order=hue_order, ax=ax, alpha=0.8) if log_log: ax.set(xscale='symlog', yscale='symlog') ax.xaxis.set_major_formatter(FormatStrFormatter('%.12g')) ax.yaxis.set_major_formatter(FormatStrFormatter('%.12g')) # Make labels nicer for tc. if true_target.name == 'true tc': plt.ylabel('pred $T_c$ (K)') plt.xlabel('true $T_c$ (K)') ticks = [0.1, 0.5, 1, 5, 10, 50, 100] ax.set_xticks(ticks) ax.set_yticks(ticks) ticklabels = ['0.1', '0.5', '1', '5', '10', '50', '100'] ax.set_xticklabels(ticklabels) ax.set_yticklabels(ticklabels) # Add vertical line of perfect fit. x_min = min(true_target) x_max = max(true_target) line = np.linspace(x_min, x_max, 300) ax.plot(line, line, '--k', label='perfect fit') ax.legend(loc='upper right') return ax def parity_plot(self, target, model, repetition, hue, duplicate_col=None, log_log=False): """Plot parity plots for all given targets, models and repetitions in run_dir. """ # When All_Data is finished this should be written to use it. data_path = os.path.join(self.run_dir, 'All_values_and_predictions.csv') df, _ = ML.load_df_and_metadata(data_path) pred_target_name = All_Data.All_Data.name_preds_col(model, repetition, target) CV = All_Data.name_CV_col(repetition) # Make df of crystals to df of superconductors. used_cols = [target, pred_target_name, CV, duplicate_col, hue] df = df[used_cols] if duplicate_col != None: df = self.reduce_duplicates(df, duplicate_col, mean_cols=[target, pred_target_name]) # Make that test data points are plotted on top of train data points. df = df.sort_values(by=CV, ascending=False) plt.figure() style = df[CV].rename('') true_target = df[target].rename(f'true {target}') pred_target = df[pred_target_name].rename(f'pred {target}') self.raw_parity_plot(df, true_target, pred_target, hue, style=style, log_log=log_log) # Save plot save_dir = os.path.join(self.run_dir, 'plots/parity_plots') if not os.path.exists(save_dir): os.mkdir(save_dir) save_name = os.path.join(save_dir, f'parity_plot_{target}_{model}_{repetition}.png') plt.savefig(save_name, dpi=300) plt.clf() return def get_loss_dfs(self, plot_models, repetitions, run_dir, smoothen=False): """Returns dataframes in long and wide format with the losses of the models and repetitions in run_dir. Smoothen means summing over one epoch for the curves with higher resolution than that. """ # Loop through all saved models and get their loss curves. loss_curves = {} for modelname in plot_models: for i in repetitions: regr = ML.get_saved_model(modelname, i, run_dir) model = regr.regressor_['model'] # Standardize loss curve to dict. if isinstance(model.loss_curve_, list): loss_curves[f'train_{i}'] = model.loss_curve_ elif isinstance(model.loss_curve_, dict): for key, vals in model.loss_curve_.items(): loss_curves[f'{key}_{i}'] = vals else: raise ValueError('model.loss_curve_ is neither list nor dict.') # Get number of epochs for each repetition. num_epochs = {} for i in repetitions: num_epochs[i] = len(loss_curves[f'train_{i}']) df = pd.DataFrame(columns=['epoch']) for curve_name, loss_curve in loss_curves.items(): df_curve = pd.DataFrame(data=loss_curve, columns=[curve_name]) df_curve = df_curve.reset_index().rename(columns={'index': 'epoch'}) df_curve['epoch'] = df_curve['epoch'] + 1 rep = int(curve_name.split('_')[-1]) norm = num_epochs[rep] / len(df_curve) df_curve['epoch'] = norm * df_curve['epoch'] df = df.merge(df_curve, on='epoch', how='outer') df = df.sort_values(by='epoch') if smoothen: # Sum curves over one epoch if resolution is higher than that. df['epoch'] = df['epoch'].apply(np.ceil) df = df.groupby(by=['epoch']).sum().reset_index() df_save = copy.deepcopy(df) # Prepare df in long format for sns.lineplot. df = pd.melt(df, ['epoch'], value_name='loss', var_name='curve') df = df[df['loss'].notna()] pattern = r'^(.*)_(\d+)$' df[['mode', 'repetition']] = df['curve'].str.extract(pattern) # Extract different metrics df.loc[df['mode'] == 'train', 'mode'] = 'loss (train)' df.loc[df['mode'] == 'valid', 'mode'] = 'loss (valid)' metrics_pattern = r'^(.*) \((.*)\)$' df[['metric', 'mode']] = df['mode'].str.extract(metrics_pattern) df = df.rename(columns={'loss': 'value'}) return(df, df_save) def plot_loss_curves(self, plot_models, repetitions, run_dir, outpath, losses, ax=None, save=True, smoothen=False, scale=False, mean=False): """Plot and save loss curves of MLPregressor and torch models. Smoothen means summing over one epoch for the curves with higher resolution than that. """ df, df_save = self.get_loss_dfs(plot_models, repetitions, run_dir) wanted_metrics = df['metric'].isin(losses) assert len(wanted_metrics) > 0, f'Attribute {losses[0]} not found.' df = df[wanted_metrics] # Define style of plot. if ax == None: ax = plt.figure().gca() modes = df['mode'].unique().tolist() dashes = {mode: (2, 2) if mode.startswith('train') else '' for mode in modes} if mean or len(repetitions) == 1: if len(wanted_metrics) > 1: hue = 'metric' style = 'mode' else: hue = 'mode' style = None ax.set_ylabel(wanted_metrics[0]) else: assert len(wanted_metrics) == 1 hue = 'repetition' style = 'mode' # Plot and save loss curves. sns.lineplot(x='epoch', y='value', hue=hue, style=style, dashes=dashes, data=df, ax=ax) ax.set_title('+'.join(plot_models)) plt.yscale('log') if scale: max_losses = df.groupby(by=['mode', 'repetition'])['value'].max() max_plot_loss = 2*max_losses.median() plt.ylim(0, max_plot_loss) if save: plt.savefig(outpath + '.png', dpi=300) df_save.to_csv(outpath + '.csv') plt.show() plt.close() return() def plot_grid_loss_curves(self, plot_models, repetitions, run_dir, outpath, losses): """Plots a grid plot of all loss curves. """ # Get number of images in rows and columns (height and width). num_models = len(plot_models) if num_models <= 2: height = num_models width = 1 else: height = int(np.ceil(num_models / 2)) width = 2 fig, axes = plt.subplots(height, width, gridspec_kw=dict(hspace=0.3), figsize=(12,9), sharex=True, squeeze=False) # Plot figures in a grid. idx = -1 for w in range(width): for h in range(height): idx += 1 if idx >= num_models: continue plot_model = plot_models[idx] ax = axes[h][w] self.plot_loss_curves([plot_model], repetitions, run_dir, outpath='', losses=losses, ax=ax, save=False) # Add ticks to shared axes. ax.xaxis.set_tick_params(labelbottom=True) # Save plot. plt.savefig(outpath, dpi=300) plt.close('all') return() def final_plots(self, plot_dir, plot_models, df_data, domain_colname, features, targets, use_models, outdir): """Do some final plots of your models. """ self.plot_dir = plot_dir n_repetitions = len([col for col in df_data.columns if col.startswith('CV_')]) repetitions = list(range(n_repetitions)) # repetitions = list(range(len(train_indices))) print(f'Plot some stuff in {self.plot_dir}...') if not os.path.exists(self.plot_dir): os.mkdir(self.plot_dir) # # Plot grid figure for loss curve. # outpath = os.path.join(self.plot_dir, 'loss_all_models.png') # losses = ['train', 'valid', 'train2', 'valid2'] # if plot_models: # try: # self.plot_grid_loss_curves(plot_models, repetitions, self.run_dir, outpath, losses) # except AttributeError: # warnings.warn('Could not make a grid plot of loss curves because one of the specified models did not have the attribute `loss_curve_`!') # losses = ['loss', 'mse', 'mae'] # for plot_model in plot_models: # # Plot individual figures for curve. # outpath = os.path.join(self.plot_dir, f'loss_{"+".join(plot_models)}') # self.plot_loss_curves([plot_model], repetitions, self.run_dir, outpath, losses) # Plot individual loss curves with mean and standard deviation. losses = ['loss', 'mse'] for plot_model in plot_models: outpath = os.path.join(self.plot_dir, f'mean_{"+".join(losses)}_{"+".join(plot_models)}') self.plot_loss_curves([plot_model], repetitions, self.run_dir, outpath, losses, mean=True) # # Plot figures for each of the minor loss curves. # losses = ['extrapol', 'oracle', 'erm', 'holdout', 'eff_regret', 'regret', 'total', 'eff_loss', 'rep_loss'] # for plot_model in ['RGM']: # if not plot_model in plot_models: # continue # if model doesn't exist # for i in repetitions: # outpath = os.path.join(self.plot_dir, f'minor_loss_RGM_{i}') # self.plot_loss_curves([plot_model], [i], self.run_dir, outpath, losses) # # Plot figures for the norm of the gradient. # losses = ['grad_norm_before', 'grad_norm_clipped'] # for plot_model in ['RGM']: # if not plot_model in plot_models: # continue # if model doesn't exist # for i in repetitions: # outpath = os.path.join(self.plot_dir, f'norm_grad_RGM_{i}') # self.plot_loss_curves([plot_model], [i], self.run_dir, outpath, losses) # Plot backwards graphs. for modelname in use_models.keys(): repetition = 0 model = ML.get_saved_model(modelname, repetition, self.run_dir) try: for i, graph in enumerate(model.backward_graphs): filename = f'Backward_{modelname}_{repetition}_{i}' graph.render(filename, self.plot_dir, cleanup=True) except AttributeError: pass # Plot prediction surface if features are 2D. try: for modelname in use_models.keys(): for repetition in repetitions: model = ML.get_saved_model(modelname, repetition, self.run_dir) feature_dict = {'x': (-1.5, 1.5), 'y': (-1.5, 1.5)} x_true = df_data['x_0'].to_numpy() y_true = df_data['x_1'].to_numpy() target_true = df_data['target'].to_numpy() outpath = os.path.join(self.plot_dir, f'Preds_surface_{modelname}_{repetition}.png') # Get color per domain. domains = df_data[domain_colname].to_numpy() if len(domains) > 0: cycle = plt.rcParams['axes.prop_cycle'].by_key()['color'] color_dict = {d: cycle[i] for i, d in enumerate(np.unique(domains))} colors = [color_dict[d] for d in domains] else: colors = None scatter_kwargs = {'c': colors} self.plot_2D_preds(model, feature_dict, outpath, x_true, y_true, target_true, scatter_kwargs) except KeyError: pass # Plot prediction line if features are 1D. models = [model for model in use_models.keys() if not (model == 'LR')] plot_1D_features = len(features) == 1 and is_numeric_dtype(df_data[features]) if plot_1D_features: try: x_true = df_data[features].to_numpy() target_true = df_data[targets[0]].to_numpy() for repetition in repetitions: model_dict = {} for modelname in models: model_dict[modelname] = ML.get_saved_model(modelname, repetition, self.run_dir) # Get test and train feature and target for this repetition. CV_col = All_Data.All_Data.name_CV_col(repetition) train_indices = df_data[CV_col] == 'train' test_indices = df_data[CV_col] == 'test' x_train = x_true[train_indices] target_train = target_true[train_indices] x_test = x_true[test_indices] target_test = target_true[test_indices] data_dict = { 'train data': (x_train, target_train), 'test data': (x_test, target_test), } scatter_kwargs = {'markersize': 5} outpath = os.path.join(self.plot_dir, f'Preds_1D_line_{modelname}_{repetition}.png') # x_limits = [-2, 2] # y_limits = [-1.5, 1.5] # add_fn={'Cbrt': np.cbrt} x_limits = [] y_limits = [] add_fn = {} self.plot_1D_preds(model_dict, data_dict, outpath, scatter_kwargs, x_limits=x_limits, y_limits=y_limits, add_fn=add_fn) except KeyError: pass # Parity plot for all test data. plot_models = use_models.keys() duplicate_col = 'formula_sc' if 'formula_sc' in df_data.columns else None log_log = True for target, model in product(targets, plot_models): self.parity_plot_all_test_data(target, model, repetitions, domain_colname, duplicate_col=duplicate_col, log_log=log_log) # Plot uncertainty parity plot. plot_models = use_models.keys() duplicate_col = 'formula_sc' if 'formula_sc' in df_data.columns else None log_log = True # The function itself checks whether the model has uncertainty and if not it returns. for target, model in product(targets, plot_models): self.parity_plot_uncertainty_all_test_data(target, model, repetitions, domain_colname, duplicate_col=duplicate_col, log_log=log_log) # Plot error over target distribution. ylim = (0, 1) duplicate_col = 'formula_sc' if 'formula_sc' in df_data.columns else None for target, model in product(targets, plot_models): x = target log_bars = True if target == 'tc' else False self.hist_error_over_x(x, target, model, repetitions, domain_colname, duplicate_col=duplicate_col, log_bars=log_bars, ylim=ylim, errortype='SMAPE', uncertainty='sem') # Plot quality plot with totreldiff. x = 'totreldiff' ylim = (0, None) duplicate_col = 'formula_sc' if 'formula_sc' in df_data.columns else None if x in df_data: varying_quality = sum(df_data[x].unique()) > 1 if varying_quality: for target, model in product(targets, plot_models): self.hist_error_over_x(x, target, model, repetitions, domain_colname, duplicate_col=duplicate_col, ylim=ylim, errortype='SMAPE', uncertainty='sem') else: print('Can\'t plot dataset quality plot.') # Plot prediction error over elemental prevalence. score = 'SMAPE' chem_formula = 'formula_sc' duplicate_col = 'formula_sc' if 'formula_sc' in df_data.columns else None log = True if chem_formula in df_data.columns: for target, modelname in product(targets, plot_models): self.score_over_elemental_prevalence(model, target, score, repetitions, domain_colname, chem_formula, duplicate_col, log) # Plot feature importances. for target, modelname in product(targets, plot_models): all_importances = [] for repetition in repetitions: model = ML.get_saved_model(modelname, repetition, self.run_dir).regressor_['model'] try: importances = Feature_Importances.get_feature_importances(model) except AttributeError: continue all_importances.append(importances) if len(all_importances) > 0: importances = np.mean(all_importances, axis=0) feat_dir = os.path.join(self.run_dir, 'plots', 'feature_importances') if not os.path.exists(feat_dir): os.mkdir(feat_dir) outpath = os.path.join(feat_dir, f'{modelname}_{target}_feature_importances') Feature_Importances.plot_feature_importances(importances, features, outpath) print('Finished with plotting.') return()
55,547
43.688656
274
py
3DSC
3DSC-main/superconductors_3D/machine_learning/own_libraries/utils/Refactoring.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Wed Sep 22 10:59:22 2021 @author: Timo Sommer This script contains a class for checking if the output of the Machine_Learning() class is the same as in one reference directory. It is particular useful for automatically checking if the code still does the same after refactoring. """ import os import superconductors_3D.machine_learning.Custom_Machine_Learning_v1_3 as ML import pandas as pd import joblib import pickle import torch import filecmp import warnings import tempfile class Refactoring(): """This class is used for refactoring. It checks if the files in one directory match the files in another comparison directory, except for some things that must change like clocktime. """ def __init__(self, cmp_dir): self.cmp_dir = os.path.abspath(cmp_dir) self.expected_files = ['All_scores.csv', 'All_values_and_predictions.csv', 'Domain_statistics.csv', 'Numerical results', 'hparams.yml', 'arguments'] self.expected_dirs = ['models', 'plots'] self.out_filename = 'Refacturing_result.txt' self.skip_files = [self.out_filename] def check_if_content_as_expected(self): """Checks if all of the expected files and directories in the comparison directory are there and if there are no unexpected files. Otherwise just prints messages. """ contents = sorted(os.listdir(self.cmp_dir)) expected_contents = sorted(self.expected_files + self.expected_dirs) too_few = [name for name in expected_contents if not name in contents] too_much = [name for name in contents if not name in expected_contents] for content in too_few: if content in self.skip_files: continue self.print(f'Expected but not found: {content}') for content in too_much: if content in self.skip_files: continue self.print(f'Unexpectedly found: {content}') return() def indent(self, text): space = ' ' text = space + text.replace('\n', '\n' + space) return(text) def print(self, text, to_console=True): """Prints text to output file. If to_console=True, also prints to console. """ outpath = os.path.join(self.new_dir, self.out_filename) with open(outpath, 'a') as f: f.write('\n' + text) if to_console: print(text) return() def check_if_df_is_close(self, filename, d1, d2): """Checks if a df is at least close to another df. """ f1 = os.path.join(d1, filename) f2 = os.path.join(d2, filename) df1, _ = ML.load_df_and_metadata(f1) df2, _ = ML.load_df_and_metadata(f2) try: pd.testing.assert_frame_equal(df1, df2) is_close = True error = '' except AssertionError as e: is_close = False error = str(e) return(is_close, error) def cmp_Numerical_results(self, filename1, filename2): """Compares the files 'Numerical results' in the two directories and returns True if they are equal except for the time. """ with open(filename1) as f1: with open(filename2) as f2: # Exclude the first line that includes the time of the run. relevant1 = f1.readlines()[1:] relevant2 = f2.readlines()[1:] equal = relevant1 == relevant2 return(equal) def cmp_pytorch_NN(self, NN1, NN2): """If the model is a pytorch model, there will be non deterministic behaviour. Therefore only compare the state_dicts in this case. """ params1 = list(NN1.state_dict().values()) params2 = list(NN2.state_dict().values()) same_n_weights = len(params1) == len(params2) equal = same_n_weights and all([torch.equal(w1, w2) for w1, w2 in zip(params1, params2)]) return equal def cmp_GP_on_NN_featurizer(self, regr1, regr2): """If the model is a GP model that loads a pytorch NN, the attribute that is a path that points to the pytorch model changes between runs. Therefore before comparing these models overwrite these attributes. """ # Check NNs of GP_on_NN for equality. NN1 = regr1.regressor_['model'].NN_featurizer NN2 = regr2.regressor_['model'].NN_featurizer equal_NNs = self.cmp_pytorch_NN(NN1, NN2) # Overwrite NN_stuff of GP_on_NN. regr1.regressor_['model'].NN_path = '' regr2.regressor_['model'].NN_path = '' regr1.regressor_['model'].NN_featurizer = '' regr2.regressor_['model'].NN_featurizer = '' regr1.regressor['model'].NN_path = '' regr2.regressor['model'].NN_path = '' # Check GP with overriden NN stuff for equality. with tempfile.NamedTemporaryFile() as tmp_file1: with tempfile.NamedTemporaryFile() as tmp_file2: pickle.dump(regr1, open(tmp_file1.name, 'wb')) pickle.dump(regr2, open(tmp_file2.name, 'wb')) equal_GPs = filecmp.cmp(tmp_file1.name, tmp_file2.name, shallow=True) equal = equal_NNs and equal_GPs return equal def cmp_joblib(self, filename1, filename2): """Compare joblib files. """ regr1 = joblib.load(filename1) regr2 = joblib.load(filename2) try: NN1 = regr1.regressor_['model'].trainer NN2 = regr2.regressor_['model'].trainer try: # pytorch model equal = self.cmp_pytorch_NN(NN1, NN2) except AttributeError: # pytorch lightning model NN1 = NN1.model NN2 = NN2.model equal = self.cmp_pytorch_NN(NN1, NN2) except AttributeError: try: equal = self.cmp_GP_on_NN_featurizer(regr1, regr2) except AttributeError: equal = filecmp.cmp(filename1, filename2, shallow=True) return(equal) def cmp_files(self, filename, d1, d2): """Compares if two files are the same. In some cases, it is clear that files cannot be the same after refacturing, in this case this function compares only what can be compared. """ pkl_files = ['.pkl', '.joblib'] f1 = os.path.join(d1, filename) f2 = os.path.join(d2, filename) if filename == 'Numerical results': equal = self.cmp_Numerical_results(f1, f2) elif any([filename.endswith(pkl_file) for pkl_file in pkl_files]): equal = self.cmp_joblib(f1, f2) else: equal = filecmp.cmp(f1, f2, shallow=False) return(equal) def not_equal(self, filename, new_path, cmp_path): """Executed if files are not equal. Prints the issues. """ self.print(f'\nFile not equal: {filename}') # If file is df, check if dfs are at least close. if filename.endswith('.csv'): is_close, error = self.check_if_df_is_close(filename, new_path, cmp_path) if is_close: self.print('--> But dfs are close.') else: self.print('--> Dataframes are not even close.') print(f'--> See file {self.out_filename} for further information.') error = self.indent(error) self.print(f'--> Error:\n{error}', to_console=False) def both_exist(self, filename, d1, d2): f1 = os.path.join(d1, filename) f2 = os.path.join(d2, filename) exist = True for f in (f1, f2): if not os.path.exists(f): self.print(f'File {f} doesn\'t exist.') exist = False return(exist) def check(self, new_dir): """Checks if the original `cmp_dir` matches the new directory `new_dir` in all important files. If yes, refacturing was successfull. """ self.new_dir = os.path.abspath(new_dir) self.print(f'\nCheck refactoring against comparison directory {self.cmp_dir}.') if not os.path.exists(self.cmp_dir): warnings.warn(f'Comparison directory {self.cmp_dir} doesn\'t exist!\n Exiting without check.') return(False) self.check_if_content_as_expected() # Check every file if they are equal and otherwise print message. dirs_equal = True for new_path, _, new_files in os.walk(self.new_dir): # New path including subdirectories of the cmp directory. cmp_path = new_path.replace(self.new_dir, self.cmp_dir) for filename in new_files: if filename in self.skip_files or not self.both_exist(filename, new_path, cmp_path): continue equal = self.cmp_files(filename, new_path, cmp_path) if equal and filename in self.expected_files: print(f'File equal: {filename}') if not equal: dirs_equal = False self.not_equal(filename, new_path, cmp_path) if dirs_equal: self.print('\nAll relevant files of the directories are equal.\nRefactoring successfull!\n') else: self.print('\nThere are some relevant files that are not equal.\nRefactoring not successfull!\n') return(dirs_equal)
9,634
42.795455
232
py
3DSC
3DSC-main/superconductors_3D/machine_learning/own_libraries/utils/Models.py
#!/usr/bin/env python # -*- coding: utf-8 -*- """ Created on Mon Nov 22 14:12:31 2021 @author: Timo Sommer This script is a collection of classes for saving and loading models of the ML script. """ from megnet.models import MEGNetModel import os import pickle import io def get_modelpath(outdir, modelname, repetition): """Returns path without extension where a specific model would be saved. Used internally and externally. Outdir is the directory of the whole run with a subdirectory 'models'. """ filename = f'{modelname}_{repetition}' path = os.path.join(outdir, 'models', filename) return(path) def regressor_from_pipeline(pipe): """Returns the ML model from a given sklearn Pipeline or TransformedTargetRegressor. """ return pipe.regressor_['model'] class Pickle_tf(): """Use this class to pickle tensorflow models that cannot be pickled normally because of some ThreadLock error. """ def __init__(self): pass def model_is_MEGNet(self, modelpath): """Checks by the modelpath if the model is a MEGNET model. """ tf_filename = self.get_tf_filepath(modelpath) try: MEGNetModel.from_file(tf_filename) is_MEGNet = True except FileNotFoundError: is_MEGNet = False return is_MEGNet def get_tf_filepath(self, modelpath): """Turns the pickle path into the tf model path. """ return os.path.splitext(modelpath)[0] + '.hdf5' def save(self, regr, modelpath): """Saves a tf or MEGNet model. Part of the sklearn pipeline is saved as pickle and the not-pickable tf part is saved as tf model. """ tf_filename = self.get_tf_filepath(modelpath) try: # MEGNet tf_model = regressor_from_pipeline(regr).model # Save non-pickable part as hd5. tf_model.save_model(filename=tf_filename) except AttributeError as e: # Other tf models raise NotImplementedError(f'Not implemented for other model than MEGNet. Error message: {e}') # Remove non-pickable part from regressor. regressor_from_pipeline(regr).model = None # Pickle regressor. with open(modelpath, 'wb') as f: pickle.dump(regr, f) # Put non-pickable model back into pipeline to not change regr from the outside. Somehow using deepcopy doesn't work. regressor_from_pipeline(regr).model = tf_model def load(self, modelpath): """Loads a tf model composed of pickled regressor/pipeline and tf_model saved as hdf5. """ regr = Models().load_pickle(modelpath) tf_filepath = self.get_tf_filepath(modelpath) try: # MEGNet tf_model = MEGNetModel.from_file(tf_filepath) regressor_from_pipeline(regr).model = tf_model except AttributeError as e: # Other tf models raise NotImplementedError(f'Not implemented for other model than MEGNet. Error message: {e}') return regr class Models(): """Class for saving and loading ML models. """ def __init__(self): pass def save(self, regr, rundir, modelname, repetition, save_torch_statedict=False): """Saves model as pickle and if possible as pytorch models. """ filename = get_modelpath(rundir, modelname, repetition) outpath = filename + '.pkl' try: with open(outpath, 'wb') as f: pickle.dump(regr, f) except (TypeError, AttributeError) as e: # tf or keras models must be saved seperately because they cannot be pickled. if 'pickle' in str(e): Pickle_tf().save(regr=regr, modelpath=outpath) else: raise TypeError(e) if save_torch_statedict: outpath = filename + '.pt' try: model = regressor_from_pipeline(regr) import torch torch.save(model.trainer.state_dict(), outpath) except AttributeError: pass return() def load_pickle(self, modelpath): """Loads a pickled model from the path. """ with open(modelpath, 'rb') as f: model = CPU_Unpickler(f).load() return model def load_from_path(self, modelpath): """Loads a model from the given path with .pkl extension and returns it. Can also load tensorflow models that were not simply pickled. """ model_is_MEGNet = Pickle_tf().model_is_MEGNet(modelpath) if model_is_MEGNet: model = Pickle_tf().load(modelpath) else: model = self.load_pickle(modelpath) return model def load(self, modelname: str, repetition: int, rundir: str, regressor: bool=False): """Loads a model with given name and repetition from rundir. If `regressor==True` the returned model is not the whole sklearn pipeline/regressor but only the fitted ML model. """ modelpath = get_modelpath(rundir, modelname, repetition) + '.pkl' model = self.load_from_path(modelpath) if regressor: # Get regressor model from pipeline model = regressor_from_pipeline(model) return(model) class CPU_Unpickler(pickle.Unpickler): """Use this for unpickling instead of pickle.load(f) to be able to load a model even if it was saved on a gpu. """ def find_class(self, module, name): if module == 'torch.storage' and name == '_load_from_bytes': return lambda b: torch.load(io.BytesIO(b), map_location='cpu') else: return super().find_class(module, name)
5,903
35.670807
182
py
MatchZoo
MatchZoo-master/setup.py
import io import os from setuptools import setup, find_packages here = os.path.abspath(os.path.dirname(__file__)) # Avoids IDE errors, but actual version is read from version.py __version__ = None exec(open('matchzoo/version.py').read()) short_description = 'Facilitating the design, comparison and sharing of deep text matching models.' # Get the long description from the README file with io.open(os.path.join(here, 'README.md'), encoding='utf-8') as f: long_description = f.read() install_requires = [ 'keras >= 2.3.0', 'nltk >= 3.2.3', 'numpy >= 1.14', 'tqdm >= 4.19.4', 'dill >= 0.2.7.1', 'pandas >= 0.23.1', 'networkx >= 2.1', 'h5py >= 2.8.0', 'hyperopt >= 0.1.1' ] extras_requires = { 'tests': [ 'coverage >= 4.3.4', 'codecov >= 2.0.15', 'pytest >= 3.0.3', 'pytest-cov >= 2.4.0', 'flake8 >= 3.6.0', 'flake8_docstrings >= 1.0.2'], } setup( name="MatchZoo", version=__version__, author="Yixing Fan, Bo Wang, Zeyi Wang, Liang Pang, Liu Yang, Qinghua Wang, etc.", author_email="fanyixing@ict.ac.cn", description=(short_description), license="Apache 2.0", keywords="text matching models", url="https://github.com/NTMC-Community/MatchZoo", packages=find_packages(), long_description=long_description, long_description_content_type='text/markdown', classifiers=[ "Development Status :: 3 - Alpha", 'Environment :: Console', 'Operating System :: POSIX :: Linux', 'Topic :: Scientific/Engineering :: Artificial Intelligence', "License :: OSI Approved :: Apache Software License", 'Programming Language :: Python :: 3.6' ], install_requires=install_requires, extras_require=extras_requires )
1,799
26.692308
99
py
MatchZoo
MatchZoo-master/matchzoo/__init__.py
from pathlib import Path USER_DIR = Path.expanduser(Path('~')).joinpath('.matchzoo') if not USER_DIR.exists(): USER_DIR.mkdir() USER_DATA_DIR = USER_DIR.joinpath('datasets') if not USER_DATA_DIR.exists(): USER_DATA_DIR.mkdir() USER_TUNED_MODELS_DIR = USER_DIR.joinpath('tuned_models') from .version import __version__ from .data_pack import DataPack from .data_pack import pack from .data_pack import load_data_pack from . import metrics from . import tasks from . import preprocessors from . import data_generator from .data_generator import DataGenerator from .data_generator import DataGeneratorBuilder from .preprocessors.chain_transform import chain_transform from . import metrics from . import losses from . import engine from . import models from . import embedding from . import datasets from . import layers from . import auto from . import contrib from .engine import hyper_spaces from .engine.base_model import load_model from .engine.base_preprocessor import load_preprocessor from .engine import callbacks from .engine.param import Param from .engine.param_table import ParamTable from .embedding.embedding import Embedding from .utils import one_hot, make_keras_optimizer_picklable from .preprocessors.build_unit_from_data_pack import build_unit_from_data_pack from .preprocessors.build_vocab_unit import build_vocab_unit # deprecated, should be removed in v2.2 from .contrib.legacy_data_generator import DPoolDataGenerator from .contrib.legacy_data_generator import DPoolPairDataGenerator from .contrib.legacy_data_generator import HistogramDataGenerator from .contrib.legacy_data_generator import HistogramPairDataGenerator from .contrib.legacy_data_generator import DynamicDataGenerator from .contrib.legacy_data_generator import PairDataGenerator
1,785
30.333333
78
py
MatchZoo
MatchZoo-master/matchzoo/models/cdssm.py
"""An implementation of CDSSM (CLSM) model.""" import typing import keras from keras.models import Model from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo import preprocessors from matchzoo.utils import TensorType class CDSSM(BaseModel): """ CDSSM Model implementation. Learning Semantic Representations Using Convolutional Neural Networks for Web Search. (2014a) A Latent Semantic Model with Convolutional-Pooling Structure for Information Retrieval. (2014b) Examples: >>> model = CDSSM() >>> model.params['optimizer'] = 'adam' >>> model.params['filters'] = 32 >>> model.params['kernel_size'] = 3 >>> model.params['conv_activation_func'] = 'relu' >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" # set :attr:`with_multi_layer_perceptron` to False to support # user-defined variable dense layer units params = super().get_default_params(with_multi_layer_perceptron=True) params.add(Param(name='filters', value=32, desc="Number of filters in the 1D convolution " "layer.")) params.add(Param(name='kernel_size', value=3, desc="Number of kernel size in the 1D " "convolution layer.")) params.add(Param(name='strides', value=1, desc="Strides in the 1D convolution layer.")) params.add(Param(name='padding', value='same', desc="The padding mode in the convolution " "layer. It should be one of `same`, " "`valid`, ""and `causal`.")) params.add(Param(name='conv_activation_func', value='relu', desc="Activation function in the convolution" " layer.")) params.add(Param(name='w_initializer', value='glorot_normal')) params.add(Param(name='b_initializer', value='zeros')) params.add(Param(name='dropout_rate', value=0.3, desc="The dropout rate.")) return params def _create_base_network(self) -> typing.Callable: """ Apply conv and maxpooling operation towards to each letter-ngram. The input shape is `fixed_text_length`*`number of letter-ngram`, as described in the paper, `n` is 3, `number of letter-trigram` is about 30,000 according to their observation. :return: Wrapped Keras `Layer` as CDSSM network, tensor in tensor out. """ def _wrapper(x: TensorType): # Apply 1d convolutional on each word_ngram (lt). # Input shape: (batch_size, num_tri_letters, 90000) # Sequence of num_tri_letters vectors of 90000d vectors. x = keras.layers.Conv1D( filters=self._params['filters'], kernel_size=self._params['kernel_size'], strides=self._params['strides'], padding=self._params['padding'], activation=self._params['conv_activation_func'], kernel_initializer=self._params['w_initializer'], bias_initializer=self._params['b_initializer'])(x) # Apply max pooling by take max at each dimension across # all word_trigram features. x = keras.layers.Dropout(self._params['dropout_rate'])(x) x = keras.layers.GlobalMaxPool1D()(x) # Apply a none-linear transformation use a tanh layer. x = self._make_multi_layer_perceptron_layer()(x) return x return _wrapper def build(self): """ Build model structure. CDSSM use Siamese architecture. """ base_network = self._create_base_network() # Left input and right input. input_left, input_right = self._make_inputs() # Process left & right input. x = [base_network(input_left), base_network(input_right)] # Dot product with cosine similarity. x = keras.layers.Dot(axes=[1, 1], normalize=True)(x) x_out = self._make_output_layer()(x) self._backend = Model(inputs=[input_left, input_right], outputs=x_out) @classmethod def get_default_preprocessor(cls): """:return: Default preprocessor.""" return preprocessors.CDSSMPreprocessor() def guess_and_fill_missing_params(self, verbose: int = 1): """ Guess and fill missing parameters in :attr:`params`. Use this method to automatically fill-in hyper parameters. This involves some guessing so the parameter it fills could be wrong. For example, the default task is `Ranking`, and if we do not set it to `Classification` manually for data packs prepared for classification, then the shape of the model output and the data will mismatch. :param verbose: Verbosity. """ self._params.get('input_shapes').set_default([(10, 30), (10, 30)], verbose) super().guess_and_fill_missing_params(verbose)
5,450
39.984962
78
py
MatchZoo
MatchZoo-master/matchzoo/models/drmm.py
"""An implementation of DRMM Model.""" import typing import keras import keras.backend as K import tensorflow as tf from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable class DRMM(BaseModel): """ DRMM Model. Examples: >>> model = DRMM() >>> model.params['mlp_num_layers'] = 1 >>> model.params['mlp_num_units'] = 5 >>> model.params['mlp_num_fan_out'] = 1 >>> model.params['mlp_activation_func'] = 'tanh' >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() >>> model.compile() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True, with_multi_layer_perceptron=True) params.add(Param(name='mask_value', value=-1, desc="The value to be masked from inputs.")) params['optimizer'] = 'adam' params['input_shapes'] = [(5,), (5, 30,)] return params def build(self): """Build model structure.""" # Scalar dimensions referenced here: # B = batch size (number of sequences) # D = embedding size # L = `input_left` sequence length # R = `input_right` sequence length # H = histogram size # K = size of top-k # Left input and right input. # query: shape = [B, L] # doc: shape = [B, L, H] # Note here, the doc is the matching histogram between original query # and original document. query = keras.layers.Input( name='text_left', shape=self._params['input_shapes'][0] ) match_hist = keras.layers.Input( name='match_histogram', shape=self._params['input_shapes'][1] ) embedding = self._make_embedding_layer() # Process left input. # shape = [B, L, D] embed_query = embedding(query) # shape = [B, L] atten_mask = tf.not_equal(query, self._params['mask_value']) # shape = [B, L] atten_mask = tf.cast(atten_mask, K.floatx()) # shape = [B, L, D] atten_mask = tf.expand_dims(atten_mask, axis=2) # shape = [B, L, D] attention_probs = self.attention_layer(embed_query, atten_mask) # Process right input. # shape = [B, L, 1] dense_output = self._make_multi_layer_perceptron_layer()(match_hist) # shape = [B, 1, 1] dot_score = keras.layers.Dot(axes=[1, 1])( [attention_probs, dense_output]) flatten_score = keras.layers.Flatten()(dot_score) x_out = self._make_output_layer()(flatten_score) self._backend = keras.Model(inputs=[query, match_hist], outputs=x_out) @classmethod def attention_layer(cls, attention_input: typing.Any, attention_mask: typing.Any = None ) -> keras.layers.Layer: """ Performs attention on the input. :param attention_input: The input tensor for attention layer. :param attention_mask: A tensor to mask the invalid values. :return: The masked output tensor. """ # shape = [B, L, 1] dense_input = keras.layers.Dense(1, use_bias=False)(attention_input) if attention_mask is not None: # Since attention_mask is 1.0 for positions we want to attend and # 0.0 for masked positions, this operation will create a tensor # which is 0.0 for positions we want to attend and -10000.0 for # masked positions. # shape = [B, L, 1] dense_input = keras.layers.Lambda( lambda x: x + (1.0 - attention_mask) * -10000.0, name="attention_mask" )(dense_input) # shape = [B, L, 1] attention_probs = keras.layers.Lambda( lambda x: tf.nn.softmax(x, axis=1), output_shape=lambda s: (s[0], s[1], s[2]), name="attention_probs" )(dense_input) return attention_probs
4,248
33.827869
78
py
MatchZoo
MatchZoo-master/matchzoo/models/duet.py
"""DUET Model.""" import keras import tensorflow as tf from matchzoo.engine import hyper_spaces from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param class DUET(BaseModel): """ DUET Model. Examples: >>> model = DUET() >>> model.params['embedding_input_dim'] = 1000 >>> model.params['embedding_output_dim'] = 300 >>> model.params['lm_filters'] = 32 >>> model.params['lm_hidden_sizes'] = [64, 32] >>> model.params['dropout_rate'] = 0.5 >>> model.params['dm_filters'] = 32 >>> model.params['dm_kernel_size'] = 3 >>> model.params['dm_d_mpool'] = 4 >>> model.params['dm_hidden_sizes'] = [64, 32] >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls): """Get default parameters.""" params = super().get_default_params(with_embedding=True) params.add(Param(name='lm_filters', value=32, desc="Filter size of 1D convolution layer in " "the local model.")) params.add(Param(name='lm_hidden_sizes', value=[32], desc="A list of hidden size of the MLP layer " "in the local model.")) params.add(Param(name='dm_filters', value=32, desc="Filter size of 1D convolution layer in " "the distributed model.")) params.add(Param(name='dm_kernel_size', value=3, desc="Kernel size of 1D convolution layer in " "the distributed model.")) params.add(Param(name='dm_q_hidden_size', value=32, desc="Hidden size of the MLP layer for the " "left text in the distributed model.")) params.add(Param(name='dm_d_mpool', value=3, desc="Max pooling size for the right text in " "the distributed model.")) params.add(Param(name='dm_hidden_sizes', value=[32], desc="A list of hidden size of the MLP layer " "in the distributed model.")) params.add(Param(name='padding', value='same', desc="The padding mode in the convolution " "layer. It should be one of `same`, " "`valid`, ""and `causal`.")) params.add(Param(name='activation_func', value='relu', desc="Activation function in the convolution" " layer.")) params.add(Param( name='dropout_rate', value=0.5, hyper_space=hyper_spaces.quniform(low=0.0, high=0.8, q=0.02), desc="The dropout rate.")) return params def build(self): """Build model.""" query, doc = self._make_inputs() embedding = self._make_embedding_layer() q_embed = embedding(query) d_embed = embedding(doc) lm_xor = keras.layers.Lambda(self._xor_match)([query, doc]) lm_conv = keras.layers.Conv1D( self._params['lm_filters'], self._params['input_shapes'][1][0], padding=self._params['padding'], activation=self._params['activation_func'] )(lm_xor) lm_conv = keras.layers.Dropout(self._params['dropout_rate'])( lm_conv) lm_feat = keras.layers.Reshape((-1,))(lm_conv) for hidden_size in self._params['lm_hidden_sizes']: lm_feat = keras.layers.Dense( hidden_size, activation=self._params['activation_func'] )(lm_feat) lm_drop = keras.layers.Dropout(self._params['dropout_rate'])( lm_feat) lm_score = keras.layers.Dense(1)(lm_drop) dm_q_conv = keras.layers.Conv1D( self._params['dm_filters'], self._params['dm_kernel_size'], padding=self._params['padding'], activation=self._params['activation_func'] )(q_embed) dm_q_conv = keras.layers.Dropout(self._params['dropout_rate'])( dm_q_conv) dm_q_mp = keras.layers.MaxPooling1D( pool_size=self._params['input_shapes'][0][0])(dm_q_conv) dm_q_rep = keras.layers.Reshape((-1,))(dm_q_mp) dm_q_rep = keras.layers.Dense(self._params['dm_q_hidden_size'])( dm_q_rep) dm_q_rep = keras.layers.Lambda(lambda x: tf.expand_dims(x, 1))( dm_q_rep) dm_d_conv1 = keras.layers.Conv1D( self._params['dm_filters'], self._params['dm_kernel_size'], padding=self._params['padding'], activation=self._params['activation_func'] )(d_embed) dm_d_conv1 = keras.layers.Dropout(self._params['dropout_rate'])( dm_d_conv1) dm_d_mp = keras.layers.MaxPooling1D( pool_size=self._params['dm_d_mpool'])(dm_d_conv1) dm_d_conv2 = keras.layers.Conv1D( self._params['dm_filters'], 1, padding=self._params['padding'], activation=self._params['activation_func'] )(dm_d_mp) dm_d_conv2 = keras.layers.Dropout(self._params['dropout_rate'])( dm_d_conv2) h_dot = keras.layers.Lambda(self._hadamard_dot)([dm_q_rep, dm_d_conv2]) dm_feat = keras.layers.Reshape((-1,))(h_dot) for hidden_size in self._params['dm_hidden_sizes']: dm_feat = keras.layers.Dense(hidden_size)(dm_feat) dm_feat_drop = keras.layers.Dropout(self._params['dropout_rate'])( dm_feat) dm_score = keras.layers.Dense(1)(dm_feat_drop) add = keras.layers.Add()([lm_score, dm_score]) out = self._make_output_layer()(add) self._backend = keras.Model(inputs=[query, doc], outputs=out) @classmethod def _xor_match(cls, x): t1 = x[0] t2 = x[1] t1_shape = t1.get_shape() t2_shape = t2.get_shape() t1_expand = tf.stack([t1] * t2_shape[1], 2) t2_expand = tf.stack([t2] * t1_shape[1], 1) out_bool = tf.equal(t1_expand, t2_expand) out = tf.cast(out_bool, tf.float32) return out @classmethod def _hadamard_dot(cls, x): x1 = x[0] x2 = x[1] out = x1 * x2 return out
6,502
39.141975
79
py
MatchZoo
MatchZoo-master/matchzoo/models/conv_knrm.py
"""ConvKNRM model.""" import keras import tensorflow as tf from .knrm import KNRM from matchzoo.engine.param import Param class ConvKNRM(KNRM): """ ConvKNRM model. Examples: >>> model = ConvKNRM() >>> model.params['embedding_input_dim'] = 10000 >>> model.params['embedding_output_dim'] = 300 >>> model.params['embedding_trainable'] = True >>> model.params['filters'] = 128 >>> model.params['conv_activation_func'] = 'tanh' >>> model.params['max_ngram'] = 3 >>> model.params['use_crossmatch'] = True >>> model.params['kernel_num'] = 11 >>> model.params['sigma'] = 0.1 >>> model.params['exact_sigma'] = 0.001 >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ def get_default_params(cls): """Get default parameters.""" params = super().get_default_params() params.add(Param(name='filters', value=128, desc="The filter size in the convolution" " layer.")) params.add(Param(name='conv_activation_func', value='relu', desc="The activation function in the " "convolution layer.")) params.add(Param(name='max_ngram', value=3, desc="The maximum length of n-grams for the " "convolution layer.")) params.add(Param(name='use_crossmatch', value=True, desc="Whether to match left n-grams and right " "n-grams of different lengths")) return params def build(self): """Build model.""" query, doc = self._make_inputs() embedding = self._make_embedding_layer() q_embed = embedding(query) d_embed = embedding(doc) q_convs = [] d_convs = [] for i in range(self._params['max_ngram']): c = keras.layers.Conv1D( self._params['filters'], i + 1, activation=self._params['conv_activation_func'], padding='same' ) q_convs.append(c(q_embed)) d_convs.append(c(d_embed)) KM = [] for qi in range(self._params['max_ngram']): for di in range(self._params['max_ngram']): # do not match n-gram with different length if use crossmatch if not self._params['use_crossmatch'] and qi != di: continue q_ngram = q_convs[qi] d_ngram = d_convs[di] mm = keras.layers.Dot(axes=[2, 2], normalize=True)([q_ngram, d_ngram]) for i in range(self._params['kernel_num']): mu = 1. / (self._params['kernel_num'] - 1) + (2. * i) / ( self._params['kernel_num'] - 1) - 1.0 sigma = self._params['sigma'] if mu > 1.0: sigma = self._params['exact_sigma'] mu = 1.0 mm_exp = self._kernel_layer(mu, sigma)(mm) mm_doc_sum = keras.layers.Lambda( lambda x: tf.reduce_sum(x, 2))( mm_exp) mm_log = keras.layers.Activation(tf.math.log1p)(mm_doc_sum) mm_sum = keras.layers.Lambda( lambda x: tf.reduce_sum(x, 1))(mm_log) KM.append(mm_sum) phi = keras.layers.Lambda(lambda x: tf.stack(x, 1))(KM) out = self._make_output_layer()(phi) self._backend = keras.Model(inputs=[query, doc], outputs=[out])
3,736
37.132653
79
py
MatchZoo
MatchZoo-master/matchzoo/models/dssm.py
"""An implementation of DSSM, Deep Structured Semantic Model.""" from keras.models import Model from keras.layers import Input, Dot from matchzoo.engine.param_table import ParamTable from matchzoo.engine.base_model import BaseModel from matchzoo import preprocessors class DSSM(BaseModel): """ Deep structured semantic model. Examples: >>> model = DSSM() >>> model.params['mlp_num_layers'] = 3 >>> model.params['mlp_num_units'] = 300 >>> model.params['mlp_num_fan_out'] = 128 >>> model.params['mlp_activation_func'] = 'relu' >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_multi_layer_perceptron=True) return params def build(self): """ Build model structure. DSSM use Siamese arthitecture. """ dim_triletter = self._params['input_shapes'][0][0] input_shape = (dim_triletter,) base_network = self._make_multi_layer_perceptron_layer() # Left input and right input. input_left = Input(name='text_left', shape=input_shape) input_right = Input(name='text_right', shape=input_shape) # Process left & right input. x = [base_network(input_left), base_network(input_right)] # Dot product with cosine similarity. x = Dot(axes=[1, 1], normalize=True)(x) x_out = self._make_output_layer()(x) self._backend = Model( inputs=[input_left, input_right], outputs=x_out) @classmethod def get_default_preprocessor(cls): """:return: Default preprocessor.""" return preprocessors.DSSMPreprocessor()
1,847
31.421053
77
py
MatchZoo
MatchZoo-master/matchzoo/models/match_pyramid.py
"""An implementation of MatchPyramid Model.""" import typing import keras import matchzoo from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo.engine import hyper_spaces class MatchPyramid(BaseModel): """ MatchPyramid Model. Examples: >>> model = MatchPyramid() >>> model.params['embedding_output_dim'] = 300 >>> model.params['num_blocks'] = 2 >>> model.params['kernel_count'] = [16, 32] >>> model.params['kernel_size'] = [[3, 3], [3, 3]] >>> model.params['dpool_size'] = [3, 10] >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params.add(Param(name='num_blocks', value=1, desc="Number of convolution blocks.")) params.add(Param(name='kernel_count', value=[32], desc="The kernel count of the 2D convolution " "of each block.")) params.add(Param(name='kernel_size', value=[[3, 3]], desc="The kernel size of the 2D convolution " "of each block.")) params.add(Param(name='activation', value='relu', desc="The activation function.")) params.add(Param(name='dpool_size', value=[3, 10], desc="The max-pooling size of each block.")) params.add(Param( name='padding', value='same', desc="The padding mode in the convolution layer." )) params.add(Param( name='dropout_rate', value=0.0, hyper_space=hyper_spaces.quniform(low=0.0, high=0.8, q=0.01), desc="The dropout rate." )) return params def build(self): """ Build model structure. MatchPyramid text matching as image recognition. """ input_left, input_right = self._make_inputs() input_dpool_index = keras.layers.Input( name='dpool_index', shape=[self._params['input_shapes'][0][0], self._params['input_shapes'][1][0], 2], dtype='int32') embedding = self._make_embedding_layer() embed_left = embedding(input_left) embed_right = embedding(input_right) # Interaction matching_layer = matchzoo.layers.MatchingLayer(matching_type='dot') embed_cross = matching_layer([embed_left, embed_right]) for i in range(self._params['num_blocks']): embed_cross = self._conv_block( embed_cross, self._params['kernel_count'][i], self._params['kernel_size'][i], self._params['padding'], self._params['activation'] ) # Dynamic Pooling dpool_layer = matchzoo.layers.DynamicPoolingLayer( *self._params['dpool_size']) embed_pool = dpool_layer([embed_cross, input_dpool_index]) embed_flat = keras.layers.Flatten()(embed_pool) x = keras.layers.Dropout(rate=self._params['dropout_rate'])(embed_flat) inputs = [input_left, input_right, input_dpool_index] x_out = self._make_output_layer()(x) self._backend = keras.Model(inputs=inputs, outputs=x_out) @classmethod def _conv_block( cls, x, kernel_count: int, kernel_size: int, padding: str, activation: str ) -> typing.Any: output = keras.layers.Conv2D(kernel_count, kernel_size, padding=padding, activation=activation)(x) return output
4,014
34.530973
79
py
MatchZoo
MatchZoo-master/matchzoo/models/arci.py
"""An implementation of ArcI Model.""" import typing import keras from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo.engine import hyper_spaces class ArcI(BaseModel): """ ArcI Model. Examples: >>> model = ArcI() >>> model.params['num_blocks'] = 1 >>> model.params['left_filters'] = [32] >>> model.params['right_filters'] = [32] >>> model.params['left_kernel_sizes'] = [3] >>> model.params['right_kernel_sizes'] = [3] >>> model.params['left_pool_sizes'] = [2] >>> model.params['right_pool_sizes'] = [4] >>> model.params['conv_activation_func'] = 'relu' >>> model.params['mlp_num_layers'] = 1 >>> model.params['mlp_num_units'] = 64 >>> model.params['mlp_num_fan_out'] = 32 >>> model.params['mlp_activation_func'] = 'relu' >>> model.params['dropout_rate'] = 0.5 >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params( with_embedding=True, with_multi_layer_perceptron=True ) params['optimizer'] = 'adam' params.add(Param(name='num_blocks', value=1, desc="Number of convolution blocks.")) params.add(Param(name='left_filters', value=[32], desc="The filter size of each convolution " "blocks for the left input.")) params.add(Param(name='left_kernel_sizes', value=[3], desc="The kernel size of each convolution " "blocks for the left input.")) params.add(Param(name='right_filters', value=[32], desc="The filter size of each convolution " "blocks for the right input.")) params.add(Param(name='right_kernel_sizes', value=[3], desc="The kernel size of each convolution " "blocks for the right input.")) params.add(Param(name='conv_activation_func', value='relu', desc="The activation function in the " "convolution layer.")) params.add(Param(name='left_pool_sizes', value=[2], desc="The pooling size of each convolution " "blocks for the left input.")) params.add(Param(name='right_pool_sizes', value=[2], desc="The pooling size of each convolution " "blocks for the right input.")) params.add(Param( name='padding', value='same', hyper_space=hyper_spaces.choice( ['same', 'valid', 'causal']), desc="The padding mode in the convolution layer. It should be one" "of `same`, `valid`, and `causal`." )) params.add(Param( 'dropout_rate', 0.0, hyper_space=hyper_spaces.quniform( low=0.0, high=0.8, q=0.01), desc="The dropout rate." )) return params def build(self): """ Build model structure. ArcI use Siamese arthitecture. """ input_left, input_right = self._make_inputs() embedding = self._make_embedding_layer() embed_left = embedding(input_left) embed_right = embedding(input_right) for i in range(self._params['num_blocks']): embed_left = self._conv_pool_block( embed_left, self._params['left_filters'][i], self._params['left_kernel_sizes'][i], self._params['padding'], self._params['conv_activation_func'], self._params['left_pool_sizes'][i] ) embed_right = self._conv_pool_block( embed_right, self._params['right_filters'][i], self._params['right_kernel_sizes'][i], self._params['padding'], self._params['conv_activation_func'], self._params['right_pool_sizes'][i] ) rep_left = keras.layers.Flatten()(embed_left) rep_right = keras.layers.Flatten()(embed_right) concat = keras.layers.Concatenate(axis=1)([rep_left, rep_right]) dropout = keras.layers.Dropout( rate=self._params['dropout_rate'])(concat) mlp = self._make_multi_layer_perceptron_layer()(dropout) inputs = [input_left, input_right] x_out = self._make_output_layer()(mlp) self._backend = keras.Model(inputs=inputs, outputs=x_out) def _conv_pool_block( self, input_: typing.Any, filters: int, kernel_size: int, padding: str, conv_activation_func: str, pool_size: int ) -> typing.Any: output = keras.layers.Conv1D( filters, kernel_size, padding=padding, activation=conv_activation_func )(input_) output = keras.layers.MaxPooling1D(pool_size=pool_size)(output) return output
5,386
37.205674
78
py
MatchZoo
MatchZoo-master/matchzoo/models/mvlstm.py
"""An implementation of MVLSTM Model.""" import keras import tensorflow as tf from matchzoo.engine import hyper_spaces from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable class MVLSTM(BaseModel): """ MVLSTM Model. Examples: >>> model = MVLSTM() >>> model.params['lstm_units'] = 32 >>> model.params['top_k'] = 50 >>> model.params['mlp_num_layers'] = 2 >>> model.params['mlp_num_units'] = 20 >>> model.params['mlp_num_fan_out'] = 10 >>> model.params['mlp_activation_func'] = 'relu' >>> model.params['dropout_rate'] = 0.5 >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params( with_embedding=True, with_multi_layer_perceptron=True) params.add(Param(name='lstm_units', value=32, desc="Integer, the hidden size in the " "bi-directional LSTM layer.")) params.add(Param(name='dropout_rate', value=0.0, desc="Float, the dropout rate.")) params.add(Param( 'top_k', value=10, hyper_space=hyper_spaces.quniform(low=2, high=100), desc="Integer, the size of top-k pooling layer." )) params['optimizer'] = 'adam' return params def build(self): """Build model structure.""" query, doc = self._make_inputs() # Embedding layer embedding = self._make_embedding_layer(mask_zero=True) embed_query = embedding(query) embed_doc = embedding(doc) # Bi-directional LSTM layer rep_query = keras.layers.Bidirectional(keras.layers.LSTM( self._params['lstm_units'], return_sequences=True, dropout=self._params['dropout_rate'] ))(embed_query) rep_doc = keras.layers.Bidirectional(keras.layers.LSTM( self._params['lstm_units'], return_sequences=True, dropout=self._params['dropout_rate'] ))(embed_doc) # Top-k matching layer matching_matrix = keras.layers.Dot( axes=[2, 2], normalize=False)([rep_query, rep_doc]) matching_signals = keras.layers.Reshape((-1,))(matching_matrix) matching_topk = keras.layers.Lambda( lambda x: tf.nn.top_k(x, k=self._params['top_k'], sorted=True)[0] )(matching_signals) # Multilayer perceptron layer. mlp = self._make_multi_layer_perceptron_layer()(matching_topk) mlp = keras.layers.Dropout( rate=self._params['dropout_rate'])(mlp) x_out = self._make_output_layer()(mlp) self._backend = keras.Model(inputs=[query, doc], outputs=x_out)
2,963
34.285714
77
py
MatchZoo
MatchZoo-master/matchzoo/models/anmm.py
"""An implementation of aNMM Model.""" import keras from keras.activations import softmax from keras.initializers import RandomUniform from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo.engine import hyper_spaces class ANMM(BaseModel): """ ANMM Model. Examples: >>> model = ANMM() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params.add(Param( name='dropout_rate', value=0.1, desc="The dropout rate.", hyper_space=hyper_spaces.quniform(0, 1, 0.05) )) params.add(Param( name='num_layers', value=2, desc="Number of hidden layers in the MLP " "layer." )) params.add(Param( name='hidden_sizes', value=[30, 30], desc="Number of hidden size for each hidden" " layer" )) return params def build(self): """ Build model structure. aNMM model based on bin weighting and query term attentions """ # query is [batch_size, left_text_len] # doc is [batch_size, right_text_len, bin_num] query, doc = self._make_inputs() embedding = self._make_embedding_layer() q_embed = embedding(query) q_attention = keras.layers.Dense( 1, kernel_initializer=RandomUniform(), use_bias=False)(q_embed) q_text_len = self._params['input_shapes'][0][0] q_attention = keras.layers.Lambda( lambda x: softmax(x, axis=1), output_shape=(q_text_len,) )(q_attention) d_bin = keras.layers.Dropout( rate=self._params['dropout_rate'])(doc) for layer_id in range(self._params['num_layers'] - 1): d_bin = keras.layers.Dense( self._params['hidden_sizes'][layer_id], kernel_initializer=RandomUniform())(d_bin) d_bin = keras.layers.Activation('tanh')(d_bin) d_bin = keras.layers.Dense( self._params['hidden_sizes'][self._params['num_layers'] - 1])( d_bin) d_bin = keras.layers.Reshape((q_text_len,))(d_bin) q_attention = keras.layers.Reshape((q_text_len,))(q_attention) score = keras.layers.Dot(axes=[1, 1])([d_bin, q_attention]) x_out = self._make_output_layer()(score) self._backend = keras.Model(inputs=[query, doc], outputs=x_out)
2,720
33.0125
75
py
MatchZoo
MatchZoo-master/matchzoo/models/drmmtks.py
"""An implementation of DRMMTKS Model.""" import typing import keras import tensorflow as tf from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo.engine import hyper_spaces class DRMMTKS(BaseModel): """ DRMMTKS Model. Examples: >>> model = DRMMTKS() >>> model.params['embedding_input_dim'] = 10000 >>> model.params['embedding_output_dim'] = 100 >>> model.params['top_k'] = 20 >>> model.params['mlp_num_layers'] = 1 >>> model.params['mlp_num_units'] = 5 >>> model.params['mlp_num_fan_out'] = 1 >>> model.params['mlp_activation_func'] = 'tanh' >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params( with_embedding=True, with_multi_layer_perceptron=True ) params.add(Param(name='mask_value', value=-1, desc="The value to be masked from inputs.")) params['input_shapes'] = [(5,), (300,)] params.add(Param( 'top_k', value=10, hyper_space=hyper_spaces.quniform(low=2, high=100), desc="Size of top-k pooling layer." )) return params def build(self): """Build model structure.""" # Scalar dimensions referenced here: # B = batch size (number of sequences) # D = embedding size # L = `input_left` sequence length # R = `input_right` sequence length # K = size of top-k # Left input and right input. # shape = [B, L] # shape = [B, R] query, doc = self._make_inputs() embedding = self._make_embedding_layer() # Process left input. # shape = [B, L, D] embed_query = embedding(query) # shape = [B, R, D] embed_doc = embedding(doc) # shape = [B, L] atten_mask = tf.not_equal(query, self._params['mask_value']) # shape = [B, L] atten_mask = tf.cast(atten_mask, keras.backend.floatx()) # shape = [B, L, 1] atten_mask = tf.expand_dims(atten_mask, axis=2) # shape = [B, L, 1] attention_probs = self.attention_layer(embed_query, atten_mask) # Matching histogram of top-k # shape = [B, L, R] matching_matrix = keras.layers.Dot(axes=[2, 2], normalize=True)( [embed_query, embed_doc]) # shape = [B, L, K] effective_top_k = min(self._params['top_k'], self.params['input_shapes'][0][0], self.params['input_shapes'][1][0]) matching_topk = keras.layers.Lambda( lambda x: tf.nn.top_k(x, k=effective_top_k, sorted=True)[0] )(matching_matrix) # Process right input. # shape = [B, L, 1] dense_output = self._make_multi_layer_perceptron_layer()(matching_topk) # shape = [B, 1, 1] dot_score = keras.layers.Dot(axes=[1, 1])( [attention_probs, dense_output]) flatten_score = keras.layers.Flatten()(dot_score) x_out = self._make_output_layer()(flatten_score) self._backend = keras.Model(inputs=[query, doc], outputs=x_out) @classmethod def attention_layer(cls, attention_input: typing.Any, attention_mask: typing.Any = None ) -> keras.layers.Layer: """ Performs attention on the input. :param attention_input: The input tensor for attention layer. :param attention_mask: A tensor to mask the invalid values. :return: The masked output tensor. """ # shape = [B, L, 1] dense_input = keras.layers.Dense(1, use_bias=False)(attention_input) if attention_mask is not None: # Since attention_mask is 1.0 for positions we want to attend and # 0.0 for masked positions, this operation will create a tensor # which is 0.0 for positions we want to attend and -10000.0 for # masked positions. # shape = [B, L, 1] dense_input = keras.layers.Lambda( lambda x: x + (1.0 - attention_mask) * -10000.0, name="attention_mask" )(dense_input) # shape = [B, L, 1] attention_probs = keras.layers.Lambda( lambda x: tf.nn.softmax(x, axis=1), output_shape=lambda s: (s[0], s[1], s[2]), name="attention_probs" )(dense_input) return attention_probs
4,766
34.311111
79
py
MatchZoo
MatchZoo-master/matchzoo/models/arcii.py
"""An implementation of ArcII Model.""" import typing import keras import matchzoo from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo.engine import hyper_spaces class ArcII(BaseModel): """ ArcII Model. Examples: >>> model = ArcII() >>> model.params['embedding_output_dim'] = 300 >>> model.params['num_blocks'] = 2 >>> model.params['kernel_1d_count'] = 32 >>> model.params['kernel_1d_size'] = 3 >>> model.params['kernel_2d_count'] = [16, 32] >>> model.params['kernel_2d_size'] = [[3, 3], [3, 3]] >>> model.params['pool_2d_size'] = [[2, 2], [2, 2]] >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params['optimizer'] = 'adam' opt_space = hyper_spaces.choice(['adam', 'rmsprop', 'adagrad']) params.get('optimizer').hyper_space = opt_space params.add(Param(name='num_blocks', value=1, desc="Number of 2D convolution blocks.")) params.add(Param(name='kernel_1d_count', value=32, desc="Kernel count of 1D convolution layer.")) params.add(Param(name='kernel_1d_size', value=3, desc="Kernel size of 1D convolution layer.")) params.add(Param(name='kernel_2d_count', value=[32], desc="Kernel count of 2D convolution layer in" "each block")) params.add(Param(name='kernel_2d_size', value=[[3, 3]], desc="Kernel size of 2D convolution layer in" " each block.")) params.add(Param(name='activation', value='relu', desc="Activation function.")) params.add(Param(name='pool_2d_size', value=[[2, 2]], desc="Size of pooling layer in each block.")) params.add(Param( name='padding', value='same', hyper_space=hyper_spaces.choice( ['same', 'valid']), desc="The padding mode in the convolution layer. It should be one" "of `same`, `valid`." )) params.add(Param( name='dropout_rate', value=0.0, hyper_space=hyper_spaces.quniform(low=0.0, high=0.8, q=0.01), desc="The dropout rate." )) return params def build(self): """ Build model structure. ArcII has the desirable property of letting two sentences meet before their own high-level representations mature. """ input_left, input_right = self._make_inputs() embedding = self._make_embedding_layer() embed_left = embedding(input_left) embed_right = embedding(input_right) # Phrase level representations conv_1d_left = keras.layers.Conv1D( self._params['kernel_1d_count'], self._params['kernel_1d_size'], padding=self._params['padding'] )(embed_left) conv_1d_right = keras.layers.Conv1D( self._params['kernel_1d_count'], self._params['kernel_1d_size'], padding=self._params['padding'] )(embed_right) # Interaction matching_layer = matchzoo.layers.MatchingLayer(matching_type='plus') embed_cross = matching_layer([conv_1d_left, conv_1d_right]) for i in range(self._params['num_blocks']): embed_cross = self._conv_pool_block( embed_cross, self._params['kernel_2d_count'][i], self._params['kernel_2d_size'][i], self._params['padding'], self._params['activation'], self._params['pool_2d_size'][i] ) embed_flat = keras.layers.Flatten()(embed_cross) x = keras.layers.Dropout(rate=self._params['dropout_rate'])(embed_flat) inputs = [input_left, input_right] x_out = self._make_output_layer()(x) self._backend = keras.Model(inputs=inputs, outputs=x_out) @classmethod def _conv_pool_block( cls, x, kernel_count: int, kernel_size: int, padding: str, activation: str, pool_size: int ) -> typing.Any: output = keras.layers.Conv2D(kernel_count, kernel_size, padding=padding, activation=activation)(x) output = keras.layers.MaxPooling2D(pool_size=pool_size)(output) return output
4,891
36.630769
79
py
MatchZoo
MatchZoo-master/matchzoo/models/knrm.py
"""KNRM model.""" import keras import tensorflow as tf from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine import hyper_spaces class KNRM(BaseModel): """ KNRM model. Examples: >>> model = KNRM() >>> model.params['embedding_input_dim'] = 10000 >>> model.params['embedding_output_dim'] = 10 >>> model.params['embedding_trainable'] = True >>> model.params['kernel_num'] = 11 >>> model.params['sigma'] = 0.1 >>> model.params['exact_sigma'] = 0.001 >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls): """Get default parameters.""" params = super().get_default_params(with_embedding=True) params.add(Param( name='kernel_num', value=11, hyper_space=hyper_spaces.quniform(low=5, high=20), desc="The number of RBF kernels." )) params.add(Param( name='sigma', value=0.1, hyper_space=hyper_spaces.quniform( low=0.01, high=0.2, q=0.01), desc="The `sigma` defines the kernel width." )) params.add(Param( name='exact_sigma', value=0.001, desc="The `exact_sigma` denotes the `sigma` " "for exact match." )) return params def build(self): """Build model.""" query, doc = self._make_inputs() embedding = self._make_embedding_layer() q_embed = embedding(query) d_embed = embedding(doc) mm = keras.layers.Dot(axes=[2, 2], normalize=True)([q_embed, d_embed]) KM = [] for i in range(self._params['kernel_num']): mu = 1. / (self._params['kernel_num'] - 1) + (2. * i) / ( self._params['kernel_num'] - 1) - 1.0 sigma = self._params['sigma'] if mu > 1.0: sigma = self._params['exact_sigma'] mu = 1.0 mm_exp = self._kernel_layer(mu, sigma)(mm) mm_doc_sum = keras.layers.Lambda( lambda x: tf.reduce_sum(x, 2))(mm_exp) mm_log = keras.layers.Activation(tf.math.log1p)(mm_doc_sum) mm_sum = keras.layers.Lambda( lambda x: tf.reduce_sum(x, 1))(mm_log) KM.append(mm_sum) phi = keras.layers.Lambda(lambda x: tf.stack(x, 1))(KM) out = self._make_output_layer()(phi) self._backend = keras.Model(inputs=[query, doc], outputs=[out]) @classmethod def _kernel_layer(cls, mu: float, sigma: float) -> keras.layers.Layer: """ Gaussian kernel layer in KNRM. :param mu: Float, mean of the kernel. :param sigma: Float, sigma of the kernel. :return: `keras.layers.Layer`. """ def kernel(x): return tf.math.exp(-0.5 * (x - mu) * (x - mu) / sigma / sigma) return keras.layers.Activation(kernel)
3,057
31.189474
78
py
MatchZoo
MatchZoo-master/matchzoo/models/naive.py
"""Naive model with a simplest structure for testing purposes.""" import keras from matchzoo.engine.base_model import BaseModel from matchzoo.engine import hyper_spaces class Naive(BaseModel): """ Naive model with a simplest structure for testing purposes. Bare minimum functioning model. The best choice to get things rolling. The worst choice to fit and evaluate performance. """ @classmethod def get_default_params(cls): """Default parameters.""" params = super().get_default_params() params.get('optimizer').hyper_space = \ hyper_spaces.choice(['adam', 'adagrad', 'rmsprop']) return params def build(self): """Build.""" x_in = self._make_inputs() x = keras.layers.concatenate(x_in) x_out = self._make_output_layer()(x) self._backend = keras.models.Model(inputs=x_in, outputs=x_out)
909
28.354839
74
py
MatchZoo
MatchZoo-master/matchzoo/models/dense_baseline.py
"""A simple densely connected baseline model.""" import keras.layers from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param_table import ParamTable from matchzoo.engine import hyper_spaces class DenseBaseline(BaseModel): """ A simple densely connected baseline model. Examples: >>> model = DenseBaseline() >>> model.params['mlp_num_layers'] = 2 >>> model.params['mlp_num_units'] = 300 >>> model.params['mlp_num_fan_out'] = 128 >>> model.params['mlp_activation_func'] = 'relu' >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() >>> model.compile() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_multi_layer_perceptron=True) params['mlp_num_units'] = 256 params.get('mlp_num_units').hyper_space = \ hyper_spaces.quniform(16, 512) params.get('mlp_num_layers').hyper_space = \ hyper_spaces.quniform(1, 5) return params def build(self): """Model structure.""" x_in = self._make_inputs() x = keras.layers.concatenate(x_in) x = self._make_multi_layer_perceptron_layer()(x) x_out = self._make_output_layer()(x) self._backend = keras.models.Model(inputs=x_in, outputs=x_out)
1,420
31.295455
77
py
MatchZoo
MatchZoo-master/matchzoo/datasets/snli/load_data.py
"""SNLI data loader.""" import typing from pathlib import Path import pandas as pd import keras import matchzoo _url = "https://nlp.stanford.edu/projects/snli/snli_1.0.zip" def load_data( stage: str = 'train', task: str = 'classification', target_label: str = 'entailment', return_classes: bool = False ) -> typing.Union[matchzoo.DataPack, tuple]: """ Load SNLI data. :param stage: One of `train`, `dev`, and `test`. (default: `train`) :param task: Could be one of `ranking`, `classification` or a :class:`matchzoo.engine.BaseTask` instance. (default: `ranking`) :param target_label: If `ranking`, chose one of `entailment`, `contradiction`, `neutral`, and `-` as the positive label. (default: `entailment`) :param return_classes: `True` to return classes for classification task, `False` otherwise. :return: A DataPack unless `task` is `classificiation` and `return_classes` is `True`: a tuple of `(DataPack, classes)` in that case. """ if stage not in ('train', 'dev', 'test'): raise ValueError(f"{stage} is not a valid stage." f"Must be one of `train`, `dev`, and `test`.") data_root = _download_data() file_path = data_root.joinpath(f'snli_1.0_{stage}.txt') data_pack = _read_data(file_path) if task == 'ranking': task = matchzoo.tasks.Ranking() if task == 'classification': task = matchzoo.tasks.Classification() if isinstance(task, matchzoo.tasks.Ranking): if target_label not in ['entailment', 'contradiction', 'neutral', '-']: raise ValueError(f"{target_label} is not a valid target label." f"Must be one of `entailment`, `contradiction`, " f"`neutral` and `-`.") binary = (data_pack.relation['label'] == target_label).astype(float) data_pack.relation['label'] = binary return data_pack elif isinstance(task, matchzoo.tasks.Classification): classes = ['entailment', 'contradiction', 'neutral', '-'] label = data_pack.relation['label'].apply(classes.index) data_pack.relation['label'] = label data_pack.one_hot_encode_label(num_classes=4, inplace=True) if return_classes: return data_pack, classes else: return data_pack else: raise ValueError(f"{task} is not a valid task." f"Must be one of `Ranking` and `Classification`.") def _download_data(): ref_path = keras.utils.data_utils.get_file( 'snli', _url, extract=True, cache_dir=matchzoo.USER_DATA_DIR, cache_subdir='snli' ) return Path(ref_path).parent.joinpath('snli_1.0') def _read_data(path): table = pd.read_csv(path, sep='\t') df = pd.DataFrame({ 'text_left': table['sentence1'], 'text_right': table['sentence2'], 'label': table['gold_label'] }) df = df.dropna(axis=0, how='any').reset_index(drop=True) return matchzoo.pack(df)
3,067
33.863636
79
py
MatchZoo
MatchZoo-master/matchzoo/datasets/quora_qp/load_data.py
"""Quora Question Pairs data loader.""" import typing from pathlib import Path import keras import pandas as pd import matchzoo _url = "https://firebasestorage.googleapis.com/v0/b/mtl-sentence" \ "-representations.appspot.com/o/data%2FQQP.zip?alt=media&" \ "token=700c6acf-160d-4d89-81d1-de4191d02cb5" def load_data( stage: str = 'train', task: str = 'classification', return_classes: bool = False, ) -> typing.Union[matchzoo.DataPack, tuple]: """ Load QuoraQP data. :param path: `None` for download from quora, specific path for downloaded data. :param stage: One of `train`, `dev`, and `test`. :param task: Could be one of `ranking`, `classification` or a :class:`matchzoo.engine.BaseTask` instance. :param return_classes: Whether return classes for classification task. :return: A DataPack if `ranking`, a tuple of (DataPack, classes) if `classification`. """ if stage not in ('train', 'dev', 'test'): raise ValueError(f"{stage} is not a valid stage." f"Must be one of `train`, `dev`, and `test`.") data_root = _download_data() file_path = data_root.joinpath(f"{stage}.tsv") data_pack = _read_data(file_path, stage) if task == 'ranking': task = matchzoo.tasks.Ranking() elif task == 'classification': task = matchzoo.tasks.Classification() if isinstance(task, matchzoo.tasks.Ranking): return data_pack elif isinstance(task, matchzoo.tasks.Classification): if stage != 'test': data_pack.one_hot_encode_label(num_classes=2, inplace=True) if return_classes: return data_pack, [False, True] else: return data_pack else: raise ValueError(f"{task} is not a valid task.") def _download_data(): ref_path = keras.utils.data_utils.get_file( 'quora_qp', _url, extract=True, cache_dir=matchzoo.USER_DATA_DIR, cache_subdir='quora_qp' ) return Path(ref_path).parent.joinpath('QQP') def _read_data(path, stage): data = pd.read_csv(path, sep='\t', error_bad_lines=False) data = data.dropna(axis=0, how='any').reset_index(drop=True) if stage in ['train', 'dev']: df = pd.DataFrame({ 'id_left': data['qid1'], 'id_right': data['qid2'], 'text_left': data['question1'], 'text_right': data['question2'], 'label': data['is_duplicate'].astype(int) }) else: df = pd.DataFrame({ 'text_left': data['question1'], 'text_right': data['question2'] }) return matchzoo.pack(df)
2,677
30.505882
74
py
MatchZoo
MatchZoo-master/matchzoo/datasets/cqa_ql_16/load_data.py
"""CQA-QL-16 data loader.""" import xml import typing from pathlib import Path import keras import pandas as pd import matchzoo _train_dev_url = "http://alt.qcri.org/semeval2016/task3/data/uploads/" \ "semeval2016-task3-cqa-ql-traindev-v3.2.zip" _test_url = "http://alt.qcri.org/semeval2016/task3/data/uploads/" \ "semeval2016_task3_test.zip" def load_data( stage: str = 'train', task: str = 'classification', target_label: str = 'PerfectMatch', return_classes: bool = False, match_type: str = 'question', mode: str = 'both', ) -> typing.Union[matchzoo.DataPack, tuple]: """ Load CQA-QL-16 data. :param stage: One of `train`, `dev`, and `test`. (default: `train`) :param task: Could be one of `ranking`, `classification` or instance of :class:`matchzoo.engine.BaseTask`. (default: `classification`) :param target_label: If `ranking`, choose one of classification label as the positive label. (default: `PerfectMatch`) :param return_classes: `True` to return classes for classification task, `False` otherwise. :param match_type: Matching text types. One of `question`, `answer`, and `external_answer`. (default: `question`) :param mode: Train data use method. One of `part1`, `part2`, and `both`. (default: `both`) :return: A DataPack unless `task` is `classification` and `return_classes` is `True`: a tuple of `(DataPack, classes)` in that case. """ if stage not in ('train', 'dev', 'test'): raise ValueError(f"{stage} is not a valid stage." f"Must be one of `train`, `dev`, and `test`.") if match_type not in ('question', 'answer', 'external_answer'): raise ValueError(f"{match_type} is not a valid method. Must be one of" f" `question`, `answer`, `external_answer`.") if mode not in ('part1', 'part2', 'both'): raise ValueError(f"{mode} is not a valid method." f"Must be one of `part1`, `part2`, `both`.") data_root = _download_data(stage) data_pack = _read_data(data_root, stage, match_type, mode) if task == 'ranking': if match_type in ('anwer', 'external_answer') and target_label not in [ 'Good', 'PotentiallyUseful', 'Bad']: raise ValueError(f"{target_label} is not a valid target label." f"Must be one of `Good`, `PotentiallyUseful`," f" `Bad`.") elif match_type == 'question' and target_label not in [ 'PerfectMatch', 'Relevant', 'Irrelevant']: raise ValueError(f"{target_label} is not a valid target label." f" Must be one of `PerfectMatch`, `Relevant`," f" `Irrelevant`.") binary = (data_pack.relation['label'] == target_label).astype(float) data_pack.relation['label'] = binary return data_pack elif task == 'classification': if match_type in ('answer', 'external_answer'): classes = ['Good', 'PotentiallyUseful', 'Bad'] else: classes = ['PerfectMatch', 'Relevant', 'Irrelevant'] label = data_pack.relation['label'].apply(classes.index) data_pack.relation['label'] = label data_pack.one_hot_encode_label(num_classes=3, inplace=True) if return_classes: return data_pack, classes else: return data_pack else: raise ValueError(f"{task} is not a valid task." f"Must be one of `Ranking` and `Classification`.") def _download_data(stage): if stage in ['train', 'dev']: return _download_train_dev_data() else: return _download_test_data() def _download_train_dev_data(): ref_path = keras.utils.data_utils.get_file( 'semeval_train', _train_dev_url, extract=True, cache_dir=matchzoo.USER_DATA_DIR, cache_subdir='semeval_train' ) return Path(ref_path).parent.joinpath('v3.2') def _download_test_data(): ref_path = keras.utils.data_utils.get_file( 'semeval_test', _test_url, extract=True, cache_dir=matchzoo.USER_DATA_DIR, cache_subdir='semeval_test' ) return Path(ref_path).parent.joinpath('SemEval2016_task3_test/English') def _read_data(path, stage, match_type, mode='both'): if stage == 'train': if mode == 'part1': path = path.joinpath( 'train/SemEval2016-Task3-CQA-QL-train-part1.xml') data = _load_data_by_type(path, match_type) elif mode == 'part2': path = path.joinpath( 'train/SemEval2016-Task3-CQA-QL-train-part2.xml') data = _load_data_by_type(path, match_type) else: part1 = path.joinpath( 'train/SemEval2016-Task3-CQA-QL-train-part1.xml') p1 = _load_data_by_type(part1, match_type) part2 = path.joinpath( 'train/SemEval2016-Task3-CQA-QL-train-part1.xml') p2 = _load_data_by_type(part2, match_type) data = pd.concat([p1, p2], ignore_index=True) return matchzoo.pack(data) elif stage == 'dev': path = path.joinpath('dev/SemEval2016-Task3-CQA-QL-dev.xml') data = _load_data_by_type(path, match_type) return matchzoo.pack(data) else: path = path.joinpath('SemEval2016-Task3-CQA-QL-test.xml') data = _load_data_by_type(path, match_type) return matchzoo.pack(data) def _load_data_by_type(path, match_type): if match_type == 'question': return _load_question(path) elif match_type == 'answer': return _load_answer(path) else: return _load_external_answer(path) def _load_question(path): doc = xml.etree.ElementTree.parse(path) dataset = [] for question in doc.iterfind('OrgQuestion'): qid = question.attrib['ORGQ_ID'] query = question.findtext('OrgQBody') rel_question = question.find('Thread').find('RelQuestion') question = rel_question.findtext('RelQBody') question_id = rel_question.attrib['RELQ_ID'] dataset.append([qid, question_id, query, question, rel_question.attrib['RELQ_RELEVANCE2ORGQ']]) df = pd.DataFrame(dataset, columns=[ 'id_left', 'id_right', 'text_left', 'text_right', 'label']) return df def _load_answer(path): doc = xml.etree.ElementTree.parse(path) dataset = [] for org_q in doc.iterfind('OrgQuestion'): for thread in org_q.iterfind('Thread'): ques = thread.find('RelQuestion') qid = ques.attrib['RELQ_ID'] question = ques.findtext('RelQBody') for comment in thread.iterfind('RelComment'): aid = comment.attrib['RELC_ID'] answer = comment.findtext('RelCText') dataset.append([qid, aid, question, answer, comment.attrib['RELC_RELEVANCE2RELQ']]) df = pd.DataFrame(dataset, columns=[ 'id_left', 'id_right', 'text_left', 'text_right', 'label']) return df def _load_external_answer(path): doc = xml.etree.ElementTree.parse(path) dataset = [] for question in doc.iterfind('OrgQuestion'): qid = question.attrib['ORGQ_ID'] query = question.findtext('OrgQBody') thread = question.find('Thread') for comment in thread.iterfind('RelComment'): answer = comment.findtext('RelCText') aid = comment.attrib['RELC_ID'] dataset.append([qid, aid, query, answer, comment.attrib['RELC_RELEVANCE2ORGQ']]) df = pd.DataFrame(dataset, columns=[ 'id_left', 'id_right', 'text_left', 'text_right', 'label']) return df
7,865
37.558824
79
py
MatchZoo
MatchZoo-master/matchzoo/datasets/embeddings/load_glove_embedding.py
"""Embedding data loader.""" from pathlib import Path import keras import matchzoo as mz _glove_embedding_url = "http://nlp.stanford.edu/data/glove.6B.zip" def load_glove_embedding(dimension: int = 50) -> mz.embedding.Embedding: """ Return the pretrained glove embedding. :param dimension: the size of embedding dimension, the value can only be 50, 100, or 300. :return: The :class:`mz.embedding.Embedding` object. """ file_name = 'glove.6B.' + str(dimension) + 'd.txt' file_path = (Path(mz.USER_DATA_DIR) / 'glove').joinpath(file_name) if not file_path.exists(): keras.utils.data_utils.get_file('glove_embedding', _glove_embedding_url, extract=True, cache_dir=mz.USER_DATA_DIR, cache_subdir='glove') return mz.embedding.load_from_file(file_path=str(file_path), mode='glove')
995
33.344828
78
py
MatchZoo
MatchZoo-master/matchzoo/datasets/wiki_qa/load_data.py
"""WikiQA data loader.""" import typing import csv from pathlib import Path import keras import pandas as pd import matchzoo _url = "https://download.microsoft.com/download/E/5/F/" \ "E5FCFCEE-7005-4814-853D-DAA7C66507E0/WikiQACorpus.zip" def load_data( stage: str = 'train', task: str = 'ranking', filtered: bool = False, return_classes: bool = False ) -> typing.Union[matchzoo.DataPack, tuple]: """ Load WikiQA data. :param stage: One of `train`, `dev`, and `test`. :param task: Could be one of `ranking`, `classification` or a :class:`matchzoo.engine.BaseTask` instance. :param filtered: Whether remove the questions without correct answers. :param return_classes: `True` to return classes for classification task, `False` otherwise. :return: A DataPack unless `task` is `classificiation` and `return_classes` is `True`: a tuple of `(DataPack, classes)` in that case. """ if stage not in ('train', 'dev', 'test'): raise ValueError(f"{stage} is not a valid stage." f"Must be one of `train`, `dev`, and `test`.") data_root = _download_data() file_path = data_root.joinpath(f'WikiQA-{stage}.tsv') data_pack = _read_data(file_path) if filtered and stage in ('dev', 'test'): ref_path = data_root.joinpath(f'WikiQA-{stage}.ref') filter_ref_path = data_root.joinpath(f'WikiQA-{stage}-filtered.ref') with open(filter_ref_path, mode='r') as f: filtered_ids = set([line.split()[0] for line in f]) filtered_lines = [] with open(ref_path, mode='r') as f: for idx, line in enumerate(f.readlines()): if line.split()[0] in filtered_ids: filtered_lines.append(idx) data_pack = data_pack[filtered_lines] if task == 'ranking': task = matchzoo.tasks.Ranking() if task == 'classification': task = matchzoo.tasks.Classification() if isinstance(task, matchzoo.tasks.Ranking): return data_pack elif isinstance(task, matchzoo.tasks.Classification): data_pack.one_hot_encode_label(task.num_classes, inplace=True) if return_classes: return data_pack, [False, True] else: return data_pack else: raise ValueError(f"{task} is not a valid task." f"Must be one of `Ranking` and `Classification`.") def _download_data(): ref_path = keras.utils.data_utils.get_file( 'wikiqa', _url, extract=True, cache_dir=matchzoo.USER_DATA_DIR, cache_subdir='wiki_qa' ) return Path(ref_path).parent.joinpath('WikiQACorpus') def _read_data(path): table = pd.read_csv(path, sep='\t', header=0, quoting=csv.QUOTE_NONE) df = pd.DataFrame({ 'text_left': table['Question'], 'text_right': table['Sentence'], 'id_left': table['QuestionID'], 'id_right': table['SentenceID'], 'label': table['Label'] }) return matchzoo.pack(df)
3,045
32.472527
79
py
MatchZoo
MatchZoo-master/matchzoo/layers/matching_layer.py
"""An implementation of Matching Layer.""" import typing import tensorflow as tf from keras.engine import Layer class MatchingLayer(Layer): """ Layer that computes a matching matrix between samples in two tensors. :param normalize: Whether to L2-normalize samples along the dot product axis before taking the dot product. If set to True, then the output of the dot product is the cosine proximity between the two samples. :param matching_type: the similarity function for matching :param kwargs: Standard layer keyword arguments. Examples: >>> import matchzoo as mz >>> layer = mz.layers.MatchingLayer(matching_type='dot', ... normalize=True) >>> num_batch, left_len, right_len, num_dim = 5, 3, 2, 10 >>> layer.build([[num_batch, left_len, num_dim], ... [num_batch, right_len, num_dim]]) """ def __init__(self, normalize: bool = False, matching_type: str = 'dot', **kwargs): """:class:`MatchingLayer` constructor.""" super().__init__(**kwargs) self._normalize = normalize self._validate_matching_type(matching_type) self._matching_type = matching_type self._shape1 = None self._shape2 = None @classmethod def _validate_matching_type(cls, matching_type: str = 'dot'): valid_matching_type = ['dot', 'mul', 'plus', 'minus', 'concat'] if matching_type not in valid_matching_type: raise ValueError(f"{matching_type} is not a valid matching type, " f"{valid_matching_type} expected.") def build(self, input_shape: list): """ Build the layer. :param input_shape: the shapes of the input tensors, for MatchingLayer we need tow input tensors. """ # Used purely for shape validation. if not isinstance(input_shape, list) or len(input_shape) != 2: raise ValueError('A `MatchingLayer` layer should be called ' 'on a list of 2 inputs.') self._shape1 = input_shape[0] self._shape2 = input_shape[1] for idx in 0, 2: if self._shape1[idx] != self._shape2[idx]: raise ValueError( 'Incompatible dimensions: ' f'{self._shape1[idx]} != {self._shape2[idx]}.' f'Layer shapes: {self._shape1}, {self._shape2}.' ) def call(self, inputs: list, **kwargs) -> typing.Any: """ The computation logic of MatchingLayer. :param inputs: two input tensors. """ x1 = inputs[0] x2 = inputs[1] if self._matching_type == 'dot': if self._normalize: x1 = tf.math.l2_normalize(x1, axis=2) x2 = tf.math.l2_normalize(x2, axis=2) return tf.expand_dims(tf.einsum('abd,acd->abc', x1, x2), 3) else: if self._matching_type == 'mul': def func(x, y): return x * y elif self._matching_type == 'plus': def func(x, y): return x + y elif self._matching_type == 'minus': def func(x, y): return x - y elif self._matching_type == 'concat': def func(x, y): return tf.concat([x, y], axis=3) else: raise ValueError(f"Invalid matching type." f"{self._matching_type} received." f"Mut be in `dot`, `mul`, `plus`, " f"`minus` and `concat`.") x1_exp = tf.stack([x1] * self._shape2[1], 2) x2_exp = tf.stack([x2] * self._shape1[1], 1) return func(x1_exp, x2_exp) def compute_output_shape(self, input_shape: list) -> tuple: """ Calculate the layer output shape. :param input_shape: the shapes of the input tensors, for MatchingLayer we need tow input tensors. """ if not isinstance(input_shape, list) or len(input_shape) != 2: raise ValueError('A `MatchingLayer` layer should be called ' 'on a list of 2 inputs.') shape1 = list(input_shape[0]) shape2 = list(input_shape[1]) if len(shape1) != 3 or len(shape2) != 3: raise ValueError('A `MatchingLayer` layer should be called ' 'on 2 inputs with 3 dimensions.') if shape1[0] != shape2[0] or shape1[2] != shape2[2]: raise ValueError('A `MatchingLayer` layer should be called ' 'on 2 inputs with same 0,2 dimensions.') if self._matching_type in ['mul', 'plus', 'minus']: return shape1[0], shape1[1], shape2[1], shape1[2] elif self._matching_type == 'dot': return shape1[0], shape1[1], shape2[1], 1 elif self._matching_type == 'concat': return shape1[0], shape1[1], shape2[1], shape1[2] + shape2[2] else: raise ValueError(f"Invalid `matching_type`." f"{self._matching_type} received." f"Must be in `mul`, `plus`, `minus` " f"`dot` and `concat`.") def get_config(self) -> dict: """Get the config dict of MatchingLayer.""" config = { 'normalize': self._normalize, 'matching_type': self._matching_type, } base_config = super(MatchingLayer, self).get_config() return dict(list(base_config.items()) + list(config.items()))
5,753
39.808511
78
py
MatchZoo
MatchZoo-master/matchzoo/layers/dynamic_pooling_layer.py
"""An implementation of Dynamic Pooling Layer.""" import typing import tensorflow as tf from keras.engine import Layer class DynamicPoolingLayer(Layer): """ Layer that computes dynamic pooling of one tensor. :param psize1: pooling size of dimension 1 :param psize2: pooling size of dimension 2 :param kwargs: Standard layer keyword arguments. Examples: >>> import matchzoo as mz >>> layer = mz.layers.DynamicPoolingLayer(3, 2) >>> num_batch, left_len, right_len, num_dim = 5, 3, 2, 10 >>> layer.build([[num_batch, left_len, right_len, num_dim], ... [num_batch, left_len, right_len, 3]]) """ def __init__(self, psize1: int, psize2: int, **kwargs): """:class:`DynamicPoolingLayer` constructor.""" super().__init__(**kwargs) self._psize1 = psize1 self._psize2 = psize2 def build(self, input_shape: typing.List[int]): """ Build the layer. :param input_shape: the shapes of the input tensors, for DynamicPoolingLayer we need tow input tensors. """ super().build(input_shape) input_shape_one = input_shape[0] self._msize1 = input_shape_one[1] self._msize2 = input_shape_one[2] def call(self, inputs: list, **kwargs) -> typing.Any: """ The computation logic of DynamicPoolingLayer. :param inputs: two input tensors. """ self._validate_dpool_size() x, dpool_index = inputs dpool_shape = tf.shape(dpool_index) batch_index_one = tf.expand_dims( tf.expand_dims( tf.range(dpool_shape[0]), axis=-1), axis=-1) batch_index = tf.expand_dims( tf.tile(batch_index_one, [1, self._msize1, self._msize2]), axis=-1) dpool_index_ex = tf.concat([batch_index, dpool_index], axis=3) x_expand = tf.gather_nd(x, dpool_index_ex) stride1 = self._msize1 // self._psize1 stride2 = self._msize2 // self._psize2 x_pool = tf.nn.max_pool(x_expand, [1, stride1, stride2, 1], [1, stride1, stride2, 1], "VALID") return x_pool def compute_output_shape(self, input_shape: list) -> tuple: """ Calculate the layer output shape. :param input_shape: the shapes of the input tensors, for DynamicPoolingLayer we need tow input tensors. """ input_shape_one = input_shape[0] return (None, self._psize1, self._psize2, input_shape_one[3]) def get_config(self) -> dict: """Get the config dict of DynamicPoolingLayer.""" config = { 'psize1': self._psize1, 'psize2': self._psize2 } base_config = super(DynamicPoolingLayer, self).get_config() return dict(list(base_config.items()) + list(config.items())) def _validate_dpool_size(self): suggestion = self.get_size_suggestion( self._msize1, self._msize2, self._psize1, self._psize2 ) if suggestion != (self._psize1, self._psize2): raise ValueError( "DynamicPooling Layer can not " f"generate ({self._psize1} x {self._psize2}) output " f"feature map, please use ({suggestion[0]} x {suggestion[1]})" f" instead. `model.params['dpool_size'] = {suggestion}` " ) @classmethod def get_size_suggestion( cls, msize1: int, msize2: int, psize1: int, psize2: int ) -> typing.Tuple[int, int]: """ Get `dpool_size` suggestion for a given shape. Returns the nearest legal `dpool_size` for the given combination of `(psize1, psize2)`. :param msize1: size of the left text. :param msize2: size of the right text. :param psize1: base size of the pool. :param psize2: base size of the pool. :return: """ stride1 = msize1 // psize1 stride2 = msize2 // psize2 suggestion1 = msize1 // stride1 suggestion2 = msize2 // stride2 return (suggestion1, suggestion2)
4,315
32.457364
78
py
MatchZoo
MatchZoo-master/matchzoo/data_generator/data_generator.py
"""Base generator.""" import math import typing import keras import numpy as np import pandas as pd import matchzoo as mz from matchzoo.data_generator.callbacks import Callback class DataGenerator(keras.utils.Sequence): """ Data Generator. Used to divide a :class:`matchzoo.DataPack` into batches. This is helpful for generating batch-wise features and delaying data preprocessing to the `fit` time. See `tutorials/data_handling.ipynb` for a walkthrough. :param data_pack: DataPack to generator data from. :param mode: One of "point", "pair", and "list". (default: "point") :param num_dup: Number of duplications per instance, only effective when `mode` is "pair". (default: 1) :param num_neg: Number of negative samples per instance, only effective when `mode` is "pair". (default: 1) :param resample: Either to resample for each epoch, only effective when `mode` is "pair". (default: `True`) :param batch_size: Batch size. (default: 128) :param shuffle: Either to shuffle the samples/instances. (default: `True`) :param callbacks: Callbacks. See `matchzoo.data_generator.callbacks` for more details. Examples:: >>> import numpy as np >>> import matchzoo as mz >>> np.random.seed(0) >>> data_pack = mz.datasets.toy.load_data() >>> batch_size = 8 To generate data points: >>> point_gen = mz.DataGenerator( ... data_pack=data_pack, ... batch_size=batch_size ... ) >>> len(point_gen) 13 >>> x, y = point_gen[0] >>> for key, value in sorted(x.items()): ... print(key, str(value)[:30]) id_left ['Q6' 'Q17' 'Q1' 'Q13' 'Q16' ' id_right ['D6-6' 'D17-1' 'D1-2' 'D13-3' text_left ['how long is the term for fed text_right ['See Article I and Article II To generate data pairs: >>> pair_gen = mz.DataGenerator( ... data_pack=data_pack, ... mode='pair', ... num_dup=4, ... num_neg=4, ... batch_size=batch_size, ... shuffle=False ... ) >>> len(pair_gen) 3 >>> x, y = pair_gen[0] >>> for key, value in sorted(x.items()): ... print(key, str(value)[:30]) id_left ['Q1' 'Q1' 'Q1' 'Q1' 'Q1' 'Q1' id_right ['D1-3' 'D1-4' 'D1-0' 'D1-1' ' text_left ['how are glacier caves formed text_right ['A glacier cave is a cave for To generate data lists: # TODO: """ def __init__( self, data_pack: mz.DataPack, mode='point', num_dup: int = 1, num_neg: int = 1, resample: bool = True, batch_size: int = 128, shuffle: bool = True, callbacks: typing.List[Callback] = None ): """Init.""" if callbacks is None: callbacks = [] if mode not in ('point', 'pair', 'list'): raise ValueError(f"{mode} is not a valid mode type." f"Must be one of `point`, `pair` or `list`.") self._mode = mode self._num_dup = num_dup self._num_neg = num_neg self._batch_size = batch_size self._shuffle = shuffle self._resample = resample self._orig_relation = data_pack.relation self._callbacks = callbacks if mode == 'pair': data_pack.relation = self._reorganize_pair_wise( data_pack.relation, num_dup=num_dup, num_neg=num_neg ) self._data_pack = data_pack self._batch_indices = None self.reset_index() def __getitem__(self, item: int) -> typing.Tuple[dict, np.ndarray]: """Get a batch from index idx. :param item: the index of the batch. """ if isinstance(item, slice): indices = sum(self._batch_indices[item], []) else: indices = self._batch_indices[item] batch_data_pack = self._data_pack[indices] self._handle_callbacks_on_batch_data_pack(batch_data_pack) x, y = batch_data_pack.unpack() self._handle_callbacks_on_batch_unpacked(x, y) return x, y def __len__(self) -> int: """Get the total number of batches.""" return len(self._batch_indices) def on_epoch_end(self): """Reorganize the index array while epoch is ended.""" if self._mode == 'pair' and self._resample: self._data_pack.relation = self._reorganize_pair_wise( relation=self._orig_relation, num_dup=self._num_dup, num_neg=self._num_neg ) self.reset_index() def reset_index(self): """ Set the :attr:`index_array`. Here the :attr:`index_array` records the index of all the instances. """ # index pool: index -> instance index if self._mode == 'point': num_instances = len(self._data_pack) index_pool = list(range(num_instances)) elif self._mode == 'pair': index_pool = [] step_size = self._num_neg + 1 num_instances = int(len(self._data_pack) / step_size) for i in range(num_instances): lower = i * step_size upper = (i + 1) * step_size indices = list(range(lower, upper)) if indices: index_pool.append(indices) elif self._mode == 'list': raise NotImplementedError( f'{self._mode} data generator not implemented.') else: raise ValueError(f"{self._mode} is not a valid mode type" f"Must be one of `point`, `pair` or `list`.") if self._shuffle: np.random.shuffle(index_pool) # batch_indices: index -> batch of indices self._batch_indices = [] for i in range(math.ceil(num_instances / self._batch_size)): lower = self._batch_size * i upper = self._batch_size * (i + 1) candidates = index_pool[lower:upper] if self._mode == 'pair': candidates = sum(candidates, []) if candidates: self._batch_indices.append(candidates) def _handle_callbacks_on_batch_data_pack(self, batch_data_pack): for callback in self._callbacks: callback.on_batch_data_pack(batch_data_pack) def _handle_callbacks_on_batch_unpacked(self, x, y): for callback in self._callbacks: callback.on_batch_unpacked(x, y) @property def callbacks(self): """`callbacks` getter.""" return self._callbacks @callbacks.setter def callbacks(self, value): """`callbacks` setter.""" self._callbacks = value @property def num_neg(self): """`num_neg` getter.""" return self._num_neg @num_neg.setter def num_neg(self, value): """`num_neg` setter.""" self._num_neg = value self.reset_index() @property def num_dup(self): """`num_dup` getter.""" return self._num_dup @num_dup.setter def num_dup(self, value): """`num_dup` setter.""" self._num_dup = value self.reset_index() @property def mode(self): """`mode` getter.""" return self._mode @mode.setter def mode(self, value): """`mode` setter.""" self._mode = value self.reset_index() @property def batch_size(self): """`batch_size` getter.""" return self._batch_size @batch_size.setter def batch_size(self, value): """`batch_size` setter.""" self._batch_size = value self.reset_index() @property def shuffle(self): """`shuffle` getter.""" return self._shuffle @shuffle.setter def shuffle(self, value): """`shuffle` setter.""" self._shuffle = value self.reset_index() @property def batch_indices(self): """`batch_indices` getter.""" return self._batch_indices @classmethod def _reorganize_pair_wise( cls, relation: pd.DataFrame, num_dup: int = 1, num_neg: int = 1 ): """Re-organize the data pack as pair-wise format.""" pairs = [] groups = relation.sort_values( 'label', ascending=False).groupby('id_left') for idx, group in groups: labels = group.label.unique() for label in labels[:-1]: pos_samples = group[group.label == label] pos_samples = pd.concat([pos_samples] * num_dup) neg_samples = group[group.label < label] for _, pos_sample in pos_samples.iterrows(): pos_sample = pd.DataFrame([pos_sample]) neg_sample = neg_samples.sample(num_neg, replace=True) pairs.extend((pos_sample, neg_sample)) new_relation = pd.concat(pairs, ignore_index=True) return new_relation
9,187
30.251701
78
py
MatchZoo
MatchZoo-master/matchzoo/engine/base_model.py
"""Base Model.""" import abc import typing from pathlib import Path import dill import numpy as np import keras import keras.backend as K import pandas as pd import matchzoo from matchzoo import DataGenerator from matchzoo.engine import hyper_spaces from matchzoo.engine.base_preprocessor import BasePreprocessor from matchzoo.engine.base_metric import BaseMetric from matchzoo.engine.param_table import ParamTable from matchzoo.engine.param import Param from matchzoo import tasks class BaseModel(abc.ABC): """ Abstract base class of all MatchZoo models. MatchZoo models are wrapped over keras models, and the actual keras model built can be accessed by `model.backend`. `params` is a set of model hyper-parameters that deterministically builds a model. In other words, `params['model_class'](params=params)` of the same `params` always create models with the same structure. :param params: Model hyper-parameters. (default: return value from :meth:`get_default_params`) :param backend: A keras model as the model backend. Usually not passed as an argument. Example: >>> BaseModel() # doctest: +ELLIPSIS Traceback (most recent call last): ... TypeError: Can't instantiate abstract class BaseModel ... >>> class MyModel(BaseModel): ... def build(self): ... pass >>> isinstance(MyModel(), BaseModel) True """ BACKEND_WEIGHTS_FILENAME = 'backend_weights.h5' PARAMS_FILENAME = 'params.dill' def __init__( self, params: typing.Optional[ParamTable] = None, backend: typing.Optional[keras.models.Model] = None ): """Init.""" self._params = params or self.get_default_params() self._backend = backend @classmethod def get_default_params( cls, with_embedding=False, with_multi_layer_perceptron=False ) -> ParamTable: """ Model default parameters. The common usage is to instantiate :class:`matchzoo.engine.ModelParams` first, then set the model specific parametrs. Examples: >>> class MyModel(BaseModel): ... def build(self): ... print(self._params['num_eggs'], 'eggs') ... print('and', self._params['ham_type']) ... ... @classmethod ... def get_default_params(cls): ... params = ParamTable() ... params.add(Param('num_eggs', 512)) ... params.add(Param('ham_type', 'Parma Ham')) ... return params >>> my_model = MyModel() >>> my_model.build() 512 eggs and Parma Ham Notice that all parameters must be serialisable for the entire model to be serialisable. Therefore, it's strongly recommended to use python native data types to store parameters. :return: model parameters """ params = ParamTable() params.add(Param( name='model_class', value=cls, desc="Model class. Used internally for save/load. " "Changing this may cause unexpected behaviors." )) params.add(Param( name='input_shapes', desc="Dependent on the model and data. Should be set manually." )) params.add(Param( name='task', desc="Decides model output shape, loss, and metrics." )) params.add(Param( name='optimizer', value='adam', )) if with_embedding: params.add(Param( name='with_embedding', value=True, desc="A flag used help `auto` module. Shouldn't be changed." )) params.add(Param( name='embedding_input_dim', desc='Usually equals vocab size + 1. Should be set manually.' )) params.add(Param( name='embedding_output_dim', desc='Should be set manually.' )) params.add(Param( name='embedding_trainable', value=True, desc='`True` to enable embedding layer training, ' '`False` to freeze embedding parameters.' )) if with_multi_layer_perceptron: params.add(Param( name='with_multi_layer_perceptron', value=True, desc="A flag of whether a multiple layer perceptron is used. " "Shouldn't be changed." )) params.add(Param( name='mlp_num_units', value=128, desc="Number of units in first `mlp_num_layers` layers.", hyper_space=hyper_spaces.quniform(8, 256, 8) )) params.add(Param( name='mlp_num_layers', value=3, desc="Number of layers of the multiple layer percetron.", hyper_space=hyper_spaces.quniform(1, 6) )) params.add(Param( name='mlp_num_fan_out', value=64, desc="Number of units of the layer that connects the multiple " "layer percetron and the output.", hyper_space=hyper_spaces.quniform(4, 128, 4) )) params.add(Param( name='mlp_activation_func', value='relu', desc='Activation function used in the multiple ' 'layer perceptron.' )) return params @classmethod def get_default_preprocessor(cls) -> BasePreprocessor: """ Model default preprocessor. The preprocessor's transform should produce a correctly shaped data pack that can be used for training. Some extra configuration (e.g. setting `input_shapes` in :class:`matchzoo.models.DSSMModel` may be required on the user's end. :return: Default preprocessor. """ return matchzoo.preprocessors.BasicPreprocessor() @property def params(self) -> ParamTable: """:return: model parameters.""" return self._params @params.setter def params(self, val): self._params = val @property def backend(self) -> keras.models.Model: """:return model backend, a keras model instance.""" if not self._backend: raise ValueError("Backend not found." "Please build the model first.") else: return self._backend @abc.abstractmethod def build(self): """Build model, each subclass need to impelemnt this method.""" def compile(self): """ Compile model for training. Only `keras` native metrics are compiled together with backend. MatchZoo metrics are evaluated only through :meth:`evaluate`. Notice that `keras` count `loss` as one of the metrics while MatchZoo :class:`matchzoo.engine.BaseTask` does not. Examples: >>> from matchzoo import models >>> model = models.Naive() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.params['task'].metrics = ['mse', 'map'] >>> model.params['task'].metrics ['mse', mean_average_precision(0.0)] >>> model.build() >>> model.compile() """ self._backend.compile(optimizer=self._params['optimizer'], loss=self._params['task'].loss) def fit( self, x: typing.Union[np.ndarray, typing.List[np.ndarray], dict], y: np.ndarray, batch_size: int = 128, epochs: int = 1, verbose: int = 1, **kwargs ) -> keras.callbacks.History: """ Fit the model. See :meth:`keras.models.Model.fit` for more details. :param x: input data. :param y: labels. :param batch_size: number of samples per gradient update. :param epochs: number of epochs to train the model. :param verbose: 0, 1, or 2. Verbosity mode. 0 = silent, 1 = verbose, 2 = one log line per epoch. Key word arguments not listed above will be propagated to keras's fit. :return: A `keras.callbacks.History` instance. Its history attribute contains all information collected during training. """ return self._backend.fit(x=x, y=y, batch_size=batch_size, epochs=epochs, verbose=verbose, **kwargs) def fit_generator( self, generator: matchzoo.DataGenerator, epochs: int = 1, verbose: int = 1, **kwargs ) -> keras.callbacks.History: """ Fit the model with matchzoo `generator`. See :meth:`keras.models.Model.fit_generator` for more details. :param generator: A generator, an instance of :class:`engine.DataGenerator`. :param epochs: Number of epochs to train the model. :param verbose: 0, 1, or 2. Verbosity mode. 0 = silent, 1 = verbose, 2 = one log line per epoch. :return: A `keras.callbacks.History` instance. Its history attribute contains all information collected during training. """ return self._backend.fit_generator( generator=generator, epochs=epochs, verbose=verbose, **kwargs ) def evaluate( self, x: typing.Dict[str, np.ndarray], y: np.ndarray, batch_size: int = 128 ) -> typing.Dict[BaseMetric, float]: """ Evaluate the model. :param x: Input data. :param y: Labels. :param batch_size: Number of samples when `predict` for evaluation. (default: 128) Examples:: >>> import matchzoo as mz >>> data_pack = mz.datasets.toy.load_data() >>> preprocessor = mz.preprocessors.NaivePreprocessor() >>> data_pack = preprocessor.fit_transform(data_pack, verbose=0) >>> m = mz.models.DenseBaseline() >>> m.params['task'] = mz.tasks.Ranking() >>> m.params['task'].metrics = [ ... 'acc', 'mse', 'mae', 'ce', ... 'average_precision', 'precision', 'dcg', 'ndcg', ... 'mean_reciprocal_rank', 'mean_average_precision', 'mrr', ... 'map', 'MAP', ... mz.metrics.AveragePrecision(threshold=1), ... mz.metrics.Precision(k=2, threshold=2), ... mz.metrics.DiscountedCumulativeGain(k=2), ... mz.metrics.NormalizedDiscountedCumulativeGain( ... k=3, threshold=-1), ... mz.metrics.MeanReciprocalRank(threshold=2), ... mz.metrics.MeanAveragePrecision(threshold=3) ... ] >>> m.guess_and_fill_missing_params(verbose=0) >>> m.build() >>> m.compile() >>> x, y = data_pack.unpack() >>> evals = m.evaluate(x, y) >>> type(evals) <class 'dict'> """ result = dict() matchzoo_metrics, keras_metrics = self._separate_metrics() y_pred = self.predict(x, batch_size) for metric in keras_metrics: metric_func = keras.metrics.get(metric) result[metric] = K.eval(K.mean( metric_func(K.variable(y), K.variable(y_pred)))) if matchzoo_metrics: if not isinstance(self.params['task'], tasks.Ranking): raise ValueError("Matchzoo metrics only works on ranking.") for metric in matchzoo_metrics: result[metric] = self._eval_metric_on_data_frame( metric, x['id_left'], y, y_pred) return result def evaluate_generator( self, generator: DataGenerator, batch_size: int = 128 ) -> typing.Dict['BaseMetric', float]: """ Evaluate the model. :param generator: DataGenerator to evluate. :param batch_size: Batch size. (default: 128) """ x, y = generator[:] return self.evaluate(x, y, batch_size=batch_size) def _separate_metrics(self): matchzoo_metrics = [] keras_metrics = [] for metric in self._params['task'].metrics: if isinstance(metric, BaseMetric): matchzoo_metrics.append(metric) else: keras_metrics.append(metric) return matchzoo_metrics, keras_metrics @classmethod def _eval_metric_on_data_frame( cls, metric: BaseMetric, id_left: typing.Union[list, np.array], y: typing.Union[list, np.array], y_pred: typing.Union[list, np.array] ): eval_df = pd.DataFrame(data={ 'id': id_left, 'true': y.squeeze(), 'pred': y_pred.squeeze() }) assert isinstance(metric, BaseMetric) val = eval_df.groupby(by='id').apply( lambda df: metric(df['true'].values, df['pred'].values) ).mean() return val def predict( self, x: typing.Dict[str, np.ndarray], batch_size=128 ) -> np.ndarray: """ Generate output predictions for the input samples. See :meth:`keras.models.Model.predict` for more details. :param x: input data :param batch_size: number of samples per gradient update :return: numpy array(s) of predictions """ return self._backend.predict(x=x, batch_size=batch_size) def save(self, dirpath: typing.Union[str, Path]): """ Save the model. A saved model is represented as a directory with two files. One is a model parameters file saved by `pickle`, and the other one is a model h5 file saved by `keras`. :param dirpath: directory path of the saved model Example: >>> import matchzoo as mz >>> model = mz.models.Naive() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() >>> model.save('temp-model') >>> import shutil >>> shutil.rmtree('temp-model') """ dirpath = Path(dirpath) params_path = dirpath.joinpath(self.PARAMS_FILENAME) weights_path = dirpath.joinpath(self.BACKEND_WEIGHTS_FILENAME) if not dirpath.exists(): dirpath.mkdir(parents=True) else: raise FileExistsError(f'{dirpath} already exist, fail to save.') self._backend.save_weights(weights_path) with open(params_path, mode='wb') as params_file: dill.dump(self._params, params_file) def get_embedding_layer( self, name: str = 'embedding' ) -> keras.layers.Layer: """ Get the embedding layer. All MatchZoo models with a single embedding layer set the embedding layer name to `embedding`, and this method should return that layer. :param name: Name of the embedding layer. (default: `embedding`) """ for layer in self._backend.layers: if layer.name == name: return layer raise ValueError(f"Layer {name} not found. Initialize your embedding " f"layer with `name='{name}'`.") def load_embedding_matrix( self, embedding_matrix: np.ndarray, name: str = 'embedding' ): """ Load an embedding matrix. Load an embedding matrix into the model's embedding layer. The name of the embedding layer is specified by `name`. For models with only one embedding layer, set `name='embedding'` when creating the keras layer, and use the default `name` when load the matrix. For models with more than one embedding layers, initialize keras layer with different layer names, and set `name` accordingly to load a matrix to a chosen layer. :param embedding_matrix: Embedding matrix to be loaded. :param name: Name of the layer. (default: 'embedding') """ self.get_embedding_layer(name).set_weights([embedding_matrix]) def guess_and_fill_missing_params(self, verbose=1): """ Guess and fill missing parameters in :attr:`params`. Use this method to automatically fill-in other hyper parameters. This involves some guessing so the parameter it fills could be wrong. For example, the default task is `Ranking`, and if we do not set it to `Classification` manaully for data packs prepared for classification, then the shape of the model output and the data will mismatch. :param verbose: Verbosity. """ self._params.get('task').set_default(tasks.Ranking(), verbose) self._params.get('input_shapes').set_default([(30,), (30,)], verbose) if 'with_embedding' in self._params: self._params.get('embedding_input_dim').set_default(300, verbose) self._params.get('embedding_output_dim').set_default(300, verbose) def _set_param_default(self, name: str, default_val: str, verbose: int = 0): if self._params[name] is None: self._params[name] = default_val if verbose: print(f"Parameter \"{name}\" set to {default_val}.") def _make_inputs(self) -> list: input_left = keras.layers.Input( name='text_left', shape=self._params['input_shapes'][0] ) input_right = keras.layers.Input( name='text_right', shape=self._params['input_shapes'][1] ) return [input_left, input_right] def _make_output_layer(self) -> keras.layers.Layer: """:return: a correctly shaped keras dense layer for model output.""" task = self._params['task'] if isinstance(task, tasks.Classification): return keras.layers.Dense(task.num_classes, activation='softmax') elif isinstance(task, tasks.Ranking): return keras.layers.Dense(1, activation='linear') else: raise ValueError(f"{task} is not a valid task type." f"Must be in `Ranking` and `Classification`.") def _make_embedding_layer( self, name: str = 'embedding', **kwargs ) -> keras.layers.Layer: return keras.layers.Embedding( self._params['embedding_input_dim'], self._params['embedding_output_dim'], trainable=self._params['embedding_trainable'], name=name, **kwargs ) def _make_multi_layer_perceptron_layer(self) -> keras.layers.Layer: # TODO: do not create new layers for a second call if not self._params['with_multi_layer_perceptron']: raise AttributeError( 'Parameter `with_multi_layer_perception` not set.') def _wrapper(x): activation = self._params['mlp_activation_func'] for _ in range(self._params['mlp_num_layers']): x = keras.layers.Dense(self._params['mlp_num_units'], activation=activation)(x) return keras.layers.Dense(self._params['mlp_num_fan_out'], activation=activation)(x) return _wrapper def load_model(dirpath: typing.Union[str, Path]) -> BaseModel: """ Load a model. The reverse function of :meth:`BaseModel.save`. :param dirpath: directory path of the saved model :return: a :class:`BaseModel` instance Example: >>> import matchzoo as mz >>> model = mz.models.Naive() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() >>> model.save('my-model') >>> model.params.keys() == mz.load_model('my-model').params.keys() True >>> import shutil >>> shutil.rmtree('my-model') """ dirpath = Path(dirpath) params_path = dirpath.joinpath(BaseModel.PARAMS_FILENAME) weights_path = dirpath.joinpath(BaseModel.BACKEND_WEIGHTS_FILENAME) with open(params_path, mode='rb') as params_file: params = dill.load(params_file) model_instance = params['model_class'](params=params) model_instance.build() model_instance.compile() model_instance.backend.load_weights(weights_path) return model_instance
20,711
34.587629
79
py
MatchZoo
MatchZoo-master/matchzoo/engine/callbacks.py
"""Callbacks.""" import typing from pathlib import Path import numpy as np import keras import matchzoo from matchzoo.engine.base_model import BaseModel class EvaluateAllMetrics(keras.callbacks.Callback): """ Callback to evaluate all metrics. MatchZoo metrics can not be evaluated batch-wise since they require dataset-level information. As a result, MatchZoo metrics are not evaluated automatically when a Model `fit`. When this callback is used, all metrics, including MatchZoo metrics and Keras metrics, are evluated once every `once_every` epochs. :param model: Model to evaluate. :param x: X. :param y: y. :param once_every: Evaluation only triggers when `epoch % once_every == 0`. (default: 1, i.e. evaluate on every epoch's end) :param batch_size: Number of samples per evaluation. This only affects the evaluation of Keras metrics, since MatchZoo metrics are always evaluated using the full data. :param model_save_path: Directory path to save the model after each evaluate callback, (default: None, i.e., no saving.) :param verbose: Verbosity. """ def __init__( self, model: 'BaseModel', x: typing.Union[np.ndarray, typing.List[np.ndarray]], y: np.ndarray, once_every: int = 1, batch_size: int = 128, model_save_path: str = None, verbose=1 ): """Initializer.""" super().__init__() self._model = model self._dev_x = x self._dev_y = y self._valid_steps = once_every self._batch_size = batch_size self._model_save_path = model_save_path self._verbose = verbose def on_epoch_end(self, epoch: int, logs: dict = None): """ Called at the end of en epoch. :param epoch: integer, index of epoch. :param logs: dictionary of logs. :return: dictionary of logs. """ if (epoch + 1) % self._valid_steps == 0: val_logs = self._model.evaluate(self._dev_x, self._dev_y, self._batch_size) if self._verbose: print('Validation: ' + ' - '.join( f'{k}: {v}' for k, v in val_logs.items())) for k, v in val_logs.items(): logs[k] = v if self._model_save_path: curr_path = self._model_save_path + str('%d/' % (epoch + 1)) self._model.save(curr_path)
2,513
32.972973
79
py
MatchZoo
MatchZoo-master/matchzoo/engine/parse_metric.py
import typing import matchzoo from matchzoo.engine.base_metric import BaseMetric from matchzoo.engine import base_task def parse_metric( metric: typing.Union[str, typing.Type[BaseMetric], BaseMetric], task: 'base_task.BaseTask' = None ) -> typing.Union['BaseMetric', str]: """ Parse input metric in any form into a :class:`BaseMetric` instance. :param metric: Input metric in any form. :param task: Task type for determining specific metric. :return: A :class:`BaseMetric` instance Examples:: >>> from matchzoo import metrics >>> from matchzoo.engine.parse_metric import parse_metric Use `str` as keras native metrics: >>> parse_metric('mse') 'mse' Use `str` as MatchZoo metrics: >>> mz_metric = parse_metric('map') >>> type(mz_metric) <class 'matchzoo.metrics.mean_average_precision.MeanAveragePrecision'> Use :class:`matchzoo.engine.BaseMetric` subclasses as MatchZoo metrics: >>> type(parse_metric(metrics.AveragePrecision)) <class 'matchzoo.metrics.average_precision.AveragePrecision'> Use :class:`matchzoo.engine.BaseMetric` instances as MatchZoo metrics: >>> type(parse_metric(metrics.AveragePrecision())) <class 'matchzoo.metrics.average_precision.AveragePrecision'> """ if task is None: task = matchzoo.tasks.Ranking() if isinstance(metric, str): metric = metric.lower() # ignore case # matchzoo metrics in str form for subclass in BaseMetric.__subclasses__(): if metric == subclass.ALIAS or metric in subclass.ALIAS: return subclass() # keras native metrics return _remap_keras_metric(metric, task) elif isinstance(metric, BaseMetric): return metric elif issubclass(metric, BaseMetric): return metric() else: raise ValueError(metric) def _remap_keras_metric(metric: str, task) -> str: # we do not support sparse label in classification. lookup = { matchzoo.tasks.Ranking: { 'acc': 'binary_accuracy', 'accuracy': 'binary_accuracy', 'crossentropy': 'binary_crossentropy', 'ce': 'binary_crossentropy', }, matchzoo.tasks.Classification: { 'acc': 'categorical_accuracy', 'accuracy': 'categorical_accuracy', 'crossentropy': 'categorical_crossentropy', 'ce': 'categorical_crossentropy', } } return lookup[type(task)].get(metric, metric)
2,559
31.405063
78
py
MatchZoo
MatchZoo-master/matchzoo/utils/make_keras_optimizer_picklable.py
import keras def make_keras_optimizer_picklable(): """ Fix https://github.com/NTMC-Community/MatchZoo/issues/726. This function changes how keras behaves, use with caution. """ def __getstate__(self): return keras.optimizers.serialize(self) def __setstate__(self, state): optimizer = keras.optimizers.deserialize(state) self.__dict__ = optimizer.__dict__ cls = keras.optimizers.Optimizer cls.__getstate__ = __getstate__ cls.__setstate__ = __setstate__
517
24.9
62
py
MatchZoo
MatchZoo-master/matchzoo/utils/__init__.py
from .one_hot import one_hot from .tensor_type import TensorType from .list_recursive_subclasses import list_recursive_concrete_subclasses from .make_keras_optimizer_picklable import make_keras_optimizer_picklable
214
42
74
py
MatchZoo
MatchZoo-master/matchzoo/contrib/models/esim.py
"""ESIM model.""" import keras import keras.backend as K import tensorflow as tf import matchzoo as mz from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable class ESIM(BaseModel): """ ESIM model. Examples: >>> model = ESIM() >>> task = classification_task = mz.tasks.Classification(num_classes=2) >>> model.params['task'] = task >>> model.params['input_shapes'] = [(20, ), (40, )] >>> model.params['lstm_dim'] = 300 >>> model.params['mlp_num_units'] = 300 >>> model.params['embedding_input_dim'] = 5000 >>> model.params['embedding_output_dim'] = 10 >>> model.params['embedding_trainable'] = False >>> model.params['mlp_num_layers'] = 0 >>> model.params['mlp_num_fan_out'] = 300 >>> model.params['mlp_activation_func'] = 'tanh' >>> model.params['mask_value'] = 0 >>> model.params['dropout_rate'] = 0.5 >>> model.params['optimizer'] = keras.optimizers.Adam(lr=4e-4) >>> model.guess_and_fill_missing_params() >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """Get default parameters.""" params = super().get_default_params(with_embedding=True, with_multi_layer_perceptron=True) params.add(Param( name='dropout_rate', value=0.5, desc="The dropout rate for all fully-connected layer" )) params.add(Param( name='lstm_dim', value=8, desc="The dimension of LSTM layer." )) params.add(Param( name='mask_value', value=0, desc="The value would be regarded as pad" )) return params def _expand_dim(self, inp: tf.Tensor, axis: int) -> keras.layers.Layer: """ Wrap keras.backend.expand_dims into a Lambda layer. :param inp: input tensor to expand the dimension :param axis: the axis of new dimension """ return keras.layers.Lambda(lambda x: tf.expand_dims(x, axis=axis))(inp) def _make_atten_mask_layer(self) -> keras.layers.Layer: """ Make mask layer for attention weight matrix so that each word won't pay attention to <PAD> timestep. """ return keras.layers.Lambda( lambda weight_mask: weight_mask[0] + (1.0 - weight_mask[1]) * -1e7, name="atten_mask") def _make_bilstm_layer(self, lstm_dim: int) -> keras.layers.Layer: """ Bidirectional LSTM layer in ESIM. :param lstm_dim: int, dimension of LSTM layer :return: `keras.layers.Layer`. """ return keras.layers.Bidirectional( layer=keras.layers.LSTM(lstm_dim, return_sequences=True), merge_mode='concat') def _max(self, texts: tf.Tensor, mask: tf.Tensor) -> tf.Tensor: """ Compute the max of each text according to their real length :param texts: np.array with shape [B, T, H] :param lengths: np.array with shape [B, T, ], where 1 means valid, 0 means pad """ mask = self._expand_dim(mask, axis=2) new_texts = keras.layers.Multiply()([texts, mask]) text_max = keras.layers.Lambda( lambda x: tf.reduce_max(x, axis=1), )(new_texts) return text_max def _avg(self, texts: tf.Tensor, mask: tf.Tensor) -> tf.Tensor: """ Compute the mean of each text according to their real length :param texts: np.array with shape [B, T, H] :param lengths: np.array with shape [B, T, ], where 1 means valid, 0 means pad """ mask = self._expand_dim(mask, axis=2) new_texts = keras.layers.Multiply()([texts, mask]) # timestep-wise division, exclude the PAD number when calc avg text_avg = keras.layers.Lambda( lambda text_mask: tf.reduce_sum(text_mask[0], axis=1) / tf.reduce_sum(text_mask[1], axis=1), )([new_texts, mask]) return text_avg def build(self): """Build model.""" # parameters lstm_dim = self._params['lstm_dim'] dropout_rate = self._params['dropout_rate'] # layers create_mask = keras.layers.Lambda( lambda x: tf.cast(tf.not_equal(x, self._params['mask_value']), K.floatx()) ) embedding = self._make_embedding_layer() lstm_compare = self._make_bilstm_layer(lstm_dim) lstm_compose = self._make_bilstm_layer(lstm_dim) dense_compare = keras.layers.Dense(units=lstm_dim, activation='relu', use_bias=True) dropout = keras.layers.Dropout(dropout_rate) # model a, b = self._make_inputs() # [B, T_a], [B, T_b] a_mask = create_mask(a) # [B, T_a] b_mask = create_mask(b) # [B, T_b] # encoding a_emb = dropout(embedding(a)) # [B, T_a, E_dim] b_emb = dropout(embedding(b)) # [B, T_b, E_dim] a_ = lstm_compare(a_emb) # [B, T_a, H*2] b_ = lstm_compare(b_emb) # [B, T_b, H*2] # mask a_ and b_, since the <pad> position is no more zero a_ = keras.layers.Multiply()([a_, self._expand_dim(a_mask, axis=2)]) b_ = keras.layers.Multiply()([b_, self._expand_dim(b_mask, axis=2)]) # local inference e = keras.layers.Dot(axes=-1)([a_, b_]) # [B, T_a, T_b] _ab_mask = keras.layers.Multiply()( # _ab_mask: [B, T_a, T_b] [self._expand_dim(a_mask, axis=2), # [B, T_a, 1] self._expand_dim(b_mask, axis=1)]) # [B, 1, T_b] pm = keras.layers.Permute((2, 1)) mask_layer = self._make_atten_mask_layer() softmax_layer = keras.layers.Softmax(axis=-1) e_a = softmax_layer(mask_layer([e, _ab_mask])) # [B, T_a, T_b] e_b = softmax_layer(mask_layer([pm(e), pm(_ab_mask)])) # [B, T_b, T_a] # alignment (a_t = a~, b_t = b~ ) a_t = keras.layers.Dot(axes=(2, 1))([e_a, b_]) # [B, T_a, H*2] b_t = keras.layers.Dot(axes=(2, 1))([e_b, a_]) # [B, T_b, H*2] # local inference info enhancement m_a = keras.layers.Concatenate(axis=-1)([ a_, a_t, keras.layers.Subtract()([a_, a_t]), keras.layers.Multiply()([a_, a_t])]) # [B, T_a, H*2*4] m_b = keras.layers.Concatenate(axis=-1)([ b_, b_t, keras.layers.Subtract()([b_, b_t]), keras.layers.Multiply()([b_, b_t])]) # [B, T_b, H*2*4] # project m_a and m_b from 4*H*2 dim to H dim m_a = dropout(dense_compare(m_a)) # [B, T_a, H] m_b = dropout(dense_compare(m_b)) # [B, T_a, H] # inference composition v_a = lstm_compose(m_a) # [B, T_a, H*2] v_b = lstm_compose(m_b) # [B, T_b, H*2] # pooling v_a = keras.layers.Concatenate(axis=-1)( [self._avg(v_a, a_mask), self._max(v_a, a_mask)]) # [B, H*4] v_b = keras.layers.Concatenate(axis=-1)( [self._avg(v_b, b_mask), self._max(v_b, b_mask)]) # [B, H*4] v = keras.layers.Concatenate(axis=-1)([v_a, v_b]) # [B, H*8] # mlp (multilayer perceptron) classifier output = self._make_multi_layer_perceptron_layer()(v) # [B, H] output = dropout(output) output = self._make_output_layer()(output) # [B, #classes] self._backend = keras.Model(inputs=[a, b], outputs=output)
7,807
35.657277
90
py
MatchZoo
MatchZoo-master/matchzoo/contrib/models/match_lstm.py
"""Match LSTM model.""" import keras import keras.backend as K import tensorflow as tf from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine import hyper_spaces class MatchLSTM(BaseModel): """ Match LSTM model. Examples: >>> model = MatchLSTM() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.params['embedding_input_dim'] = 10000 >>> model.params['embedding_output_dim'] = 100 >>> model.params['embedding_trainable'] = True >>> model.params['fc_num_units'] = 200 >>> model.params['lstm_num_units'] = 256 >>> model.params['dropout_rate'] = 0.5 >>> model.build() """ @classmethod def get_default_params(cls): """Get default parameters.""" params = super().get_default_params(with_embedding=True) params.add(Param( 'lstm_num_units', 256, hyper_space=hyper_spaces.quniform(low=128, high=384, q=32), desc="The hidden size in the LSTM layer." )) params.add(Param( 'fc_num_units', 200, hyper_space=hyper_spaces.quniform( low=100, high=300, q=20), desc="The hidden size in the full connection layer." )) params.add(Param( 'dropout_rate', 0.0, hyper_space=hyper_spaces.quniform( low=0.0, high=0.9, q=0.01), desc="The dropout rate." )) return params def build(self): """Build model.""" input_left, input_right = self._make_inputs() len_left = input_left.shape[1] len_right = input_right.shape[1] embedding = self._make_embedding_layer() embed_left = embedding(input_left) embed_right = embedding(input_right) lstm_left = keras.layers.LSTM(self._params['lstm_num_units'], return_sequences=True, name='lstm_left') lstm_right = keras.layers.LSTM(self._params['lstm_num_units'], return_sequences=True, name='lstm_right') encoded_left = lstm_left(embed_left) encoded_right = lstm_right(embed_right) def attention(tensors): """Attention layer.""" left, right = tensors tensor_left = tf.expand_dims(left, axis=2) tensor_right = tf.expand_dims(right, axis=1) tensor_left = K.repeat_elements(tensor_left, len_right, 2) tensor_right = K.repeat_elements(tensor_right, len_left, 1) tensor_merged = tf.concat([tensor_left, tensor_right], axis=-1) middle_output = keras.layers.Dense(self._params['fc_num_units'], activation='tanh')( tensor_merged) attn_scores = keras.layers.Dense(1)(middle_output) attn_scores = tf.squeeze(attn_scores, axis=3) exp_attn_scores = tf.math.exp( attn_scores - tf.reduce_max(attn_scores, axis=-1, keepdims=True)) exp_sum = tf.reduce_sum(exp_attn_scores, axis=-1, keepdims=True) attention_weights = exp_attn_scores / exp_sum return K.batch_dot(attention_weights, right) attn_layer = keras.layers.Lambda(attention) left_attn_vec = attn_layer([encoded_left, encoded_right]) concat = keras.layers.Concatenate(axis=1)( [left_attn_vec, encoded_right]) lstm_merge = keras.layers.LSTM(self._params['lstm_num_units'] * 2, return_sequences=False, name='lstm_merge') merged = lstm_merge(concat) dropout = keras.layers.Dropout( rate=self._params['dropout_rate'])(merged) phi = keras.layers.Dense(self._params['fc_num_units'], activation='tanh')(dropout) inputs = [input_left, input_right] out = self._make_output_layer()(phi) self._backend = keras.Model(inputs=inputs, outputs=[out])
4,182
39.221154
81
py
MatchZoo
MatchZoo-master/matchzoo/contrib/models/hbmp.py
"""HBMP model.""" import keras import typing from matchzoo.engine import hyper_spaces from matchzoo.engine.param_table import ParamTable from matchzoo.engine.param import Param from matchzoo.engine.base_model import BaseModel class HBMP(BaseModel): """ HBMP model. Examples: >>> model = HBMP() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.params['embedding_input_dim'] = 200 >>> model.params['embedding_output_dim'] = 100 >>> model.params['embedding_trainable'] = True >>> model.params['alpha'] = 0.1 >>> model.params['mlp_num_layers'] = 3 >>> model.params['mlp_num_units'] = [10, 10] >>> model.params['lstm_num_units'] = 5 >>> model.params['dropout_rate'] = 0.1 >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params['optimizer'] = 'adam' params.add(Param(name='alpha', value=0.1, desc="Negative slope coefficient of LeakyReLU " "function.")) params.add(Param(name='mlp_num_layers', value=3, desc="The number of layers of mlp.")) params.add(Param(name='mlp_num_units', value=[10, 10], desc="The hidden size of the FC layers, but not " "include the final layer.")) params.add(Param(name='lstm_num_units', value=5, desc="The hidden size of the LSTM layer.")) params.add(Param( name='dropout_rate', value=0.1, hyper_space=hyper_spaces.quniform( low=0.0, high=0.8, q=0.01), desc="The dropout rate." )) return params def build(self): """Build model structure.""" input_left, input_right = self._make_inputs() embedding = self._make_embedding_layer() embed_left = embedding(input_left) embed_right = embedding(input_right) # Get sentence embedding embed_sen_left = self._sentence_encoder( embed_left, lstm_num_units=self._params['lstm_num_units'], drop_rate=self._params['dropout_rate']) embed_sen_right = self._sentence_encoder( embed_right, lstm_num_units=self._params['lstm_num_units'], drop_rate=self._params['dropout_rate']) # Concatenate two sentence embedding: [embed_sen_left, embed_sen_right, # |embed_sen_left-embed_sen_right|, embed_sen_left*embed_sen_right] embed_minus = keras.layers.Subtract()( [embed_sen_left, embed_sen_right]) embed_minus_abs = keras.layers.Lambda(lambda x: abs(x))(embed_minus) embed_multiply = keras.layers.Multiply()( [embed_sen_left, embed_sen_right]) concat = keras.layers.Concatenate(axis=1)( [embed_sen_left, embed_sen_right, embed_minus_abs, embed_multiply]) # Multiply perception layers to classify mlp_out = self._classifier( concat, mlp_num_layers=self._params['mlp_num_layers'], mlp_num_units=self._params['mlp_num_units'], drop_rate=self._params['dropout_rate'], leaky_relu_alpah=self._params['alpha']) out = self._make_output_layer()(mlp_out) self._backend = keras.Model( inputs=[input_left, input_right], outputs=out) def _classifier( self, input_: typing.Any, mlp_num_layers: int, mlp_num_units: list, drop_rate: float, leaky_relu_alpah: float ) -> typing.Any: for i in range(mlp_num_layers - 1): input_ = keras.layers.Dropout(rate=drop_rate)(input_) input_ = keras.layers.Dense(mlp_num_units[i])(input_) input_ = keras.layers.LeakyReLU(alpha=leaky_relu_alpah)(input_) return input_ def _sentence_encoder( self, input_: typing.Any, lstm_num_units: int, drop_rate: float ) -> typing.Any: """ Stack three BiLSTM MaxPooling blocks as a hierarchical structure. Concatenate the output of three blocs as the input sentence embedding. Each BiLSTM layer reads the input sentence as the input. Each BiLSTM layer except the first one is initialized(the initial hidden state and the cell state) with the final state of the previous layer. """ emb1 = keras.layers.Bidirectional( keras.layers.LSTM( units=lstm_num_units, return_sequences=True, return_state=True, dropout=drop_rate, recurrent_dropout=drop_rate), merge_mode='concat')(input_) emb1_maxpooling = keras.layers.GlobalMaxPooling1D()(emb1[0]) emb2 = keras.layers.Bidirectional( keras.layers.LSTM( units=lstm_num_units, return_sequences=True, return_state=True, dropout=drop_rate, recurrent_dropout=drop_rate), merge_mode='concat')(input_, initial_state=emb1[1:5]) emb2_maxpooling = keras.layers.GlobalMaxPooling1D()(emb2[0]) emb3 = keras.layers.Bidirectional( keras.layers.LSTM( units=lstm_num_units, return_sequences=True, return_state=True, dropout=drop_rate, recurrent_dropout=drop_rate), merge_mode='concat')(input_, initial_state=emb2[1:5]) emb3_maxpooling = keras.layers.GlobalMaxPooling1D()(emb3[0]) emb = keras.layers.Concatenate(axis=1)( [emb1_maxpooling, emb2_maxpooling, emb3_maxpooling]) return emb
5,896
37.045161
79
py
MatchZoo
MatchZoo-master/matchzoo/contrib/models/diin.py
"""DIIN model.""" import typing import keras import keras.backend as K import tensorflow as tf from matchzoo import preprocessors from matchzoo.contrib.layers import DecayingDropoutLayer from matchzoo.contrib.layers import EncodingLayer from matchzoo.engine import hyper_spaces from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable class DIIN(BaseModel): """ DIIN model. Examples: >>> model = DIIN() >>> model.guess_and_fill_missing_params() >>> model.params['embedding_input_dim'] = 10000 >>> model.params['embedding_output_dim'] = 300 >>> model.params['embedding_trainable'] = True >>> model.params['optimizer'] = 'adam' >>> model.params['dropout_initial_keep_rate'] = 1.0 >>> model.params['dropout_decay_interval'] = 10000 >>> model.params['dropout_decay_rate'] = 0.977 >>> model.params['char_embedding_input_dim'] = 100 >>> model.params['char_embedding_output_dim'] = 8 >>> model.params['char_conv_filters'] = 100 >>> model.params['char_conv_kernel_size'] = 5 >>> model.params['first_scale_down_ratio'] = 0.3 >>> model.params['nb_dense_blocks'] = 3 >>> model.params['layers_per_dense_block'] = 8 >>> model.params['growth_rate'] = 20 >>> model.params['transition_scale_down_ratio'] = 0.5 >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params['optimizer'] = 'adam' params.add(Param(name='dropout_decay_interval', value=10000, desc="The decay interval of decaying_dropout.")) params.add(Param(name='char_embedding_input_dim', value=100, desc="The input dimension of character embedding " "layer.")) params.add(Param(name='char_embedding_output_dim', value=2, desc="The output dimension of character embedding " "layer.")) params.add(Param(name='char_conv_filters', value=8, desc="The filter size of character convolution " "layer.")) params.add(Param(name='char_conv_kernel_size', value=2, desc="The kernel size of character convolution " "layer.")) params.add(Param(name='first_scale_down_ratio', value=0.3, desc="The channel scale down ratio of the " "convolution layer before densenet.")) params.add(Param(name='nb_dense_blocks', value=1, desc="The number of blocks in densenet.")) params.add(Param(name='layers_per_dense_block', value=2, desc="The number of convolution layers in dense " "block.")) params.add(Param(name='growth_rate', value=2, desc="The filter size of each convolution layer in " "dense block.")) params.add(Param(name='transition_scale_down_ratio', value=0.5, desc="The channel scale down ratio of the " "convolution layer in transition block.")) params.add(Param( name='dropout_initial_keep_rate', value=1.0, hyper_space=hyper_spaces.quniform( low=0.8, high=1.0, q=0.02), desc="The initial keep rate of decaying_dropout." )) params.add(Param( name='dropout_decay_rate', value=0.97, hyper_space=hyper_spaces.quniform( low=0.90, high=0.99, q=0.01), desc="The decay rate of decaying_dropout." )) return params @classmethod def get_default_preprocessor(cls): """:return: Default preprocessor.""" return preprocessors.DIINPreprocessor() def guess_and_fill_missing_params(self, verbose: int = 1): """ Guess and fill missing parameters in :attr:'params'. Use this method to automatically fill-in hyper parameters. This involves some guessing so the parameter it fills could be wrong. For example, the default task is 'Ranking', and if we do not set it to 'Classification' manually for data packs prepared for classification, then the shape of the model output and the data will mismatch. :param verbose: Verbosity. """ self._params.get('input_shapes').set_default([(32,), (32,), (32, 16), (32, 16), (32,), (32,)], verbose) super().guess_and_fill_missing_params(verbose) def _make_inputs(self) -> list: text_left = keras.layers.Input( name='text_left', shape=self._params['input_shapes'][0] ) text_right = keras.layers.Input( name='text_right', shape=self._params['input_shapes'][1] ) char_left = keras.layers.Input( name='char_left', shape=self._params['input_shapes'][2] ) char_right = keras.layers.Input( name='char_right', shape=self._params['input_shapes'][3] ) match_left = keras.layers.Input( name='match_left', shape=self._params['input_shapes'][4] ) match_right = keras.layers.Input( name='match_right', shape=self._params['input_shapes'][5] ) return [text_left, text_right, char_left, char_right, match_left, match_right] def build(self): """Build model structure.""" # Scalar dimensions referenced here: # B = batch size (number of sequences) # D = word embedding size # L = 'input_left' sequence length # R = 'input_right' sequence length # C = fixed word length inputs = self._make_inputs() # Left text and right text. # shape = [B, L] # shape = [B, R] text_left, text_right = inputs[0:2] # Left character and right character. # shape = [B, L, C] # shape = [B, R, C] char_left, char_right = inputs[2:4] # Left exact match and right exact match. # shape = [B, L] # shape = [B, R] match_left, match_right = inputs[4:6] # Embedding module left_embeddings = [] right_embeddings = [] # Word embedding feature word_embedding = self._make_embedding_layer() # shape = [B, L, D] left_word_embedding = word_embedding(text_left) # shape = [B, R, D] right_word_embedding = word_embedding(text_right) left_word_embedding = DecayingDropoutLayer( initial_keep_rate=self._params['dropout_initial_keep_rate'], decay_interval=self._params['dropout_decay_interval'], decay_rate=self._params['dropout_decay_rate'] )(left_word_embedding) right_word_embedding = DecayingDropoutLayer( initial_keep_rate=self._params['dropout_initial_keep_rate'], decay_interval=self._params['dropout_decay_interval'], decay_rate=self._params['dropout_decay_rate'] )(right_word_embedding) left_embeddings.append(left_word_embedding) right_embeddings.append(right_word_embedding) # Exact match feature # shape = [B, L, 1] left_exact_match = keras.layers.Reshape( target_shape=(K.int_shape(match_left)[1], 1,) )(match_left) # shape = [B, R, 1] right_exact_match = keras.layers.Reshape( target_shape=(K.int_shape(match_left)[1], 1,) )(match_right) left_embeddings.append(left_exact_match) right_embeddings.append(right_exact_match) # Char embedding feature char_embedding = self._make_char_embedding_layer() char_embedding.build( input_shape=(None, None, K.int_shape(char_left)[-1])) left_char_embedding = char_embedding(char_left) right_char_embedding = char_embedding(char_right) left_embeddings.append(left_char_embedding) right_embeddings.append(right_char_embedding) # Concatenate left_embedding = keras.layers.Concatenate()(left_embeddings) right_embedding = keras.layers.Concatenate()(right_embeddings) d = K.int_shape(left_embedding)[-1] # Encoding module left_encoding = EncodingLayer( initial_keep_rate=self._params['dropout_initial_keep_rate'], decay_interval=self._params['dropout_decay_interval'], decay_rate=self._params['dropout_decay_rate'] )(left_embedding) right_encoding = EncodingLayer( initial_keep_rate=self._params['dropout_initial_keep_rate'], decay_interval=self._params['dropout_decay_interval'], decay_rate=self._params['dropout_decay_rate'] )(right_embedding) # Interaction module interaction = keras.layers.Lambda(self._make_interaction)( [left_encoding, right_encoding]) # Feature extraction module feature_extractor_input = keras.layers.Conv2D( filters=int(d * self._params['first_scale_down_ratio']), kernel_size=(1, 1), activation=None)(interaction) feature_extractor = self._create_densenet() features = feature_extractor(feature_extractor_input) # Output module features = DecayingDropoutLayer( initial_keep_rate=self._params['dropout_initial_keep_rate'], decay_interval=self._params['dropout_decay_interval'], decay_rate=self._params['dropout_decay_rate'])(features) out = self._make_output_layer()(features) self._backend = keras.Model(inputs=inputs, outputs=out) def _make_char_embedding_layer(self) -> keras.layers.Layer: """ Apply embedding, conv and maxpooling operation over time dimension for each token to obtain a vector. :return: Wrapper Keras 'Layer' as character embedding feature extractor. """ return keras.layers.TimeDistributed(keras.Sequential([ keras.layers.Embedding( input_dim=self._params['char_embedding_input_dim'], output_dim=self._params['char_embedding_output_dim'], input_length=self._params['input_shapes'][2][-1]), keras.layers.Conv1D( filters=self._params['char_conv_filters'], kernel_size=self._params['char_conv_kernel_size']), keras.layers.GlobalMaxPooling1D()])) def _make_interaction(self, inputs_) -> typing.Any: left_encoding = inputs_[0] right_encoding = inputs_[1] left_encoding = tf.expand_dims(left_encoding, axis=2) right_encoding = tf.expand_dims(right_encoding, axis=1) interaction = left_encoding * right_encoding return interaction def _create_densenet(self) -> typing.Callable: """ DenseNet is consisted of 'nb_dense_blocks' sets of Dense block and Transition block pair. :return: Wrapper Keras 'Layer' as DenseNet, tensor in tensor out. """ def _wrapper(x): for _ in range(self._params['nb_dense_blocks']): # Dense block # Apply 'layers_per_dense_block' convolution layers. for _ in range(self._params['layers_per_dense_block']): out_conv = keras.layers.Conv2D( filters=self._params['growth_rate'], kernel_size=(3, 3), padding='same', activation='relu')(x) x = keras.layers.Concatenate(axis=-1)([x, out_conv]) # Transition block # Apply a convolution layer and a maxpooling layer. scale_down_ratio = self._params['transition_scale_down_ratio'] nb_filter = int(K.int_shape(x)[-1] * scale_down_ratio) x = keras.layers.Conv2D( filters=nb_filter, kernel_size=(1, 1), padding='same', activation=None)(x) x = keras.layers.MaxPool2D(strides=(2, 2))(x) out_densenet = keras.layers.Flatten()(x) return out_densenet return _wrapper
12,960
40.27707
78
py
MatchZoo
MatchZoo-master/matchzoo/contrib/models/match_srnn.py
"""An implementation of Match-SRNN Model.""" import keras from matchzoo.contrib.layers import MatchingTensorLayer from matchzoo.contrib.layers import SpatialGRU from matchzoo.engine import hyper_spaces from matchzoo.engine.base_model import BaseModel from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable class MatchSRNN(BaseModel): """ Match-SRNN Model. Examples: >>> model = MatchSRNN() >>> model.params['channels'] = 4 >>> model.params['units'] = 10 >>> model.params['dropout_rate'] = 0.0 >>> model.params['direction'] = 'lt' >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params.add(Param(name='channels', value=4, desc="Number of word interaction tensor channels")) params.add(Param(name='units', value=10, desc="Number of SpatialGRU units")) params.add(Param(name='direction', value='lt', desc="Direction of SpatialGRU scanning")) params.add(Param( name='dropout_rate', value=0.0, hyper_space=hyper_spaces.quniform(low=0.0, high=0.8, q=0.01), desc="The dropout rate." )) return params def build(self): """ Build model structure. Match-SRNN: Modeling the Recursive Matching Structure with Spatial RNN """ # Scalar dimensions referenced here: # B = batch size (number of sequences) # D = embedding size # L = `input_left` sequence length # R = `input_right` sequence length # C = number of channels # Left input and right input. # query = [B, L] # doc = [B, R] query, doc = self._make_inputs() # Process left and right input. # embed_query = [B, L, D] # embed_doc = [B, R, D] embedding = self._make_embedding_layer() embed_query = embedding(query) embed_doc = embedding(doc) # Get matching tensor # matching_tensor = [B, C, L, R] matching_tensor_layer = MatchingTensorLayer( channels=self._params['channels']) matching_tensor = matching_tensor_layer([embed_query, embed_doc]) # Apply spatial GRU to the word level interaction tensor # h_ij = [B, U] spatial_gru = SpatialGRU( units=self._params['units'], direction=self._params['direction']) h_ij = spatial_gru(matching_tensor) # Apply Dropout x = keras.layers.Dropout( rate=self._params['dropout_rate'])(h_ij) # Make output layer x_out = self._make_output_layer()(x) self._backend = keras.Model(inputs=[query, doc], outputs=x_out)
3,058
31.542553
76
py
MatchZoo
MatchZoo-master/matchzoo/contrib/models/bimpm.py
"""BiMPM.""" from keras.models import Model from keras.layers import Dense, Concatenate, Dropout from keras.layers import Bidirectional, LSTM from matchzoo.engine.param import Param from matchzoo.engine.param_table import ParamTable from matchzoo.engine.base_model import BaseModel from matchzoo.contrib.layers import MultiPerspectiveLayer class BiMPM(BaseModel): """ BiMPM. Reference: https://github.com/zhiguowang/BiMPM/blob/master/src/SentenceMatchModelGraph.py#L43-L186 Examples: >>> import matchzoo as mz >>> model = mz.contrib.models.BiMPM() >>> model.guess_and_fill_missing_params(verbose=0) >>> model.build() """ @classmethod def get_default_params(cls) -> ParamTable: """:return: model default parameters.""" params = super().get_default_params(with_embedding=True) params['optimizer'] = 'adam' # params.add(Param('dim_word_embedding', 50)) # TODO(tjf): remove unused params in the final version # params.add(Param('dim_char_embedding', 50)) # params.add(Param('word_embedding_mat')) # params.add(Param('char_embedding_mat')) # params.add(Param('embedding_random_scale', 0.2)) # params.add(Param('activation_embedding', 'softmax')) # BiMPM Setting params.add(Param('perspective', {'full': True, 'max-pooling': True, 'attentive': True, 'max-attentive': True})) params.add(Param('mp_dim', 3)) params.add(Param('att_dim', 3)) params.add(Param('hidden_size', 4)) params.add(Param('dropout_rate', 0.0)) params.add(Param('w_initializer', 'glorot_uniform')) params.add(Param('b_initializer', 'zeros')) params.add(Param('activation_hidden', 'linear')) params.add(Param('with_match_highway', False)) params.add(Param('with_aggregation_highway', False)) return params def build(self): """Build model structure.""" # ~ Input Layer input_left, input_right = self._make_inputs() # Word Representation Layer # TODO: concatenate word level embedding and character level embedding. embedding = self._make_embedding_layer() embed_left = embedding(input_left) embed_right = embedding(input_right) # L119-L121 # https://github.com/zhiguowang/BiMPM/blob/master/src/SentenceMatchModelGraph.py#L119-L121 embed_left = Dropout(self._params['dropout_rate'])(embed_left) embed_right = Dropout(self._params['dropout_rate'])(embed_right) # ~ Word Level Matching Layer # Reference: # https://github.com/zhiguowang/BiMPM/blob/master/src/match_utils.py#L207-L223 # TODO pass # ~ Encoding Layer # Note: When merge_mode = None, output will be [forward, backward], # The default merge_mode is concat, and the output will be [lstm]. # If with return_state, then the output would append [h,c,h,c]. bi_lstm = Bidirectional( LSTM(self._params['hidden_size'], return_sequences=True, return_state=True, dropout=self._params['dropout_rate'], kernel_initializer=self._params['w_initializer'], bias_initializer=self._params['b_initializer']), merge_mode='concat') # x_left = [lstm_lt, forward_h_lt, _, backward_h_lt, _ ] x_left = bi_lstm(embed_left) x_right = bi_lstm(embed_right) # ~ Multi-Perspective Matching layer. # Output is two sequence of vectors. # Cons: Haven't support multiple context layer multi_perspective = MultiPerspectiveLayer(self._params['att_dim'], self._params['mp_dim'], self._params['perspective']) # Note: input to `keras layer` must be list of tensors. mp_left = multi_perspective(x_left + x_right) mp_right = multi_perspective(x_right + x_left) # ~ Dropout Layer mp_left = Dropout(self._params['dropout_rate'])(mp_left) mp_right = Dropout(self._params['dropout_rate'])(mp_right) # ~ Highway Layer # reference: # https://github.com/zhiguowang/BiMPM/blob/master/src/match_utils.py#L289-L295 if self._params['with_match_highway']: # the input is left matching representations (question / passage) pass # ~ Aggregation layer # TODO: mask the above layer aggregation = Bidirectional( LSTM(self._params['hidden_size'], return_sequences=False, return_state=False, dropout=self._params['dropout_rate'], kernel_initializer=self._params['w_initializer'], bias_initializer=self._params['b_initializer']), merge_mode='concat') rep_left = aggregation(mp_left) rep_right = aggregation(mp_right) # Concatenate the concatenated vector of left and right. x = Concatenate()([rep_left, rep_right]) # ~ Highway Network # reference: # https://github.com/zhiguowang/BiMPM/blob/master/src/match_utils.py#L289-L295 if self._params['with_aggregation_highway']: pass # ~ Prediction layer. # reference: # https://github.com/zhiguowang/BiMPM/blob/master/src/SentenceMatchModelGraph.py#L140-L153 x = Dense(self._params['hidden_size'], activation=self._params['activation_hidden'])(x) x = Dense(self._params['hidden_size'], activation=self._params['activation_hidden'])(x) x_out = self._make_output_layer()(x) self._backend = Model(inputs=[input_left, input_right], outputs=x_out)
6,010
39.073333
98
py
MatchZoo
MatchZoo-master/matchzoo/contrib/layers/decaying_dropout_layer.py
"""An implementation of Decaying Dropout Layer.""" import tensorflow as tf from keras import backend as K from keras.engine import Layer class DecayingDropoutLayer(Layer): """ Layer that processes dropout with exponential decayed keep rate during training. :param initial_keep_rate: the initial keep rate of decaying dropout. :param decay_interval: the decay interval of decaying dropout. :param decay_rate: the decay rate of decaying dropout. :param noise_shape: a 1D integer tensor representing the shape of the binary dropout mask that will be multiplied with the input. :param seed: a python integer to use as random seed. :param kwargs: standard layer keyword arguments. Examples: >>> import matchzoo as mz >>> layer = mz.contrib.layers.DecayingDropoutLayer( ... initial_keep_rate=1.0, ... decay_interval=10000, ... decay_rate=0.977, ... ) >>> num_batch, num_dim =5, 10 >>> layer.build([num_batch, num_dim]) """ def __init__(self, initial_keep_rate: float = 1.0, decay_interval: int = 10000, decay_rate: float = 0.977, noise_shape=None, seed=None, **kwargs): """:class: 'DecayingDropoutLayer' constructor.""" super(DecayingDropoutLayer, self).__init__(**kwargs) self._iterations = None self._initial_keep_rate = initial_keep_rate self._decay_interval = decay_interval self._decay_rate = min(1.0, max(0.0, decay_rate)) self._noise_shape = noise_shape self._seed = seed def _get_noise_shape(self, inputs): if self._noise_shape is None: return self._noise_shape symbolic_shape = tf.shape(inputs) noise_shape = [symbolic_shape[axis] if shape is None else shape for axis, shape in enumerate(self._noise_shape)] return tuple(noise_shape) def build(self, input_shape): """ Build the layer. :param input_shape: the shape of the input tensor, for DecayingDropoutLayer we need one input tensor. """ self._iterations = self.add_weight(name='iterations', shape=(1,), dtype=K.floatx(), initializer='zeros', trainable=False) super(DecayingDropoutLayer, self).build(input_shape) def call(self, inputs, training=None): """ The computation logic of DecayingDropoutLayer. :param inputs: an input tensor. """ noise_shape = self._get_noise_shape(inputs) t = tf.cast(self._iterations, K.floatx()) + 1 p = t / float(self._decay_interval) keep_rate = self._initial_keep_rate * tf.pow(self._decay_rate, p) def dropped_inputs(): update_op = self._iterations.assign_add([1]) with tf.control_dependencies([update_op]): return tf.nn.dropout(inputs, 1 - keep_rate[0], noise_shape, seed=self._seed) return K.in_train_phase(dropped_inputs, inputs, training=training) def get_config(self): """Get the config dict of DecayingDropoutLayer.""" config = {'initial_keep_rate': self._initial_keep_rate, 'decay_interval': self._decay_interval, 'decay_rate': self._decay_rate, 'noise_shape': self._noise_shape, 'seed': self._seed} base_config = super(DecayingDropoutLayer, self).get_config() return dict(list(base_config.items()) + list(config.items()))
3,805
37.06
75
py
MatchZoo
MatchZoo-master/matchzoo/contrib/layers/spatial_gru.py
"""An implementation of Spatial GRU Layer.""" import typing import tensorflow as tf from keras import backend as K from keras.engine import Layer from keras.layers import Permute from keras.layers import Reshape from keras import activations from keras import initializers class SpatialGRU(Layer): """ Spatial GRU layer. :param units: Number of SpatialGRU units. :param activation: Activation function to use. Default: hyperbolic tangent (`tanh`). If you pass `None`, no activation is applied (ie. "linear" activation: `a(x) = x`). :param recurrent_activation: Activation function to use for the recurrent step. Default: sigmoid (`sigmoid`). If you pass `None`, no activation is applied (ie. "linear" activation: `a(x) = x`). :param kernel_initializer: Initializer for the `kernel` weights matrix, used for the linear transformation of the inputs. :param recurrent_initializer: Initializer for the `recurrent_kernel` weights matrix, used for the linear transformation of the recurrent state. :param direction: Scanning direction. `lt` (i.e., left top) indicates the scanning from left top to right bottom, and `rb` (i.e., right bottom) indicates the scanning from right bottom to left top. :param kwargs: Standard layer keyword arguments. Examples: >>> import matchzoo as mz >>> layer = mz.contrib.layers.SpatialGRU(units=10, ... direction='lt') >>> num_batch, channel, left_len, right_len = 5, 5, 3, 2 >>> layer.build([num_batch, channel, left_len, right_len]) """ def __init__( self, units: int = 10, activation: str = 'tanh', recurrent_activation: str = 'sigmoid', kernel_initializer: str = 'glorot_uniform', recurrent_initializer: str = 'orthogonal', direction: str = 'lt', **kwargs ): """:class:`SpatialGRU` constructor.""" super().__init__(**kwargs) self._units = units self._activation = activations.get(activation) self._recurrent_activation = activations.get(recurrent_activation) self._kernel_initializer = initializers.get(kernel_initializer) self._recurrent_initializer = initializers.get(recurrent_initializer) self._direction = direction def build(self, input_shape: typing.Any): """ Build the layer. :param input_shape: the shapes of the input tensors. """ # Scalar dimensions referenced here: # B = batch size (number of sequences) # L = `input_left` sequence length # R = `input_right` sequence length # C = number of channels # U = number of units # input_shape = [B, C, L, R] self._batch_size = input_shape[0] self._channel = input_shape[1] self._input_dim = self._channel + 3 * self._units self._text1_maxlen = input_shape[2] self._text2_maxlen = input_shape[3] self._recurrent_step = self._text1_maxlen * self._text2_maxlen # W = [3*U+C, 7*U] self._W = self.add_weight( name='W', shape=(self._input_dim, self._units * 7), initializer=self._kernel_initializer, trainable=True ) # U = [3*U, U] self._U = self.add_weight( name='U', shape=(self._units * 3, self._units), initializer=self._recurrent_initializer, trainable=True ) # bias = [8*U,] self._bias = self.add_weight( name='bias', shape=(self._units * 8,), initializer='zeros', trainable=True ) # w_rl, w_rt, w_rd = [B, 3*U] self._wr = self._W[:, :self._units * 3] # b_rl, b_rt, b_rd = [B, 3*U] self._br = self._bias[:self._units * 3] # w_zi, w_zl, w_zt, w_zd = [B, 4*U] self._wz = self._W[:, self._units * 3: self._units * 7] # b_zi, b_zl, b_zt, b_zd = [B, 4*U] self._bz = self._bias[self._units * 3: self._units * 7] # w_ij = [C, U] self._w_ij = self.add_weight( name='W_ij', shape=(self._channel, self._units), initializer=self._recurrent_initializer, trainable=True ) # b_ij = [7*U] self._b_ij = self._bias[self._units * 7:] super(SpatialGRU, self).build(input_shape) def softmax_by_row(self, z: typing.Any) -> tuple: """Conduct softmax on each dimension across the four gates.""" # z_transform: [B, U, 4] z_transform = Permute((2, 1))(Reshape((4, self._units))(z)) size = [-1, 1, -1] # Perform softmax on each slice for i in range(0, self._units): begin = [0, i, 0] # z_slice: [B, 1, 4] z_slice = tf.slice(z_transform, begin, size) if i == 0: z_s = tf.nn.softmax(z_slice) else: z_s = tf.concat([z_s, tf.nn.softmax(z_slice)], 1) # zi, zl, zt, zd: [B, U] zi, zl, zt, zd = tf.unstack(z_s, axis=2) return zi, zl, zt, zd def calculate_recurrent_unit( self, inputs: typing.Any, states: typing.Any, step: int, h: typing.Any, ) -> tuple: """ Calculate recurrent unit. :param inputs: A TensorArray which contains interaction between left text and right text. :param states: A TensorArray which stores the hidden state of every step. :param step: Recurrent step. :param h: Hidden state from last operation. """ # Get index i, j i = tf.math.floordiv(step, tf.constant(self._text2_maxlen)) j = tf.math.mod(step, tf.constant(self._text2_maxlen)) # Get hidden state h_diag, h_top, h_left # h_diag, h_top, h_left = [B, U] h_diag = states.read(i * (self._text2_maxlen + 1) + j) h_top = states.read(i * (self._text2_maxlen + 1) + j + 1) h_left = states.read((i + 1) * (self._text2_maxlen + 1) + j) # Get interaction between word i, j: s_ij # s_ij = [B, C] s_ij = inputs.read(step) # Concatenate h_top, h_left, h_diag, s_ij # q = [B, 3*U+C] q = tf.concat([tf.concat([h_top, h_left], 1), tf.concat([h_diag, s_ij], 1)], 1) # Calculate reset gate # r = [B, 3*U] r = self._recurrent_activation( self._time_distributed_dense(self._wr, q, self._br)) # Calculate updating gate # z: [B, 4*U] z = self._time_distributed_dense(self._wz, q, self._bz) # Perform softmax # zi, zl, zt, zd: [B, U] zi, zl, zt, zd = self.softmax_by_row(z) # Get h_ij_ # h_ij_ = [B, U] h_ij_l = self._time_distributed_dense(self._w_ij, s_ij, self._b_ij) h_ij_r = K.dot(r * (tf.concat([h_left, h_top, h_diag], 1)), self._U) h_ij_ = self._activation(h_ij_l + h_ij_r) # Calculate h_ij # h_ij = [B, U] h_ij = zl * h_left + zt * h_top + zd * h_diag + zi * h_ij_ # Write h_ij to states states = states.write(((i + 1) * (self._text2_maxlen + 1) + j + 1), h_ij) h_ij.set_shape(h_top.get_shape()) return inputs, states, step + 1, h_ij def call(self, inputs: list, **kwargs) -> typing.Any: """ The computation logic of SpatialGRU. :param inputs: input tensors. """ batch_size = tf.shape(inputs)[0] # h0 = [B, U] self._bounder_state_h0 = tf.zeros([batch_size, self._units]) # input_x = [L, R, B, C] input_x = tf.transpose(inputs, [2, 3, 0, 1]) if self._direction == 'rb': # input_x: [R, L, B, C] input_x = tf.reverse(input_x, [0, 1]) elif self._direction != 'lt': raise ValueError(f"Invalid direction. " f"`{self._direction}` received. " f"Must be in `lt`, `rb`.") # input_x = [L*R*B, C] input_x = tf.reshape(input_x, [-1, self._channel]) # input_x = L*R * [B, C] input_x = tf.split( axis=0, num_or_size_splits=self._text1_maxlen * self._text2_maxlen, value=input_x ) # inputs = L*R * [B, C] inputs = tf.TensorArray( dtype=tf.float32, size=self._text1_maxlen * self._text2_maxlen, name='inputs' ) inputs = inputs.unstack(input_x) # states = (L+1)*(R+1) * [B, U] states = tf.TensorArray( dtype=tf.float32, size=(self._text1_maxlen + 1) * (self._text2_maxlen + 1), name='states', clear_after_read=False ) # Initialize states for i in range(self._text2_maxlen + 1): states = states.write(i, self._bounder_state_h0) for i in range(1, self._text1_maxlen + 1): states = states.write(i * (self._text2_maxlen + 1), self._bounder_state_h0) # Calculate h_ij # h_ij = [B, U] _, _, _, h_ij = tf.while_loop( cond=lambda _0, _1, i, _3: tf.less(i, self._recurrent_step), body=self.calculate_recurrent_unit, loop_vars=( inputs, states, tf.constant(0, dtype=tf.int32), self._bounder_state_h0 ), parallel_iterations=1, swap_memory=True ) return h_ij def compute_output_shape(self, input_shape: typing.Any) -> tuple: """ Calculate the layer output shape. :param input_shape: the shapes of the input tensors. """ output_shape = [input_shape[0], self._units] return tuple(output_shape) @classmethod def _time_distributed_dense(cls, w, x, b): x = K.dot(x, w) x = K.bias_add(x, b) return x
10,175
33.969072
77
py
MatchZoo
MatchZoo-master/matchzoo/contrib/layers/attention_layer.py
"""An implementation of Attention Layer for Bimpm model.""" import tensorflow as tf from keras import backend as K from keras.engine import Layer class AttentionLayer(Layer): """ Layer that compute attention for BiMPM model. For detailed information, see Bilateral Multi-Perspective Matching for Natural Language Sentences, section 3.2. Reference: https://github.com/zhiguowang/BiMPM/blob/master/src/layer_utils.py#L145-L196 Examples: >>> import matchzoo as mz >>> layer = mz.contrib.layers.AttentionLayer(att_dim=50) >>> layer.compute_output_shape([(32, 10, 100), (32, 40, 100)]) (32, 10, 40) """ def __init__(self, att_dim: int, att_type: str = 'default', dropout_rate: float = 0.0): """ class: `AttentionLayer` constructor. :param att_dim: int :param att_type: int """ super(AttentionLayer, self).__init__() self._att_dim = att_dim self._att_type = att_type self._dropout_rate = dropout_rate @property def att_dim(self): """Get the attention dimension.""" return self._att_dim @property def att_type(self): """Get the attention type.""" return self._att_type def build(self, input_shapes): """ Build the layer. :param input_shapes: input_shape_lt, input_shape_rt """ if not isinstance(input_shapes, list): raise ValueError('A attention layer should be called ' 'on a list of inputs.') hidden_dim_lt = input_shapes[0][2] hidden_dim_rt = input_shapes[1][2] self.attn_w1 = self.add_weight(name='attn_w1', shape=(hidden_dim_lt, self._att_dim), initializer='uniform', trainable=True) if hidden_dim_lt == hidden_dim_rt: self.attn_w2 = self.attn_w1 else: self.attn_w2 = self.add_weight(name='attn_w2', shape=(hidden_dim_rt, self._att_dim), initializer='uniform', trainable=True) # diagonal_W: (1, 1, a) self.diagonal_W = self.add_weight(name='diagonal_W', shape=(1, 1, self._att_dim), initializer='uniform', trainable=True) self.built = True def call(self, x: list, **kwargs): """ Calculate attention. :param x: [reps_lt, reps_rt] :return attn_prob: [b, s_lt, s_rt] """ if not isinstance(x, list): raise ValueError('A attention layer should be called ' 'on a list of inputs.') reps_lt, reps_rt = x attn_w1 = self.attn_w1 attn_w1 = tf.expand_dims(tf.expand_dims(attn_w1, axis=0), axis=0) # => [1, 1, d, a] reps_lt = tf.expand_dims(reps_lt, axis=-1) attn_reps_lt = tf.reduce_sum(reps_lt * attn_w1, axis=2) # => [b, s_lt, d, -1] attn_w2 = self.attn_w2 attn_w2 = tf.expand_dims(tf.expand_dims(attn_w2, axis=0), axis=0) # => [1, 1, d, a] reps_rt = tf.expand_dims(reps_rt, axis=-1) attn_reps_rt = tf.reduce_sum(reps_rt * attn_w2, axis=2) # [b, s_rt, d, -1] attn_reps_lt = tf.tanh(attn_reps_lt) # [b, s_lt, a] attn_reps_rt = tf.tanh(attn_reps_rt) # [b, s_rt, a] # diagonal_W attn_reps_lt = attn_reps_lt * self.diagonal_W # [b, s_lt, a] attn_reps_rt = tf.transpose(attn_reps_rt, (0, 2, 1)) # => [b, a, s_rt] attn_value = K.batch_dot(attn_reps_lt, attn_reps_rt) # [b, s_lt, s_rt] # Softmax operation attn_prob = tf.nn.softmax(attn_value) # [b, s_lt, s_rt] # TODO(tjf) remove diagonal or not for normalization # if remove_diagonal: attn_value = attn_value * diagonal if len(x) == 4: mask_lt, mask_rt = x[2], x[3] attn_prob *= tf.expand_dims(mask_lt, axis=2) attn_prob *= tf.expand_dims(mask_rt, axis=1) return attn_prob def compute_output_shape(self, input_shapes): """Calculate the layer output shape.""" if not isinstance(input_shapes, list): raise ValueError('A attention layer should be called ' 'on a list of inputs.') input_shape_lt, input_shape_rt = input_shapes[0], input_shapes[1] return input_shape_lt[0], input_shape_lt[1], input_shape_rt[1]
4,960
33.213793
83
py
MatchZoo
MatchZoo-master/matchzoo/contrib/layers/semantic_composite_layer.py
"""An implementation of EncodingModule for DIIN model.""" import tensorflow as tf from keras import backend as K from keras.engine import Layer from matchzoo.contrib.layers import DecayingDropoutLayer class EncodingLayer(Layer): """ Apply a self-attention layer and a semantic composite fuse gate to compute the encoding result of one tensor. :param initial_keep_rate: the initial_keep_rate parameter of DecayingDropoutLayer. :param decay_interval: the decay_interval parameter of DecayingDropoutLayer. :param decay_rate: the decay_rate parameter of DecayingDropoutLayer. :param kwargs: standard layer keyword arguments. Example: >>> import matchzoo as mz >>> layer = mz.contrib.layers.EncodingLayer(1.0, 10000, 0.977) >>> num_batch, left_len, num_dim = 5, 32, 10 >>> layer.build([num_batch, left_len, num_dim]) """ def __init__(self, initial_keep_rate: float, decay_interval: int, decay_rate: float, **kwargs): """:class: 'EncodingLayer' constructor.""" super(EncodingLayer, self).__init__(**kwargs) self._initial_keep_rate = initial_keep_rate self._decay_interval = decay_interval self._decay_rate = decay_rate self._w_itr_att = None self._w1 = None self._w2 = None self._w3 = None self._b1 = None self._b2 = None self._b3 = None def build(self, input_shape): """ Build the layer. :param input_shape: the shape of the input tensor, for EncodingLayer we need one input tensor. """ d = input_shape[-1] self._w_itr_att = self.add_weight( name='w_itr_att', shape=(3 * d,), initializer='glorot_uniform') self._w1 = self.add_weight( name='w1', shape=(2 * d, d,), initializer='glorot_uniform') self._w2 = self.add_weight( name='w2', shape=(2 * d, d,), initializer='glorot_uniform') self._w3 = self.add_weight( name='w3', shape=(2 * d, d,), initializer='glorot_uniform') self._b1 = self.add_weight( name='b1', shape=(d,), initializer='zeros') self._b2 = self.add_weight( name='b2', shape=(d,), initializer='zeros') self._b3 = self.add_weight( name='b3', shape=(d,), initializer='zeros') super(EncodingLayer, self).build(input_shape) def call(self, inputs, **kwargs): """ The computation logic of EncodingLayer. :param inputs: an input tensor. """ # Scalar dimensions referenced here: # b = batch size # p = inputs.shape()[1] # d = inputs.shape()[2] # The input shape is [b, p, d] # shape = [b, 1, p, d] x = tf.expand_dims(inputs, 1) * 0 # shape = [b, 1, d, p] x = tf.transpose(x, (0, 1, 3, 2)) # shape = [b, p, d, p] mid = x + tf.expand_dims(inputs, -1) # shape = [b, p, d, p] up = tf.transpose(mid, (0, 3, 2, 1)) # shape = [b, p, 3d, p] inputs_concat = tf.concat([up, mid, up * mid], axis=2) # Self-attention layer. # shape = [b, p, p] A = K.dot(self._w_itr_att, inputs_concat) # shape = [b, p, p] SA = tf.nn.softmax(A, axis=2) # shape = [b, p, d] itr_attn = K.batch_dot(SA, inputs) # Semantic composite fuse gate. # shape = [b, p, 2d] inputs_attn_concat = tf.concat([inputs, itr_attn], axis=2) concat_dropout = DecayingDropoutLayer( initial_keep_rate=self._initial_keep_rate, decay_interval=self._decay_interval, decay_rate=self._decay_rate )(inputs_attn_concat) # shape = [b, p, d] z = tf.tanh(K.dot(concat_dropout, self._w1) + self._b1) # shape = [b, p, d] r = tf.sigmoid(K.dot(concat_dropout, self._w2) + self._b2) # shape = [b, p, d] f = tf.sigmoid(K.dot(concat_dropout, self._w3) + self._b3) # shape = [b, p, d] encoding = r * inputs + f * z return encoding
4,198
33.418033
75
py
MatchZoo
MatchZoo-master/matchzoo/contrib/layers/multi_perspective_layer.py
"""An implementation of MultiPerspectiveLayer for Bimpm model.""" import tensorflow as tf from keras import backend as K from keras.engine import Layer from matchzoo.contrib.layers.attention_layer import AttentionLayer class MultiPerspectiveLayer(Layer): """ A keras implementation of multi-perspective layer of BiMPM. For detailed information, see Bilateral Multi-Perspective Matching for Natural Language Sentences, section 3.2. Examples: >>> import matchzoo as mz >>> perspective={'full': True, 'max-pooling': True, ... 'attentive': True, 'max-attentive': True} >>> layer = mz.contrib.layers.MultiPerspectiveLayer( ... att_dim=50, mp_dim=20, perspective=perspective) >>> layer.compute_output_shape( ... [(32, 10, 100), (32, 50), None, (32, 50), None, ... [(32, 40, 100), (32, 50), None, (32, 50), None]]) (32, 10, 83) """ def __init__(self, att_dim: int, mp_dim: int, perspective: dict): """Class initialization.""" super(MultiPerspectiveLayer, self).__init__() self._att_dim = att_dim self._mp_dim = mp_dim self._perspective = perspective @classmethod def list_available_perspectives(cls) -> list: """List available strategy for multi-perspective matching.""" return ['full', 'max-pooling', 'attentive', 'max-attentive'] @property def num_perspective(self): """Get the number of perspectives that is True.""" return sum(self._perspective.values()) def build(self, input_shape: list): """Input shape.""" # The shape of the weights is l * d. if self._perspective.get('full'): self.full_match = MpFullMatch(self._mp_dim) if self._perspective.get('max-pooling'): self.max_pooling_match = MpMaxPoolingMatch(self._mp_dim) if self._perspective.get('attentive'): self.attentive_match = MpAttentiveMatch(self._att_dim, self._mp_dim) if self._perspective.get('max-attentive'): self.max_attentive_match = MpMaxAttentiveMatch(self._att_dim) self.built = True def call(self, x: list, **kwargs): """Call.""" seq_lt, seq_rt = x[:5], x[5:] # unpack seq_left and seq_right # all hidden states, last hidden state of forward pass, # last cell state of forward pass, last hidden state of # backward pass, last cell state of backward pass. lstm_reps_lt, forward_h_lt, _, backward_h_lt, _ = seq_lt lstm_reps_rt, forward_h_rt, _, backward_h_rt, _ = seq_rt match_tensor_list = [] match_dim = 0 if self._perspective.get('full'): # Each forward & backward contextual embedding compare # with the last step of the last time step of the other sentence. h_lt = tf.concat([forward_h_lt, backward_h_lt], axis=-1) full_match_tensor = self.full_match([h_lt, lstm_reps_rt]) match_tensor_list.append(full_match_tensor) match_dim += self._mp_dim + 1 if self._perspective.get('max-pooling'): # Each contextual embedding compare with each contextual embedding. # retain the maximum of each dimension. max_match_tensor = self.max_pooling_match([lstm_reps_lt, lstm_reps_rt]) match_tensor_list.append(max_match_tensor) match_dim += self._mp_dim if self._perspective.get('attentive'): # Each contextual embedding compare with each contextual embedding. # retain sum of weighted mean of each dimension. attentive_tensor = self.attentive_match([lstm_reps_lt, lstm_reps_rt]) match_tensor_list.append(attentive_tensor) match_dim += self._mp_dim + 1 if self._perspective.get('max-attentive'): # Each contextual embedding compare with each contextual embedding. # retain max of weighted mean of each dimension. relevancy_matrix = _calc_relevancy_matrix(lstm_reps_lt, lstm_reps_rt) max_attentive_tensor = self.max_attentive_match([lstm_reps_lt, lstm_reps_rt, relevancy_matrix]) match_tensor_list.append(max_attentive_tensor) match_dim += self._mp_dim + 1 mp_tensor = tf.concat(match_tensor_list, axis=-1) return mp_tensor def compute_output_shape(self, input_shape: list): """Compute output shape.""" shape_a = input_shape[0] match_dim = 0 if self._perspective.get('full'): match_dim += self._mp_dim + 1 if self._perspective.get('max-pooling'): match_dim += self._mp_dim if self._perspective.get('attentive'): match_dim += self._mp_dim + 1 if self._perspective.get('max-attentive'): match_dim += self._mp_dim + 1 return shape_a[0], shape_a[1], match_dim class MpFullMatch(Layer): """Mp Full Match Layer.""" def __init__(self, mp_dim): """Init.""" super(MpFullMatch, self).__init__() self.mp_dim = mp_dim def build(self, input_shapes): """Build.""" # input_shape = input_shapes[0] self.built = True def call(self, x, **kwargs): """Call. """ rep_lt, reps_rt = x att_lt = tf.expand_dims(rep_lt, 1) match_tensor, match_dim = _multi_perspective_match(self.mp_dim, reps_rt, att_lt) # match_tensor => [b, len_rt, mp_dim+1] return match_tensor def compute_output_shape(self, input_shape): """Compute output shape.""" return input_shape[1][0], input_shape[1][1], self.mp_dim + 1 class MpMaxPoolingMatch(Layer): """MpMaxPoolingMatch.""" def __init__(self, mp_dim): """Init.""" super(MpMaxPoolingMatch, self).__init__() self.mp_dim = mp_dim def build(self, input_shapes): """Build.""" d = input_shapes[0][-1] self.kernel = self.add_weight(name='kernel', shape=(1, 1, 1, self.mp_dim, d), initializer='uniform', trainable=True) self.built = True def call(self, x, **kwargs): """Call.""" reps_lt, reps_rt = x # kernel: [1, 1, 1, mp_dim, d] # lstm_lt => [b, len_lt, 1, 1, d] reps_lt = tf.expand_dims(reps_lt, axis=2) reps_lt = tf.expand_dims(reps_lt, axis=2) reps_lt = reps_lt * self.kernel # lstm_rt -> [b, 1, len_rt, 1, d] reps_rt = tf.expand_dims(reps_rt, axis=2) reps_rt = tf.expand_dims(reps_rt, axis=1) match_tensor = _cosine_distance(reps_lt, reps_rt, cosine_norm=False) max_match_tensor = tf.reduce_max(match_tensor, axis=1) # match_tensor => [b, len_rt, m] return max_match_tensor def compute_output_shape(self, input_shape): """Compute output shape.""" return input_shape[1][0], input_shape[1][1], self.mp_dim class MpAttentiveMatch(Layer): """ MpAttentiveMatch Layer. Reference: https://github.com/zhiguowang/BiMPM/blob/master/src/match_utils.py#L188-L193 Examples: >>> import matchzoo as mz >>> layer = mz.contrib.layers.multi_perspective_layer.MpAttentiveMatch( ... att_dim=30, mp_dim=20) >>> layer.compute_output_shape([(32, 10, 100), (32, 40, 100)]) (32, 40, 20) """ def __init__(self, att_dim, mp_dim): """Init.""" super(MpAttentiveMatch, self).__init__() self.att_dim = att_dim self.mp_dim = mp_dim def build(self, input_shapes): """Build.""" # input_shape = input_shapes[0] self.built = True def call(self, x, **kwargs): """Call.""" reps_lt, reps_rt = x[0], x[1] # attention prob matrix attention_layer = AttentionLayer(self.att_dim) attn_prob = attention_layer([reps_rt, reps_lt]) # attention reps att_lt = K.batch_dot(attn_prob, reps_lt) # mp match attn_match_tensor, match_dim = _multi_perspective_match(self.mp_dim, reps_rt, att_lt) return attn_match_tensor def compute_output_shape(self, input_shape): """Compute output shape.""" return input_shape[1][0], input_shape[1][1], self.mp_dim class MpMaxAttentiveMatch(Layer): """MpMaxAttentiveMatch.""" def __init__(self, mp_dim): """Init.""" super(MpMaxAttentiveMatch, self).__init__() self.mp_dim = mp_dim def build(self, input_shapes): """Build.""" # input_shape = input_shapes[0] self.built = True def call(self, x): """Call.""" reps_lt, reps_rt = x[0], x[1] relevancy_matrix = x[2] max_att_lt = cal_max_question_representation(reps_lt, relevancy_matrix) max_attentive_tensor, match_dim = _multi_perspective_match(self.mp_dim, reps_rt, max_att_lt) return max_attentive_tensor def cal_max_question_representation(reps_lt, attn_scores): """ Calculate max_question_representation. :param reps_lt: [batch_size, passage_len, hidden_size] :param attn_scores: [] :return: [batch_size, passage_len, hidden_size]. """ attn_positions = tf.argmax(attn_scores, axis=2) max_reps_lt = collect_representation(reps_lt, attn_positions) return max_reps_lt def collect_representation(representation, positions): """ Collect_representation. :param representation: [batch_size, node_num, feature_dim] :param positions: [batch_size, neighbour_num] :return: [batch_size, neighbour_num]? """ return collect_probs(representation, positions) def collect_final_step_of_lstm(lstm_representation, lengths): """ Collect final step of lstm. :param lstm_representation: [batch_size, len_rt, dim] :param lengths: [batch_size] :return: [batch_size, dim] """ lengths = tf.maximum(lengths, K.zeros_like(lengths)) batch_size = tf.shape(lengths)[0] # shape (batch_size) batch_nums = tf.range(0, limit=batch_size) # shape (batch_size, 2) indices = tf.stack((batch_nums, lengths), axis=1) result = tf.gather_nd(lstm_representation, indices, name='last-forwar-lstm') # [batch_size, dim] return result def collect_probs(probs, positions): """ Collect Probabilities. Reference: https://github.com/zhiguowang/BiMPM/blob/master/src/layer_utils.py#L128-L140 :param probs: [batch_size, chunks_size] :param positions: [batch_size, pair_size] :return: [batch_size, pair_size] """ batch_size = tf.shape(probs)[0] pair_size = tf.shape(positions)[1] # shape (batch_size) batch_nums = K.arange(0, batch_size) # [batch_size, 1] batch_nums = tf.reshape(batch_nums, shape=[-1, 1]) # [batch_size, pair_size] batch_nums = K.tile(batch_nums, [1, pair_size]) # shape (batch_size, pair_size, 2) # Alert: to solve error message positions = tf.cast(positions, tf.int32) indices = tf.stack([batch_nums, positions], axis=2) pair_probs = tf.gather_nd(probs, indices) # pair_probs = tf.reshape(pair_probs, shape=[batch_size, pair_size]) return pair_probs def _multi_perspective_match(mp_dim, reps_rt, att_lt, with_cosine=True, with_mp_cosine=True): """ The core function of zhiguowang's implementation. reference: https://github.com/zhiguowang/BiMPM/blob/master/src/match_utils.py#L207-L223 :param mp_dim: about 20 :param reps_rt: [batch, len_rt, dim] :param att_lt: [batch, len_rt, dim] :param with_cosine: True :param with_mp_cosine: True :return: [batch, len, 1 + mp_dim] """ shape_rt = tf.shape(reps_rt) batch_size = shape_rt[0] len_lt = shape_rt[1] match_dim = 0 match_result_list = [] if with_cosine: cosine_tensor = _cosine_distance(reps_rt, att_lt, False) cosine_tensor = tf.reshape(cosine_tensor, [batch_size, len_lt, 1]) match_result_list.append(cosine_tensor) match_dim += 1 if with_mp_cosine: mp_cosine_layer = MpCosineLayer(mp_dim) mp_cosine_tensor = mp_cosine_layer([reps_rt, att_lt]) mp_cosine_tensor = tf.reshape(mp_cosine_tensor, [batch_size, len_lt, mp_dim]) match_result_list.append(mp_cosine_tensor) match_dim += mp_cosine_layer.mp_dim match_result = tf.concat(match_result_list, 2) return match_result, match_dim class MpCosineLayer(Layer): """ Implementation of Multi-Perspective Cosine Distance. Reference: https://github.com/zhiguowang/BiMPM/blob/master/src/match_utils.py#L121-L129 Examples: >>> import matchzoo as mz >>> layer = mz.contrib.layers.multi_perspective_layer.MpCosineLayer( ... mp_dim=50) >>> layer.compute_output_shape([(32, 10, 100), (32, 10, 100)]) (32, 10, 50) """ def __init__(self, mp_dim, **kwargs): """Init.""" self.mp_dim = mp_dim super(MpCosineLayer, self).__init__(**kwargs) def build(self, input_shape): """Build.""" self.kernel = self.add_weight(name='kernel', shape=(1, 1, self.mp_dim, input_shape[0][-1]), initializer='uniform', trainable=True) super(MpCosineLayer, self).build(input_shape) def call(self, x, **kwargs): """Call.""" v1, v2 = x v1 = tf.expand_dims(v1, 2) * self.kernel # [b, s_lt, m, d] v2 = tf.expand_dims(v2, 2) # [b, s_lt, 1, d] return _cosine_distance(v1, v2, False) def compute_output_shape(self, input_shape): """Compute output shape.""" return input_shape[0][0], input_shape[0][1], self.mp_dim def _calc_relevancy_matrix(reps_lt, reps_rt): reps_lt = tf.expand_dims(reps_lt, 1) # [b, 1, len_lt, d] reps_rt = tf.expand_dims(reps_rt, 2) # [b, len_rt, 1, d] relevancy_matrix = _cosine_distance(reps_lt, reps_rt) # => [b, len_rt, len_lt, d] return relevancy_matrix def _mask_relevancy_matrix(relevancy_matrix, mask_lt, mask_rt): """ Mask relevancy matrix. :param relevancy_matrix: [b, len_rt, len_lt] :param mask_lt: [b, len_lt] :param mask_rt: [b, len_rt] :return: masked_matrix: [b, len_rt, len_lt] """ if mask_lt is not None: relevancy_matrix = relevancy_matrix * tf.expand_dims(mask_lt, 1) relevancy_matrix = relevancy_matrix * tf.expand_dims(mask_rt, 2) return relevancy_matrix def _cosine_distance(v1, v2, cosine_norm=True, eps=1e-6): """ Only requires `tf.reduce_sum(v1 * v2, axis=-1)`. :param v1: [batch, time_steps(v1), 1, m, d] :param v2: [batch, 1, time_steps(v2), m, d] :param cosine_norm: True :param eps: 1e-6 :return: [batch, time_steps(v1), time_steps(v2), m] """ cosine_numerator = tf.reduce_sum(v1 * v2, axis=-1) if not cosine_norm: return K.tanh(cosine_numerator) v1_norm = K.sqrt(tf.maximum(tf.reduce_sum(tf.square(v1), axis=-1), eps)) v2_norm = K.sqrt(tf.maximum(tf.reduce_sum(tf.square(v2), axis=-1), eps)) return cosine_numerator / v1_norm / v2_norm
16,251
33.652452
80
py
MatchZoo
MatchZoo-master/matchzoo/contrib/layers/matching_tensor_layer.py
"""An implementation of Matching Tensor Layer.""" import typing import numpy as np import tensorflow as tf from keras import backend as K from keras.engine import Layer from keras.initializers import constant class MatchingTensorLayer(Layer): """ Layer that captures the basic interactions between two tensors. :param channels: Number of word interaction tensor channels :param normalize: Whether to L2-normalize samples along the dot product axis before taking the dot product. If set to True, then the output of the dot product is the cosine proximity between the two samples. :param init_diag: Whether to initialize the diagonal elements of the matrix. :param kwargs: Standard layer keyword arguments. Examples: >>> import matchzoo as mz >>> layer = mz.contrib.layers.MatchingTensorLayer(channels=4, ... normalize=True, ... init_diag=True) >>> num_batch, left_len, right_len, num_dim = 5, 3, 2, 10 >>> layer.build([[num_batch, left_len, num_dim], ... [num_batch, right_len, num_dim]]) """ def __init__(self, channels: int = 4, normalize: bool = True, init_diag: bool = True, **kwargs): """:class:`MatchingTensorLayer` constructor.""" super().__init__(**kwargs) self._channels = channels self._normalize = normalize self._init_diag = init_diag self._shape1 = None self._shape2 = None def build(self, input_shape: list): """ Build the layer. :param input_shape: the shapes of the input tensors, for MatchingTensorLayer we need two input tensors. """ # Used purely for shape validation. if not isinstance(input_shape, list) or len(input_shape) != 2: raise ValueError('A `MatchingTensorLayer` layer should be called ' 'on a list of 2 inputs.') self._shape1 = input_shape[0] self._shape2 = input_shape[1] for idx in (0, 2): if self._shape1[idx] != self._shape2[idx]: raise ValueError( 'Incompatible dimensions: ' f'{self._shape1[idx]} != {self._shape2[idx]}.' f'Layer shapes: {self._shape1}, {self._shape2}.' ) if self._init_diag: interaction_matrix = np.float32( np.random.uniform( -0.05, 0.05, [self._channels, self._shape1[2], self._shape2[2]] ) ) for channel_index in range(self._channels): np.fill_diagonal(interaction_matrix[channel_index], 0.1) self.interaction_matrix = self.add_weight( name='interaction_matrix', shape=(self._channels, self._shape1[2], self._shape2[2]), initializer=constant(interaction_matrix), trainable=True ) else: self.interaction_matrix = self.add_weight( name='interaction_matrix', shape=(self._channels, self._shape1[2], self._shape2[2]), initializer='uniform', trainable=True ) super(MatchingTensorLayer, self).build(input_shape) def call(self, inputs: list, **kwargs) -> typing.Any: """ The computation logic of MatchingTensorLayer. :param inputs: two input tensors. """ x1 = inputs[0] x2 = inputs[1] # Normalize x1 and x2 if self._normalize: x1 = K.l2_normalize(x1, axis=2) x2 = K.l2_normalize(x2, axis=2) # b = batch size # l = length of `x1` # r = length of `x2` # d, e = embedding size # c = number of channels # output = [b, c, l, r] output = tf.einsum( 'bld,cde,bre->bclr', x1, self.interaction_matrix, x2 ) return output def compute_output_shape(self, input_shape: list) -> tuple: """ Calculate the layer output shape. :param input_shape: the shapes of the input tensors, for MatchingTensorLayer we need two input tensors. """ if not isinstance(input_shape, list) or len(input_shape) != 2: raise ValueError('A `MatchingTensorLayer` layer should be called ' 'on a list of 2 inputs.') shape1 = list(input_shape[0]) shape2 = list(input_shape[1]) if len(shape1) != 3 or len(shape2) != 3: raise ValueError('A `MatchingTensorLayer` layer should be called ' 'on 2 inputs with 3 dimensions.') if shape1[0] != shape2[0] or shape1[2] != shape2[2]: raise ValueError('A `MatchingTensorLayer` layer should be called ' 'on 2 inputs with same 0,2 dimensions.') output_shape = [shape1[0], self._channels, shape1[1], shape2[1]] return tuple(output_shape)
5,169
37.014706
78
py
MatchZoo
MatchZoo-master/matchzoo/losses/rank_hinge_loss.py
"""The rank hinge loss.""" import numpy as np import tensorflow as tf from keras import layers, backend as K from keras.losses import Loss from keras.utils import losses_utils class RankHingeLoss(Loss): """ Rank hinge loss. Examples: >>> from keras import backend as K >>> x_pred = K.variable(np.array([[1.0], [1.2], [0.8], [1.4]])) >>> x_true = K.variable(np.array([[1], [0], [1], [0]])) >>> expect = ((1.0 + 1.2 - 1.0) + (1.0 + 1.4 - 0.8)) / 2 >>> expect 1.4 >>> loss = K.eval(RankHingeLoss(num_neg=1, margin=1.0)(x_true, x_pred)) >>> np.isclose(loss, expect) True """ def __init__(self, num_neg: int = 1, margin: float = 1.0): """ :class:`RankHingeLoss` constructor. :param num_neg: number of negative instances in hinge loss. :param margin: the margin between positive and negative scores. """ super().__init__(reduction=losses_utils.Reduction.SUM_OVER_BATCH_SIZE, name="rank_hinge") self._num_neg = num_neg self._margin = margin def call(self, y_true: np.array, y_pred: np.array, sample_weight=None) -> np.array: """ Calculate rank hinge loss. :param y_true: Label. :param y_pred: Predicted result. :return: Hinge loss computed by user-defined margin. """ y_pos = layers.Lambda(lambda a: a[::(self._num_neg + 1), :], output_shape=(1,))(y_pred) y_neg = [] for neg_idx in range(self._num_neg): y_neg.append( layers.Lambda( lambda a: a[(neg_idx + 1)::(self._num_neg + 1), :], output_shape=(1,))(y_pred)) y_neg = tf.concat(y_neg, axis=-1) y_neg = tf.reduce_mean(y_neg, axis=-1, keepdims=True) loss = tf.maximum(0., self._margin + y_neg - y_pos) return losses_utils.compute_weighted_loss( loss, sample_weight, reduction=self.reduction) @property def num_neg(self): """`num_neg` getter.""" return self._num_neg @property def margin(self): """`margin` getter.""" return self._margin
2,253
30.305556
79
py
MatchZoo
MatchZoo-master/matchzoo/losses/rank_cross_entropy_loss.py
"""The rank cross entropy loss.""" import numpy as np import tensorflow as tf from keras import layers, backend as K from keras.losses import Loss from keras.utils import losses_utils class RankCrossEntropyLoss(Loss): """ Rank cross entropy loss. Examples: >>> from keras import backend as K >>> softmax = lambda x: np.exp(x)/np.sum(np.exp(x), axis=0) >>> x_pred = K.variable(np.array([[1.0], [1.2], [0.8]])) >>> x_true = K.variable(np.array([[1], [0], [0]])) >>> expect = -np.log(softmax(np.array([[1.0], [1.2], [0.8]]))) >>> loss = K.eval(RankCrossEntropyLoss(num_neg=2)(x_true, x_pred)) >>> np.isclose(loss, expect[0]).all() True """ def __init__(self, num_neg: int = 1): """ :class:`RankCrossEntropyLoss` constructor. :param num_neg: number of negative instances in cross entropy loss. """ super().__init__(reduction=losses_utils.Reduction.SUM_OVER_BATCH_SIZE, name="rank_crossentropy") self._num_neg = num_neg def call(self, y_true: np.array, y_pred: np.array, sample_weight=None) -> np.array: """ Calculate rank cross entropy loss. :param y_true: Label. :param y_pred: Predicted result. :return: Crossentropy loss computed by user-defined negative number. """ logits = layers.Lambda(lambda a: a[::(self._num_neg + 1), :])(y_pred) labels = layers.Lambda(lambda a: a[::(self._num_neg + 1), :])(y_true) logits, labels = [logits], [labels] for neg_idx in range(self._num_neg): neg_logits = layers.Lambda( lambda a: a[neg_idx + 1::(self._num_neg + 1), :])(y_pred) neg_labels = layers.Lambda( lambda a: a[neg_idx + 1::(self._num_neg + 1), :])(y_true) logits.append(neg_logits) labels.append(neg_labels) logits = tf.concat(logits, axis=-1) labels = tf.concat(labels, axis=-1) smoothed_prob = tf.nn.softmax(logits) + np.finfo(float).eps loss = -(tf.reduce_sum(labels * tf.math.log(smoothed_prob), axis=-1)) return losses_utils.compute_weighted_loss( loss, sample_weight, reduction=self.reduction) @property def num_neg(self): """`num_neg` getter.""" return self._num_neg
2,389
35.212121
78
py
MatchZoo
MatchZoo-master/tests/unit_test/test_layers.py
import numpy as np import pytest from keras import backend as K from matchzoo import layers from matchzoo.contrib.layers import SpatialGRU from matchzoo.contrib.layers import MatchingTensorLayer def test_matching_layers(): s1_value = np.array([[[1, 2], [2, 3], [3, 4]], [[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]] ]) s2_value = np.array([[[1, 2], [2, 3]], [[0.1, 0.2], [0.2, 0.3]] ]) s3_value = np.array([[[1, 2], [2, 3]], [[0.1, 0.2], [0.2, 0.3]], [[0.1, 0.2], [0.2, 0.3]] ]) s1_tensor = K.variable(s1_value) s2_tensor = K.variable(s2_value) s3_tensor = K.variable(s3_value) for matching_type in ['dot', 'mul', 'plus', 'minus', 'concat']: model = layers.MatchingLayer(matching_type=matching_type)([s1_tensor, s2_tensor]) ret = K.eval(model) with pytest.raises(ValueError): layers.MatchingLayer(matching_type='error') with pytest.raises(ValueError): layers.MatchingLayer()([s1_tensor, s3_tensor]) def test_spatial_gru(): s_value = K.variable(np.array([[[[1, 2], [2, 3], [3, 4]], [[4, 5], [5, 6], [6, 7]]], [[[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]], [[0.4, 0.5], [0.5, 0.6], [0.6, 0.7]]]])) for direction in ['lt', 'rb']: model = SpatialGRU(direction=direction) _ = K.eval(model(s_value)) with pytest.raises(ValueError): SpatialGRU(direction='lr')(s_value) def test_matching_tensor_layer(): s1_value = np.array([[[1, 2], [2, 3], [3, 4]], [[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]]]) s2_value = np.array([[[1, 2], [2, 3]], [[0.1, 0.2], [0.2, 0.3]]]) s3_value = np.array([[[1, 2], [2, 3]], [[0.1, 0.2], [0.2, 0.3]], [[0.1, 0.2], [0.2, 0.3]]]) s1_tensor = K.variable(s1_value) s2_tensor = K.variable(s2_value) s3_tensor = K.variable(s3_value) for init_diag in [True, False]: model = MatchingTensorLayer(init_diag=init_diag) _ = K.eval(model([s1_tensor, s2_tensor])) with pytest.raises(ValueError): MatchingTensorLayer()([s1_tensor, s3_tensor])
2,369
37.852459
89
py
MatchZoo
MatchZoo-master/tests/unit_test/test_losses.py
import numpy as np from keras import backend as K from matchzoo import losses def test_hinge_loss(): true_value = K.variable(np.array([[1.2], [1], [1], [1]])) pred_value = K.variable(np.array([[1.2], [0.1], [0], [-0.3]])) expected_loss = (0 + 1 - 0.3 + 0) / 2.0 loss = K.eval(losses.RankHingeLoss()(true_value, pred_value)) assert np.isclose(expected_loss, loss) expected_loss = (2 + 0.1 - 1.2 + 2 - 0.3 + 0) / 2.0 loss = K.eval(losses.RankHingeLoss(margin=2)(true_value, pred_value)) assert np.isclose(expected_loss, loss) true_value = K.variable(np.array([[1.2], [1], [0.8], [1], [1], [0.8]])) pred_value = K.variable(np.array([[1.2], [0.1], [-0.5], [0], [0], [-0.3]])) expected_loss = (0 + 1 - 0.15) / 2.0 loss = K.eval(losses.RankHingeLoss(num_neg=2, margin=1)( true_value, pred_value)) assert np.isclose(expected_loss, loss) def test_rank_crossentropy_loss(): losses.neg_num = 1 def softmax(x): return np.exp(x) / np.sum(np.exp(x), axis=0) true_value = K.variable(np.array([[1.], [0.], [0.], [1.]])) pred_value = K.variable(np.array([[0.8], [0.1], [0.8], [0.1]])) expected_loss = (-np.log(softmax([0.8, 0.1])[0]) - np.log( softmax([0.8, 0.1])[1])) / 2 loss = K.eval(losses.RankCrossEntropyLoss()(true_value, pred_value)) assert np.isclose(expected_loss, loss) true_value = K.variable(np.array([[1.], [0.], [0.], [0.], [1.], [0.]])) pred_value = K.variable(np.array([[0.8], [0.1], [0.1], [0.8], [0.1], [0.1]])) expected_loss = (-np.log(softmax([0.8, 0.1, 0.1])[0]) - np.log( softmax([0.8, 0.1, 0.1])[1])) / 2 loss = K.eval(losses.RankCrossEntropyLoss(num_neg=2)( true_value, pred_value)) assert np.isclose(expected_loss, loss)
2,082
39.843137
73
py
MatchZoo
MatchZoo-master/tests/unit_test/test_data_generator.py
import copy import pytest import keras import matchzoo as mz @pytest.fixture(scope='module') def data_gen(): return mz.DataGenerator(mz.datasets.toy.load_data()) @pytest.mark.parametrize('attr', [ 'callbacks', 'num_neg', 'num_dup', 'mode', 'batch_size', 'shuffle', ]) def test_data_generator_getters_setters(data_gen, attr): assert hasattr(data_gen, attr) val = getattr(data_gen, attr) setattr(data_gen, attr, val) assert getattr(data_gen, attr) == val def test_resample(): model = mz.models.Naive() prpr = model.get_default_preprocessor() data_raw = mz.datasets.toy.load_data() data = prpr.fit_transform(data_raw) model.params.update(prpr.context) model.params['task'] = mz.tasks.Ranking() model.build() model.compile() data_gen = mz.DataGenerator( data_pack=data, mode='pair', resample=True, batch_size=4 ) class CheckResample(keras.callbacks.Callback): def __init__(self, data_gen): super().__init__() self._data_gen = data_gen self._orig_indices = None self._flags = [] def on_epoch_end(self, epoch, logs=None): curr_indices = self._data_gen.batch_indices if not self._orig_indices: self._orig_indices = copy.deepcopy(curr_indices) else: self._flags.append(self._orig_indices != curr_indices) self._orig_indices = curr_indices check_resample = CheckResample(data_gen) model.fit_generator(data_gen, epochs=5, callbacks=[check_resample]) assert check_resample._flags assert all(check_resample._flags)
1,700
24.772727
71
py
MatchZoo
MatchZoo-master/tests/unit_test/models/test_models.py
""" These tests are simplied because the original verion takes too much time to run, making CI fails as it reaches the time limit. """ import pytest import copy from pathlib import Path import shutil import matchzoo as mz from keras.backend import clear_session @pytest.fixture(scope='module', params=[ mz.tasks.Ranking(loss=mz.losses.RankCrossEntropyLoss(num_neg=2)), mz.tasks.Classification(num_classes=2), ]) def task(request): return request.param @pytest.fixture(scope='module') def train_raw(task): return mz.datasets.toy.load_data('train', task)[:5] @pytest.fixture(scope='module', params=mz.models.list_available()) def model_class(request): return request.param @pytest.fixture(scope='module') def embedding(): return mz.datasets.toy.load_embedding() @pytest.fixture(scope='module') def setup(task, model_class, train_raw, embedding): clear_session() # prevent OOM during CI tests return mz.auto.prepare( task=task, model_class=model_class, data_pack=train_raw, embedding=embedding ) @pytest.fixture(scope='module') def model(setup): return setup[0] @pytest.fixture(scope='module') def preprocessor(setup): return setup[1] @pytest.fixture(scope='module') def gen_builder(setup): return setup[2] @pytest.fixture(scope='module') def embedding_matrix(setup): return setup[3] @pytest.fixture(scope='module') def data(train_raw, preprocessor, gen_builder): return gen_builder.build(preprocessor.transform(train_raw))[0] @pytest.mark.slow def test_model_fit_eval_predict(model, data): x, y = data batch_size = len(x['id_left']) assert model.fit(x, y, batch_size=batch_size, verbose=0) assert model.evaluate(x, y, batch_size=batch_size) assert model.predict(x, batch_size=batch_size) is not None @pytest.mark.cron def test_save_load_model(model): tmpdir = '.matchzoo_test_save_load_tmpdir' if Path(tmpdir).exists(): shutil.rmtree(tmpdir) try: model.save(tmpdir) assert mz.load_model(tmpdir) with pytest.raises(FileExistsError): model.save(tmpdir) finally: if Path(tmpdir).exists(): shutil.rmtree(tmpdir) @pytest.mark.cron def test_hyper_space(model): for _ in range(2): new_params = copy.deepcopy(model.params) sample = mz.hyper_spaces.sample(new_params.hyper_space) for key, value in sample.items(): new_params[key] = value new_model = new_params['model_class'](params=new_params) new_model.build() new_model.compile()
2,603
22.889908
75
py
MatchZoo
MatchZoo-master/docs/source/conf.py
# -*- coding: utf-8 -*- # # Configuration file for the Sphinx documentation builder. # # This file does only contain 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('../..')) sys.path.insert(0, os.path.abspath('../../matchzoo')) sys.path.insert(0, os.path.abspath('../../matchzoo/auto')) sys.path.insert(0, os.path.abspath('../../matchzoo/data_generator')) sys.path.insert(0, os.path.abspath('../../matchzoo/data_pack')) sys.path.insert(0, os.path.abspath('../../matchzoo/datasets')) sys.path.insert(0, os.path.abspath('../../matchzoo/embedding')) sys.path.insert(0, os.path.abspath('../../matchzoo/engine')) sys.path.insert(0, os.path.abspath('../../matchzoo/layers')) sys.path.insert(0, os.path.abspath('../../matchzoo/losses')) sys.path.insert(0, os.path.abspath('../../matchzoo/metrics')) sys.path.insert(0, os.path.abspath('../../matchzoo/models')) sys.path.insert(0, os.path.abspath('../../matchzoo/preprocessors')) sys.path.insert(0, os.path.abspath('../../matchzoo/tasks')) sys.path.insert(0, os.path.abspath('../../matchzoo/utils')) # -- Project information ----------------------------------------------------- project = 'MatchZoo' copyright = '2018, MatchZoo' author = 'MatchZoo' # The short X.Y version version = '' # The full version, including alpha/beta/rc tags release = '2.1' # -- General configuration --------------------------------------------------- # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # 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.autodoc', 'sphinx.ext.mathjax', 'sphinx.ext.napoleon', 'sphinx_autodoc_typehints', 'nbsphinx', ] # Add any paths that contain templates here, relative to this directory. templates_path = ['_templates'] # The suffix(es) of source filenames. # You can specify multiple suffix as a list of string: # # source_suffix = ['.rst', '.md'] source_suffix = '.rst' # from recommonmark.parser import CommonMarkParser # source_parsers = { # '.md':CommonMarkParser # } # The master toctree document. master_doc = 'index' # 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 = ['_build', 'Thumbs.db', '.DS_Store','**.ipynb_checkpoints'] # The name of the Pygments (syntax highlighting) style to use. pygments_style = 'sphinx' # -- 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' # 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 = {} # 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'] # Custom sidebar templates, must be a dictionary that maps document names # to template names. # # The default sidebars (for documents that don't match any pattern) are # defined by theme itself. Builtin themes are using these templates by # default: ``['localtoc.html', 'relations.html', 'sourcelink.html', # 'searchbox.html']``. # # html_sidebars = {} # -- Options for HTMLHelp output --------------------------------------------- # Output file base name for HTML help builder. htmlhelp_basename = 'MatchZoodoc' # -- Options for LaTeX output ------------------------------------------------ latex_elements = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # # 'pointsize': '10pt', # Additional stuff for the LaTeX preamble. # # 'preamble': '', # Latex figure (float) alignment # # 'figure_align': 'htbp', } # Grouping the document tree into LaTeX files. List of tuples # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ (master_doc, 'MatchZoo.tex', 'MatchZoo Documentation', 'MatchZoo', 'manual'), ] # -- Options for manual page output ------------------------------------------ # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ (master_doc, 'matchzoo', 'MatchZoo Documentation', [author], 1) ] # -- Options for Texinfo output ---------------------------------------------- # Grouping the document tree into Texinfo files. List of tuples # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ (master_doc, 'MatchZoo', 'MatchZoo Documentation', author, 'MatchZoo', 'One line description of project.', 'Miscellaneous'), ] # -- Extension configuration -------------------------------------------------
5,927
31.751381
79
py
Analyzing-the-Generalization-Capability-of-SGLD-Using-Properties-of-Gaussian-Channels
Analyzing-the-Generalization-Capability-of-SGLD-Using-Properties-of-Gaussian-Channels-main/code/main.py
import numpy as np import torch import math from torch import nn, optim from torch.utils.data import DataLoader from torch.utils.data import SubsetRandomSampler import importlib import copy import argparse from torchvision import transforms, datasets from torch.autograd import Variable from torch.optim import Optimizer from torch.optim.sgd import SGD from torch.nn.utils import clip_grad_norm_ class CIFAR10RandomLabels(datasets.CIFAR10): def __init__(self, corrupt_prob=0.0, num_classes=10, **kwargs): super(CIFAR10RandomLabels, self).__init__(**kwargs) self.n_classes = num_classes if corrupt_prob > 0: self.corrupt_labels(corrupt_prob) def corrupt_labels(self, corrupt_prob): labels = np.array(self.targets) np.random.seed(12345) mask = np.random.rand(len(labels)) <= corrupt_prob rnd_labels = np.random.choice(self.n_classes, mask.sum()) labels[mask] = rnd_labels labels = [int(x) for x in labels] self.targets = labels class MNISTRandomLabels(datasets.MNIST): def __init__(self, corrupt_prob=0.0, num_classes=10, **kwargs): super(MNISTRandomLabels, self).__init__(**kwargs) self.n_classes = num_classes if corrupt_prob > 0: self.corrupt_labels(corrupt_prob) def corrupt_labels(self, corrupt_prob): labels = np.array(self.targets) np.random.seed(12345) mask = np.random.rand(len(labels)) <= corrupt_prob rnd_labels = np.random.choice(self.n_classes, mask.sum()) labels[mask] = rnd_labels labels = [int(x) for x in labels] self.targets = labels class Langevin_SGD(Optimizer): def __init__(self, params, lr, weight_decay=0, nesterov=False, beta=1, K=100, D=50, sigma=0.5, decay_rate = 0.96, decay_steps=2000): if lr < 0.0: raise ValueError("Invalid learning rate: {}".format(lr)) if weight_decay < 0.0: raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) defaults = dict(lr=lr, weight_decay=weight_decay) self.beta = beta self.K = K self.D = D self.lr = lr self.sigma = sigma self.decay_rate = decay_rate self.decay_steps = decay_steps self.steps = 0 super(Langevin_SGD, self).__init__(params, defaults) def __setstate__(self, state): super(SGD, self).__setstate__(state) for group in self.param_groups: group.setdefault('nesterov', False) def step(self, closure=None): self.beta = 4 * self.lr / ((0.002*math.sqrt(2) * self.lr)**2) gradient = [] for group in self.param_groups: weight_decay = group['weight_decay'] clip_grad_norm_(group['params'], self.K, norm_type=2) for p in group['params']: if p.grad is None: continue d_p = p.grad.data if len(p.shape) == 1 and p.shape[0] == 1: p.data.add_(-self.lr, d_p) else: if weight_decay != 0: d_p.add_(weight_decay, p.data) unit_noise = Variable(p.data.new(p.size()).normal_()) p.data.add_(-self.lr, d_p) p.data.add_((2*self.lr/self.beta)**0.5, unit_noise) if torch.norm(p.data).item() >= self.D/2: p.data = p.data / torch.norm(p.data) * (self.D/2) gradient = gradient + (d_p.cpu().numpy().flatten().tolist()) if (self.steps > 0 and self.steps % self.decay_steps==0): self.lr = self.lr * self.decay_rate self.steps = self.steps + 1 if self.lr < 0.0005: self.lr = 0.0005 return (np.array(gradient)).flatten() def train(args, model, device, train_loader, criterion, optimizer, epoch, batchsize, num_batches): sum_loss, sum_correct = 0, 0 model.train() gradient_array = np.zeros((num_batches, count_parameters(model))) for i, (data, target) in enumerate(train_loader): data, target = data.to(device), target.to(device) output = model(data) loss = criterion(output, target) pred = output.max(1)[1] sum_correct += pred.eq(target).sum().item() sum_loss += len(data) * loss.item() optimizer.zero_grad() loss.backward() gradient = optimizer.step() gradient_array[i] = gradient return 1 - (sum_correct / len(train_loader.dataset)), sum_loss / len(train_loader.dataset), np.array(gradient_array) def validate(args, model, device, val_loader, criterion, optimizer, length=0): sum_loss, sum_correct = 0, 0 model.eval() total_grad = [] count = 0 for i, (data, target) in enumerate(val_loader): count = count + 1 data, target = data.to(device), target.to(device) output = model(data) loss = criterion(output, target) pred = output.max(1)[1] sum_correct += pred.eq(target).sum().item() sum_loss += len(data) * criterion(output, target).item() optimizer.zero_grad() loss.backward() gradient = [] params = list(model.parameters()) for p in params: if p.grad is None: continue d_p = p.grad.data gradient = gradient + (d_p.cpu().numpy().flatten().tolist()) gradient = (np.array(gradient)).flatten() if (total_grad == []): total_grad = gradient else: total_grad = total_grad + gradient if (length == 0): length = len(val_loader.dataset) return 1 - (sum_correct / length), sum_loss / length, total_grad / count def count_parameters(model): return sum(p.numel() for p in model.parameters() if p.requires_grad) def main(): parser = argparse.ArgumentParser(description='Training a fully connected NN with one hidden layer') parser.add_argument('--no-cuda', default=False, action='store_true', help='disables CUDA training') parser.add_argument('--datadir', default='datasets', type=str, help='path to the directory that contains the datasets (default: datasets)') parser.add_argument('--dataset', default='CIFAR10', type=str, help='name of the dataset (options: MNIST | CIFAR10 | CIFAR100 | SVHN, default: CIFAR10)') parser.add_argument('--model', default='vgg', type=str, help='architecture (options: fc | vgg, default: vgg)') parser.add_argument('--epochs', default=1000, type=int, help='number of epochs to train (default: 1000)') parser.add_argument('--stopcond', default=0.01, type=float, help='stopping condtion based on the cross-entropy loss (default: 0.01)') parser.add_argument('--batchsize', default=64, type=int, help='input batch size (default: 64)') parser.add_argument('--learningrate', default=0.05, type=float, help='learning rate (default: 0.01)') parser.add_argument('--momentum', default=0.9, type=float, help='momentum (default: 0.9)') parser.add_argument('--label_corrupt_prob', default=0, type=float, help='label_corrupt_prob (default: 0)') parser.add_argument('--num_sample_path', default=1, type=float, help='num_sample_path (default: 0)') args = parser.parse_args() use_cuda = torch.cuda.is_available() device = torch.device("cuda" if use_cuda else "cpu") print(device) kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {} nchannels, nclasses, img_dim, = 3, 10, 32 if args.dataset == 'MNIST': nchannels = 1 if args.dataset == 'CIFAR100': nclasses = 100 num_sample_path = int(args.num_sample_path) size_of_training_set = 5000 num_batches = size_of_training_set // args.batchsize tr_err_list = np.empty((args.epochs, num_sample_path)) tr_loss_list = np.empty((args.epochs, num_sample_path)) val_err_list = np.empty((args.epochs, num_sample_path)) val_loss_list = np.empty((args.epochs, num_sample_path)) variance_list = np.empty((num_batches, args.epochs)) optimizer = None subset_indices = np.random.choice(50000,size_of_training_set, replace=False) if args.dataset == 'MNIST': normalize = transforms.Normalize(mean=[0.131], std=[0.289]) else: normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) transform_train = transforms.Compose([transforms.Resize(32), transforms.ToTensor(), normalize]) transform_test = transforms.Compose([transforms.Resize(32), transforms.ToTensor(), normalize]) if args.dataset == 'MNIST': train_dataset = MNISTRandomLabels(root='./data', train=True, download=True, transform=transform_train, num_classes=10, corrupt_prob=args.label_corrupt_prob) train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=args.batchsize, sampler=subset_indices, shuffle=False, **kwargs) val_loader = torch.utils.data.DataLoader( MNISTRandomLabels(root='./data', train=False, transform=transform_test, num_classes=10, corrupt_prob=args.label_corrupt_prob), batch_size=args.batchsize, shuffle=False, **kwargs) else: train_dataset = CIFAR10RandomLabels(root='./data', train=True, download=True, transform=transform_train, num_classes=10, corrupt_prob=args.label_corrupt_prob) train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=args.batchsize, sampler=subset_indices, shuffle=False, **kwargs) val_loader = torch.utils.data.DataLoader( CIFAR10RandomLabels(root='./data', train=False, transform=transform_test, num_classes=10, corrupt_prob=args.label_corrupt_prob), batch_size=args.batchsize, shuffle=False, **kwargs) for i in range(num_sample_path): model = getattr(importlib.import_module('models.{}'.format(args.model)), 'Network')(nchannels, nclasses) model = model.to(device) criterion = nn.CrossEntropyLoss().to(device) optimizer = Langevin_SGD(model.parameters(), args.learningrate) for epoch in range(0, args.epochs): tr_err, tr_loss, gradient = train(args, model, device, train_loader, criterion, optimizer, epoch, args.batchsize, num_batches) tr_err, tr_loss, _ = validate(args, model, device, train_loader, criterion, optimizer, length=size_of_training_set) val_err, val_loss, _ = validate(args, model, device, val_loader, criterion, optimizer) tr_err_list[epoch, i] = tr_err tr_loss_list[epoch, i] = tr_loss val_err_list[epoch, i] = val_err val_loss_list[epoch, i] = val_loss for t in range(gradient.shape[0]): filename = "gradient/" + str(t) + "_" + str(epoch) + "_" + str(i) np.save(filename, gradient[t]) print("epoch " + str(epoch+1) + " completed") print("Sample path " + str(i+1) + " completed") average_tr_err = np.mean(tr_err_list, axis=1) average_tr_loss = np.mean(tr_loss_list, axis=1) average_val_err = np.mean(val_err_list, axis=1) average_val_loss = np.mean(val_loss_list, axis=1) for i in range(num_batches): for j in range(args.epochs): sum_var = 0 bar = 0 for k in range(num_sample_path): filename = "gradient/" + str(i) + "_" + str(j) + "_" + str(k) + ".npy" bar = bar + np.load(filename) bar = bar / num_sample_path for k in range(num_sample_path): filename = "gradient/" + str(i) + "_" + str(j) + "_" + str(k) + ".npy" foo = np.load(filename) sum_var = sum_var + np.linalg.norm(foo-bar)**2 sum_var = sum_var/num_sample_path variance_list[i,j] = sum_var print('Print the average result from multiple sample path:') left_hand = [] right_hand = [] surrogate_loss = [] for epoch in range(0, args.epochs): tr_err = average_tr_err[epoch] tr_loss = average_tr_loss[epoch] val_err = average_val_err[epoch] val_loss = average_val_loss[epoch] bound_var = 0 for m in range(num_batches): sum_var = 0 for t in range(m, num_batches * (epoch+1), num_batches): eta = args.learningrate * (optimizer.decay_rate ** (t // optimizer.decay_steps)) beta = 4 * eta / ((0.002*math.sqrt(2) * eta)**2) product_var = eta * beta * variance_list[m, t // num_batches] for t_q in range(t+1, (epoch+1) * num_batches): if not (t_q % num_batches == m): product_var = product_var sum_var = sum_var + product_var bound_var = bound_var + math.sqrt(sum_var) bound_var = bound_var * math.sqrt(2 * args.batchsize) * 0.5 / (2.0 * size_of_training_set) left_hand.append(abs(average_tr_err[epoch] - average_val_err[epoch])) surrogate_loss.append(abs(average_tr_loss[epoch] - average_val_loss[epoch])) right_hand.append(bound_var) print(f'Epoch: {epoch + 1}/{args.epochs}\t Average Training loss: {tr_loss:.8f}', f'Average Training error: {tr_err:.8f}\t Average Validation error: {val_err:.8f}', f'Average Validation loss: {val_loss:.8f}\t Average Bound: {bound_var:.8f}\t') acc_hand = average_tr_err np.save("left_hand", left_hand) np.save("right_hand", right_hand) np.save("acc", acc_hand) if __name__ == '__main__': main()
14,060
39.059829
251
py
Analyzing-the-Generalization-Capability-of-SGLD-Using-Properties-of-Gaussian-Channels
Analyzing-the-Generalization-Capability-of-SGLD-Using-Properties-of-Gaussian-Channels-main/code/models/fc.py
import torch.nn as nn class Network(nn.Module): def __init__(self, nchannels, nclasses): super(Network, self).__init__() self.classifier = nn.Sequential(nn.Linear( nchannels * 32 * 32, 32, bias=True), nn.ReLU(inplace=True), nn.Linear( 32, 32, bias=True), nn.ReLU(inplace=True), nn.Linear( 32, nclasses, bias=True)) def forward(self, x): x = x.view(x.size(0), -1) x = self.classifier(x) return x
525
36.571429
110
py