import math import logging import torch import torch.nn.functional as F from timm.models.layers import DropPath, to_2tuple, trunc_normal_ from torch import nn import torch.utils.checkpoint as checkpoint from functools import partial from einops import rearrange logger = logging.getLogger(__name__) # try: # from .flash_attention_class import FlashAttention # except: # logger.warn(f'flash_attn is not installed, you can install it by `pip install flash_attn` ') # try: # from flash_attn.modules.mlp import FusedMLP # except: # logger.warn(f'FusedMLP of flash_attn is not installed!!!') # try: # from flash_attn.ops.rms_norm import DropoutAddRMSNorm # except: # logger.warn(f'DropoutAddRMSNorm of flash_attn is not installed!!!') import numpy as np import torch import logging logger = logging.getLogger(__name__) # -------------------------------------------------------- # 3D sine-cosine position embedding # References: # MVD: https://github.com/ruiwang2021/mvd/blob/main/modeling_finetune.py # -------------------------------------------------------- def get_3d_sincos_pos_embed(embed_dim, grid_size, t_size, cls_token=False): """ grid_size: int of the grid height and width t_size: int of the temporal size return: pos_embed: [t_size*grid_size*grid_size, embed_dim] or [1+t_size*grid_size*grid_size, embed_dim] (w/ or w/o cls_token) """ assert embed_dim % 4 == 0 embed_dim_spatial = embed_dim // 4 * 3 embed_dim_temporal = embed_dim // 4 # spatial grid_h = np.arange(grid_size, dtype=np.float32) grid_w = np.arange(grid_size, dtype=np.float32) grid = np.meshgrid(grid_w, grid_h) # here w goes first grid = np.stack(grid, axis=0) grid = grid.reshape([2, 1, grid_size, grid_size]) pos_embed_spatial = get_2d_sincos_pos_embed_from_grid( embed_dim_spatial, grid ) # temporal grid_t = np.arange(t_size, dtype=np.float32) pos_embed_temporal = get_1d_sincos_pos_embed_from_grid( embed_dim_temporal, grid_t ) # concate: [T, H, W] order pos_embed_temporal = pos_embed_temporal[:, np.newaxis, :] pos_embed_temporal = np.repeat( pos_embed_temporal, grid_size**2, axis=1 ) # [T, H*W, D // 4] pos_embed_spatial = pos_embed_spatial[np.newaxis, :, :] pos_embed_spatial = np.repeat( pos_embed_spatial, t_size, axis=0 ) # [T, H*W, D // 4 * 3] pos_embed = np.concatenate([pos_embed_temporal, pos_embed_spatial], axis=-1) pos_embed = pos_embed.reshape([-1, embed_dim]) # [T*H*W, D] if cls_token: pos_embed = np.concatenate( [np.zeros([1, embed_dim]), pos_embed], axis=0 ) return pos_embed # -------------------------------------------------------- # 2D sine-cosine position embedding # References: # Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py # MoCo v3: https://github.com/facebookresearch/moco-v3 # -------------------------------------------------------- def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False): """ grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) """ grid_h = np.arange(grid_size, dtype=np.float32) grid_w = np.arange(grid_size, dtype=np.float32) grid = np.meshgrid(grid_w, grid_h) # here w goes first grid = np.stack(grid, axis=0) grid = grid.reshape([2, 1, grid_size, grid_size]) pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) if cls_token: pos_embed = np.concatenate( [np.zeros([1, embed_dim]), pos_embed], axis=0 ) return pos_embed def get_1d_sincos_pos_embed(embed_dim, t_size, cls_token=False): """ t_size: int of the temporal size return: pos_embed: [t_size, embed_dim] or [1+t_size, embed_dim] (w/ or w/o cls_token) """ grid_t = np.arange(t_size, dtype=np.float32) pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, grid_t) if cls_token: pos_embed = np.concatenate( [np.zeros([1, embed_dim]), pos_embed], axis=0 ) return pos_embed def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): assert embed_dim % 2 == 0 # use half of dimensions to encode grid_h emb_h = get_1d_sincos_pos_embed_from_grid( embed_dim // 2, grid[0] ) # (H*W, D/2) emb_w = get_1d_sincos_pos_embed_from_grid( embed_dim // 2, grid[1] ) # (H*W, D/2) emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) return emb def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): """ embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D) """ assert embed_dim % 2 == 0 omega = np.arange(embed_dim // 2, dtype=np.float32) omega /= embed_dim / 2.0 omega = 1.0 / 10000**omega # (D/2,) pos = pos.reshape(-1) # (M,) out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product emb_sin = np.sin(out) # (M, D/2) emb_cos = np.cos(out) # (M, D/2) emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) return emb def interpolate_pos_embed_internvideo2(checkpoint_model, model, orig_t_size = 8): # interpolate position embedding for pos_name in ['pos_embed', 'clip_pos_embed']: if pos_name in checkpoint_model: pos_embed_checkpoint = checkpoint_model[pos_name] embedding_size = pos_embed_checkpoint.shape[-1] # channel dim num_patches = model.patch_embed.num_patches # num_extra_tokens = model.pos_embed.shape[-2] - num_patches # 0/1 # we use 8 frames for pretraining # new_t_size = args.num_frames * args.num_segments // model.patch_embed.tubelet_size new_t_size = model.num_frames // model.tubelet_size # height (== width) for the checkpoint position embedding orig_size = int(((pos_embed_checkpoint.shape[-2] - num_extra_tokens)//(orig_t_size)) ** 0.5) # height (== width) for the new position embedding new_size = int((num_patches // (new_t_size))** 0.5) # class_token and dist_token are kept unchanged if orig_t_size != new_t_size: logger.info(f"Temporal interpolate from {orig_t_size} to {new_t_size} ({pos_name})") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> B， T, HW, C -> BHW, C, T (B = 1) pos_tokens = pos_tokens.view(1, orig_t_size, -1, embedding_size) pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, embedding_size, orig_t_size) pos_tokens = torch.nn.functional.interpolate(pos_tokens, size=new_t_size, mode='linear') pos_tokens = pos_tokens.view(1, -1, embedding_size, new_t_size) pos_tokens = pos_tokens.permute(0, 3, 1, 2).reshape(1, -1, embedding_size) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model[pos_name] = new_pos_embed pos_embed_checkpoint = new_pos_embed # class_token and dist_token are kept unchanged if orig_size != new_size: logger.info(f"Position interpolate from {orig_size}x{orig_size} to {new_size}x{new_size} ({pos_name})") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> BT, H, W, C -> BT, C, H, W pos_tokens = pos_tokens.reshape(-1, new_t_size, orig_size, orig_size, embedding_size) pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) # BT, C, H, W -> BT, H, W, C -> B, T, H, W, C pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, new_t_size, new_size, new_size, embedding_size) pos_tokens = pos_tokens.flatten(1, 3) # B, L, C new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model[pos_name] = new_pos_embed if 'pos_embed_spatial' in checkpoint_model or 'pos_embed_temporal' in checkpoint_model: raise NotImplementedError def interpolate_pos_embed_internvideo2_new(checkpoint_model, model, orig_t_size = 8): pos_names = [] for k in checkpoint_model.keys(): if ('pos_embed' in k or 'clip_pos_embed' in k) and 'img_pos_embed' not in k: # NOTE 暂时不插值img_pos，高分辨率时可能需要再加 pos_names.append(k) logger.info(f"pos names list for interpolating: {pos_names}") assert len(pos_names) > 0, checkpoint_model.keys() if 'pos_embed_spatial' in checkpoint_model.keys() or 'pos_embed_temporal' in checkpoint_model.keys(): raise NotImplementedError # interpolate position embedding for pos_name in pos_names: pos_embed_checkpoint = checkpoint_model[pos_name] embedding_size = pos_embed_checkpoint.shape[-1] # channel dim num_patches = model.patch_embed.num_patches # num_extra_tokens = model.pos_embed.shape[-2] - num_patches # 0/1 # we use 8 frames for pretraining # new_t_size = args.num_frames * args.num_segments // model.patch_embed.tubelet_size new_t_size = model.num_frames // model.tubelet_size # height (== width) for the checkpoint position embedding orig_size = int(((pos_embed_checkpoint.shape[-2] - num_extra_tokens)//(orig_t_size)) ** 0.5) # height (== width) for the new position embedding new_size = int((num_patches // (new_t_size))** 0.5) # class_token and dist_token are kept unchanged if orig_t_size != new_t_size: logger.info(f"Temporal interpolate from {orig_t_size} to {new_t_size} ({pos_name})") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> B， T, HW, C -> BHW, C, T (B = 1) pos_tokens = pos_tokens.view(1, orig_t_size, -1, embedding_size) pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, embedding_size, orig_t_size) pos_tokens = torch.nn.functional.interpolate(pos_tokens, size=new_t_size, mode='linear') pos_tokens = pos_tokens.view(1, -1, embedding_size, new_t_size) pos_tokens = pos_tokens.permute(0, 3, 1, 2).reshape(1, -1, embedding_size) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model[pos_name] = new_pos_embed pos_embed_checkpoint = new_pos_embed # class_token and dist_token are kept unchanged if orig_size != new_size: logger.info(f"Position interpolate from {orig_size}x{orig_size} to {new_size}x{new_size} ({pos_name})") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> BT, H, W, C -> BT, C, H, W pos_tokens = pos_tokens.reshape(-1, new_t_size, orig_size, orig_size, embedding_size) pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) # BT, C, H, W -> BT, H, W, C -> B, T, H, W, C pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, new_t_size, new_size, new_size, embedding_size) pos_tokens = pos_tokens.flatten(1, 3) # B, L, C new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model[pos_name] = new_pos_embed def interpolate_pos_embed(checkpoint_model, model, orig_t_size=4, pos_name='vision_encoder.pos_embed'): if pos_name in checkpoint_model: pos_embed_checkpoint = checkpoint_model[pos_name] embedding_size = pos_embed_checkpoint.shape[-1] # channel dim num_patches = model.patch_embed.num_patches # num_extra_tokens = model.pos_embed.shape[-2] - num_patches # 0/1 # we use 4 frames for pretraining new_t_size = model.T # height (== width) for the checkpoint position embedding orig_size = int(((pos_embed_checkpoint.shape[-2] - num_extra_tokens)//(orig_t_size)) ** 0.5) # height (== width) for the new position embedding new_size = int((num_patches // (new_t_size))** 0.5) # class_token and dist_token are kept unchanged if orig_t_size != new_t_size: print(f"Temporal interpolate from {orig_t_size} to {new_t_size} ({pos_name})") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> B， T, HW, C -> BHW, C, T (B = 1) pos_tokens = pos_tokens.view(1, orig_t_size, -1, embedding_size) pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, embedding_size, orig_t_size) pos_tokens = torch.nn.functional.interpolate(pos_tokens, size=new_t_size, mode='linear') pos_tokens = pos_tokens.view(1, -1, embedding_size, new_t_size) pos_tokens = pos_tokens.permute(0, 3, 1, 2).reshape(1, -1, embedding_size) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model[pos_name] = new_pos_embed pos_embed_checkpoint = new_pos_embed # class_token and dist_token are kept unchanged if orig_size != new_size: print(f"Position interpolate from {orig_size}x{orig_size} to {new_size}x{new_size} ({pos_name})") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> BT, H, W, C -> BT, C, H, W pos_tokens = pos_tokens.reshape(-1, new_t_size, orig_size, orig_size, embedding_size) pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) # BT, C, H, W -> BT, H, W, C -> B, T, H, W, C pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, new_t_size, new_size, new_size, embedding_size) pos_tokens = pos_tokens.flatten(1, 3) # B, L, C new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model[pos_name] = new_pos_embed else: raise NotImplementedError class CrossAttention(nn.Module): def __init__( self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., attn_head_dim=None, out_dim=None): super().__init__() if out_dim is None: out_dim = dim self.num_heads = num_heads head_dim = dim // num_heads if attn_head_dim is not None: head_dim = attn_head_dim all_head_dim = head_dim * self.num_heads self.scale = qk_scale or head_dim ** -0.5 assert all_head_dim == dim self.q = nn.Linear(dim, all_head_dim, bias=False) self.k = nn.Linear(dim, all_head_dim, bias=False) self.v = nn.Linear(dim, all_head_dim, bias=False) if qkv_bias: self.q_bias = nn.Parameter(torch.zeros(all_head_dim)) self.k_bias = nn.Parameter(torch.zeros(all_head_dim)) self.v_bias = nn.Parameter(torch.zeros(all_head_dim)) else: self.q_bias = None self.k_bias = None self.v_bias = None self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(all_head_dim, out_dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, k=None, v=None): B, N, C = x.shape N_k = k.shape[1] N_v = v.shape[1] q_bias, k_bias, v_bias = None, None, None if self.q_bias is not None: q_bias = self.q_bias k_bias = self.k_bias v_bias = self.v_bias q = F.linear(input=x, weight=self.q.weight, bias=q_bias) q = q.reshape(B, N, 1, self.num_heads, -1).permute(2, 0, 3, 1, 4).squeeze(0) # (B, N_head, N_q, dim) k = F.linear(input=k, weight=self.k.weight, bias=k_bias) k = k.reshape(B, N_k, 1, self.num_heads, -1).permute(2, 0, 3, 1, 4).squeeze(0) v = F.linear(input=v, weight=self.v.weight, bias=v_bias) v = v.reshape(B, N_v, 1, self.num_heads, -1).permute(2, 0, 3, 1, 4).squeeze(0) q = q * self.scale attn = (q @ k.transpose(-2, -1)) # (B, N_head, N_q, N_k) attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, -1) x = self.proj(x) x = self.proj_drop(x) return x class AttentiveBlock(nn.Module): def __init__(self, dim, num_heads, qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., norm_layer=nn.LayerNorm, attn_head_dim=None, out_dim=None): super().__init__() self.norm1_q = norm_layer(dim) self.norm1_k = norm_layer(dim) self.norm1_v = norm_layer(dim) self.cross_attn = CrossAttention( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim, out_dim=out_dim) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() def forward(self, x_q, x_kv, pos_q, pos_k, bool_masked_pos, rel_pos_bias=None): x_q = self.norm1_q(x_q + pos_q) x_k = self.norm1_k(x_kv + pos_k) x_v = self.norm1_v(x_kv) x = self.cross_attn(x_q, k=x_k, v=x_v) return x class AttentionPoolingBlock(AttentiveBlock): def forward(self, x): x_q = x.mean(1, keepdim=True) x_kv, pos_q, pos_k = x, 0, 0 x = super().forward(x_q, x_kv, pos_q, pos_k, bool_masked_pos=None, rel_pos_bias=None) x = x.squeeze(1) return x class RMSNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states): input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) class Attention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0., use_flash_attn=False, causal=False, norm_layer=nn.LayerNorm, qk_normalization=False, use_fused_rmsnorm=False): super().__init__() assert dim % num_heads == 0, 'dim should be divisible by num_heads' self.num_heads = num_heads head_dim = dim // num_heads self.scale = 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) self.use_flash_attn = use_flash_attn if use_flash_attn: self.causal = causal self.inner_attn = FlashAttention(attention_dropout=attn_drop) self.qk_normalization = qk_normalization self.q_norm = norm_layer(dim) if qk_normalization else nn.Identity() self.k_norm = norm_layer(dim) if qk_normalization else nn.Identity() self.use_fused_rmsnorm = use_fused_rmsnorm def _naive_attn(self, x): B, N, C = x.shape # print(x.shape, torch.cuda.memory_allocated(), torch.cuda.memory_allocated()) 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.unbind(0) # make torchscript happy (cannot use tensor as tuple) if self.qk_normalization: B_, H_, N_, D_ = q.shape q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2) k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2) attn = ((q * self.scale) @ k.transpose(-2, -1)) # attn = attn - attn.max(-1)[0].unsqueeze(-1) # in case of overflow for fp16 attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) # print(torch.cuda.memory_allocated(), torch.cuda.memory_allocated()) x = (attn @ v).transpose(1, 2).reshape(B, N, C) # print(f"\033[31m这{x.device}是{self.proj.weight.device} {self.proj.bias.device}\033[0m") # print(f"\033[31m类型{x.dtype}是{self.proj.weight.dtype} {self.proj.bias.dtype}\033[0m") x = self.proj(x) x = self.proj_drop(x) return x def _flash_attn(self, x, key_padding_mask=None, need_weights=False): qkv = self.qkv(x) qkv = rearrange(qkv, "b s (three h d) -> b s three h d", three=3, h=self.num_heads) if self.qk_normalization: q, k, v = qkv.unbind(2) if self.use_fused_rmsnorm: q = self.q_norm(q.flatten(-2, -1))[0].view(q.shape) k = self.k_norm(k.flatten(-2, -1))[0].view(k.shape) else: q = self.q_norm(q.flatten(-2, -1)).view(q.shape) k = self.k_norm(k.flatten(-2, -1)).view(k.shape) qkv = torch.stack([q, k, v], dim=2) context, _ = self.inner_attn( qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=self.causal ) outs = self.proj(rearrange(context, "b s h d -> b s (h d)")) outs = self.proj_drop(outs) return outs def forward(self, x): x = self._naive_attn(x) if not self.use_flash_attn else self._flash_attn(x) return x class Mlp(nn.Module): """ MLP as used in Vision Transformer, MLP-Mixer and related networks """ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, bias=True, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features bias = to_2tuple(bias) drop_probs = to_2tuple(drop) self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0]) self.act = act_layer() self.drop1 = nn.Dropout(drop_probs[0]) self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1]) self.drop2 = nn.Dropout(drop_probs[1]) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop1(x) x = self.fc2(x) x = self.drop2(x) return x class Block(nn.Module): def __init__( self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0., init_values=None, drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_flash_attn=False, use_fused_mlp=False, fused_mlp_heuristic=1, with_cp=False, qk_normalization=False, layerscale_no_force_fp32=False, use_fused_rmsnorm=False): super().__init__() self.norm1 = norm_layer(dim) self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop, use_flash_attn=use_flash_attn, causal=False, norm_layer=norm_layer, qk_normalization=qk_normalization, use_fused_rmsnorm=use_fused_rmsnorm) self.ls1 = nn.Parameter(init_values * torch.ones(dim)) # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) if use_fused_mlp: self.mlp = FusedMLP(in_features=dim, hidden_features=mlp_hidden_dim, heuristic=fused_mlp_heuristic) else: self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) self.ls2 = nn.Parameter(init_values * torch.ones(dim)) self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.with_cp = with_cp self.use_fused_rmsnorm = use_fused_rmsnorm def forward(self, x, residual=None): def _inner_forward(x, residual=None): if self.use_fused_rmsnorm: x, residual = self.norm1(x, residual) x = self.drop_path1(self.ls1 * self.attn(x) ) x, residual = self.norm2(x, residual) x = self.drop_path2(self.ls2 * self.mlp(x) ) return x, residual else: assert residual is None x = x + self.drop_path1(self.ls1 * self.attn(self.norm1(x))) x = x + self.drop_path2(self.ls2 * self.mlp(self.norm2(x))) return x if self.with_cp: # print(f"\033[31m use_checkpoint [0m") return checkpoint.checkpoint(_inner_forward, x, residual) else: return _inner_forward(x, residual=residual) class PatchEmbed(nn.Module): """ 3D Image to Patch Embedding """ def __init__( self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, num_frames=8, tubelet_size=1, norm_layer=None ): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) self.tubelet_size = tubelet_size self.img_size = img_size self.patch_size = patch_size self.grid_size = ( num_frames // tubelet_size, img_size[0] // patch_size[0], img_size[1] // patch_size[1] ) # (T, H, W) self.num_patches = self.grid_size[0] * self.grid_size[1] * self.grid_size[2] self.num_img_patches = self.grid_size[1] * self.grid_size[2] self.proj = nn.Conv3d( in_channels=in_chans, out_channels=embed_dim, kernel_size=(tubelet_size, patch_size[0], patch_size[1]), stride=(tubelet_size, patch_size[0], patch_size[1]) ) self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity() def forward(self, x): x = self.proj(x) x = x.flatten(3).permute(0, 2, 3, 1) # B x C x T x HW => B x T x HW x C x = self.norm(x) return x class PretrainVisionTransformer_clean(nn.Module): def __init__( self, in_chans: int = 3, patch_size: int = 14, img_size: int = 224, qkv_bias: bool = False, # follow internvl_clip to set False drop_path_rate: float = 0.25, # may need ablation embed_dim: int = 1408, num_heads: int = 16, mlp_ratio: float = 48/11, init_values: float = 1e-5, # may need ablation qk_normalization: bool = True, depth: int = 40, use_flash_attn: bool = True, use_fused_rmsnorm: bool = True, use_fused_mlp: bool = True, fused_mlp_heuristic: int = 1, attn_pool_num_heads: int = 16, clip_embed_dim: int = 768, layerscale_no_force_fp32: bool = False, # whether True for training? num_frames: int = 8, tubelet_size: int = 1, sep_pos_embed: bool = False, sep_image_video_pos_embed: bool = False, use_checkpoint: bool = False, checkpoint_num: int = 0, # for unmasked teacher x_vis_return_idx=-1, x_vis_only=False ): super().__init__() self.num_frames = num_frames self.tubelet_size = tubelet_size assert use_flash_attn == use_fused_rmsnorm == use_fused_mlp, 'use_flash_attn, use_fused_rmsnorm and use_fused_mlp should be consistent' self.use_flash_attn = use_flash_attn self.embed_dim = embed_dim logger.info(f"Origin depth: {depth}") depth = depth + x_vis_return_idx + 1 logger.info(f"New depth: {depth}") self.depth = depth self.x_vis_only = x_vis_only if use_fused_rmsnorm: norm_layer_for_blocks = partial(DropoutAddRMSNorm, eps=1e-6, prenorm=True) else: norm_layer_for_blocks = partial(RMSNorm, eps=1e-6) self.norm_layer_for_blocks = norm_layer_for_blocks self.patch_embed = PatchEmbed( img_size, patch_size, in_chans, embed_dim, num_frames=num_frames, tubelet_size=tubelet_size, ) num_patches = self.patch_embed.num_patches num_img_patches = self.patch_embed.num_img_patches # print(f"num_patches: {num_patches}, num_img_patches: {num_img_patches}") self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) # stolen from https://github.com/facebookresearch/mae_st/blob/dc072aaaf640d06892e23a33b42223a994efe272/models_vit.py#L65-L73C17 self.sep_pos_embed = sep_pos_embed self.sep_image_video_pos_embed = sep_image_video_pos_embed if sep_pos_embed: raise NotImplementedError else: if sep_image_video_pos_embed: logger.info("Use joint position embedding, for image and video we use different pos_embed.") self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) self.img_pos_embed = nn.Parameter(torch.zeros(1, num_img_patches + 1, embed_dim)) else: logger.info("Use joint position embedding, for image and video we use same pos_embed.") self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # choose which layer to use checkpoint with_cp_list = [False] * depth if use_checkpoint: for idx in range(depth): if idx < checkpoint_num: with_cp_list[idx] = True logger.info(f"Droppath rate: {dpr}") logger.info(f"Checkpoint list: {with_cp_list}") self.blocks = nn.ModuleList([ Block(embed_dim, num_heads, mlp_ratio, qkv_bias=qkv_bias, norm_layer=norm_layer_for_blocks, drop_path=dpr[i], init_values=init_values, attn_drop=0., use_flash_attn=use_flash_attn, use_fused_mlp=use_fused_mlp, fused_mlp_heuristic=fused_mlp_heuristic, with_cp=with_cp_list[i], qk_normalization=qk_normalization, layerscale_no_force_fp32=layerscale_no_force_fp32, use_fused_rmsnorm=use_fused_rmsnorm) for i in range(depth)]) if not self.x_vis_only: self.clip_projector = AttentionPoolingBlock( dim=embed_dim, num_heads=attn_pool_num_heads, qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., norm_layer=partial(nn.LayerNorm, eps=1e-5), out_dim=clip_embed_dim) self.init_pos_embed() # trunc_normal_(self.cls_token, std=.02) # self.apply(self._init_weights) # self.fix_init_weight() def init_pos_embed(self): logger.info("Init pos_embed from sincos pos_embed") if self.sep_pos_embed: raise NotImplementedError else: pos_embed = get_3d_sincos_pos_embed( self.pos_embed.shape[-1], self.patch_embed.grid_size[1], # height & weight self.patch_embed.grid_size[0], # t_size cls_token=True ) self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) if self.sep_image_video_pos_embed: img_pos_embed = get_3d_sincos_pos_embed( self.pos_embed.shape[-1], self.patch_embed.grid_size[1], # height & weight 1, cls_token=True ) self.img_pos_embed.data.copy_(torch.from_numpy(img_pos_embed).float().unsqueeze(0)) 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) def fix_init_weight(self): def rescale(param, layer_id): param.div_(math.sqrt(2.0 * layer_id)) for layer_id, layer in enumerate(self.blocks): rescale(layer.attn.proj.weight.data, layer_id + 1) rescale(layer.mlp.fc2.weight.data, layer_id + 1) @property def dtype(self): return self.patch_embed.proj.weight.dtype def get_num_layers(self): return len(self.blocks) @torch.jit.ignore def no_weight_decay(self): return { 'pos_embed', 'pos_embed_spatial', 'pos_embed_temporal', 'pos_embed_cls', 'img_pos_embed', 'cls_token' } def expand_pos_embed(self, pos_embed, new_t_size, L, use_vitar_fuzzing=False): ''' @param: pos_embed: original pos_embed, (1, T*L + 1, embed_dim) T: frames L: w * h method: interpolation method ''' pos_embed_checkpoint = pos_embed embedding_size = pos_embed_checkpoint.shape[-1] num_extra_tokens = 1 # height (== width) for the checkpoint position embedding orig_size = int(((pos_embed_checkpoint.shape[-2] - num_extra_tokens)//(self.num_frames / self.patch_embed.tubelet_size)) ** 0.5) # height (== width) for the new position embedding new_size = int(L ** 0.5) # class_token and dist_token are kept unchanged if self.num_frames != new_t_size: logger.info(f"Temporal interpolate from {self.num_frames} to {new_t_size} ") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> B， T, HW, C -> BHW, C, T (B = 1) pos_tokens = pos_tokens.view(1, self.num_frames, -1, embedding_size) pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, embedding_size, self.num_frames) pos_tokens = torch.nn.functional.interpolate(pos_tokens.cpu(), size=new_t_size, mode='linear').cuda() pos_tokens = pos_tokens.view(1, -1, embedding_size, new_t_size) pos_tokens = pos_tokens.permute(0, 3, 1, 2).reshape(1, -1, embedding_size) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) pos_embed_checkpoint = new_pos_embed # class_token and dist_token are kept unchanged if orig_size != new_size: logger.info(f"Position interpolate from {orig_size}x{orig_size} to {new_size}x{new_size}") extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] # B, L, C -> BT, H, W, C -> BT, C, H, W pos_tokens = pos_tokens.reshape(-1, new_t_size, orig_size, orig_size, embedding_size) pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens.cpu(), size=(new_size, new_size), mode='bicubic', align_corners=False).cuda() # BT, C, H, W -> BT, H, W, C -> B, T, H, W, C pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, new_t_size, new_size, new_size, embedding_size) pos_tokens = pos_tokens.flatten(1, 3) # B, L, C new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) if use_vitar_fuzzing: ... return new_pos_embed # @torch.cuda.amp.autocast(enabled=False) def forward(self, x, mask=None, use_image=False): x = self.patch_embed(x.type(self.dtype)) # print(f"x.shape: {x.shape} x.dtype: {x.dtype}, model.dtype: {self.dtype}") B, T, L, C = x.shape # T: temporal; L: spatial x = x.view([B, T * L, C]) # append cls token cls_tokens = self.cls_token.expand(B, -1, -1) x = torch.cat((cls_tokens, x), dim=1) # add pos_embed if self.sep_pos_embed: raise NotImplementedError else: if use_image: if self.sep_image_video_pos_embed: pos_embed = self.img_pos_embed else: # (1, num_img_patches + 1, embed_dim) # print('origin pos_embed.shape:', self.pos_embed.shape) cls_pos_embed = self.pos_embed[:, 0:1, :] # print('cls_pos_embed.shape:', cls_pos_embed.shape) img_pos_embed = self.pos_embed[:, 1:, :].view(1, self.num_frames, self.patch_embed.num_patches // self.num_frames, self.embed_dim).mean(dim=1) # print('img_pos_embed.shape:', img_pos_embed.shape) pos_embed = torch.cat([cls_pos_embed, img_pos_embed], dim=1) # print('final img_pos_embed.shape:', pos_embed.shape) else: pos_embed = self.pos_embed if pos_embed[0].shape != x[0].shape: # print(f'pos embed shape {pos_embed.shape} does not match x[0].shape {x[0].shape}') pos_embed = self.expand_pos_embed(pos_embed, T, L) # can accelerate here assert pos_embed[0].shape == x[0].shape, f'pos embed shape: {pos_embed.shape} not match x[0].shape {x[0].shape}' # print("pos_embed.shape:", pos_embed.shape) x = x + pos_embed # mask tokens, ~mask means visible if mask is not None: x = x[~mask].reshape(B, -1, C) else: x = x.reshape(B, -1, C) residual = None for idx, blk in enumerate(self.blocks): if isinstance(x, tuple) and len(x) == 2: x, residual = x x = blk(x, residual=residual) if isinstance(x, tuple) and len(x) == 2: x, residual = x if residual is not None: x = x + residual x_vis = x if self.x_vis_only: return x_vis else: x_pool_vis = self.clip_projector(x_vis) return x_vis, x_pool_vis, None, None def pretrain_internvideo2_giant_patch14_224_clean(config): model = PretrainVisionTransformer_clean( in_chans=3, img_size=224, patch_size=14, embed_dim=1408, depth=40, num_heads=16, mlp_ratio=48/11, attn_pool_num_heads=16, qkv_bias=False, drop_path_rate=0.25, init_values=0.00001, qk_normalization=True, use_flash_attn=config.vision_encoder.get('use_flash_attn', False), use_fused_rmsnorm=config.vision_encoder.get('use_fused_rmsnorm', False), use_fused_mlp=config.vision_encoder.get('use_fused_mlp', False), fused_mlp_heuristic=1, layerscale_no_force_fp32=True, num_frames=config.vision_encoder.num_frames, tubelet_size=config.vision_encoder.tubelet_size, sep_pos_embed=False, sep_image_video_pos_embed=config.vision_encoder.sep_image_video_pos_embed, use_checkpoint=config.vision_encoder.use_checkpoint, checkpoint_num=config.vision_encoder.checkpoint_num, x_vis_return_idx=config.vision_encoder.x_vis_return_idx, x_vis_only=config.vision_encoder.x_vis_only, ) if config.vision_encoder.pretrained is not None: logger.info(f"Loading pretrained weights from {config.vision_encoder.pretrained}") state_dict = torch.load(config.vision_encoder.pretrained, map_location='cpu') interpolate_pos_embed_internvideo2(state_dict, model, orig_t_size=4) # NOTE 8f for stage1 message = model.load_state_dict(state_dict, strict=False) logger.info(message) else: logger.info("No pretrained weights!!!") return model def pretrain_internvideo2_6b_patch14_224_clean(config): model = PretrainVisionTransformer_clean( in_chans=3, img_size=224, patch_size=14, embed_dim=3200, depth=48, num_heads=25, mlp_ratio=4, clip_embed_dim=config.vision_encoder.clip_embed_dim, attn_pool_num_heads=16, qkv_bias=False, drop_path_rate=0.3, init_values=0.00001, qk_normalization=True, use_flash_attn=config.vision_encoder.get('use_flash_attn', True), use_fused_rmsnorm=config.vision_encoder.get('use_fused_rmsnorm', True), use_fused_mlp=config.vision_encoder.get('use_fused_mlp', True), fused_mlp_heuristic=1, layerscale_no_force_fp32=True, num_frames=config.vision_encoder.num_frames, tubelet_size=config.vision_encoder.tubelet_size, sep_pos_embed=False, sep_image_video_pos_embed=config.vision_encoder.sep_image_video_pos_embed, use_checkpoint=config.vision_encoder.use_checkpoint, checkpoint_num=config.vision_encoder.checkpoint_num, x_vis_return_idx=config.vision_encoder.x_vis_return_idx, x_vis_only=config.vision_encoder.x_vis_only ) if config.vision_encoder.pretrained is not None: logger.info(f"Loading pretrained weights from {config.vision_encoder.pretrained}") state_dict = torch.load(config.vision_encoder.pretrained, map_location='cpu') interpolate_pos_embed_internvideo2(state_dict, model, orig_t_size=8) # NOTE 8f for stage1 msg = model.load_state_dict(state_dict, strict=False) logger.info(msg) else: logger.info("No pretrained weights!!!") return model