projects/umt/models/vit.py

from functools import partial

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from mmcv.cnn.bricks import DropPath
from mmengine import to_2tuple

from mmaction.registry import MODELS


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 original 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 = 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]

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))

        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 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)
        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):

    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):
        B, C, T, H, W = x.shape
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model " \
            f'({self.img_size[0]}*{self.img_size[1]}).'
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x


# sin-cos position encoding
def get_sinusoid_encoding_table(n_position,
                                d_hid,
                                cur_frame=-1,
                                pre_n_position=1568):
    """Sinusoid position encoding table."""

    def get_position_angle_vec(position):
        return [
            position / np.power(10000, 2 * (hid_j // 2) / d_hid)
            for hid_j in range(d_hid)
        ]

    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('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)
    if cur_frame != -1 and cur_frame != 8:
        print(f'Pretraining uses 8 frames, but current frame is {cur_frame}')
        print('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)
    if n_position == pre_n_position:
        return sinusoid_table
    else:
        print('Use learnable position embedding')
        return nn.Parameter(sinusoid_table, requires_grad=True)


@MODELS.register_module()
class UMTViT(nn.Module):

    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=partial(nn.LayerNorm, eps=1e-6),
                 init_values=0.,
                 use_learnable_pos_emb=False,
                 all_frames=16,
                 tubelet_size=1,
                 use_checkpoint=False,
                 checkpoint_num=0,
                 use_mean_pooling=True):
        super().__init__()
        self.num_features = self.embed_dim = embed_dim
        self.tubelet_size = tubelet_size
        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
        self.use_checkpoint = use_checkpoint
        self.checkpoint_num = checkpoint_num
        print(f'Use checkpoint: {use_checkpoint}')
        print(f'Checkpoint number: {checkpoint_num}')

        if use_learnable_pos_emb:
            self.pos_embed = nn.Parameter(
                torch.zeros(1, num_patches, embed_dim))
        else:
            # 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)]
        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 = nn.Identity() if use_mean_pooling else norm_layer(
            embed_dim)
        self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None

    def forward_features(self, x):
        x = self.patch_embed(x)
        B, _, _ = 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)

        for idx, blk in enumerate(self.blocks):
            if self.use_checkpoint and idx < self.checkpoint_num:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)

        x = self.norm(x)
        if self.fc_norm is not None:
            return self.fc_norm(x.mean(1))
        else:
            return x[:, 0]

    def forward(self, x):
        x = self.forward_features(x)
        return x