vit_tweaks.py

# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from collections.abc import Callable

import numpy as np
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddle import ParamAttr
from paddle.nn.initializer import Constant
from paddle.regularizer import L2Decay

from ...utils import load_ckpt
from ..registry import BACKBONES
from ..weight_init import trunc_normal_

__all__ = ['VisionTransformer_tweaks']

zeros_ = Constant(value=0.)
ones_ = Constant(value=1.)


def to_2tuple(x):
    return tuple([x] * 2)


def rand_bbox(size, lam):
    """ rand_bbox """
    w = size[2]
    h = size[3]
    cut_rat = np.sqrt(1. - lam)
    cut_w = np.int(w * cut_rat)
    cut_h = np.int(h * cut_rat)

    # uniform
    cx = np.random.randint(w)
    cy = np.random.randint(h)

    bbx1 = np.clip(cx - cut_w // 2, 0, w)
    bby1 = np.clip(cy - cut_h // 2, 0, h)
    bbx2 = np.clip(cx + cut_w // 2, 0, w)
    bby2 = np.clip(cy + cut_h // 2, 0, h)

    return bbx1, bby1, bbx2, bby2


def drop_path(x, drop_prob=0., training=False):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    # issuecomment-532968956 ...
    See discussion: https://github.com/tensorflow/tpu/issues/494
    """
    if drop_prob == 0. or not training:
        return x
    keep_prob = paddle.to_tensor(1 - drop_prob)
    shape = (paddle.shape(x)[0], ) + (1, ) * (x.ndim - 1)
    random_tensor = keep_prob + paddle.rand(shape, dtype=x.dtype)
    random_tensor = paddle.floor(random_tensor)  # binarize
    output = x.divide(keep_prob) * random_tensor

    return output


class DropPath(nn.Layer):
    """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)


class Identity(nn.Layer):
    def __init__(self):
        super(Identity, self).__init__()

    def forward(self, input):
        return input


class Mlp(nn.Layer):
    def __init__(self,
                 in_features,
                 hidden_features=None,
                 out_features=None,
                 act_layer=nn.GELU,
                 drop=0.,
                 wd_bias=True,
                 lr_mult=1.0):
        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)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class Attention(nn.Layer):
    def __init__(self,
                 dim,
                 num_heads=8,
                 qkv_bias=False,
                 qk_scale=None,
                 attn_drop=0.,
                 proj_drop=0.,
                 wd_bias=True,
                 lr_mult=1.0):
        super().__init__()

        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim**-0.5
        self.qkv = nn.Linear(dim, dim * 3, bias_attr=qkv_bias)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.attn_drop = nn.Dropout(attn_drop)

    def forward(self, x):
        N, C = x.shape[1:]
        qkv = self.qkv(x).reshape(
            (-1, N, 3, self.num_heads, C // self.num_heads)).transpose(
                (2, 0, 3, 1, 4))
        q, k, v = qkv[0], qkv[1], qkv[2]

        attn = (q.matmul(k.transpose((0, 1, 3, 2)))) * self.scale
        attn = nn.functional.softmax(attn, axis=-1)
        attn = self.attn_drop(attn)

        x = (attn.matmul(v)).transpose((0, 2, 1, 3)).reshape((-1, N, C))
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class Block(nn.Layer):
    def __init__(self,
                 dim,
                 num_heads,
                 mlp_ratio=4.0,
                 qkv_bias=False,
                 qk_scale=None,
                 drop=0.0,
                 attn_drop=0.0,
                 drop_path=0.1,
                 act_layer=nn.GELU,
                 norm_layer='nn.LayerNorm',
                 epsilon=1e-5,
                 attention_type='divided_space_time',
                 wd_bias=True,
                 lr_mult=1.0):

        super().__init__()
        self.attention_type = attention_type
        if isinstance(norm_layer, str):
            self.norm1 = eval(norm_layer)(dim, epsilon=epsilon)
        elif isinstance(norm_layer, Callable):
            self.norm1 = norm_layer(dim, epsilon=epsilon)
        else:
            raise TypeError(
                "The norm_layer must be str or paddle.nn.layer.Layer class")

        self.attn = Attention(dim,
                              num_heads=num_heads,
                              qkv_bias=qkv_bias,
                              qk_scale=qk_scale,
                              attn_drop=attn_drop,
                              proj_drop=drop,
                              wd_bias=wd_bias,
                              lr_mult=lr_mult)

        # Temporal Attention Parameters
        if self.attention_type == 'divided_space_time':
            if isinstance(norm_layer, str):
                self.temporal_norm1 = eval(norm_layer)(dim, epsilon=epsilon)
            elif isinstance(norm_layer, Callable):
                self.temporal_norm1 = norm_layer(dim, epsilon=epsilon)
            else:
                raise TypeError(
                    "The norm_layer must be str or paddle.nn.layer.Layer class")
            self.temporal_attn = Attention(dim,
                                           num_heads=num_heads,
                                           qkv_bias=qkv_bias,
                                           qk_scale=qk_scale,
                                           attn_drop=attn_drop,
                                           proj_drop=drop,
                                           wd_bias=wd_bias,
                                           lr_mult=lr_mult)
            self.temporal_fc = nn.Linear(dim, 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 Identity()
        if isinstance(norm_layer, str):
            self.norm2 = eval(norm_layer)(dim, epsilon=epsilon)
        elif isinstance(norm_layer, Callable):
            self.norm2 = norm_layer(dim, epsilon=epsilon)
        else:
            raise TypeError(
                "The norm_layer must be str or paddle.nn.layer.Layer class")

        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim,
                       hidden_features=mlp_hidden_dim,
                       act_layer=act_layer,
                       drop=drop,
                       wd_bias=wd_bias,
                       lr_mult=lr_mult)

    def forward(self, x, B, T, W):
        num_spatial_tokens = (x.shape[1] - 1) // T
        H = num_spatial_tokens // W
        if self.attention_type in ['space_only', 'joint_space_time']:
            x = x + self.drop_path(self.attn(self.norm1(x)))
            x = x + self.drop_path(self.mlp(self.norm2(x)))
            return x
        elif self.attention_type == 'divided_space_time':
            ########## Temporal ##########
            xt = x[:, 1:, :]
            _, _, _, _t, _m = B, H, W, T, xt.shape[-1]
            xt = xt.reshape([-1, _t, _m])

            res_temporal = self.drop_path(
                self.temporal_attn(self.temporal_norm1(xt)))

            _, _h, _w, _t, _m = B, H, W, T, res_temporal.shape[-1]
            res_temporal = res_temporal.reshape([-1, _h * _w * _t, _m])

            res_temporal = self.temporal_fc(res_temporal)
            xt = x[:, 1:, :] + res_temporal

            ########## Spatial ##########
            init_cls_token = x[:, 0, :].unsqueeze(1)
            cls_token = init_cls_token.tile((1, T, 1))
            _b, _t, _m = cls_token.shape
            cls_token = cls_token.reshape([-1, _m]).unsqueeze(1)

            xs = xt
            _, _h, _w, _t, _m = B, H, W, T, xs.shape[-1]
            xs = xs.reshape([-1, _h, _w, _t, _m]).transpose(
                (0, 3, 1, 2, 4)).reshape([-1, _h * _w, _m])
            xs = paddle.concat((cls_token, xs), axis=1)
            res_spatial = self.drop_path(self.attn(self.norm1(xs)))

            # Taking care of CLS token
            cls_token = res_spatial[:, 0, :]
            _, _t, _m = B, T, cls_token.shape[-1]
            cls_token = cls_token.reshape([-1, _t, _m])
            # averaging for every frame
            cls_token = paddle.mean(cls_token, axis=1, keepdim=True)

            res_spatial = res_spatial[:, 1:, :]
            _, _t, _h, _w, _m = B, T, H, W, res_spatial.shape[-1]
            res_spatial = res_spatial.reshape([-1, _t, _h, _w, _m]).transpose(
                (0, 2, 3, 1, 4)).reshape([-1, _h * _w * _t, _m])

            res = res_spatial
            x = xt
            x = paddle.concat((init_cls_token, x), axis=1) + paddle.concat(
                (cls_token, res), axis=1)

            # Mlp
            x = x + self.drop_path(self.mlp(self.norm2(x)))
            return x
        else:
            raise NotImplementedError


class PatchEmbed(nn.Layer):
    """ Image to Patch Embedding
    """
    def __init__(self,
                 img_size=224,
                 patch_size=16,
                 in_channels=3,
                 embed_dim=768,
                 wd_bias=True,
                 lr_mult=1.0):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] //
                                                        patch_size[0])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.proj = nn.Conv2D(in_channels,
                              embed_dim,
                              kernel_size=patch_size,
                              stride=patch_size)

    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 ({self.img_size[0]}*{self.img_size[1]})."
        x = x.transpose((0, 2, 1, 3, 4))  # [B,T,C,H,W]
        x = x.reshape([-1, C, H, W])  # [BT,C,H,W]
        x = self.proj(x)  # [BT,F,nH,nW]
        W = x.shape[-1]
        x = x.flatten(2).transpose((0, 2, 1))  # [BT,F,nHnW]
        return x, T, W


@BACKBONES.register()
class VisionTransformer_tweaks(nn.Layer):
    """ Vision Transformer with support for patch input
    """
    def __init__(self,
                 pretrained=None,
                 img_size=224,
                 patch_size=16,
                 in_channels=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=0.1,
                 norm_layer='nn.LayerNorm',
                 epsilon=1e-5,
                 num_seg=8,
                 attention_type='divided_space_time',
                 wd_bias=True,
                 lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0],
                 **args):
        super().__init__()
        self.pretrained = pretrained
        self.num_seg = num_seg
        self.attention_type = attention_type
        self.lr_mult_list = lr_mult_list
        self.num_features = self.embed_dim = embed_dim

        self.patch_embed = PatchEmbed(img_size=img_size,
                                      patch_size=patch_size,
                                      in_channels=in_channels,
                                      embed_dim=embed_dim,
                                      wd_bias=wd_bias,
                                      lr_mult=self.lr_mult_list[0])
        num_patches = self.patch_embed.num_patches

        # Positional Embeddings
        self.cls_token = self.create_parameter(
            shape=(1, 1, embed_dim),
            default_initializer=zeros_,
            attr=ParamAttr(regularizer=L2Decay(0.0)))
        self.pos_embed = self.create_parameter(
            shape=(1, num_patches + 1, embed_dim),
            default_initializer=zeros_,
            attr=ParamAttr(regularizer=L2Decay(0.0)))
        self.pos_drop = nn.Dropout(p=drop_rate)

        if self.attention_type != 'space_only':
            self.time_embed = self.create_parameter(
                shape=(1, num_seg, embed_dim),
                default_initializer=zeros_,
                attr=ParamAttr(regularizer=L2Decay(0.0)))
            self.time_drop = nn.Dropout(p=drop_rate)

        self.add_parameter("pos_embed", self.pos_embed)
        self.add_parameter("cls_token", self.cls_token)

        dpr = np.linspace(0, drop_path_rate, depth)

        self.blocks = nn.LayerList([
            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,
                  epsilon=epsilon,
                  attention_type=self.attention_type,
                  wd_bias=wd_bias,
                  lr_mult=self.lr_mult_list[(i // 4) + 1]) for i in range(depth)
        ])

        self.norm = eval(norm_layer)(embed_dim, epsilon=epsilon)

    def init_weights(self):
        """First init model's weight"""
        trunc_normal_(self.pos_embed, std=0.02)
        trunc_normal_(self.cls_token, std=0.02)
        self.apply(self._init_fn)

        if self.attention_type == 'divided_space_time':
            i = 0
            for m in self.blocks.sublayers(include_self=True):
                m_str = str(m)
                if 'Block' in m_str:
                    if i > 0:
                        zeros_(m.temporal_fc.weight)
                        zeros_(m.temporal_fc.bias)
                    i += 1
        """Second, if provide pretrained ckpt, load it"""
        if isinstance(
                self.pretrained, str
        ) and self.pretrained.strip() != "":  # load pretrained weights
            load_ckpt(self,
                      self.pretrained,
                      num_patches=self.patch_embed.num_patches,
                      num_seg=self.num_seg,
                      attention_type=self.attention_type)
        elif self.pretrained is None or self.pretrained.strip() == "":
            pass
        else:
            raise NotImplementedError

    def _init_fn(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight)
            if m.bias is not None:
                zeros_(m.bias)
        elif isinstance(m, nn.LayerNorm):
            ones_(m.weight)
            zeros_(m.bias)

    def forward_features(self, x):
        # B = x.shape[0]
        B = paddle.shape(x)[0]
        x, T, W = self.patch_embed(x)  # [BT,nH*nW,F]
        cls_tokens = self.cls_token.expand((B * T, -1, -1))  # [1,1,F]->[BT,1,F]
        x = paddle.concat((cls_tokens, x), axis=1)
        pos_interp = (x.shape[1] != self.pos_embed.shape[1])
        if pos_interp:
            pos_embed = self.pos_embed
            cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)
            other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(
                (0, 2, 1))
            P = int(other_pos_embed.shape[2]**0.5)
            H = x.shape[1] // W
            other_pos_embed = other_pos_embed.reshape([1, x.shape[2], P, P])
            new_pos_embed = F.interpolate(other_pos_embed,
                                          size=(H, W),
                                          mode='nearest')
            new_pos_embed = new_pos_embed.flatten(2)
            new_pos_embed = new_pos_embed.transpose((0, 2, 1))
            new_pos_embed = paddle.concat((cls_pos_embed, new_pos_embed),
                                          axis=1)
            x = x + new_pos_embed
        else:
            x = x + self.pos_embed

        x = self.pos_drop(x)

        # Time Embeddings
        if self.attention_type != 'space_only':
            cls_tokens = x[:B, 0, :].unsqueeze(1) if B > 0 else x.split(
                T)[0].index_select(paddle.to_tensor([0]), axis=1)
            x = x[:, 1:]
            _, _n, _m = x.shape
            _t = T
            x = x.reshape([-1, _t, _n, _m]).transpose(
                (0, 2, 1, 3)).reshape([-1, _t, _m])
            # Resizing time embeddings in case they don't match
            time_interp = (T != self.time_embed.shape[1])
            if time_interp:  # T' != T
                time_embed = self.time_embed.transpose((0, 2, 1)).unsqueeze(0)
                new_time_embed = F.interpolate(time_embed,
                                               size=(T, x.shape[-1]),
                                               mode='nearest').squeeze(0)
                new_time_embed = new_time_embed.transpose((0, 2, 1))
                x = x + new_time_embed
            else:
                x = x + self.time_embed

            x = self.time_drop(x)
            _, _t, _m = x.shape
            x = x.reshape([-1, W * W * T, _m])
            x = paddle.concat((cls_tokens, x), axis=1)

        # Attention blocks
        for blk in self.blocks:
            x = blk(x, B, T, W)

        # Predictions for space-only baseline
        if self.attention_type == 'space_only':
            _, _n, _m = x.shape
            _t = T
            x = x.reshape([-1, _t, _n, _m])
            x = paddle.mean(x, 1)  # averaging predictions for every frame

        x = self.norm(x)
        return x[:, 0]  # [B,  embed_dim]

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