EE292D_Final_Project/model.py at master · ncomly/EE292D_Final_Project · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
# ported to TF from https://github.com/midas-research/mobile-vsr/blob/master/model.py
import math
import numpy as np

import tensorflow as tf
#import keras
from depthwise import *
from tensorflow.keras.layers import Input, Dense, ReLU, Flatten, Permute
from tensorflow.keras.layers import Conv1D, Conv2D, Conv3D, ZeroPadding3D
from tensorflow.keras.layers import BatchNormalization, AveragePooling2D, MaxPool1D
from tensorflow.keras.models import Model
from DepthwiseConv3D import DepthwiseConv3D

def LipRes(alpha=2, reduction=1, num_classes=256):
    block = lambda in_planes, planes, stride: \
        LipResBlock(in_planes, planes, stride, reduction=reduction)

    return ResNet(block, [alpha, alpha, alpha, alpha], reduction, num_classes) # TODO tunable alpha param + # alpha blocks

class ResNet(tf.keras.Model):
    def __init__(self, block, num_blocks, reduction=1, num_classes=256):
        super(ResNet, self).__init__()

        self.reduction = float(reduction) ** 0.5
        self.num_classes = num_classes
        self.in_planes = 64 #int(16 / self.reduction)

        self.layer1   = self._make_layer(block, self.in_planes, num_blocks[0], stride=1)
        self.layer2   = self._make_layer(block, 128, num_blocks[1], stride=2)
        self.layer3   = self._make_layer(block, 256, num_blocks[2], stride=2)
        self.layer4   = self._make_layer(block, 512, num_blocks[3], stride=2)
        self.flatten  = Flatten()
        self.fc       = Dense(num_classes)
        self.bnfc     = BatchNormalization(momentum=0.1, epsilon=1e-5)
        self.avgpool = AveragePooling2D()

        # TODO: weight initialization port -> done in depthwise.py
        # for m in self.modules():
        #     if isinstance(m, Conv2d):
        #         torch.nn.init.kaiming_uniform_(m.weight)
        #     elif isinstance(m, BatchNormalization):
        #         m.weight.data.fill_(1)
        #         m.bias.data.zero_()


    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        planes = int(planes)
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes
        return Sequential(layers)


    def call(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        # 464 512 1 1
        x = self.flatten(x)
        # 464 512
        # x = self.inp(x)
        x = self.fc(x)
        x = self.bnfc(x)
        return x

class LipNext(tf.keras.Model):
    def __init__(self, inputDim=256, hiddenDim=512, nClasses=500, frameLen=29, alpha=2):
        super(LipNext, self).__init__()

#        initializer = tf.initializers.VarianceScaling(scale=2.0) # added for initialization
        initializer = 'glorot_uniform'

        self.inputDim = inputDim
        self.hiddenDim = hiddenDim
        self.nClasses = nClasses
        self.frameLen = frameLen
        self.nLayers = 2
        self.alpha = alpha
        # frontend3D
        #self.frontend3D = Sequential ( [
        #        ZeroPadding3D(padding=(1,1,1),), #input_shape=(1,29,96,96)), # double check channel placement
        #        Conv3D(64, kernel_size=(3,3,3), strides=(1,2,2), use_bias=False, kernel_initializer=initializer, padding='valid'),
        #        BatchNormalization(momentum=.1, epsilon=1e-5), # should this be .9 instead?
        #        ReLU(), # check in place?
        #        # group convolution - TODO: THIS IS NOT RIGHT
        #        ZeroPadding3D(padding=(1,1,1)),
        #        Conv3D(64, kernel_size=(3,3,3), strides=(1,2,2), use_bias=False, kernel_initializer=initializer, padding='valid'),
        #        ZeroPadding3D(padding=(1,0,0)), # double check channel placement
        #        Conv3D(64, kernel_size=(3,1,1), strides=(1,1,1), use_bias=False, kernel_initializer=initializer, padding='valid')
        #    ] )
        self.frontend3D = Sequential ( [
                ZeroPadding3D(padding=(1,1,1),), #input_shape=(1,29,96,96)), # double check channel placement
                Conv3D(64, kernel_size=(3,3,3), strides=(1,2,2), use_bias=False, kernel_initializer=initializer, padding='valid'),
                BatchNormalization(momentum=.1, epsilon=1e-5), # should this be .9 instead?
                ReLU(), # check in place?
                # group convolution - TODO: THIS IS NOT RIGHT
                ZeroPadding3D(padding=(1,1,1)),])
        #        Conv3D(64, kernel_size=(3,3,3), strides=(1,2,2), use_bias=False, kernel_initializer=initializer, padding='valid'),
        #        ZeroPadding3D(padding=(1,0,0)), # double check channel placement
        #        Conv3D(64, kernel_size=(3,1,1), strides=(1,1,1), use_bias=False, kernel_initializer=initializer, padding='valid')
        #    ] )

        self.perm1 = Permute((4,1,2,3))
        self.DConv3D = DepthwiseConv3D(kernel_size=(3,3,3), depth_multiplier=1, strides=(1,2,2), use_bias=False, data_format='channels_last')
        self.perm2 = Permute((2,3,4,1))
        self.pad1 = ZeroPadding3D(padding=(1,0,0)) # double check channel placement
        self.front_conv3d = Conv3D(64, kernel_size=(3,1,1), strides=(1,1,1), use_bias=False, kernel_initializer=initializer, padding='valid')
        # resnet
        self.permute1 = Permute((1,4,2,3))
        self.resnet34 = LipRes(self.alpha)
        # backend
        self.backend_conv1 = Sequential ( [
                Conv1D(2*self.inputDim, kernel_size=5, strides=2, use_bias=False, kernel_initializer=initializer),
                BatchNormalization(momentum=0.1, epsilon=1e-5),
                ReLU(),
                MaxPool1D(2,2),
                Conv1D(4*self.inputDim, kernel_size=5, strides=2, use_bias=False, kernel_initializer=initializer),
                BatchNormalization(momentum=0.1, epsilon=1e-5),
                ReLU()
            ] )
        self.permute2 = Permute((2,1))
        self.backend_conv2 = Sequential ( [
                Dense(self.inputDim, input_shape=(4 * self.inputDim, )),
                BatchNormalization(momentum=0.1, epsilon=1e-5),
                ReLU(),
                Dense(self.nClasses)
            ] )

        # now ignored due to initializer
        # self._initialize_weights()

    def call(self, x):
        # Shape: None, 29, 88, 88, 1

        x = self.frontend3D(x)
        #print(f'frontend3D {x.shape}')
        # Shape: None, 29, 22, 22, 64
        # depthwise conv3d
        #x = self.perm1(x)
        #print(x)
        #print(x.shape)
        x = self.DConv3D(x)
        #print(x.shape)
        #print(x)
        #x = self.perm2(x)
        x = self.pad1(x)
        x = self.front_conv3d(x)
        #print(x.shape)
        #print(f'post front {x.shape}')
        # 29, 22, 22, 64

        x = tf.reshape(x, [-1,  x.shape[2], x.shape[3], 64])
        # Shape: None, 22, 22, 64

        x = self.resnet34(x)
        # Shape: None, 256

        x = tf.reshape(x, [-1, self.frameLen, self.inputDim])
        # Shape: None, 29, 256

        x = self.backend_conv1(x)
        # Shape: None, 1, 1024

        x = tf.math.reduce_mean(x, axis=1)
        # Shape: None, 1024

        x = self.backend_conv2(x)
        # Shape: None, nClasses

        return x

    def _initialize_weights(self):
        raise NotImplementedError

def lipnext(inputDim=256, hiddenDim=512, nClasses=500, frameLen=29, alpha=2):
    model = LipNext(inputDim=inputDim, hiddenDim=hiddenDim, nClasses=nClasses, frameLen=frameLen, alpha=alpha)
    return model