# 2022.06.08-Changed for implementation of TokenFusion # Huawei Technologies Co., Ltd. # --------------------------------------------------------------- # Copyright (c) 2021, NVIDIA Corporation. All rights reserved. # # This work is licensed under the NVIDIA Source Code License # --------------------------------------------------------------- import math import torch import torch.nn as nn import torch.nn.functional as F from functools import partial from timm.models.layers import DropPath, to_2tuple, trunc_normal_ from .modules import ModuleParallel, LayerNormParallel, num_parallel, LinearFuse 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 = ModuleParallel(nn.Linear(in_features, hidden_features)) self.dwconv = DWConv(hidden_features) self.act = ModuleParallel(act_layer()) self.fc2 = ModuleParallel(nn.Linear(hidden_features, out_features)) self.drop = ModuleParallel(nn.Dropout(drop)) 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) elif isinstance(m, nn.Conv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels fan_out //= m.groups m.weight.data.normal_(0, math.sqrt(2.0 / fan_out)) if m.bias is not None: m.bias.data.zero_() def forward(self, x, H, W): x = self.fc1(x) x = [self.dwconv(x[0], H, W), self.dwconv(x[1], H, W)] 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, ratio, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1): super().__init__() assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}." self.dim = dim self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.q = ModuleParallel(nn.Linear(dim, dim, bias=qkv_bias)) self.kv = ModuleParallel(nn.Linear(dim, dim * 2, bias=qkv_bias)) self.attn_drop = ModuleParallel(nn.Dropout(attn_drop)) self.proj = ModuleParallel(nn.Linear(dim, dim)) self.proj_drop = ModuleParallel(nn.Dropout(proj_drop)) self.sr_ratio = sr_ratio if sr_ratio > 1: self.sr = ModuleParallel(nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)) self.norm = LayerNormParallel(dim) self.exchange = LinearFuse(ratio) 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) elif isinstance(m, nn.Conv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels fan_out //= m.groups m.weight.data.normal_(0, math.sqrt(2.0 / fan_out)) if m.bias is not None: m.bias.data.zero_() def forward(self, x, H, W): B, N, C = x[0].shape q = self.q(x) q = [q_.reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) for q_ in q] if self.sr_ratio > 1: x = [x_.permute(0, 2, 1).reshape(B, C, H, W) for x_ in x] x = self.sr(x) x = [x_.reshape(B, C, -1).permute(0, 2, 1) for x_ in x] x = self.norm(x) kv = self.kv(x) kv = [kv_.reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) for kv_ in kv] else: kv = self.kv(x) kv = [kv_.reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) for kv_ in kv] k, v = [kv[0][0], kv[1][0]], [kv[0][1], kv[1][1]] attn = [(q_ @ k_.transpose(-2, -1)) * self.scale for (q_, k_) in zip(q, k)] attn = [attn_.softmax(dim=-1) for attn_ in attn] attn = self.attn_drop(attn) x = [(attn_ @ v_).transpose(1, 2).reshape(B, N, C) for (attn_, v_) in zip(attn, v)] x = self.proj(x) x = self.proj_drop(x) # x = [x_ * mask_.unsqueeze(2) for (x_, mask_) in zip(x, mask)] x = self.exchange(x) return x # class PredictorLG(nn.Module): # """ Image to Patch Embedding from DydamicVit # """ # def __init__(self, embed_dim=384): # super().__init__() # self.score_nets = nn.ModuleList([nn.Sequential( # nn.LayerNorm(embed_dim), # nn.Linear(embed_dim, embed_dim), # nn.GELU(), # nn.Linear(embed_dim, embed_dim // 2), # nn.GELU(), # nn.Linear(embed_dim // 2, embed_dim // 4), # nn.GELU(), # nn.Linear(embed_dim // 4, 2), # nn.LogSoftmax(dim=-1) # ) for _ in range(num_parallel)]) # def forward(self, x): # x = [self.score_nets[i](x[i]) for i in range(num_parallel)] # return x class Block(nn.Module): def __init__(self, dim, ratio, 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=LayerNormParallel, sr_ratio=1): super().__init__() self.norm1 = norm_layer(dim) # self.score = PredictorLG(dim) self.attn = Attention( dim, ratio, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio) # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here self.drop_path = ModuleParallel(DropPath(drop_path)) if drop_path > 0. else ModuleParallel(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) # self.exchange = TokenExchange() 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) elif isinstance(m, nn.Conv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels fan_out //= m.groups m.weight.data.normal_(0, math.sqrt(2.0 / fan_out)) if m.bias is not None: m.bias.data.zero_() def forward(self, x, H, W, mask=None): B = x[0].shape[0] # norm1 = self.norm1(x) # score = self.score(norm1) # mask = [F.gumbel_softmax(score_.reshape(B, -1, 2), hard=True)[:, :, 0] for score_ in score] # if mask is not None: # norm = [norm_ * mask_.unsqueeze(2) for (norm_, mask_) in zip(norm, mask)] f = self.drop_path(self.attn(self.norm1(x), H, W)) x = [x_ + f_ for (x_, f_) in zip (x, f)] f = self.drop_path(self.mlp(self.norm2(x), H, W)) x = [x_ + f_ for (x_, f_) in zip (x, f)] # if mask is not None: # x = self.exchange(x, mask, mask_threshold=0.02) return x class OverlapPatchEmbedAndMask(nn.Module): """ Image to Patch Embedding """ def __init__(self, masking_ratio = 0.25, img_size=224, patch_size=7, stride=4, in_chans=3, embed_dim=768): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) self.img_size = img_size self.patch_size = patch_size self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1] self.num_patches = self.H * self.W self.proj = ModuleParallel(nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride, padding=(patch_size[0] // 2, patch_size[1] // 2))) self.norm = LayerNormParallel(embed_dim) self.masking_ratio = masking_ratio self.embed_dim = embed_dim self.mask_token = nn.parameter.Parameter(torch.randn(self.embed_dim), requires_grad = True)#None #When training in the SupOnly loop, unused params raise error in DDP. Hence instantiating mask_token only when masked training begins 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) elif isinstance(m, nn.Conv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels fan_out //= m.groups m.weight.data.normal_(0, math.sqrt(2.0 / fan_out)) if m.bias is not None: m.bias.data.zero_() def mask_with_mean(self, x): # print(x.shape) avg = torch.mean(x, dim = 1) avg = avg.clone().detach().requires_grad_(False) N, L, D = x.shape # batch, length, dim len_keep = int(L * (1 - self.masking_ratio)) noise = torch.rand(N, L, device=x.device) # noise in [0, 1] # sort noise for each sample ids_shuffle = torch.argsort(noise, dim=1) # ascend: small is keep, large is remove # keep the first subset ids_mask = ids_shuffle[:, len_keep:] # for i in range(N): # x[i][ids_mask[i]] = avg[i] # return x avg = avg.unsqueeze(dim = 1) avg = avg.repeat(1, L, 1) mask = ids_mask.unsqueeze(dim = 2) mask = mask.repeat(1, 1, D) masked = torch.scatter(x, dim = 1, index = mask, src = avg) # self[i] [index[i][j][k]] [k] = src[i][j][k] # if dim == 1 # self.printcheck(x[0], masked[0], avg[0]) return masked def mask_with_learnt_mask(self, x): # if self.mask_token is None: #When training in the SupOnly loop, unused params raise error in DDP. Hence instantiating mask_token only when masked training begins # self.mask_token = nn.parameter.Parameter(torch.randn(self.embed_dim, device=x.device), requires_grad = True) # print(self.mask_token[:10], x.device, "token") N, L, D = x.shape # batch, length, dim indicies = torch.FloatTensor(N, L).uniform_() <= self.masking_ratio x[indicies] = self.mask_token return x def printcheck(self, x, masked, avg): L, D = x.shape same = 0 avgsame = 0 for i in range(L): if (x[i] == masked[i]).all(): same += 1 # else: # print(i, x[i]) if (masked[i].data == avg[i].data).all(): avgsame += 1 print(same, avgsame) return def forward(self, x, masking_branch = -1, range_batches_to_mask = None): assert masking_branch < num_parallel and masking_branch >= -1 sum_mask = torch.sum(self.mask_token) x = self.proj(x) _, _, H, W = x[0].shape x = [x_.flatten(2).transpose(1, 2) for x_ in x] x = self.norm(x) if not masking_branch == -1: assert range_batches_to_mask is not None, "expected the range of batches to mask to not mask the labeled images" # x[masking_branch][range_batches_to_mask[0]:range_batches_to_mask[1]] = self.mask_with_mean(x[masking_branch][range_batches_to_mask[0]:range_batches_to_mask[1]]) x[masking_branch][range_batches_to_mask[0]:range_batches_to_mask[1]] = self.mask_with_learnt_mask(x[masking_branch][range_batches_to_mask[0]:range_batches_to_mask[1]]) # masking_branch = 1 # x[masking_branch] = self.mask_with_mean(x[masking_branch]) # x[masking_branch] = self.mask_with_learnt_mask(x[masking_branch]) else: x[0] = x[0] + 0*sum_mask #So that when training with SupOnly (and not using any masking), DDP doesn't raise an error that you have unused parameters. return x, H, W class MixVisionTransformer(nn.Module): def __init__(self, ratio, masking_ratio, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512], num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=LayerNormParallel, depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]): super().__init__() self.num_classes = num_classes self.depths = depths self.embed_dims = embed_dims # patch_embed self.patch_embed1 = OverlapPatchEmbedAndMask(masking_ratio = masking_ratio, img_size=img_size, patch_size=7, stride=4, in_chans=in_chans, embed_dim=embed_dims[0]) self.patch_embed2 = OverlapPatchEmbedAndMask(masking_ratio = masking_ratio, img_size=img_size // 4, patch_size=3, stride=2, in_chans=embed_dims[0], embed_dim=embed_dims[1]) self.patch_embed3 = OverlapPatchEmbedAndMask(masking_ratio = masking_ratio, img_size=img_size // 8, patch_size=3, stride=2, in_chans=embed_dims[1], embed_dim=embed_dims[2]) self.patch_embed4 = OverlapPatchEmbedAndMask(masking_ratio = masking_ratio, img_size=img_size // 16, patch_size=3, stride=2, in_chans=embed_dims[2], embed_dim=embed_dims[3]) # predictor_list = [PredictorLG(embed_dims[i]) for i in range(len(depths))] # self.score_predictor = nn.ModuleList(predictor_list) # transformer encoder dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule cur = 0 self.block1 = nn.ModuleList([Block( dim=embed_dims[0], ratio = ratio, num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, sr_ratio=sr_ratios[0]) for i in range(depths[0])]) self.norm1 = norm_layer(embed_dims[0]) cur += depths[0] self.block2 = nn.ModuleList([Block( dim=embed_dims[1], ratio = ratio, num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, sr_ratio=sr_ratios[1]) for i in range(depths[1])]) self.norm2 = norm_layer(embed_dims[1]) cur += depths[1] self.block3 = nn.ModuleList([Block( dim=embed_dims[2], ratio = ratio, num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, sr_ratio=sr_ratios[2]) for i in range(depths[2])]) self.norm3 = norm_layer(embed_dims[2]) cur += depths[2] self.block4 = nn.ModuleList([Block( dim=embed_dims[3], ratio = ratio, num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, sr_ratio=sr_ratios[3]) for i in range(depths[3])]) self.norm4 = norm_layer(embed_dims[3]) # classification head # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity() 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) elif isinstance(m, nn.Conv2d): fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels fan_out //= m.groups m.weight.data.normal_(0, math.sqrt(2.0 / fan_out)) if m.bias is not None: m.bias.data.zero_() ''' def init_weights(self, pretrained=None): if isinstance(pretrained, str): logger = get_root_logger() load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger) ''' def reset_drop_path(self, drop_path_rate): dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))] cur = 0 for i in range(self.depths[0]): self.block1[i].drop_path.drop_prob = dpr[cur + i] cur += self.depths[0] for i in range(self.depths[1]): self.block2[i].drop_path.drop_prob = dpr[cur + i] cur += self.depths[1] for i in range(self.depths[2]): self.block3[i].drop_path.drop_prob = dpr[cur + i] cur += self.depths[2] for i in range(self.depths[3]): self.block4[i].drop_path.drop_prob = dpr[cur + i] def freeze_patch_emb(self): self.patch_embed1.requires_grad = False @torch.jit.ignore def no_weight_decay(self): return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'} # has pos_embed may be better 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, x, masking_branch, range_batches_to_mask): B = x[0].shape[0] outs0, outs1 = [], [] # masks = [] # stage 1 x, H, W = self.patch_embed1(x, masking_branch = masking_branch, range_batches_to_mask = range_batches_to_mask) for i, blk in enumerate(self.block1): # score = self.score_predictor[0](x) # mask = [F.softmax(score_.reshape(B, -1, 2), dim=2)[:, :, 0] for score_ in score] # mask_: [B, N] # masks.append(mask) x = blk(x, H, W) x = self.norm1(x) x = [x_.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() for x_ in x] outs0.append(x[0]) outs1.append(x[1]) # stage 2 x, H, W = self.patch_embed2(x, masking_branch = -1) # x, H, W = self.patch_embed2(x, masking_branch = masking_branch, range_batches_to_mask = range_batches_to_mask) for i, blk in enumerate(self.block2): # score = self.score_predictor[1](x) # mask = [F.softmax(score_.reshape(B, -1, 2), dim=2)[:, :, 0] for score_ in score] # mask_: [B, N] # masks.append(mask) x = blk(x, H, W) x = self.norm2(x) x = [x_.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() for x_ in x] outs0.append(x[0]) outs1.append(x[1]) # stage 3 x, H, W = self.patch_embed3(x, masking_branch = -1) # x, H, W = self.patch_embed3(x, masking_branch = masking_branch, range_batches_to_mask = range_batches_to_mask) for i, blk in enumerate(self.block3): # score = self.score_predictor[2](x) # mask = [F.softmax(score_.reshape(B, -1, 2), dim=2)[:, :, 0] for score_ in score] # mask_: [B, N] # masks.append(mask) x = blk(x, H, W) x = self.norm3(x) x = [x_.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() for x_ in x] outs0.append(x[0]) outs1.append(x[1]) # stage 4 x, H, W = self.patch_embed4(x, masking_branch = -1) # x, H, W = self.patch_embed4(x, masking_branch = masking_branch, range_batches_to_mask = range_batches_to_mask) for i, blk in enumerate(self.block4): # score = self.score_predictor[3](x) # mask = [F.softmax(score_.reshape(B, -1, 2), dim=2)[:, :, 0] for score_ in score] # mask_: [B, N] # masks.append(mask) x = blk(x, H, W) x = self.norm4(x) x = [x_.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() for x_ in x] outs0.append(x[0]) outs1.append(x[1]) return [outs0, outs1] def forward(self, x, masking_branch, range_batches_to_mask): x = self.forward_features(x, masking_branch, range_batches_to_mask) return x class DWConv(nn.Module): def __init__(self, dim=768): super(DWConv, self).__init__() self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim) def forward(self, x, H, W): B, N, C = x.shape x = x.transpose(1, 2).contiguous().view(B, C, H, W) x = self.dwconv(x) x = x.flatten(2).transpose(1, 2).contiguous() return x class mit_b0(MixVisionTransformer): def __init__(self, ratio, masking_ratio, **kwargs): super(mit_b0, self).__init__(ratio = ratio, masking_ratio = masking_ratio, patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=LayerNormParallel, depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1) class mit_b1(MixVisionTransformer): def __init__(self, ratio, masking_ratio, **kwargs): super(mit_b1, self).__init__(ratio = ratio, masking_ratio = masking_ratio, patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=LayerNormParallel, depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1) class mit_b2(MixVisionTransformer): def __init__(self, ratio, masking_ratio, **kwargs): super(mit_b2, self).__init__(ratio = ratio, masking_ratio = masking_ratio, patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=LayerNormParallel, depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1) class mit_b3(MixVisionTransformer): def __init__(self, ratio, masking_ratio, **kwargs): super(mit_b3, self).__init__(ratio = ratio, masking_ratio = masking_ratio, patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=LayerNormParallel, depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1) class mit_b4(MixVisionTransformer): def __init__(self, ratio, masking_ratio, **kwargs): super(mit_b4, self).__init__(ratio = ratio, masking_ratio = masking_ratio, patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=LayerNormParallel, depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1) class mit_b5(MixVisionTransformer): def __init__(self, ratio, masking_ratio, **kwargs): super(mit_b5, self).__init__(ratio = ratio, masking_ratio = masking_ratio, patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=LayerNormParallel, depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1)