""" Vision Transformer (ViT) in PyTorch A PyTorch implement of Vision Transformers as described in 'An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale' - https://arxiv.org/abs/2010.11929 The official jax code is released and available at https://github.com/google-research/vision_transformer Status/TODO: * Models updated to be compatible with official impl. Args added to support backward compat for old PyTorch weights. * Weights ported from official jax impl for 384x384 base and small models, 16x16 and 32x32 patches. * Trained (supervised on ImageNet-1k) my custom 'small' patch model to 77.9, 'base' to 79.4 top-1 with this code. * Hopefully find time and GPUs for SSL or unsupervised pretraining on OpenImages w/ ImageNet fine-tune in future. Acknowledgments: * The paper authors for releasing code and weights, thanks! * I fixed my class token impl based on Phil Wang's https://github.com/lucidrains/vit-pytorch ... check it out for some einops/einsum fun * Simple transformer style inspired by Andrej Karpathy's https://github.com/karpathy/minGPT * Bert reference code checks against Huggingface Transformers and Tensorflow Bert Hacked together by / Copyright 2020 Ross Wightman """ import torch import torch.nn as nn import torch.utils.model_zoo as model_zoo import torch.utils.checkpoint as checkpoint import torch.nn.functional as F import math from functools import partial from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from timm.models.layers import DropPath, to_2tuple, trunc_normal_ from timm.models.registry import register_model def _cfg(url='', **kwargs): return { 'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None, 'crop_pct': .9, 'interpolation': 'bicubic', 'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, 'first_conv': 'patch_embed.proj', 'classifier': 'head', **kwargs } default_cfgs = { # patch models # mae ViT-B/16-224 pre-trained model 'vit_base_patch16_224_mae': _cfg( url='https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_base.pth', input_size=(3, 224, 224), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_base_patch16_224_default': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_384-83fb41ba.pth', input_size=(3, 224, 224), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), # mae ViT-L/16-224 pre-trained model 'vit_large_patch16_224_mae': _cfg( url='https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_large.pth', input_size=(3, 224, 224), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), # mae ViT-H/14-224 pre-trained model 'vit_huge_patch14_224_mae': _cfg( url='https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_huge.pth', input_size=(3, 224, 224), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_small_patch16_224': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_small_p16_224-15ec54c9.pth', ), 'vit_base_patch16_224': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), ), 'vit_base_patch16_384': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_384-83fb41ba.pth', input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_base_patch32_384': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p32_384-830016f5.pth', input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_large_patch16_224': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_224-4ee7a4dc.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)), 'vit_large_patch16_384': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_384-b3be5167.pth', input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_large_patch32_384': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p32_384-9b920ba8.pth', input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_huge_patch16_224': _cfg(), 'vit_huge_patch32_384': _cfg(input_size=(3, 384, 384)), # hybrid models 'vit_small_resnet26d_224': _cfg(), 'vit_small_resnet50d_s3_224': _cfg(), 'vit_base_resnet26d_224': _cfg(), 'vit_base_resnet50d_224': _cfg(), } class Mlp(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.fc2 = nn.Linear(hidden_features, out_features) self.drop = nn.Dropout(drop) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) x = self.drop(x) return x class Attention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Block(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm): super().__init__() self.norm1 = norm_layer(dim) self.attn = Attention( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) def forward(self, x): x = x + self.drop_path(self.attn(self.norm1(x))) x = x + self.drop_path(self.mlp(self.norm2(x))) return x class PatchEmbed(nn.Module): """ Image to Patch Embedding """ def __init__(self, patch_size=16, in_chans=3, embed_dim=768): super().__init__() self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) def forward(self, x): # B, C, H, W = x.shape x = self.proj(x).flatten(2).transpose(1, 2) return x class HybridEmbed(nn.Module): """ CNN Feature Map Embedding Extract feature map from CNN, flatten, project to embedding dim. """ def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768): super().__init__() assert isinstance(backbone, nn.Module) img_size = to_2tuple(img_size) self.img_size = img_size self.backbone = backbone if feature_size is None: with torch.no_grad(): # FIXME this is hacky, but most reliable way of determining the exact dim of the output feature # map for all networks, the feature metadata has reliable channel and stride info, but using # stride to calc feature dim requires info about padding of each stage that isn't captured. training = backbone.training if training: backbone.eval() o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1] feature_size = o.shape[-2:] feature_dim = o.shape[1] backbone.train(training) else: feature_size = to_2tuple(feature_size) feature_dim = self.backbone.feature_info.channels()[-1] self.num_patches = feature_size[0] * feature_size[1] self.proj = nn.Linear(feature_dim, embed_dim) def forward(self, x): x = self.backbone(x)[-1] x = x.flatten(2).transpose(1, 2) x = self.proj(x) return x class VisionTransformer(nn.Module): """ Vision Transformer with support for patch or hybrid CNN input stage """ def __init__(self, search_size=384, template_size=192, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=nn.LayerNorm, search_number=1, template_number=1, use_checkpoint=False): super().__init__() self.use_checkpoint = use_checkpoint self.num_classes = num_classes self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models self.embed_dim_list = [embed_dim] self.num_search = search_number self.num_template = template_number self.patch_embed = PatchEmbed( patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) self.num_patches_search = (search_size // patch_size) * (search_size // patch_size) self.num_patches_template = (template_size // patch_size) * (template_size // patch_size) self.pos_embed = nn.Parameter(torch.zeros(1, self.num_patches_search + self.num_patches_template, embed_dim)) # self.pos_embed_search = nn.Parameter(torch.zeros(1, self.num_patches_search, embed_dim)) # self.pos_embed_template = nn.Parameter(torch.zeros(1, self.num_patches_template, embed_dim)) self.pos_drop = nn.Dropout(p=drop_rate) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule self.blocks = nn.ModuleList([ Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer) for i in range(depth)]) self.norm = norm_layer(embed_dim) trunc_normal_(self.pos_embed, std=.02) # trunc_normal_(self.pos_embed_search, std=.02) # trunc_normal_(self.pos_embed_template, std=.02) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) @torch.jit.ignore def no_weight_decay(self): return {'pos_embed', 'cls_token'} def get_classifier(self): return self.head def reset_classifier(self, num_classes, global_pool=''): self.num_classes = num_classes self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity() def forward_features(self, template_list, search_list): num_template = len(template_list) num_search = len(search_list) z = torch.stack(template_list, dim=1)#(b,n,c,h,w) z = z.view(-1, *z.size()[2:])#(bn,c,h,w) z = self.patch_embed(z) z = z + self.pos_embed[:, self.num_patches_search:, :] z = z.view(-1,num_template,z.size(-2),z.size(-1)) z = z.reshape(z.size(0),-1,z.size(-1)) x = torch.stack(search_list, dim=1)#(b,n,c,h,w) x = x.view(-1, *x.size()[2:])#(bn,c,h,w) x = self.patch_embed(x) x = x + self.pos_embed[:, :self.num_patches_search, :] x = x.view(-1,num_search,x.size(-2),x.size(-1)) x = x.reshape(x.size(0),-1,x.size(-1)) xz = torch.cat([x, z], dim=1) xz = self.pos_drop(xz) for blk in self.blocks: #batch is the first dimension. if self.use_checkpoint: xz = checkpoint.checkpoint(blk, xz) else: xz = blk(xz) xz = self.norm(xz) # B,N,C return xz def forward(self, template_list, search_list): xz = self.forward_features(template_list, search_list) out=[xz] return out def _conv_filter(state_dict, patch_size=16): """ convert patch embedding weight from manual patchify + linear proj to conv""" out_dict = {} for k, v in state_dict.items(): if 'patch_embed.proj.weight' in k: v = v.reshape((v.shape[0], 3, patch_size, patch_size)) out_dict[k] = v return out_dict @register_model def vit_base_patch16(pretrained=False, pretrain_type='default', search_size=384, template_size=192, **kwargs): patch_size = 16 model = VisionTransformer( search_size=search_size, template_size=template_size, patch_size=patch_size, num_classes=0, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) cfg_type = 'vit_base_patch16_224_' + pretrain_type if pretrain_type == 'scratch': pretrained = False return model model.default_cfg = default_cfgs[cfg_type] if pretrained: load_pretrained(model, pretrain_type, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3)) return model @register_model def vit_large_patch16(pretrained=False, pretrain_type='default', search_size=384, template_size=192, **kwargs): patch_size = 16 model = VisionTransformer( search_size=search_size, template_size=template_size, patch_size=patch_size, num_classes=0, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) cfg_type = 'vit_large_patch16_224_' + pretrain_type if pretrain_type == 'scratch': pretrained = False return model model.default_cfg = default_cfgs[cfg_type] if pretrained: load_pretrained(model, pretrain_type, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3)) return model @register_model def vit_huge_patch14(pretrained=False, pretrain_type='default', search_size=364, template_size=182, **kwargs): patch_size = 14 model = VisionTransformer( search_size=search_size, template_size=template_size, patch_size=patch_size, num_classes=0, embed_dim=1280, depth=32, num_heads=16, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) cfg_type = 'vit_huge_patch14_224_' + pretrain_type if pretrain_type == 'scratch': pretrained = False return model model.default_cfg = default_cfgs[cfg_type] if pretrained: load_pretrained(model, pretrain_type, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3)) return model def load_pretrained(model, pretrain_type='default', cfg=None, num_classes=1000, in_chans=3, filter_fn=None, strict=True): if cfg is None: cfg = getattr(model, 'default_cfg') if cfg is None or 'url' not in cfg or not cfg['url']: print("Pretrained model URL is invalid, using random initialization.") return state_dict = model_zoo.load_url(cfg['url'], progress=False, map_location='cpu') if pretrain_type == 'mae': state_dict = state_dict['model'] if filter_fn is not None: state_dict = filter_fn(state_dict) if in_chans == 1: conv1_name = cfg['first_conv'] print('Converting first conv (%s) pretrained weights from 3 to 1 channel' % conv1_name) conv1_weight = state_dict[conv1_name + '.weight'] # Some weights are in torch.half, ensure it's float for sum on CPU conv1_type = conv1_weight.dtype conv1_weight = conv1_weight.float() O, I, J, K = conv1_weight.shape if I > 3: assert conv1_weight.shape[1] % 3 == 0 # For models with space2depth stems conv1_weight = conv1_weight.reshape(O, I // 3, 3, J, K) conv1_weight = conv1_weight.sum(dim=2, keepdim=False) else: conv1_weight = conv1_weight.sum(dim=1, keepdim=True) conv1_weight = conv1_weight.to(conv1_type) state_dict[conv1_name + '.weight'] = conv1_weight elif in_chans != 3: conv1_name = cfg['first_conv'] conv1_weight = state_dict[conv1_name + '.weight'] conv1_type = conv1_weight.dtype conv1_weight = conv1_weight.float() O, I, J, K = conv1_weight.shape if I != 3: print('Deleting first conv (%s) from pretrained weights.' % conv1_name) del state_dict[conv1_name + '.weight'] strict = False else: # NOTE this strategy should be better than random init, but there could be other combinations of # the original RGB input layer weights that'd work better for specific cases. print('Repeating first conv (%s) weights in channel dim.' % conv1_name) repeat = int(math.ceil(in_chans / 3)) conv1_weight = conv1_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :] conv1_weight *= (3 / float(in_chans)) conv1_weight = conv1_weight.to(conv1_type) state_dict[conv1_name + '.weight'] = conv1_weight classifier_name = cfg['classifier'] if pretrain_type == "mae": pass elif num_classes == 1000 and cfg['num_classes'] == 1001: # special case for imagenet trained models with extra background class in pretrained weights classifier_weight = state_dict[classifier_name + '.weight'] state_dict[classifier_name + '.weight'] = classifier_weight[1:] classifier_bias = state_dict[classifier_name + '.bias'] state_dict[classifier_name + '.bias'] = classifier_bias[1:] elif num_classes != cfg['num_classes']: # completely discard fully connected for all other differences between pretrained and created model del state_dict[classifier_name + '.weight'] del state_dict[classifier_name + '.bias'] # adjust position encoding pe = state_dict['pos_embed'][:,1:,:] b_pe, hw_pe, c_pe = pe.shape side_pe = int(math.sqrt(hw_pe)) side_num_patches_search = int(math.sqrt(model.num_patches_search)) side_num_patches_template = int(math.sqrt(model.num_patches_template)) pe_2D = pe.reshape([b_pe, side_pe, side_pe, c_pe]).permute([0,3,1,2]) #b,c,h,w if side_pe != side_num_patches_search: pe_s_2D = nn.functional.interpolate(pe_2D, [side_num_patches_search, side_num_patches_search], align_corners=True, mode='bicubic') pe_s = torch.flatten(pe_s_2D.permute([0,2,3,1]),1,2) else: pe_s = pe if side_pe != side_num_patches_template: pe_t_2D = nn.functional.interpolate(pe_2D, [side_num_patches_template, side_num_patches_template], align_corners=True, mode='bicubic') pe_t = torch.flatten(pe_t_2D.permute([0, 2, 3, 1]), 1, 2) else: pe_t = pe pe_xz = torch.cat((pe_s, pe_t), dim=1) state_dict['pos_embed'] = pe_xz # state_dict['pos_embed_template'] = pe_t # state_dict['pos_embed_search'] = pe_s # del state_dict['pos_embed'] del state_dict['cls_token'] model.load_state_dict(state_dict, strict=strict)