Update app.py
Browse files
app.py
CHANGED
@@ -85,6 +85,536 @@ parser.add_argument('--data_threads', type=int, default=5, help='number of data
|
|
85 |
opt = parser.parse_args(args=[])
|
86 |
|
87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
def display_gif(file_name, save_name):
|
89 |
images = []
|
90 |
|
@@ -175,6 +705,12 @@ def run(domain_source, action_source, hair_source, top_source, bottom_source, do
|
|
175 |
|
176 |
gif_target = display_gif_pad(file_name_target, 'avatar_target.gif')
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
return 'demo.gif'
|
179 |
|
180 |
|
|
|
85 |
opt = parser.parse_args(args=[])
|
86 |
|
87 |
|
88 |
+
class GradReverse(Function):
|
89 |
+
@staticmethod
|
90 |
+
def forward(ctx, x, beta):
|
91 |
+
ctx.beta = beta
|
92 |
+
return x.view_as(x)
|
93 |
+
|
94 |
+
@staticmethod
|
95 |
+
def backward(ctx, grad_output):
|
96 |
+
grad_input = grad_output.neg() * ctx.beta
|
97 |
+
return grad_input, None
|
98 |
+
|
99 |
+
|
100 |
+
class TransferVAE_Video(nn.Module):
|
101 |
+
|
102 |
+
def __init__(self, opt):
|
103 |
+
super(TransferVAE_Video, self).__init__()
|
104 |
+
self.f_dim = opt.f_dim
|
105 |
+
self.z_dim = opt.z_dim
|
106 |
+
self.fc_dim = opt.fc_dim
|
107 |
+
self.channels = opt.channels
|
108 |
+
self.input_type = opt.input_type
|
109 |
+
self.frames = opt.num_segments
|
110 |
+
self.use_bn = opt.use_bn
|
111 |
+
self.frame_aggregation = opt.frame_aggregation
|
112 |
+
self.batch_size = opt.batch_size
|
113 |
+
self.use_attn = opt.use_attn
|
114 |
+
self.dropout_rate = opt.dropout_rate
|
115 |
+
self.num_class = opt.num_class
|
116 |
+
self.prior_sample = opt.prior_sample
|
117 |
+
|
118 |
+
if self.input_type == 'image':
|
119 |
+
import dcgan_64
|
120 |
+
self.encoder = dcgan_64.encoder(self.fc_dim, self.channels)
|
121 |
+
self.decoder = dcgan_64.decoder_woSkip(self.z_dim + self.f_dim, self.channels)
|
122 |
+
self.fc_output_dim = self.fc_dim
|
123 |
+
elif self.input_type == 'feature':
|
124 |
+
if opt.backbone == 'resnet101':
|
125 |
+
model_backnone = getattr(torchvision.models, opt.backbone)(True) # model_test is only used for getting the dim #
|
126 |
+
self.input_dim = model_backnone.fc.in_features
|
127 |
+
elif opt.backbone == 'I3Dpretrain':
|
128 |
+
self.input_dim = 2048
|
129 |
+
elif opt.backbone == 'I3Dfinetune':
|
130 |
+
self.input_dim = 2048
|
131 |
+
self.add_fc = opt.add_fc
|
132 |
+
self.enc_fc_layer1 = nn.Linear(self.input_dim, self.fc_dim)
|
133 |
+
self.dec_fc_layer1 = nn.Linear(self.fc_dim, self.input_dim)
|
134 |
+
self.fc_output_dim = self.fc_dim
|
135 |
+
|
136 |
+
if self.use_bn == 'shared':
|
137 |
+
self.bn_enc_layer1 = nn.BatchNorm1d(self.fc_output_dim)
|
138 |
+
self.bn_dec_layer1 = nn.BatchNorm1d(self.input_dim)
|
139 |
+
elif self.use_bn == 'separated':
|
140 |
+
self.bn_S_enc_layer1 = nn.BatchNorm1d(self.fc_output_dim)
|
141 |
+
self.bn_T_enc_layer1 = nn.BatchNorm1d(self.fc_output_dim)
|
142 |
+
self.bn_S_dec_layer1 = nn.BatchNorm1d(self.input_dim)
|
143 |
+
self.bn_T_dec_layer1 = nn.BatchNorm1d(self.input_dim)
|
144 |
+
|
145 |
+
if self.add_fc > 1:
|
146 |
+
self.enc_fc_layer2 = nn.Linear(self.fc_dim, self.fc_dim)
|
147 |
+
self.dec_fc_layer2 = nn.Linear(self.fc_dim, self.fc_dim)
|
148 |
+
self.fc_output_dim = self.fc_dim
|
149 |
+
## use batchnormalization or not (if yes whether the source and target share the same batchnormalization)
|
150 |
+
if self.use_bn == 'shared':
|
151 |
+
self.bn_enc_layer2 = nn.BatchNorm1d(self.fc_output_dim)
|
152 |
+
self.bn_dec_layer2 = nn.BatchNorm1d(self.fc_dim)
|
153 |
+
elif self.use_bn == 'separated':
|
154 |
+
self.bn_S_enc_layer2 = nn.BatchNorm1d(self.fc_output_dim)
|
155 |
+
self.bn_T_enc_layer2 = nn.BatchNorm1d(self.fc_output_dim)
|
156 |
+
self.bn_S_dec_layer2 = nn.BatchNorm1d(self.fc_dim)
|
157 |
+
self.bn_T_dec_layer2 = nn.BatchNorm1d(self.fc_dim)
|
158 |
+
|
159 |
+
if self.add_fc > 2:
|
160 |
+
self.enc_fc_layer3 = nn.Linear(self.fc_dim, self.fc_dim)
|
161 |
+
self.dec_fc_layer3 = nn.Linear(self.fc_dim, self.fc_dim)
|
162 |
+
self.fc_output_dim = self.fc_dim
|
163 |
+
## use batchnormalization or not (if yes whether the source and target share the same batchnormalization)
|
164 |
+
if self.use_bn == 'shared':
|
165 |
+
self.bn_enc_layer3 = nn.BatchNorm1d(self.fc_output_dim)
|
166 |
+
self.bn_dec_layer3 = nn.BatchNorm1d(self.fc_dim)
|
167 |
+
elif self.use_bn == 'separated':
|
168 |
+
self.bn_S_enc_layer3 = nn.BatchNorm1d(self.fc_output_dim)
|
169 |
+
self.bn_T_enc_layer3 = nn.BatchNorm1d(self.fc_output_dim)
|
170 |
+
self.bn_S_dec_layer3 = nn.BatchNorm1d(self.fc_dim)
|
171 |
+
self.bn_T_dec_layer3 = nn.BatchNorm1d(self.fc_dim)
|
172 |
+
|
173 |
+
self.z_2_out = nn.Linear(self.z_dim + self.f_dim, self.fc_output_dim)
|
174 |
+
|
175 |
+
|
176 |
+
## nonlinearity and dropout
|
177 |
+
self.relu = nn.LeakyReLU(0.1)
|
178 |
+
self.dropout_f = nn.Dropout(p=self.dropout_rate)
|
179 |
+
self.dropout_v = nn.Dropout(p=self.dropout_rate)
|
180 |
+
# -------------------------------
|
181 |
+
|
182 |
+
## Disentangle strcuture
|
183 |
+
# -------------------------------
|
184 |
+
#self.hidden_dim = opt.rnn_size
|
185 |
+
self.hidden_dim = opt.z_dim
|
186 |
+
self.f_rnn_layers = opt.f_rnn_layers
|
187 |
+
|
188 |
+
# Prior of content is a uniform Gaussian and prior of the dynamics is an LSTM
|
189 |
+
self.z_prior_lstm_ly1 = nn.LSTMCell(self.z_dim, self.hidden_dim)
|
190 |
+
self.z_prior_lstm_ly2 = nn.LSTMCell(self.hidden_dim, self.hidden_dim)
|
191 |
+
|
192 |
+
self.z_prior_mean = nn.Linear(self.hidden_dim, self.z_dim)
|
193 |
+
self.z_prior_logvar = nn.Linear(self.hidden_dim, self.z_dim)
|
194 |
+
|
195 |
+
# POSTERIOR DISTRIBUTION NETWORKS
|
196 |
+
# content and motion features share one lstm
|
197 |
+
self.z_lstm = nn.LSTM(self.fc_output_dim, self.hidden_dim, self.f_rnn_layers, bidirectional=True, batch_first=True)
|
198 |
+
self.f_mean = nn.Linear(self.hidden_dim * 2, self.f_dim)
|
199 |
+
self.f_logvar = nn.Linear(self.hidden_dim * 2, self.f_dim)
|
200 |
+
|
201 |
+
self.z_rnn = nn.RNN(self.hidden_dim * 2, self.hidden_dim, batch_first=True)
|
202 |
+
# Each timestep is for each z so no reshaping and feature mixing
|
203 |
+
self.z_mean = nn.Linear(self.hidden_dim, self.z_dim)
|
204 |
+
self.z_logvar = nn.Linear(self.hidden_dim, self.z_dim)
|
205 |
+
# -------------------------------
|
206 |
+
|
207 |
+
## z_t constraints
|
208 |
+
# -------------------------------
|
209 |
+
## adversarial loss for frame features z_t
|
210 |
+
self.fc_feature_domain_frame = nn.Linear(self.z_dim, self.z_dim)
|
211 |
+
self.fc_classifier_domain_frame = nn.Linear(self.z_dim, 2)
|
212 |
+
|
213 |
+
## #------ aggregate frame-based features (frame feature --> video feature) ------#
|
214 |
+
if self.frame_aggregation == 'rnn':
|
215 |
+
self.bilstm = nn.LSTM(self.z_dim, self.z_dim * 2, self.f_rnn_layers, bidirectional=True, batch_first=True)
|
216 |
+
self.feat_aggregated_dim = self.z_dim * 2
|
217 |
+
elif self.frame_aggregation == 'trn': # 4. TRN (ECCV 2018) ==> fix segment # for both train/val
|
218 |
+
self.num_bottleneck = 256 # 256
|
219 |
+
self.TRN = RelationModuleMultiScale(self.z_dim, self.num_bottleneck, self.frames)
|
220 |
+
self.bn_trn_S = nn.BatchNorm1d(self.num_bottleneck)
|
221 |
+
self.bn_trn_T = nn.BatchNorm1d(self.num_bottleneck)
|
222 |
+
self.feat_aggregated_dim = self.num_bottleneck
|
223 |
+
|
224 |
+
## adversarial loss for video features
|
225 |
+
self.fc_feature_domain_video = nn.Linear(self.feat_aggregated_dim, self.feat_aggregated_dim)
|
226 |
+
self.fc_classifier_domain_video = nn.Linear(self.feat_aggregated_dim, 2)
|
227 |
+
|
228 |
+
## adversarial loss for each relation of features
|
229 |
+
if self.frame_aggregation == 'trn':
|
230 |
+
self.relation_domain_classifier_all = nn.ModuleList()
|
231 |
+
for i in range(self.frames-1):
|
232 |
+
relation_domain_classifier = nn.Sequential(
|
233 |
+
nn.Linear(self.feat_aggregated_dim, self.feat_aggregated_dim),
|
234 |
+
nn.ReLU(),
|
235 |
+
nn.Linear(self.feat_aggregated_dim, 2)
|
236 |
+
)
|
237 |
+
self.relation_domain_classifier_all += [relation_domain_classifier]
|
238 |
+
|
239 |
+
## classifier for action prediction task
|
240 |
+
self.pred_classifier_video = nn.Linear(self.feat_aggregated_dim, self.num_class)
|
241 |
+
|
242 |
+
## classifier for prediction domains
|
243 |
+
self.fc_feature_domain_latent = nn.Linear(self.f_dim, self.f_dim)
|
244 |
+
self.fc_classifier_doamin_latent = nn.Linear(self.f_dim, 2)
|
245 |
+
|
246 |
+
## attention option
|
247 |
+
if self.use_attn == 'general':
|
248 |
+
self.attn_layer = nn.Sequential(
|
249 |
+
nn.Linear(self.feat_aggregated_dim, self.feat_aggregated_dim),
|
250 |
+
nn.Tanh(),
|
251 |
+
nn.Linear(self.feat_aggregated_dim, 1)
|
252 |
+
)
|
253 |
+
|
254 |
+
def domain_classifier_frame(self, feat, beta):
|
255 |
+
feat_fc_domain_frame = GradReverse.apply(feat, beta)
|
256 |
+
feat_fc_domain_frame = self.fc_feature_domain_frame(feat_fc_domain_frame)
|
257 |
+
feat_fc_domain_frame = self.relu(feat_fc_domain_frame)
|
258 |
+
pred_fc_domain_frame = self.fc_classifier_domain_frame(feat_fc_domain_frame)
|
259 |
+
return pred_fc_domain_frame
|
260 |
+
|
261 |
+
def domain_classifier_video(self, feat_video, beta):
|
262 |
+
feat_fc_domain_video = GradReverse.apply(feat_video, beta)
|
263 |
+
feat_fc_domain_video = self.fc_feature_domain_video(feat_fc_domain_video)
|
264 |
+
feat_fc_domain_video = self.relu(feat_fc_domain_video)
|
265 |
+
pred_fc_domain_video = self.fc_classifier_domain_video(feat_fc_domain_video)
|
266 |
+
return pred_fc_domain_video
|
267 |
+
|
268 |
+
def domain_classifier_latent(self, f):
|
269 |
+
feat_fc_domain_latent = self.fc_feature_domain_latent(f)
|
270 |
+
feat_fc_domain_latent = self.relu(feat_fc_domain_latent)
|
271 |
+
pred_fc_domain_latent = self.fc_classifier_doamin_latent(feat_fc_domain_latent)
|
272 |
+
return pred_fc_domain_latent
|
273 |
+
|
274 |
+
def domain_classifier_relation(self, feat_relation, beta):
|
275 |
+
pred_fc_domain_relation_video = None
|
276 |
+
for i in range(len(self.relation_domain_classifier_all)):
|
277 |
+
feat_relation_single = feat_relation[:,i,:].squeeze(1) # 128x1x256 --> 128x256
|
278 |
+
feat_fc_domain_relation_single = GradReverse.apply(feat_relation_single, beta) # the same beta for all relations (for now)
|
279 |
+
|
280 |
+
pred_fc_domain_relation_single = self.relation_domain_classifier_all[i](feat_fc_domain_relation_single)
|
281 |
+
|
282 |
+
if pred_fc_domain_relation_video is None:
|
283 |
+
pred_fc_domain_relation_video = pred_fc_domain_relation_single.view(-1,1,2)
|
284 |
+
else:
|
285 |
+
pred_fc_domain_relation_video = torch.cat((pred_fc_domain_relation_video, pred_fc_domain_relation_single.view(-1,1,2)), 1)
|
286 |
+
|
287 |
+
pred_fc_domain_relation_video = pred_fc_domain_relation_video.view(-1,2)
|
288 |
+
|
289 |
+
return pred_fc_domain_relation_video
|
290 |
+
|
291 |
+
def get_trans_attn(self, pred_domain):
|
292 |
+
softmax = nn.Softmax(dim=1)
|
293 |
+
logsoftmax = nn.LogSoftmax(dim=1)
|
294 |
+
entropy = torch.sum(-softmax(pred_domain) * logsoftmax(pred_domain), 1)
|
295 |
+
weights = 1 - entropy
|
296 |
+
return weights
|
297 |
+
|
298 |
+
def get_general_attn(self, feat):
|
299 |
+
num_segments = feat.size()[1]
|
300 |
+
feat = feat.view(-1, feat.size()[-1]) # reshape features: 128x4x256 --> (128x4)x256
|
301 |
+
weights = self.attn_layer(feat) # e.g. (128x4)x1
|
302 |
+
weights = weights.view(-1, num_segments, weights.size()[-1]) # reshape attention weights: (128x4)x1 --> 128x4x1
|
303 |
+
weights = F.softmax(weights, dim=1) # softmax over segments ==> 128x4x1
|
304 |
+
return weights
|
305 |
+
|
306 |
+
def get_attn_feat_relation(self, feat_fc, pred_domain, num_segments):
|
307 |
+
if self.use_attn == 'TransAttn':
|
308 |
+
weights_attn = self.get_trans_attn(pred_domain)
|
309 |
+
elif self.use_attn == 'general':
|
310 |
+
weights_attn = self.get_general_attn(feat_fc)
|
311 |
+
|
312 |
+
weights_attn = weights_attn.view(-1, num_segments-1, 1).repeat(1,1,feat_fc.size()[-1]) # reshape & repeat weights (e.g. 16 x 4 x 256)
|
313 |
+
feat_fc_attn = (weights_attn+1) * feat_fc
|
314 |
+
|
315 |
+
return feat_fc_attn, weights_attn[:,:,0]
|
316 |
+
|
317 |
+
|
318 |
+
def encode_and_sample_post(self, x):
|
319 |
+
if isinstance(x, list):
|
320 |
+
conv_x = self.encoder_frame(x[0])
|
321 |
+
else:
|
322 |
+
conv_x = self.encoder_frame(x)
|
323 |
+
|
324 |
+
# pass the bidirectional lstm
|
325 |
+
lstm_out, _ = self.z_lstm(conv_x)
|
326 |
+
|
327 |
+
# get f:
|
328 |
+
backward = lstm_out[:, 0, self.hidden_dim:2 * self.hidden_dim]
|
329 |
+
frontal = lstm_out[:, self.frames - 1, 0:self.hidden_dim]
|
330 |
+
lstm_out_f = torch.cat((frontal, backward), dim=1)
|
331 |
+
f_mean = self.f_mean(lstm_out_f)
|
332 |
+
f_logvar = self.f_logvar(lstm_out_f)
|
333 |
+
f_post = self.reparameterize(f_mean, f_logvar, random_sampling=False)
|
334 |
+
|
335 |
+
# pass to one direction rnn
|
336 |
+
features, _ = self.z_rnn(lstm_out)
|
337 |
+
z_mean = self.z_mean(features)
|
338 |
+
z_logvar = self.z_logvar(features)
|
339 |
+
z_post = self.reparameterize(z_mean, z_logvar, random_sampling=False)
|
340 |
+
|
341 |
+
if isinstance(x, list):
|
342 |
+
f_mean_list = [f_mean]
|
343 |
+
f_post_list = [f_post]
|
344 |
+
for t in range(1,3,1):
|
345 |
+
conv_x = self.encoder_frame(x[t])
|
346 |
+
lstm_out, _ = self.z_lstm(conv_x)
|
347 |
+
# get f:
|
348 |
+
backward = lstm_out[:, 0, self.hidden_dim:2 * self.hidden_dim]
|
349 |
+
frontal = lstm_out[:, self.frames - 1, 0:self.hidden_dim]
|
350 |
+
lstm_out_f = torch.cat((frontal, backward), dim=1)
|
351 |
+
f_mean = self.f_mean(lstm_out_f)
|
352 |
+
f_logvar = self.f_logvar(lstm_out_f)
|
353 |
+
f_post = self.reparameterize(f_mean, f_logvar, random_sampling=False)
|
354 |
+
f_mean_list.append(f_mean)
|
355 |
+
f_post_list.append(f_post)
|
356 |
+
f_mean = f_mean_list
|
357 |
+
f_post = f_post_list
|
358 |
+
# f_mean and f_post are list if triple else not
|
359 |
+
return f_mean, f_logvar, f_post, z_mean, z_logvar, z_post
|
360 |
+
|
361 |
+
def decoder_frame(self,zf):
|
362 |
+
if self.input_type == 'image':
|
363 |
+
recon_x = self.decoder(zf)
|
364 |
+
return recon_x
|
365 |
+
|
366 |
+
if self.input_type == 'feature':
|
367 |
+
zf = self.z_2_out(zf) # batch,frames,(z_dim+f_dim) -> batch,frames,fc_output_dim
|
368 |
+
zf = self.relu(zf)
|
369 |
+
|
370 |
+
if self.add_fc > 2:
|
371 |
+
zf = self.dec_fc_layer3(zf)
|
372 |
+
if self.use_bn == 'shared':
|
373 |
+
zf = self.bn_dec_layer3(zf)
|
374 |
+
elif self.use_bn == 'separated':
|
375 |
+
zf_src = self.bn_S_dec_layer3(zf[:self.batchsize,:,:])
|
376 |
+
zf_tar = self.bn_T_dec_layer3(zf[self.batchsize:,:,:])
|
377 |
+
zf = torch.cat([zf_src,zf_tar],axis=0)
|
378 |
+
zf = self.relu(zf)
|
379 |
+
|
380 |
+
if self.add_fc > 1:
|
381 |
+
zf = self.dec_fc_layer2(zf)
|
382 |
+
if self.use_bn == 'shared':
|
383 |
+
zf = self.bn_dec_layer2(zf)
|
384 |
+
elif self.use_bn == 'separated':
|
385 |
+
zf_src = self.bn_S_dec_layer2(zf[:self.batchsize,:,:])
|
386 |
+
zf_tar = self.bn_T_dec_layer2(zf[self.batchsize:,:,:])
|
387 |
+
zf = torch.cat([zf_src,zf_tar],axis=0)
|
388 |
+
zf = self.relu(zf)
|
389 |
+
|
390 |
+
|
391 |
+
zf = self.dec_fc_layer1(zf)
|
392 |
+
if self.use_bn == 'shared':
|
393 |
+
zf = self.bn_dec_layer2(zf)
|
394 |
+
elif self.use_bn == 'separated':
|
395 |
+
zf_src = self.bn_S_dec_layer2(zf[:self.batchsize,:,:])
|
396 |
+
zf_tar = self.bn_T_dec_layer2(zf[self.batchsize:,:,:])
|
397 |
+
zf = torch.cat([zf_src,zf_tar],axis=0)
|
398 |
+
recon_x = self.relu(zf)
|
399 |
+
return recon_x
|
400 |
+
|
401 |
+
def encoder_frame(self, x):
|
402 |
+
if self.input_type == 'image':
|
403 |
+
# input x is list of length Frames [batchsize, channels, size, size]
|
404 |
+
# convert it to [batchsize, frames, channels, size, size]
|
405 |
+
# [batch_size, frames, channels, size, size] to [batch_size * frames, channels, size, size]
|
406 |
+
x_shape = x.shape
|
407 |
+
x = x.view(-1, x_shape[-3], x_shape[-2], x_shape[-1])
|
408 |
+
x_embed = self.encoder(x)[0]
|
409 |
+
# to [batch_size,frames,embed_dim]
|
410 |
+
|
411 |
+
return x_embed.view(x_shape[0], x_shape[1], -1)
|
412 |
+
|
413 |
+
|
414 |
+
if self.input_type == 'feature':
|
415 |
+
# input is [batchsize, framew, input_dim]
|
416 |
+
x_embed = self.enc_fc_layer1(x)
|
417 |
+
## use batchnormalization or not (if yes whether the source and target share the same batchnormalization)
|
418 |
+
if self.use_bn == 'shared':
|
419 |
+
x_embed = self.bn_enc_layer1(x_embed)
|
420 |
+
elif self.use_bn == 'separated':
|
421 |
+
x_embed_src = self.bn_S_enc_layer1(x_embed[:self.batchsize,:,:])
|
422 |
+
x_embed_tar = self.bn_T_enc_layer1(x_embed[self.batchsize:,:,:])
|
423 |
+
x_embed = torch.cat([x_embed_src,x_embed_tar],axis=0)
|
424 |
+
x_embed = self.relu(x_embed)
|
425 |
+
|
426 |
+
if self.add_fc > 1:
|
427 |
+
x_embed = self.enc_fc_layer2(x_embed)
|
428 |
+
if self.use_bn == 'shared':
|
429 |
+
x_embed = self.bn_enc_layer2(x_embed)
|
430 |
+
elif self.use_bn == 'separated':
|
431 |
+
x_embed_src = self.bn_S_enc_layer2(x_embed[:self.batchsize,:,:])
|
432 |
+
x_embed_tar = self.bn_T_enc_layer2(x_embed[self.batchsize:,:,:])
|
433 |
+
x_embed = torch.cat([x_embed_src,x_embed_tar],axis=0)
|
434 |
+
x_embed = self.relu(x_embed)
|
435 |
+
|
436 |
+
if self.add_fc > 2:
|
437 |
+
x_embed = self.enc_fc_layer3(x_embed)
|
438 |
+
if self.use_bn == 'shared':
|
439 |
+
x_embed = self.bn_enc_layer3(x_embed)
|
440 |
+
elif self.use_bn == 'separated':
|
441 |
+
x_embed_src = self.bn_S_enc_layer3(x_embed[:self.batchsize,:,:])
|
442 |
+
x_embed_tar = self.bn_T_enc_layer3(x_embed[self.batchsize:,:,:])
|
443 |
+
x_embed = torch.cat([x_embed_src,x_embed_tar],axis=0)
|
444 |
+
x_embed = self.relu(x_embed)
|
445 |
+
|
446 |
+
## [batchsize, frame, output_dim]
|
447 |
+
return x_embed
|
448 |
+
|
449 |
+
|
450 |
+
def reparameterize(self, mean, logvar, random_sampling=True):
|
451 |
+
# Reparametrization occurs only if random sampling is set to true, otherwise mean is returned
|
452 |
+
if random_sampling is True:
|
453 |
+
eps = torch.randn_like(logvar)
|
454 |
+
std = torch.exp(0.5 * logvar)
|
455 |
+
z = mean + eps * std
|
456 |
+
return z
|
457 |
+
else:
|
458 |
+
return mean
|
459 |
+
|
460 |
+
def sample_z_prior_train(self, z_post, random_sampling=True):
|
461 |
+
z_out = None # This will ultimately store all z_s in the format [batch_size, frames, z_dim]
|
462 |
+
z_means = None
|
463 |
+
z_logvars = None
|
464 |
+
batch_size = z_post.shape[0]
|
465 |
+
|
466 |
+
z_t = torch.zeros(batch_size, self.z_dim).cpu()
|
467 |
+
h_t_ly1 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
468 |
+
c_t_ly1 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
469 |
+
h_t_ly2 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
470 |
+
c_t_ly2 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
471 |
+
|
472 |
+
for i in range(self.frames):
|
473 |
+
# two layer LSTM and two one-layer FC
|
474 |
+
h_t_ly1, c_t_ly1 = self.z_prior_lstm_ly1(z_t, (h_t_ly1, c_t_ly1))
|
475 |
+
h_t_ly2, c_t_ly2 = self.z_prior_lstm_ly2(h_t_ly1, (h_t_ly2, c_t_ly2))
|
476 |
+
|
477 |
+
z_mean_t = self.z_prior_mean(h_t_ly2)
|
478 |
+
z_logvar_t = self.z_prior_logvar(h_t_ly2)
|
479 |
+
z_prior = self.reparameterize(z_mean_t, z_logvar_t, random_sampling)
|
480 |
+
if z_out is None:
|
481 |
+
# If z_out is none it means z_t is z_1, hence store it in the format [batch_size, 1, z_dim]
|
482 |
+
z_out = z_prior.unsqueeze(1)
|
483 |
+
z_means = z_mean_t.unsqueeze(1)
|
484 |
+
z_logvars = z_logvar_t.unsqueeze(1)
|
485 |
+
else:
|
486 |
+
# If z_out is not none, z_t is not the initial z and hence append it to the previous z_ts collected in z_out
|
487 |
+
z_out = torch.cat((z_out, z_prior.unsqueeze(1)), dim=1)
|
488 |
+
z_means = torch.cat((z_means, z_mean_t.unsqueeze(1)), dim=1)
|
489 |
+
z_logvars = torch.cat((z_logvars, z_logvar_t.unsqueeze(1)), dim=1)
|
490 |
+
z_t = z_post[:,i,:]
|
491 |
+
return z_means, z_logvars, z_out
|
492 |
+
|
493 |
+
# If random sampling is true, reparametrization occurs else z_t is just set to the mean
|
494 |
+
def sample_z(self, batch_size, random_sampling=True):
|
495 |
+
z_out = None # This will ultimately store all z_s in the format [batch_size, frames, z_dim]
|
496 |
+
z_means = None
|
497 |
+
z_logvars = None
|
498 |
+
|
499 |
+
# All states are initially set to 0, especially z_0 = 0
|
500 |
+
z_t = torch.zeros(batch_size, self.z_dim).cpu()
|
501 |
+
# z_mean_t = torch.zeros(batch_size, self.z_dim)
|
502 |
+
# z_logvar_t = torch.zeros(batch_size, self.z_dim)
|
503 |
+
h_t_ly1 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
504 |
+
c_t_ly1 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
505 |
+
h_t_ly2 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
506 |
+
c_t_ly2 = torch.zeros(batch_size, self.hidden_dim).cpu()
|
507 |
+
for _ in range(self.frames):
|
508 |
+
# h_t, c_t = self.z_prior_lstm(z_t, (h_t, c_t))
|
509 |
+
# two layer LSTM and two one-layer FC
|
510 |
+
h_t_ly1, c_t_ly1 = self.z_prior_lstm_ly1(z_t, (h_t_ly1, c_t_ly1))
|
511 |
+
h_t_ly2, c_t_ly2 = self.z_prior_lstm_ly2(h_t_ly1, (h_t_ly2, c_t_ly2))
|
512 |
+
|
513 |
+
z_mean_t = self.z_prior_mean(h_t_ly2)
|
514 |
+
z_logvar_t = self.z_prior_logvar(h_t_ly2)
|
515 |
+
z_t = self.reparameterize(z_mean_t, z_logvar_t, random_sampling)
|
516 |
+
if z_out is None:
|
517 |
+
# If z_out is none it means z_t is z_1, hence store it in the format [batch_size, 1, z_dim]
|
518 |
+
z_out = z_t.unsqueeze(1)
|
519 |
+
z_means = z_mean_t.unsqueeze(1)
|
520 |
+
z_logvars = z_logvar_t.unsqueeze(1)
|
521 |
+
else:
|
522 |
+
# If z_out is not none, z_t is not the initial z and hence append it to the previous z_ts collected in z_out
|
523 |
+
z_out = torch.cat((z_out, z_t.unsqueeze(1)), dim=1)
|
524 |
+
z_means = torch.cat((z_means, z_mean_t.unsqueeze(1)), dim=1)
|
525 |
+
z_logvars = torch.cat((z_logvars, z_logvar_t.unsqueeze(1)), dim=1)
|
526 |
+
return z_means, z_logvars, z_out
|
527 |
+
|
528 |
+
def forward(self, x, beta):
|
529 |
+
# beta [beta_relation, beta_video, beta_frame]
|
530 |
+
f_mean, f_logvar, f_post, z_mean_post, z_logvar_post, z_post = self.encode_and_sample_post(x)
|
531 |
+
if self.prior_sample == 'random':
|
532 |
+
z_mean_prior, z_logvar_prior, z_prior = self.sample_z(z_post.size(0),random_sampling=False)
|
533 |
+
elif self.prior_sample == 'post':
|
534 |
+
z_mean_prior, z_logvar_prior, z_prior = self.sample_z_prior_train(z_post, random_sampling=False)
|
535 |
+
|
536 |
+
|
537 |
+
if isinstance(f_post, list):
|
538 |
+
f_expand = f_post[0].unsqueeze(1).expand(-1, self.frames, self.f_dim)
|
539 |
+
else:
|
540 |
+
f_expand = f_post.unsqueeze(1).expand(-1, self.frames, self.f_dim)
|
541 |
+
zf = torch.cat((z_post, f_expand), dim=2) # batch,frames,(z_dim+f_dim)
|
542 |
+
|
543 |
+
## reconcstruct x
|
544 |
+
recon_x = self.decoder_frame(zf)
|
545 |
+
|
546 |
+
## For constraints on z_post [batch,frame,z_dim] and f_post [batch,f_dim]
|
547 |
+
pred_domain_all = [] # list save domain predictions (1) z_post (frame level) (2) each z_post_relation (if trn) (3) z_post (video level) (4)f_post
|
548 |
+
|
549 |
+
#1. adversarial on z_post (frame level)
|
550 |
+
z_post_feat = z_post.view(-1, z_post.size()[-1]) # e.g. 32 x 5 x 2048 --> 160 x 2048
|
551 |
+
z_post_feat = self.dropout_f(z_post_feat)
|
552 |
+
pred_fc_domain_frame = self.domain_classifier_frame(z_post_feat, beta[2])
|
553 |
+
pred_fc_domain_frame = pred_fc_domain_frame.view((z_post.size(0), self.frames) + pred_fc_domain_frame.size()[-1:])
|
554 |
+
pred_domain_all.append(pred_fc_domain_frame)
|
555 |
+
|
556 |
+
#2 adversarial on z_post (video level, relation level if trn is used)
|
557 |
+
|
558 |
+
if self.frame_aggregation == 'rnn':
|
559 |
+
self.bilstm.flatten_parameters()
|
560 |
+
z_post_video_feat, _ = self.bilstm(z_post)
|
561 |
+
backward = z_post_video_feat[:, 0, self.z_dim:2 * self.z_dim]
|
562 |
+
frontal = z_post_video_feat[:, self.frames - 1, 0:self.z_dim]
|
563 |
+
z_post_video_feat = torch.cat((frontal, backward), dim=1)
|
564 |
+
pred_fc_domain_relation = []
|
565 |
+
pred_domain_all.append(pred_fc_domain_relation)
|
566 |
+
|
567 |
+
elif self.frame_aggregation == 'trn':
|
568 |
+
z_post_video_relation = self.TRN(z_post) ## [batch, frame-1, self.feat_aggregated_dim]
|
569 |
+
|
570 |
+
# adversarial branch for each relation
|
571 |
+
pred_fc_domain_relation = self.domain_classifier_relation(z_post_video_relation, beta[0])
|
572 |
+
pred_domain_all.append(pred_fc_domain_relation.view((z_post.size(0), z_post_video_relation.size()[1]) + pred_fc_domain_relation.size()[-1:]))
|
573 |
+
|
574 |
+
# transferable attention
|
575 |
+
if self.use_attn != 'none': # get the attention weighting
|
576 |
+
z_post_video_relation_attn, _ = self.get_attn_feat_relation(z_post_video_relation, pred_fc_domain_relation, self.frames)
|
577 |
+
|
578 |
+
# sum up relation features (ignore 1-relation)
|
579 |
+
z_post_video_feat = torch.sum(z_post_video_relation_attn, 1)
|
580 |
+
|
581 |
+
|
582 |
+
z_post_video_feat = self.dropout_v(z_post_video_feat)
|
583 |
+
|
584 |
+
pred_fc_domain_video = self.domain_classifier_video(z_post_video_feat, beta[1])
|
585 |
+
pred_fc_domain_video = pred_fc_domain_video.view((z_post.size(0),) + pred_fc_domain_video.size()[-1:])
|
586 |
+
pred_domain_all.append(pred_fc_domain_video)
|
587 |
+
|
588 |
+
|
589 |
+
#3. video prediction
|
590 |
+
pred_video_class = self.pred_classifier_video(z_post_video_feat)
|
591 |
+
|
592 |
+
#4. domain prediction on f
|
593 |
+
if isinstance(f_post, list):
|
594 |
+
pred_fc_domain_latent = self.domain_classifier_latent(f_post[0])
|
595 |
+
else:
|
596 |
+
pred_fc_domain_latent = self.domain_classifier_latent(f_post)
|
597 |
+
pred_domain_all.append(pred_fc_domain_latent)
|
598 |
+
|
599 |
+
return f_mean, f_logvar, f_post, z_mean_post, z_logvar_post, z_post, z_mean_prior, z_logvar_prior, z_prior, recon_x, pred_domain_all, pred_video_class
|
600 |
+
|
601 |
+
|
602 |
+
def name2seq(file_name):
|
603 |
+
images = []
|
604 |
+
|
605 |
+
for frame in range(8):
|
606 |
+
frame_name = '%d' % (frame)
|
607 |
+
image_filename = file_name + frame_name + '.png'
|
608 |
+
image = imageio.imread(image_filename)
|
609 |
+
images.append(image[:, :, :3])
|
610 |
+
|
611 |
+
images = np.asarray(images, dtype='f') / 256.0
|
612 |
+
images = images.transpose((0, 3, 1, 2))
|
613 |
+
print(images.shape)
|
614 |
+
images = torch.Tensor(images).unsqueeze(dim=0)
|
615 |
+
return images
|
616 |
+
|
617 |
+
|
618 |
def display_gif(file_name, save_name):
|
619 |
images = []
|
620 |
|
|
|
705 |
|
706 |
gif_target = display_gif_pad(file_name_target, 'avatar_target.gif')
|
707 |
|
708 |
+
|
709 |
+
# == Load Model ==
|
710 |
+
model = TransferVAE_Video(opt)
|
711 |
+
model.load_state_dict(torch.load('TransferVAE.pth.tar', map_location=torch.device('cpu'))['state_dict'])
|
712 |
+
model.eval()
|
713 |
+
|
714 |
return 'demo.gif'
|
715 |
|
716 |
|