third_party/InternVideo/InternVideo2/single_modality/models/videomae.py

import os
from functools import partial
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.layers import drop_path, to_2tuple, trunc_normal_
from flash_attn import flash_attn_func


MODEL_PATH = 'your_model_path/videomae'
_MODELS = {
    # see videomaev2
    "vit_g14_hybrid": os.path.join(MODEL_PATH, "vit_g_hybrid_1200e_pre.pth"),
}


def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 400, 'input_size': (3, 224, 224), 'pool_size': None,
        'crop_pct': .9, 'interpolation': 'bicubic',
        'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),
        **kwargs
    }


class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    """
    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training)
    
    def extra_repr(self) -> str:
        return 'p={}'.format(self.drop_prob)


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        # x = self.drop(x)
        # commit this for the orignal BERT implement 
        x = self.fc2(x)
        x = self.drop(x)
        return x


class Attention(nn.Module):
    def __init__(
            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
            proj_drop=0., attn_head_dim=None):
        super().__init__()
        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

        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
        if qkv_bias:
            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
        else:
            self.q_bias = None
            self.v_bias = None

        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(all_head_dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x):
        B, N, C = x.shape
        qkv_bias = None
        if self.q_bias is not None:
            qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
        # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        x = flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=self.scale, causal=False).reshape(B, N, -1)

        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class Block(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,
                 attn_head_dim=None):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
            attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

        if init_values > 0:
            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        else:
            self.gamma_1, self.gamma_2 = None, None

    def forward(self, x):
        if self.gamma_1 is None:
            x = x + self.drop_path(self.attn(self.norm1(x)))
            x = x + self.drop_path(self.mlp(self.norm2(x)))
        else:
            x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x)))
            x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
        return x


class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, num_frames=16, tubelet_size=2):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        self.tubelet_size = int(tubelet_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) * (num_frames // self.tubelet_size)
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches
        self.proj = nn.Conv3d(in_channels=in_chans, out_channels=embed_dim, 
                            kernel_size=(self.tubelet_size, patch_size[0], patch_size[1]), 
                            stride=(self.tubelet_size, patch_size[0], patch_size[1]))

    def forward(self, x, **kwargs):
        B, C, T, H, W = x.shape
        # FIXME look at relaxing size constraints
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x
    
# sin-cos position encoding
# https://github.com/jadore801120/attention-is-all-you-need-pytorch/blob/master/transformer/Models.py#L31
def get_sinusoid_encoding_table(n_position, d_hid, cur_frame=-1, pre_n_position=1568): 
    ''' Sinusoid position encoding table ''' 
    # TODO: make it with torch instead of numpy 
    def get_position_angle_vec(position): 
        return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)] 
    
    # generate checkpoint position embedding
    sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(pre_n_position)]) 
    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i 
    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1 
    sinusoid_table = torch.tensor(sinusoid_table, dtype=torch.float, requires_grad=False).unsqueeze(0)
    print(f"n_position: {n_position}")
    print(f"pre_n_position: {pre_n_position}")
    if n_position // cur_frame * 8 != pre_n_position and cur_frame != -1:
        T = 8 # checkpoint frame
        P = 14 # checkpoint size
        C = d_hid
        new_P = int((n_position // cur_frame) ** 0.5) # testing size
        print(f'Pretraining uses 14x14, but current version is {new_P}x{new_P}')
        print(f'Interpolate the position embedding')
        sinusoid_table = sinusoid_table.reshape(-1, T, P, P, C)
        sinusoid_table = sinusoid_table.reshape(-1, P, P, C).permute(0, 3, 1, 2)
        sinusoid_table = torch.nn.functional.interpolate(
            sinusoid_table, size=(new_P, new_P), mode='bicubic', align_corners=False)
        # BT, C, H, W -> BT, H, W, C ->  B, T, H, W, C
        sinusoid_table = sinusoid_table.permute(0, 2, 3, 1).reshape(-1, T, new_P, new_P, C)
        sinusoid_table = sinusoid_table.flatten(1, 3)  # B, THW, C
    if cur_frame != -1 and cur_frame != 8:
        print(f'Pretraining uses 8 frames, but current frame is {cur_frame}')
        print(f'Interpolate the position embedding')
        T = 8 # checkpoint frame
        new_T = cur_frame # testing frame
        # interpolate
        P = int((n_position // cur_frame) ** 0.5) # testing size
        C = d_hid
        sinusoid_table = sinusoid_table.reshape(-1, T, P, P, C)
        sinusoid_table = sinusoid_table.permute(0, 2, 3, 4, 1).reshape(-1, C, T)  # BHW, C, T
        sinusoid_table = torch.nn.functional.interpolate(sinusoid_table, size=new_T, mode='linear')
        sinusoid_table = sinusoid_table.reshape(1, P, P, C, new_T).permute(0, 4, 1, 2, 3) # B, T, H, W, C
        sinusoid_table = sinusoid_table.flatten(1, 3)  # B, THW, C
    if n_position == pre_n_position:
        return sinusoid_table
    else:
        print("Use learnable position embedding")
        return nn.Parameter(sinusoid_table, requires_grad=True)


class VisionTransformer(nn.Module):
    """ Vision Transformer with support for patch or hybrid CNN input stage
    """
    def __init__(
            self, 
            img_size=224, 
            patch_size=16, 
            in_chans=3, 
            embed_dim=768, 
            depth=12,
            num_heads=12, 
            mlp_ratio=4., 
            qkv_bias=False, 
            qk_scale=None, 
            drop_rate=0., 
            attn_drop_rate=0.,
            drop_path_rate=0., 
            norm_layer=nn.LayerNorm, 
            init_values=0.,
            all_frames=16,
            tubelet_size=2,
            mae_norm_type='l2',
            mae_return_layer=1,
            mae_return_interval=1,
        ):
        super().__init__()
        self.mae_norm_type = mae_norm_type
        self.return_index = []
        for i in range(mae_return_layer):
            self.return_index.append(depth - int(i * mae_return_interval) - 1)
        print(f'Normalization Type: {mae_norm_type}')
        print(f'MAE Teacher return index: : {self.return_index}')

        self.tubelet_size = tubelet_size
        self.depth = depth
        self.patch_embed = PatchEmbed(
            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, num_frames=all_frames, tubelet_size=self.tubelet_size)
        num_patches = self.patch_embed.num_patches

        # sine-cosine positional embeddings is on the way
        if patch_size == 14:
            pre_n_position = 2048
        else:
            pre_n_position = 1568
        self.pos_embed = get_sinusoid_encoding_table(
            num_patches, embed_dim, all_frames // tubelet_size,
            pre_n_position=pre_n_position
        )

        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        self.blocks = nn.ModuleList([
            Block(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
                init_values=init_values)
            for i in range(depth)])
        self.norm = norm_layer(embed_dim)

        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)

    def get_num_layers(self):
        return len(self.blocks)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token'}

    def forward(self, x, mask=None):
        x = self.patch_embed(x)
        B, _, C = x.size()

        if self.pos_embed is not None:
            x = x + self.pos_embed.expand(B, -1, -1).type_as(x).to(x.device).clone().detach()
        x = self.pos_drop(x)

        if mask is not None:
            x = x[~mask].reshape(B, -1, C) # ~mask means visible

        z = []
        for idx, blk in enumerate(self.blocks):
            x = blk(x)
            if idx == self.depth - 1:
                x = self.norm(x)
            if idx in self.return_index:
                z.append(x)
        x = torch.stack(z)

        if self.mae_norm_type == 'l2':
            x = x / x.norm(dim=-1, keepdim=True)
        elif self.mae_norm_type == 'none':
            pass
        else:
            raise NotImplementedError
        
        return x
    

def load_state_dict(model, state_dict):
    from collections import OrderedDict
    new_state_dict = OrderedDict()
    for k, v in state_dict.items():
        if k.startswith('encoder.'):
            new_k = k[8:]
            if new_k == "patch_embed.proj.weight" and model.tubelet_size == 1:
                print("Kernel pooling")
                v = v.mean(dim=2, keepdim=True)
            new_state_dict[new_k] = v
    msg = model.load_state_dict(new_state_dict)
    print(msg)


def mae_g14_hybrid(pretrained=True, **kwargs):
    model = VisionTransformer(
        patch_size=14, embed_dim=1408, depth=40, num_heads=16, mlp_ratio=48/11, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    if pretrained:
        print('load MAE pretrained weights')
        state_dict = torch.load(_MODELS["vit_g14_hybrid"], map_location='cpu')
        load_state_dict(model, state_dict['model'])
    return model


if __name__ == '__main__':
    import time
    from fvcore.nn import FlopCountAnalysis
    from fvcore.nn import flop_count_table
    import numpy as np

    seed = 4217
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    num_frames = 16

    model = mae_g14_hybrid(all_frames=num_frames, tubelet_size=2).cuda().half()
    # print(model)

    flops = FlopCountAnalysis(model, torch.rand(1, 3, num_frames, 224, 224).cuda().half())
    s = time.time()
    print(flop_count_table(flops, max_depth=1))
    print(time.time()-s)
    # print(model(torch.rand(1, 3, num_frames, 224, 224)).shape)