Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,085 Bytes
aa5ee46 |
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 |
import torch
import torch.nn as nn
from sync_models.modules import *
class Transformer_RGB(nn.Module):
def __init__(self):
super().__init__()
self.net_vid = self.build_net_vid()
self.ff_vid = nn.Sequential(
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, 1024)
)
self.pos_encoder = PositionalEncoding_RGB(d_model=512)
encoder_layer = nn.TransformerEncoderLayer(d_model=512, nhead=8, batch_first=True)
self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=6)
self.net_aud = self.build_net_aud()
self.lstm = nn.LSTM(512, 256, num_layers=1, bidirectional=True, batch_first=True)
self.ff_aud = NetFC_2D(input_dim=512, hidden_dim=512, embed_dim=1024)
self.logits_scale = nn.Linear(1, 1, bias=False)
torch.nn.init.ones_(self.logits_scale.weight)
self.fc = nn.Linear(1,1)
def build_net_vid(self):
layers = [
{
'type': 'conv3d',
'n_channels': 64,
'kernel_size': (5, 7, 7),
'stride': (1, 3, 3),
'padding': (0),
'maxpool': {
'kernel_size': (1, 3, 3),
'stride': (1, 2, 2)
}
},
{
'type': 'conv3d',
'n_channels': 128,
'kernel_size': (1, 5, 5),
'stride': (1, 2, 2),
'padding': (0, 0, 0),
},
{
'type': 'conv3d',
'n_channels': 256,
'kernel_size': (1, 3, 3),
'stride': (1, 2, 2),
'padding': (0, 1, 1),
},
{
'type': 'conv3d',
'n_channels': 256,
'kernel_size': (1, 3, 3),
'stride': (1, 1, 2),
'padding': (0, 1, 1),
},
{
'type': 'conv3d',
'n_channels': 256,
'kernel_size': (1, 3, 3),
'stride': (1, 1, 1),
'padding': (0, 1, 1),
'maxpool': {
'kernel_size': (1, 3, 3),
'stride': (1, 2, 2)
}
},
{
'type': 'fc3d',
'n_channels': 512,
'kernel_size': (1, 4, 4),
'stride': (1, 1, 1),
'padding': (0),
},
]
return VGGNet(n_channels_in=3, layers=layers)
def build_net_aud(self):
layers = [
{
'type': 'conv2d',
'n_channels': 64,
'kernel_size': (3, 3),
'stride': (2, 2),
'padding': (1, 1),
'maxpool': {
'kernel_size': (3, 3),
'stride': (2, 2)
}
},
{
'type': 'conv2d',
'n_channels': 192,
'kernel_size': (3, 3),
'stride': (1, 2),
'padding': (1, 1),
'maxpool': {
'kernel_size': (3, 3),
'stride': (2, 2)
}
},
{
'type': 'conv2d',
'n_channels': 384,
'kernel_size': (3, 3),
'stride': (1, 1),
'padding': (1, 1),
},
{
'type': 'conv2d',
'n_channels': 256,
'kernel_size': (3, 3),
'stride': (1, 1),
'padding': (1, 1),
},
{
'type': 'conv2d',
'n_channels': 256,
'kernel_size': (3, 3),
'stride': (1, 1),
'padding': (1, 1),
'maxpool': {
'kernel_size': (2, 3),
'stride': (2, 2)
}
},
{
'type': 'fc2d',
'n_channels': 512,
'kernel_size': (4, 2),
'stride': (1, 1),
'padding': (0, 0),
},
]
return VGGNet(n_channels_in=1, layers=layers)
def forward_vid(self, x, return_feats=False):
out_conv = self.net_vid(x).squeeze(-1).squeeze(-1)
# print("Conv: ", out_conv.shape) # Bx1024x21x1x1
out = self.pos_encoder(out_conv.transpose(1,2))
out_trans = self.transformer_encoder(out)
# print("Transformer: ", out_trans.shape) # Bx21x1024
out = self.ff_vid(out_trans).transpose(1,2)
# print("MLP output: ", out.shape) # Bx1024
if return_feats:
return out, out_conv
else:
return out
def forward_aud(self, x):
out = self.net_aud(x)
out = self.ff_aud(out)
out = out.squeeze(-1)
return out
|