# Copyright (c) 2022 Graphcore Ltd. All rights reserved. # Copyright (c) Meta Platforms, Inc. and affiliates. # All rights reserved. # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # -------------------------------------------------------- # References: # timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm # DeiT: https://github.com/facebookresearch/deit # -------------------------------------------------------- # # This file has been modified by Graphcore Ltd. from functools import partial import torch import torch.nn as nn import poptorch from core.vision_transformer import PatchEmbed, Block from util.pos_embed import get_2d_sincos_pos_embed from core.gelu import ERF_GELU from util.log import logger def remap_tensor( x, fwd_grain_size=8, bwd_grain_size=0, fwd_clone_layout=False, bwd_clone_layout=False, fwd_after_matmul=False, bwd_after_matmul=False, debug_str="", ): if 0 == bwd_grain_size: bwd_grain_size = fwd_grain_size return poptorch.custom_op( [x], "RemapCE", "ai.graphcore", 1, example_outputs=[x], attributes={ "fwd_grain_size": fwd_grain_size, "bwd_grain_size": bwd_grain_size, "fwd_clone_layout": 1 if fwd_clone_layout is True else 0, "bwd_clone_layout": 1 if bwd_clone_layout is True else 0, "fwd_after_matmul": 1 if fwd_after_matmul is True else 0, "bwd_after_matmul": 1 if bwd_after_matmul is True else 0, "debug_str": debug_str, }, )[0] class MaeLoss(nn.Module): def __init__(self, norm_pix_loss, p): super().__init__() self.p = p self.norm_pix_loss = norm_pix_loss self.mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1) self.var = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1) def forward(self, imgs, pred, mask): """ imgs: [N, 3, H, W] pred: [N, L, p*p*3] mask: [N, L], 0 is keep, 1 is remove, """ # mean and variance should be on the same device self.mean = self.mean.to(imgs.device) self.var = self.var.to(imgs.device) imgs = imgs.detach() / 255.0 imgs = (imgs - self.mean) / self.var pred = pred.float() target = self.patchify(imgs) if self.norm_pix_loss: mean = target.mean(dim=-1, keepdim=True) var = target.var(dim=-1, keepdim=True) target = (target - mean) / (var + 1.0e-6) ** 0.5 loss = (pred - target) ** 2 loss = loss.mean(dim=-1) # [N, L], mean loss per patch loss = (loss * mask).sum() / mask.sum() # mean loss on removed patches return loss def patchify(self, imgs): """ imgs: (N, 3, H, W) x: (N, L, patch_size**2 *3) """ p = self.p assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0 h = w = imgs.shape[2] // p x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p)) x = x.permute(0, 2, 4, 3, 5, 1) x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3)) return x class MaskedAutoencoderViT(nn.Module): """Masked Autoencoder with VisionTransformer backbone""" 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=4.0, gelu_type="erf", norm_layer=nn.LayerNorm, norm_pix_loss=False, pipeline=None, device="ipu", mask_ratio=0.75, half=False, ): super().__init__() # MAE encoder specifics self.use_half = half gelu = nn.GELU if gelu_type == "erf": gelu = ERF_GELU self.mask_ratio = mask_ratio self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim) num_patches = self.patch_embed.num_patches self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) self.pos_embed = nn.Parameter( torch.zeros(1, num_patches + 1, embed_dim), requires_grad=False ) # fixed sin-cos embedding self.blocks = nn.ModuleList( [ Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer, act_layer=gelu) for i in range(depth) ] ) self.norm = norm_layer(embed_dim) # MAE decoder specifics self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True) self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim)) self.decoder_pos_embed = nn.Parameter( torch.zeros(1, num_patches + 1, decoder_embed_dim), requires_grad=False ) # fixed sin-cos embedding self.decoder_blocks = nn.ModuleList( [ Block( decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer, act_layer=gelu, ) for i in range(decoder_depth) ] ) self.decoder_norm = norm_layer(decoder_embed_dim) self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size**2 * in_chans, bias=True) # decoder to patch self.initialize_weights() self.loss = MaeLoss(norm_pix_loss, self.patch_embed.patch_size) # ipu set self.device = device self.pipeline = pipeline if self.pipeline is not None: self._set_pipeline() def _get_layer_ipu(self, pipeline): layer_ipu = [] for ipu, n_layers in enumerate(pipeline): layer_ipu += [ipu] * n_layers return layer_ipu def _set_mae(self, embed, blocks, norm, layer_ipu, first_id, block_count, stage): embed = poptorch.BeginBlock(embed, f"ipu{stage}", ipu_id=first_id) logger.info(f"embed on ipu {first_id}, stage:{stage}") for i, block in enumerate(blocks): index = block_count + i logger.info(index) ipu = layer_ipu[index] logger.info(f"begin block at ({i}) on ipu:{ipu}, stage:{stage}") blocks[i] = poptorch.BeginBlock(block, f"ipu{stage}", ipu_id=ipu) stage += 1 norm = poptorch.BeginBlock(norm, f"ipu{stage-1}", ipu_id=ipu) logger.info(f"norm on ipu {ipu}, stage:{stage-1}") stage += 1 return ipu, index + 1, stage def _set_pipeline(self): stage = 0 num = len(self.pipeline) / 2 pipeline = self.pipeline[: int(num)] layer_ipu = self._get_layer_ipu(pipeline) last_id, block_count, stage = self._set_mae(self.patch_embed, self.blocks, self.norm, layer_ipu, 0, 0, stage) pipeline = self.pipeline[int(num) :] layer_ipu = self._get_layer_ipu(pipeline) last_id, _, stage = self._set_mae( self.decoder_embed, self.decoder_blocks, self.decoder_norm, layer_ipu, 0, 0, stage ) self.decoder_pred = poptorch.BeginBlock(self.decoder_pred, f"ipu{stage}", ipu_id=last_id) logger.info(f"decoder_pred on ipu {last_id}, stage: {stage}") self.loss = poptorch.BeginBlock(self.loss, f"ipu{stage}", ipu_id=last_id) logger.info(f"loss on ipu {last_id}, stage: {stage}") def initialize_weights(self): # initialization # initialize (and freeze) pos_embed by sin-cos embedding pos_embed = get_2d_sincos_pos_embed( self.pos_embed.shape[-1], int(self.patch_embed.num_patches**0.5), cls_token=True ) self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) decoder_pos_embed = get_2d_sincos_pos_embed( self.decoder_pos_embed.shape[-1], int(self.patch_embed.num_patches**0.5), 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])) # 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=0.02) torch.nn.init.normal_(self.mask_token, std=0.02) # initialize nn.Linear and nn.LayerNorm self.apply(self._init_weights) def _init_weights(self, m): 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): """ imgs: (N, 3, H, W) x: (N, L, patch_size**2 *3) """ p = self.patch_embed.patch_size[0] assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0 h = w = imgs.shape[2] // p x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p)) x = x.permute(0, 2, 4, 3, 5, 1) x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3)) return x def random_masking(self, x, mask_ratio, ids_shuffle, keep_mat): """ Perform per-sample random masking by per-sample shuffling. Per-sample shuffling is done by argsort random noise. x: [N, L, D], sequence """ N, L, D = x.shape # batch, length, dim len_keep = int(L * (1 - mask_ratio)) # keep the first subset # workaround for gather if self.device == "ipu": x_masked = x.permute(0, 2, 1).matmul(keep_mat).permute(0, 2, 1) else: ids_keep = ids_shuffle[:, :len_keep] x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D)) return x_masked def forward_encoder(self, x, mask_ratio, ids_shuffle, keep_mat): # embed patches x = self.patch_embed(x) # add pos embed w/o cls token x = x + self.pos_embed[:, 1:, :] # masking: length -> length * mask_ratio x = self.random_masking(x, mask_ratio, ids_shuffle, keep_mat) # append cls token cls_token = self.cls_token + self.pos_embed[:, :1, :] cls_tokens = cls_token.expand(x.shape[0], -1, -1) if self.use_half: x = torch.cat((cls_tokens.half(), x), dim=1) else: x = torch.cat((cls_tokens, x), dim=1) # apply Transformer blocks for blk in self.blocks: if self.device == "ipu": x = remap_tensor(x) x = blk(x) x = remap_tensor(x) else: x = blk(x) x = self.norm(x) return x def forward_decoder(self, x, ids_restore, restore_mat): # embed tokens x = self.decoder_embed(x) # append mask tokens to sequence mask_tokens = self.mask_token.repeat(x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1) if self.use_half: x_ = torch.cat([x[:, 1:, :], mask_tokens.half()], dim=1) # no cls token else: x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1) if self.device == "ipu": x_ = x_.permute(0, 2, 1).matmul(restore_mat).permute(0, 2, 1) else: x_ = torch.gather(x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])) # unshuffle x = torch.cat([x[:, :1, :], x_], dim=1) # append cls token # add pos embed x = x + self.decoder_pos_embed.detach() # apply Transformer blocks for blk in self.decoder_blocks: if self.device == "ipu": x = remap_tensor(x) x = blk(x) x = remap_tensor(x) else: x = blk(x) x = self.decoder_norm(x) # predictor projection x = self.decoder_pred(x) # remove cls token x = x[:, 1:, :] return x def forward(self, imgs, target, ids_restore, keep_mat, restore_mat, mask, ids_shuffle=None): imgs = imgs / 255.0 mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1) var = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1) imgs = (imgs - mean) / var if self.use_half: imgs = imgs.half() latent = self.forward_encoder(imgs, self.mask_ratio, ids_shuffle, keep_mat) pred = self.forward_decoder(latent, ids_restore, restore_mat) # [N, L, p*p*3] loss = self.loss(target, pred, mask) return pred, poptorch.identity_loss(loss, reduction="mean") def mae_vit_align_patch16_dec512d4b(**kwargs): model = MaskedAutoencoderViT( patch_size=16, embed_dim=768, depth=3, num_heads=12, decoder_embed_dim=512, decoder_depth=2, decoder_num_heads=16, mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs, ) return model def mae_vit_base_patch16_dec512d8b(**kwargs): model = MaskedAutoencoderViT( 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 mae_vit_large_patch16_dec512d8b(**kwargs): model = MaskedAutoencoderViT( 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 mae_vit_huge_patch14_dec512d8b(**kwargs): model = MaskedAutoencoderViT( 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 # set recommended archs mae_vit_align_patch16 = mae_vit_align_patch16_dec512d4b mae_vit_base_patch16 = mae_vit_base_patch16_dec512d8b # decoder: 512 dim, 8 blocks mae_vit_large_patch16 = mae_vit_large_patch16_dec512d8b # decoder: 512 dim, 8 blocks mae_vit_huge_patch14 = mae_vit_huge_patch14_dec512d8b # decoder: 512 dim, 8 blocks