""" SAE Model Script ver： Oct 28th 2023 15:30 SAE stands for shuffled autoencoder, designed for PuzzleTuning # References: Based on MAE code. https://github.com/facebookresearch/mae """ from functools import partial import torch import torch.nn as nn from timm.models.vision_transformer import PatchEmbed, Block from SSL_structures.pos_embed import get_2d_sincos_pos_embed from Backbone.VPT_structure import VPT_ViT class ShuffledAutoEncoderViT(VPT_ViT): """ Shuffled Autoencoder with VisionTransformer backbone prompt_mode: "Deep" / "Shallow" by default None """ def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=1024, depth=24, num_heads=16, decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4., norm_layer=nn.LayerNorm, norm_pix_loss=False, group_shuffle_size=-1, prompt_mode=None, Prompt_Token_num=20, basic_state_dict=None, decoder=None, decoder_rep_dim=None): if prompt_mode is None: super().__init__() # SAE encoder specifics (this part just the same as ViT) # -------------------------------------------------------------------------- self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim) # BCHW -> BNC num_patches = self.patch_embed.num_patches # learnable cls token is still used but on cls head need self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) # set and freeze encoder_pos_embed, use the fixed sin-cos embedding for tokens + mask_token self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim), requires_grad=False) # Encoder blocks self.blocks = nn.ModuleList([ # qk_scale=None fixme related to timm version Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer) for i in range(depth)]) self.norm = norm_layer(embed_dim) self.prompt_mode = prompt_mode # -------------------------------------------------------------------------- else: super().__init__(img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, depth=depth, num_heads=num_heads, mlp_ratio=mlp_ratio, norm_layer=norm_layer, Prompt_Token_num=Prompt_Token_num, VPT_type=prompt_mode, basic_state_dict=None) # Firstly, set then Encoder state_dict to none here. num_patches = self.patch_embed.num_patches # set patch_embed of VPT # set and freeze encoder_pos_embed, use the fixed sin-cos embedding for tokens + mask_token self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim), requires_grad=False) self.prompt_mode = prompt_mode # Freeze Encoder parameters except of the Prompt Tokens self.Freeze() # SAE decoder specifics todo as a low-level backbone, the explore for future segmentation is need # -------------------------------------------------------------------------- # if the feature dimension of encoder and decoder are different, use decoder_embed to align them if embed_dim != decoder_embed_dim: self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True) else: self.decoder_embed = nn.Identity() # set decoder if decoder is not None: self.decoder = decoder # Decoder use a FC to reconstruct image, unlike the Encoder which use a CNN to split patch self.decoder_pred = nn.Linear(decoder_rep_dim, patch_size ** 2 * in_chans, bias=True) # decoder to patch else: self.decoder = None # set and freeze decoder_pos_embed, use the fixed sin-cos embedding for tokens + mask_token self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, decoder_embed_dim), requires_grad=False) self.decoder_blocks = nn.ModuleList([ # qk_scale=None fixme related to timm version Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer) for i in range(decoder_depth)]) self.decoder_norm = norm_layer(decoder_embed_dim) # Decoder use a FC to reconstruct image, unlike the Encoder which use a CNN to split patch self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size ** 2 * in_chans, bias=True) # decoder to patch # -------------------------------------------------------------------------- # this controls the puzzle group self.group_shuffle_size = group_shuffle_size # wether or not to use norm_pix_loss self.norm_pix_loss = norm_pix_loss # parameter initialization self.initialize_weights() # load basic state_dict of backbone for Transfer-learning-based tuning if basic_state_dict is not None: self.load_state_dict(basic_state_dict, False) def initialize_weights(self): # initialization # initialize a 2d positional encoding of (embed_dim, grid) by sin-cos embedding pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.patch_embed.num_patches ** .5), cls_token=True) self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) if self.decoder is None: # initialize a 2d positional encoding of (embed_dim, grid) by sin-cos embedding decoder_pos_embed = get_2d_sincos_pos_embed(self.decoder_pos_embed.shape[-1], int(self.patch_embed.num_patches ** .5), cls_token=True) self.decoder_pos_embed.data.copy_(torch.from_numpy(decoder_pos_embed).float().unsqueeze(0)) # initialize patch_embed like nn.Linear (instead of nn.Conv2d) w = self.patch_embed.proj.weight.data torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1])) # xavier_uniform，让输入输出的方差相同，包括前后向传播 # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.) torch.nn.init.normal_(self.cls_token, std=.02) # torch.nn.init.normal_(self.prompt_token, std=.02) # initialize nn.Linear and nn.LayerNorm self.apply(self._init_weights) def _init_weights(self, m): # initialize nn.Linear and nn.LayerNorm if isinstance(m, nn.Linear): # we use xavier_uniform following official JAX ViT: torch.nn.init.xavier_uniform_(m.weight) 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) def patchify(self, imgs, patch_size=None): """ Break image to patch tokens input: imgs: (B, 3, H, W) output: x: (B, num_patches, patch_size**2 *3) AKA [B, num_patches, flatten_dim] """ # patch_size patch_size = self.patch_embed.patch_size[0] if patch_size is None else patch_size # assert H == W and image shape is dividedable by patch assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % patch_size == 0 # patch num in rol or column h = w = imgs.shape[2] // patch_size # use reshape to split patch [B, C, H, W] -> [B, C, h_p, patch_size, w_p, patch_size] x = imgs.reshape(shape=(imgs.shape[0], 3, h, patch_size, w, patch_size)) # ReArrange dimensions [B, C, h_p, patch_size, w_p, patch_size] -> [B, h_p, w_p, patch_size, patch_size, C] x = torch.einsum('nchpwq->nhwpqc', x) # ReArrange dimensions [B, h_p, w_p, patch_size, patch_size, C] -> [B, num_patches, flatten_dim] x = x.reshape(shape=(imgs.shape[0], h * w, patch_size ** 2 * 3)) return x def patchify_decoder(self, imgs, patch_size=None): """ Break image to patch tokens fixme,注意，这里patch_size应该是按照decoder的网络设置来作为default更合理 input: imgs: (B, CLS, H, W) output: x: (B, num_patches, -1) AKA [B, num_patches, -1] """ # patch_size patch_size = self.patch_embed.patch_size[0] if patch_size is None else patch_size # assert H == W and image shape is divided-able by patch assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % patch_size == 0 # patch num in rol or column h = w = imgs.shape[2] // patch_size # use reshape to split patch [B, CLS, H, W] -> [B, CLS, h_p, patch_size, w_p, patch_size] x = imgs.reshape(shape=(imgs.shape[0], -1, h, patch_size, w, patch_size)) # ReArrange dimensions [B, CLS, h_p, patch_size, w_p, patch_size] -> [B, h_p, w_p, patch_size, patch_size, CLS] x = torch.einsum('nchpwq->nhwpqc', x) # ReArrange dimensions [B, h_p, w_p, patch_size, patch_size, C] -> [B, num_patches, flatten_dim] x = x.reshape(shape=(imgs.shape[0], h * w, -1)) return x def unpatchify(self, x, patch_size=None): """ Decoding encoded patch tokens input: x: (B, num_patches, patch_size**2 *3) AKA [B, num_patches, flatten_dim] output: imgs: (B, 3, H, W) """ # patch_size p = self.patch_embed.patch_size[0] if patch_size is None else patch_size # squre root of num_patches (without CLS token is required) h = w = int(x.shape[1] ** .5) # assert num_patches is with out CLS token assert h * w == x.shape[1] # ReArrange dimensions [B, num_patches, flatten_dim] -> [B, h_p, w_p, patch_size, patch_size, C] x = x.reshape(shape=(x.shape[0], h, w, p, p, 3)) # ReArrange dimensions [B, h_p, w_p, patch_size, patch_size, C] -> [B, C, h_p, patch_size, w_p, patch_size] x = torch.einsum('nhwpqc->nchpwq', x) # use reshape to compose patch [B, C, h_p, patch_size, w_p, patch_size] -> [B, C, H, W] imgs = x.reshape(shape=(x.shape[0], 3, h * p, h * p)) return imgs def fix_position_shuffling(self, x, fix_position_ratio, puzzle_patch_size): """ Fix-position shuffling Randomly assign patches by per-sample shuffling. After it, the fixed patches are reserved as Positional Tokens the rest patches are batch wise randomly shuffled among the batch since they serve as Relation Tokens. Per-sample shuffling is done by argsort random noise. batch wise shuffle operation is done by shuffle all idxes input: x: [B, 3, H, W], input image tensor fix_position_ratio float puzzle_patch_size int output: x_puzzled, mask x_puzzled: [B, 3, H, W] mask: [B, 3, H, W], binary mask indicating pix position with 0 """ # Break img into puzzle patches with the size of puzzle_patch_size [B, num_puzzle_patches, D_puzzle] x = self.patchify(x, puzzle_patch_size) # output: x: (B, num_patches, patch_size**2 *3) AKA [B, num_patches, flatten_dim] B, num_puzzle_patches, D = x.shape # num of fix_position puzzle patches len_fix_position = int(num_puzzle_patches * fix_position_ratio) num_shuffled_patches = num_puzzle_patches - len_fix_position # create a noise tensor to prepare shuffle idx of puzzle patches noise = torch.rand(B, num_puzzle_patches, device=x.device) # [B,num_puzzle_patches] noise in [0, 1] # 在Batch里面每个序列上获得noise tensor经过升序排列后原本位置的idx矩阵，（各自不同） ids_shuffle = torch.argsort(noise, dim=1) # ascend: small is keep, large is remove # 再对idx矩阵继续升序排列可获得：原始noise tensor的每个位置的排序顺位 ids_restore = torch.argsort(ids_shuffle, dim=1) # keep the first subset 前面的是fix的，后面的是puzzle的 ids_fix = ids_shuffle[:, :len_fix_position] # [B,num_puzzle_patches] -> [B,fix_patches] # fix_patches=num_puzzle_patches * fix_position_ratio len_fix_position ids_puzzle = ids_shuffle[:, len_fix_position:] # [B,num_puzzle_patches] -> [B,puzzle_patches] # puzzle_patches=num_puzzle_patches*(1-fix_position_ratio) num_shuffled_patches # set puzzle patch # ids_?.unsqueeze(-1).repeat(1, 1, D) # [B,?_patches] -> [B,?_patches,1] (at each place with the idx of ori patch) -> [B,?_patches,D] # torch.gather to select patche groups x_fixed of [B,fix_patches,D] and x_puzzle of [B,puzzle_patches,D] # 要保持的，batch中每个sample不一样 x_fixed = torch.gather(x, dim=1, index=ids_fix.unsqueeze(-1).repeat(1, 1, D)) # 要shuffle的，batch中每个sample不一样 x_puzzle = torch.gather(x, dim=1, index=ids_puzzle.unsqueeze(-1).repeat(1, 1, D)) # batch&patch-wise shuffle is needed else the restore will restore all puzzles if self.group_shuffle_size == -1 or self.group_shuffle_size == B: puzzle_shuffle_indices = torch.randperm(B * num_shuffled_patches, device=x.device, requires_grad=False) else: assert B > self.group_shuffle_size > 0 and B % self.group_shuffle_size == 0 # build [B//self.group_shuffle_size, num_puzzle_patches] noise in [0, 1] group_noise = torch.rand(B // self.group_shuffle_size, num_shuffled_patches * self.group_shuffle_size, device=x.device) # get shuffled index in each (num_shuffled_patches*group_shuffle) group_ids_shuffle = torch.argsort(group_noise, dim=1) # break the dim and add the group idx(in list), stack back to tensor group_ids_shuffle = torch.stack([group_ids_shuffle[i] + num_shuffled_patches * self.group_shuffle_size * i for i in range(B // self.group_shuffle_size)]) # flattern to be idx for all (B * num_shuffled_patches) puzzle_shuffle_indices = group_ids_shuffle.view(-1) # 将0~B * num_shuffled_patches-1（包括0和B * num_shuffled_patches-1）随机打乱后获得的数字序列 x_puzzle = x_puzzle.view(B * num_shuffled_patches, D)[puzzle_shuffle_indices].view(B, num_shuffled_patches, D) # 利用randperm获得的乱序序列对应batch内所有需要shuffle的部分进行打乱顺序，之后将其恢复为原本的划分batch # pack up all puzzle patches x = torch.cat([x_fixed, x_puzzle], dim=1) # generate the binary mask: 0 is keep, 1 is remove mask = torch.ones([B, num_puzzle_patches, D], device=x.device, requires_grad=False) # no grad mask[:, :len_fix_position, :] = 0 # set the first len_fix of tokens to 0，rest to 1 # unshuffle to restore the fixed positions x = torch.gather(x, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, D)) # torch.gather to generate restored binary mask mask = torch.gather(mask, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, D)) # unpatchify to obtain puzzle images and their mask x = self.unpatchify(x, puzzle_patch_size) mask = self.unpatchify(mask, puzzle_patch_size) return x, mask # x_puzzled and mask def forward_puzzle(self, imgs, fix_position_ratio=0.25, puzzle_patch_size=32): """ Transform the input images to puzzle images input: x: [B, 3, H, W], input image tensor fix_position_ratio float puzzle_patch_size int output: x_puzzled, mask x_puzzled: [B, 3, H, W] mask: [B, 3, H, W], binary mask indicating pix position with 0 """ x_puzzled, mask = self.fix_position_shuffling(imgs, fix_position_ratio, puzzle_patch_size) return x_puzzled, mask def forward_encoder(self, imgs): """ :param imgs: [B, C, H, W], sequence of imgs :return: Encoder output: encoded tokens, mask position, restore idxs x: [B, num_patches, D], sequence of Tokens (including the cls token) CLS_token: [B, 1, D] """ if self.prompt_mode is None: # ViT # embed patches x = self.patch_embed(imgs) # add pos embed before concatenate the cls token x = x + self.pos_embed[:, 1:, :] # detatch puzzle for embed_puzzle output embed_puzzle = x.data.detach() # append cls token cls_token = self.cls_token + self.pos_embed[:, :1, :] cls_tokens = cls_token.expand(x.shape[0], -1, -1) # batch fix x = torch.cat((cls_tokens, x), dim=1) # apply Transformer blocks for blk in self.blocks: x = blk(x) else: # VPT x = self.patch_embed(imgs) # add pos embed before concatenate the cls token x = x + self.pos_embed[:, 1:, :] # detatch puzzle for embed_puzzle output embed_puzzle = x.data.detach() # copy the embed original puzzle (for illustration) # append cls token cls_token = self.cls_token + self.pos_embed[:, :1, :] cls_tokens = cls_token.expand(x.shape[0], -1, -1) # batch fix x = torch.cat((cls_tokens, x), dim=1) if self.VPT_type == "Deep": Prompt_Token_num = self.Prompt_Tokens.shape[1] for i in range(len(self.blocks)): # concatenate Prompt_Tokens Prompt_Tokens = self.Prompt_Tokens[i].unsqueeze(0) # firstly concatenate x = torch.cat((x, Prompt_Tokens.expand(x.shape[0], -1, -1)), dim=1) num_tokens = x.shape[1] # lastly remove, a good trick x = self.blocks[i](x)[:, :num_tokens - Prompt_Token_num] else: # self.VPT_type == "Shallow" Prompt_Token_num = self.Prompt_Tokens.shape[1] # concatenate Prompt_Tokens Prompt_Tokens = self.Prompt_Tokens.expand(x.shape[0], -1, -1) x = torch.cat((x, Prompt_Tokens), dim=1) num_tokens = x.shape[1] # A whole sequential process x = self.blocks(x)[:, :num_tokens - Prompt_Token_num] # last norm of Transformer x = self.norm(x) CLS_token = x[:, :1, :] x = x[:, 1:, :] # Encoder output: encoded tokens, mask position, embed original puzzle (for illustration) return x, CLS_token, embed_puzzle def forward_decoder(self, x): """ Decoder to reconstruct the puzzle image [B, 1 + num_patches, D_Encoder] -> [B, 1 + num_patches, D_Decoder] -> [B, num_patches, p*p*3] :param x: [B, 1 + num_patches, D_Encoder], sequence of Tokens (including the cls token) :return: Decoder output: reconstracted tokens x: [B, num_patches, patch_size ** 2 * in_chans], sequence of Patch Tokens """ if self.decoder is None: # embed tokens: [B, num_encoded_tokens, D_Encoder] -> [B, num_encoded_tokens, D_Decoder] x = self.decoder_embed(x) # print(x.shape) # add pos embed x = x + self.decoder_pos_embed # apply Transformer blocks for blk in self.decoder_blocks: x = blk(x) x = self.decoder_norm(x) # Reconstruction projection x = self.decoder_pred(x) # remove cls token x = x[:, 1:, :] # print("x shape: ", x.shape) # [B, N, p*p*3] else: # remove cls token x = x[:, 1:, :] # embed tokens: [B, num_encoded_tokens, D_Encoder] -> [B, num_encoded_tokens, D_Decoder] x = self.decoder_embed(x) # unpatchify to make image form [B, H, W, C] x = self.unpatchify(x) # restore image by Encoder # apply decoder module to segment the output of encoder x = self.decoder(x) # one-hot seg decoder [B, CLS, H, W] # the output of segmentation is transformed to [B, N, Dec] x = self.patchify_decoder(x) # TODO 做一个有意义的设计 # Convert the number of channels to match image for loss function x = self.decoder_pred(x) # [B, N, Dec] -> [B, N, p*p*3] # print(x.shape) return x def forward_loss(self, imgs, pred, mask): """ MSE loss for all patches towards the ori image Input: imgs: [B, 3, H, W], Encoder input image pred: [B, num_patches, p*p*3], Decoder reconstructed image mask: [B, num_patches, p*p*3], 0 is keep, 1 is puzzled """ # print("pred shape: ", pred.shape) # [64, 196, 768] # target imgs: [B, 3, H, W] -> [B, num_patches, p*p*3] target = self.patchify(imgs) # print("target shape: ", target.shape) # [64, 196, 768] # use mask as a patch indicator [B, num_patches, D] -> [B, num_patches] mask = mask[:, :, 0] # Binary mask, 1 for removed patches, 0 for reserved patches: if self.norm_pix_loss: # Normalize the target image patches mean = target.mean(dim=-1, keepdim=True) var = target.var(dim=-1, keepdim=True) target = (target - mean) / (var + 1.e-6) ** .5 # MSE loss loss = (pred - target) ** 2 loss = loss.mean(dim=-1) # [B, num_patches], mean loss on each patch pixel loss = (loss * mask).sum() / mask.sum() # mean loss on removed patches [B], scalar return loss def forward(self, imgs, fix_position_ratio=0.25, puzzle_patch_size=32, combined_pred_illustration=False): # STEP 1: Puzzle making # create puzzle images: [B, 3, H, W] imgs_puzzled, mask = self.forward_puzzle(imgs, fix_position_ratio, puzzle_patch_size) # Visualization of imgs_puzzled_patches sequence: [B, num_patches, p*p*3] imgs_puzzled_patches = self.patchify(imgs_puzzled) # here, latent crop size is automatically based on encoder embedding # STEP 2: Puzzle understanding # Encoder to obtain latent tokens and embed_puzzle: [B, num_patches, D] latent_puzzle, CLS_token, embed_puzzle = self.forward_encoder(imgs_puzzled) # VPT output size of more tokens ? currently use firstly-cat-lastly-remove so its fine # STEP 3: Puzzle restoring # step 3.(a) prepare decoder input indcator mask at the encoder output stage: mask_patches_pp3 = self.patchify(mask) # mark relation tokens with 1 [B, num_patches, p*p*3] # here, latent crop size is automatically based on encoder embedding # Reassign mask indicator shape to the encoder output dim if mask_patches_pp3.shape[-1] != latent_puzzle.shape[-1]: # [B, num_patches, p*p*3] -> [B, num_patches, 1] -> [B, num_patches, D] mask_patches = mask_patches_pp3[:, :, :1].expand(-1, -1, latent_puzzle.shape[-1]) else: mask_patches = mask_patches_pp3 # anti_mask: [B, num_patches, D], binary mask indicating fix position with 1 instead of 0 anti_mask = mask_patches * -1 + 1 # great trick to process positional operation with less calculation # Position hint # in mask, 0 is Position Tokens, therefore take only Relation Tokens latent_tokens = latent_puzzle * mask_patches # take out relation tokens（latent_tokens here) # in anti_mask, 0 is Relation Tokens, therefore take only Position Tokens hint_tokens = embed_puzzle * anti_mask # anti_mask to take hint_tokens (position tokens) # group decoder tokens: [B, num_patches, D] latent = latent_tokens + hint_tokens # append back the cls token at the first -> [B, 1+num_patches, D] x = torch.cat([CLS_token, latent], dim=1) # step 3.(b) Decoder to obtain Reconstructed image patches: # [B, 1+num_patches,D] -> [B, 1+num_patches, D_Decoder] -> [B, num_patches, p*p*3] pred = self.forward_decoder(x) # combined pred anti_mask_patches_pp3 = mask_patches_pp3 * -1 + 1 # fix position with 1, relation patches with 0 hint_img_patches = imgs_puzzled_patches * anti_mask_patches_pp3 pred_img_patches = pred * mask_patches_pp3 # mark relation tokens with 1, fix position with 0 pred_with_hint_imgs = hint_img_patches + pred_img_patches # MSE loss for all patches towards the ori image loss = self.forward_loss(imgs, pred, mask_patches) # print(loss) # check whether the loss is working if combined_pred_illustration: return loss, pred_with_hint_imgs, imgs_puzzled_patches else: return loss, pred, imgs_puzzled_patches def sae_vit_base_patch16_dec512d8b(dec_idx=None, **kwargs): print("Decoder:", dec_idx) model = ShuffledAutoEncoderViT( patch_size=16, embed_dim=768, depth=12, num_heads=12, decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) return model def sae_vit_large_patch16_dec512d8b(dec_idx=None, **kwargs): print("Decoder:", dec_idx) model = ShuffledAutoEncoderViT( patch_size=16, embed_dim=1024, depth=24, num_heads=16, decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) return model def sae_vit_huge_patch14_dec512d8b(dec_idx=None, **kwargs): print("Decoder:", dec_idx) model = ShuffledAutoEncoderViT( patch_size=14, embed_dim=1280, depth=32, num_heads=16, decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) return model # decoder def sae_vit_base_patch16_dec(dec_idx=None, num_classes=3, img_size=224, **kwargs): # num_classes做的是one-hot seg但是不是做还原，我们得设计一下如何去做这个还原才能实现预训练 if dec_idx == 'swin_unet': decoder_embed_dim = 768 decoder_rep_dim = 16 * 16 * 3 from SSL_structures.Swin_Unet_main.networks.vision_transformer import SwinUnet as ViT_seg decoder = ViT_seg(num_classes=num_classes, img_size=img_size, patch_size=16) elif dec_idx == 'transunet': decoder_embed_dim = 768 decoder_rep_dim = 16 * 16 * 3 transunet_name = 'R50-ViT-B_16' transunet_patches_size = 16 from SSL_structures.TransUNet_main.networks.vit_seg_modeling import CONFIGS as CONFIGS_Transunet_seg from SSL_structures.TransUNet_main.networks.vit_seg_modeling import VisionTransformer as Transunet_seg config_vit = CONFIGS_Transunet_seg[transunet_name] config_vit.n_classes = num_classes config_vit.n_skip = 3 if transunet_name.find('R50') != -1: config_vit.patches.grid = ( int(img_size / transunet_patches_size), int(img_size / transunet_patches_size)) decoder = Transunet_seg(config_vit, num_classes=config_vit.n_classes) elif dec_idx == 'UTNetV2': decoder_embed_dim = 768 decoder_rep_dim = 16 * 16 * 3 from SSL_structures.UtnetV2.utnetv2 import UTNetV2 as UTNetV2_seg decoder = UTNetV2_seg(in_chan=3, num_classes=num_classes) else: print('no effective decoder!') return -1 print('dec_idx: ', dec_idx) model = ShuffledAutoEncoderViT( patch_size=16, embed_dim=768, depth=12, num_heads=12, decoder_embed_dim=decoder_embed_dim, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), decoder_rep_dim=decoder_rep_dim, decoder=decoder, **kwargs) return model def sae_vit_large_patch16_dec(dec_idx=None, num_classes=3, img_size=224, **kwargs): # num_classes做的是one-hot seg但是不是做还原，我们得设计一下如何去做这个还原才能实现预训练 if dec_idx == 'swin_unet': decoder_embed_dim = 768 decoder_rep_dim = 16 * 16 * 3 from SSL_structures.Swin_Unet_main.networks.vision_transformer import SwinUnet as ViT_seg decoder = ViT_seg(num_classes=num_classes, img_size=img_size, patch_size=16) elif dec_idx == 'transunet': decoder_embed_dim = 768 decoder_rep_dim = 16 * 16 * 3 transunet_name = 'R50-ViT-B_16' transunet_patches_size = 16 from SSL_structures.TransUNet_main.networks.vit_seg_modeling import CONFIGS as CONFIGS_Transunet_seg from SSL_structures.TransUNet_main.networks.vit_seg_modeling import VisionTransformer as Transunet_seg config_vit = CONFIGS_Transunet_seg[transunet_name] config_vit.n_classes = num_classes config_vit.n_skip = 3 if transunet_name.find('R50') != -1: config_vit.patches.grid = ( int(img_size / transunet_patches_size), int(img_size / transunet_patches_size)) decoder = Transunet_seg(config_vit, num_classes=config_vit.n_classes) elif dec_idx == 'UTNetV2': decoder_embed_dim = 768 decoder_rep_dim = 16 * 16 * 3 from SSL_structures.UtnetV2.utnetv2 import UTNetV2 as UTNetV2_seg decoder = UTNetV2_seg(in_chan=3, num_classes=num_classes) else: print('no effective decoder!') return -1 print('dec_idx: ', dec_idx) model = ShuffledAutoEncoderViT( patch_size=16, embed_dim=1024, depth=24, num_heads=16, decoder_embed_dim=decoder_embed_dim, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), decoder_rep_dim=decoder_rep_dim, decoder=decoder, **kwargs) return model def sae_vit_huge_patch14_dec(dec_idx=None, num_classes=3, img_size=224, **kwargs): # num_classes做的是one-hot seg但是不是做还原，我们得设计一下如何去做这个还原才能实现预训练 if dec_idx == 'swin_unet': decoder_embed_dim = 14 * 14 * 3 decoder_rep_dim = 14 * 14 * 3 from SSL_structures.Swin_Unet_main.networks.vision_transformer import SwinUnet as ViT_seg decoder = ViT_seg(num_classes=num_classes, img_size=img_size, patch_size=16) elif dec_idx == 'transunet': decoder_embed_dim = 14 * 14 * 3 decoder_rep_dim = 14 * 14 * 3 transunet_name = 'R50-ViT-B_16' transunet_patches_size = 16 from SSL_structures.TransUNet_main.networks.vit_seg_modeling import CONFIGS as CONFIGS_Transunet_seg from SSL_structures.TransUNet_main.networks.vit_seg_modeling import VisionTransformer as Transunet_seg config_vit = CONFIGS_Transunet_seg[transunet_name] config_vit.n_classes = num_classes config_vit.n_skip = 3 if transunet_name.find('R50') != -1: config_vit.patches.grid = ( int(img_size / transunet_patches_size), int(img_size / transunet_patches_size)) decoder = Transunet_seg(config_vit, num_classes=config_vit.n_classes) elif dec_idx == 'UTNetV2': decoder_embed_dim = 14 * 14 * 3 decoder_rep_dim = 14 * 14 * 3 from SSL_structures.UtnetV2.utnetv2 import UTNetV2 as UTNetV2_seg decoder = UTNetV2_seg(in_chan=3, num_classes=num_classes) else: print('no effective decoder!') return -1 print('dec_idx: ', dec_idx) model = ShuffledAutoEncoderViT( patch_size=14, embed_dim=1280, depth=32, num_heads=16, decoder_embed_dim=decoder_embed_dim, decoder_depth=8, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), decoder_rep_dim=decoder_rep_dim, decoder=decoder, **kwargs) return model # set recommended archs following MAE sae_vit_base_patch16 = sae_vit_base_patch16_dec512d8b # decoder: 512 dim, 8 blocks sae_vit_large_patch16 = sae_vit_large_patch16_dec512d8b # decoder: 512 dim, 8 blocks sae_vit_huge_patch14 = sae_vit_huge_patch14_dec512d8b # decoder: 512 dim, 8 blocks # Equiped with decoders sae_vit_base_patch16_decoder = sae_vit_base_patch16_dec # decoder: 768 dim, HYF decoders sae_vit_large_patch16_decoder = sae_vit_large_patch16_dec # decoder: 768 dim, HYF decoders sae_vit_huge_patch14_decoder = sae_vit_huge_patch14_dec # decoder: 768 dim, HYF decoders if __name__ == '__main__': device = torch.device("cuda" if torch.cuda.is_available() else "cpu") img_size = 224 ''' num_classes = 3 # set to 3 for 3 channel x = torch.rand(2, 3, img_size, img_size, device=device) ''' image_tensor_path = './temp-tensors/color.pt' x = torch.load(image_tensor_path) x.to(device) # model = sae_vit_base_patch16(img_size=img_size, decoder=None) # model = sae_vit_huge_patch14(img_size=img_size, decoder=None) # model = sae_vit_base_patch16_decoder(prompt_mode="Deep", dec_idx='swin_unet', img_size=img_size) model = sae_vit_base_patch16(img_size=img_size, decoder=None, group_shuffle_size=2) ''' # ViT_Prompt from pprint import pprint model_names = timm.list_models('*vit*') pprint(model_names) basic_model = timm.create_model('vit_base_patch' + str(16) + '_' + str(edge_size), pretrained=True) basic_state_dict = basic_model.state_dict() model = sae_vit_base_patch16(img_size=384, prompt_mode='Deep', Prompt_Token_num=20, basic_state_dict=basic_state_dict) prompt_state_dict = model.obtain_prompt() VPT = VPT_ViT(img_size=384, VPT_type='Deep', Prompt_Token_num=20, basic_state_dict=basic_state_dict) VPT.load_prompt(prompt_state_dict) VPT.to(device) pred = VPT(x) print(pred) ''' model.to(device) loss, pred, imgs_puzzled_patches = model(x, fix_position_ratio=0.25, puzzle_patch_size=32, combined_pred_illustration=True) # combined_pred_illustration = True to add hint tokens at the pred, False to know more info # 可视化看看效果 from utils.visual_usage import * imgs_puzzled_batch = unpatchify(imgs_puzzled_patches, patch_size=16) for img_idx in range(len(imgs_puzzled_batch)): puzzled_img = imgs_puzzled_batch.cpu()[img_idx] puzzled_img = ToPILImage()(puzzled_img) puzzled_img.save(os.path.join('./temp-figs/', 'puzzled_sample_'+str(img_idx)+'.jpg')) recons_img_batch = unpatchify(pred, patch_size=16) recons_img = recons_img_batch.cpu()[img_idx] recons_img = ToPILImage()(recons_img) recons_img.save(os.path.join('./temp-figs/', 'recons_sample_'+str(img_idx)+'.jpg')) ''' print(loss, '\n') print(loss.shape, '\n') print(pred.shape, '\n') print(imgs_puzzled_patches.shape, '\n') '''