|
import torch |
|
from functools import reduce |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
from functools import partial |
|
|
|
|
|
class FeedForward(nn.Module): |
|
def __init__(self, frame_hidden, mlp_ratio, n_vecs, t2t_params, p): |
|
""" |
|
|
|
Args: |
|
frame_hidden: hidden size of frame features |
|
mlp_ratio: mlp ratio in the middle layer of the transformers |
|
n_vecs: number of vectors in the transformer |
|
t2t_params: dictionary -> {'kernel_size': kernel_size, 'stride': stride, 'padding': padding, 'output_size': output_shape} |
|
p: dropout rate, 0 by default |
|
""" |
|
super(FeedForward, self).__init__() |
|
self.conv = nn.Sequential( |
|
nn.Linear(frame_hidden, frame_hidden * mlp_ratio), |
|
nn.ReLU(inplace=True), |
|
nn.Dropout(p), |
|
nn.Linear(frame_hidden * mlp_ratio, frame_hidden), |
|
nn.Dropout(p) |
|
) |
|
|
|
def forward(self, x, n_vecs=0, output_h=0, output_w=0): |
|
x = self.conv(x) |
|
return x |
|
|
|
|
|
class FusionFeedForward(nn.Module): |
|
def __init__(self, frame_hidden, mlp_ratio, n_vecs, t2t_params, p): |
|
super(FusionFeedForward, self).__init__() |
|
self.kernel_shape = reduce((lambda x, y: x * y), t2t_params['kernel_size']) |
|
self.t2t_params = t2t_params |
|
hidden_size = self.kernel_shape * mlp_ratio |
|
self.conv1 = nn.Linear(frame_hidden, hidden_size) |
|
self.conv2 = nn.Sequential( |
|
nn.ReLU(inplace=True), |
|
nn.Dropout(p), |
|
nn.Linear(hidden_size, frame_hidden), |
|
nn.Dropout(p) |
|
) |
|
assert t2t_params is not None and n_vecs is not None |
|
tp = t2t_params.copy() |
|
self.fold = nn.Fold(**tp) |
|
del tp['output_size'] |
|
self.unfold = nn.Unfold(**tp) |
|
self.n_vecs = n_vecs |
|
|
|
def forward(self, x, n_vecs=0, output_h=0, output_w=0): |
|
x = self.conv1(x) |
|
b, n, c = x.size() |
|
if n_vecs != 0: |
|
normalizer = x.new_ones(b, n, self.kernel_shape).view(-1, n_vecs, self.kernel_shape).permute(0, 2, 1) |
|
x = self.unfold(F.fold(x.view(-1, n_vecs, c).permute(0, 2, 1), output_size=(output_h, output_w), |
|
kernel_size=self.t2t_params['kernel_size'], stride=self.t2t_params['stride'], |
|
padding=self.t2t_params['padding']) / F.fold(normalizer, |
|
output_size=(output_h, output_w), |
|
kernel_size=self.t2t_params[ |
|
'kernel_size'], |
|
stride=self.t2t_params['stride'], |
|
padding=self.t2t_params[ |
|
'padding'])).permute(0, |
|
2, |
|
1).contiguous().view( |
|
b, n, c) |
|
else: |
|
normalizer = x.new_ones(b, n, self.kernel_shape).view(-1, self.n_vecs, self.kernel_shape).permute(0, 2, 1) |
|
x = self.unfold(self.fold(x.view(-1, self.n_vecs, c).permute(0, 2, 1)) / self.fold(normalizer)).permute(0, |
|
2, |
|
1).contiguous().view( |
|
b, n, c) |
|
x = self.conv2(x) |
|
return x |
|
|
|
|
|
class ResidualBlock_noBN(nn.Module): |
|
"""Residual block w/o BN |
|
---Conv-ReLU-Conv-+- |
|
|________________| |
|
""" |
|
|
|
def __init__(self, nf=64): |
|
super(ResidualBlock_noBN, self).__init__() |
|
self.conv1 = nn.Conv2d(nf, nf, kernel_size=3, stride=1, padding=1, bias=True) |
|
self.conv2 = nn.Conv2d(nf, nf, kernel_size=3, stride=1, padding=1, bias=True) |
|
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) |
|
|
|
def forward(self, x): |
|
""" |
|
|
|
Args: |
|
x: with shape of [b, c, t, h, w] |
|
|
|
Returns: processed features with shape [b, c, t, h, w] |
|
|
|
""" |
|
identity = x |
|
out = self.lrelu(self.conv1(x)) |
|
out = self.conv2(out) |
|
out = identity + out |
|
|
|
|
|
return out |
|
|
|
|
|
def make_layer(block, n_layers): |
|
layers = [] |
|
for _ in range(n_layers): |
|
layers.append(block()) |
|
return nn.Sequential(*layers) |
|
|
