From 916cf90d04e57fee23092c966740fbe94fd92cff Mon Sep 17 00:00:00 2001 From: Jordan Gong Date: Tue, 9 Feb 2021 21:21:57 +0800 Subject: Improve performance when disentangling This is a HUGE performance optimization, up to 2x faster than before. Mainly because of the replacement of randomized for-loop with randomized tensor. --- models/auto_encoder.py | 43 +++++++++---- models/model.py | 37 +++++------ models/part_net.py | 17 +++-- models/rgb_part_net.py | 168 ++++++++++++++++--------------------------------- 4 files changed, 108 insertions(+), 157 deletions(-) diff --git a/models/auto_encoder.py b/models/auto_encoder.py index f04ffdb..a9312dd 100644 --- a/models/auto_encoder.py +++ b/models/auto_encoder.py @@ -121,23 +121,42 @@ class AutoEncoder(nn.Module): self.decoder = Decoder(embedding_dims, feature_channels, channels) def forward(self, x_c1_t2, x_c1_t1=None, x_c2_t2=None): + n, t, c, h, w = x_c1_t2.size() # x_c1_t2 is the frame for later module - (f_a_c1_t2, f_c_c1_t2, f_p_c1_t2) = self.encoder(x_c1_t2) + x_c1_t2_ = x_c1_t2.view(n * t, c, h, w) + (f_a_c1_t2_, f_c_c1_t2_, f_p_c1_t2_) = self.encoder(x_c1_t2_) if self.training: # t1 is random time step, c2 is another condition - (f_a_c1_t1, f_c_c1_t1, _) = self.encoder(x_c1_t1) - (_, f_c_c2_t2, f_p_c2_t2) = self.encoder(x_c2_t2) - - x_c1_t2_ = self.decoder(f_a_c1_t1, f_c_c1_t1, f_p_c1_t2) - xrecon_loss_t2 = F.mse_loss(x_c1_t2, x_c1_t2_) - cano_cons_loss_t2 = (F.mse_loss(f_c_c1_t1, f_c_c1_t2) - + F.mse_loss(f_c_c1_t2, f_c_c2_t2)) + x_c1_t1 = x_c1_t1.view(n * t, c, h, w) + (f_a_c1_t1_, f_c_c1_t1_, _) = self.encoder(x_c1_t1) + x_c2_t2 = x_c2_t2.view(n * t, c, h, w) + (_, f_c_c2_t2_, f_p_c2_t2_) = self.encoder(x_c2_t2) + + x_c1_t2_pred_ = self.decoder(f_a_c1_t1_, f_c_c1_t1_, f_p_c1_t2_) + x_c1_t2_pred = x_c1_t2_pred_.view(n, t, c, h, w) + + xrecon_loss = torch.stack([ + F.mse_loss(x_c1_t2[:, i, :, :, :], x_c1_t2_pred[:, i, :, :, :]) + for i in range(t) + ]).sum() + + f_c_c1_t1 = f_c_c1_t1_.view(n, t, -1) + f_c_c1_t2 = f_c_c1_t2_.view(n, t, -1) + f_c_c2_t2 = f_c_c2_t2_.view(n, t, -1) + cano_cons_loss = torch.stack([ + F.mse_loss(f_c_c1_t1[:, i, :], f_c_c1_t2[:, i, :]) + + F.mse_loss(f_c_c1_t2[:, i, :], f_c_c2_t2[:, i, :]) + for i in range(t) + ]).mean() + + f_p_c1_t2 = f_p_c1_t2_.view(n, t, -1) + f_p_c2_t2 = f_p_c2_t2_.view(n, t, -1) + pose_sim_loss = F.mse_loss(f_p_c1_t2.mean(1), f_p_c2_t2.mean(1)) return ( - (f_a_c1_t2, f_c_c1_t2, f_p_c1_t2), - (f_p_c1_t2, f_p_c2_t2), - (xrecon_loss_t2, cano_cons_loss_t2) + (f_a_c1_t2_, f_c_c1_t2_, f_p_c1_t2_), + (xrecon_loss, cano_cons_loss, pose_sim_loss * 10) ) else: # evaluating - return f_c_c1_t2, f_p_c1_t2 + return f_c_c1_t2_, f_p_c1_t2_ diff --git a/models/model.py b/models/model.py index ee07615..70c43b3 100644 --- a/models/model.py +++ b/models/model.py @@ -181,8 +181,9 @@ class Model: # Training start start_time = datetime.now() running_loss = torch.zeros(5, device=self.device) - print(f"{'Iter':^5} {'Loss':^6} {'Xrecon':^8} {'PoseSim':^8}", - f"{'CanoCons':^8} {'BATripH':^8} {'BATripP':^8} {'LRs':^19}") + print(f"{'Time':^8} {'Iter':^5} {'Loss':^6}", + f"{'Xrecon':^8} {'CanoCons':^8} {'PoseSim':^8}", + f"{'BATripH':^8} {'BATripP':^8} {'LRs':^19}") for (batch_c1, batch_c2) in dataloader: self.curr_iter += 1 # Zero the parameter gradients @@ -201,8 +202,8 @@ class Model: # Write losses to TensorBoard self.writer.add_scalar('Loss/all', loss, self.curr_iter) self.writer.add_scalars('Loss/details', dict(zip([ - 'Cross reconstruction loss', 'Pose similarity loss', - 'Canonical consistency loss', 'Batch All triplet loss (HPM)', + 'Cross reconstruction loss', 'Canonical consistency loss', + 'Pose similarity loss', 'Batch All triplet loss (HPM)', 'Batch All triplet loss (PartNet)' ], losses)), self.curr_iter) @@ -217,25 +218,25 @@ class Model: ) # Write disentangled images if self.image_log_on: - (appearance_image, canonical_image, pose_image) = images + i_a, i_c, i_p = images self.writer.add_images( - 'Canonical image', canonical_image, self.curr_iter + 'Canonical image', i_c, self.curr_iter ) - for i in range(self.pr * self.k): + for (i, (o, a, p)) in enumerate(zip(x_c1, i_a, i_p)): self.writer.add_images( - f'Original image/batch {i}', x_c1[i], self.curr_iter + f'Original image/batch {i}', o, self.curr_iter ) self.writer.add_images( - f'Appearance image/batch {i}', - appearance_image[:, i, :, :, :], - self.curr_iter + f'Appearance image/batch {i}', a, self.curr_iter ) self.writer.add_images( - f'Pose image/batch {i}', - pose_image[:, i, :, :, :], - self.curr_iter + f'Pose image/batch {i}', p, self.curr_iter ) - print(f'{self.curr_iter:5d} {running_loss.sum() / 100:6.3f}', + time_used = datetime.now() - start_time + remaining_minute, second = divmod(time_used.seconds, 60) + hour, minute = divmod(remaining_minute, 60) + print(f'{hour:02}:{minute:02}:{second:02}', + f'{self.curr_iter:5d} {running_loss.sum() / 100:6.3f}', '{:f} {:f} {:f} {:f} {:f}'.format(*running_loss / 100), '{:.3e} {:.3e}'.format(lrs[0], lrs[1])) running_loss.zero_() @@ -250,12 +251,6 @@ class Model: 'optim_state_dict': self.optimizer.state_dict(), 'loss': loss, }, self._checkpoint_name) - print(datetime.now() - start_time, 'used') - start_time = datetime.now() - print( - f"{'Iter':^5} {'Loss':^6} {'Xrecon':^8} {'PoseSim':^8}", - f"{'CanoCons':^8} {'BATripH':^8} {'BATripP':^8} {'LRs':^19}" - ) if self.curr_iter == self.total_iter: self.writer.close() diff --git a/models/part_net.py b/models/part_net.py index ac7c434..62a2bac 100644 --- a/models/part_net.py +++ b/models/part_net.py @@ -30,8 +30,8 @@ class FrameLevelPartFeatureExtractor(nn.Module): def forward(self, x): # Flatten frames in all batches - t, n, c, h, w = x.size() - x = x.view(-1, c, h, w) + n, t, c, h, w = x.size() + x = x.view(n * t, c, h, w) for fconv_block in self.fconv_blocks: x = fconv_block(x) @@ -75,8 +75,8 @@ class TemporalFeatureAggregator(nn.Module): for _ in range(self.num_part)]) def forward(self, x): - # p, t, n, c - x = x.permute(0, 2, 3, 1).contiguous() + # p, n, t, c + x = x.transpose(2, 3) p, n, c, t = x.size() feature = x.split(1, dim=0) feature = [f.squeeze(0) for f in feature] @@ -134,19 +134,18 @@ class PartNet(nn.Module): self.max_pool = nn.AdaptiveMaxPool2d(1) def forward(self, x): - t, n, _, _, _ = x.size() - # t, n, c, h, w + n, t, _, _, _ = x.size() x = self.fpfe(x) - # t_n, c, h, w + # n * t x c x h x w # Horizontal Pooling _, c, h, w = x.size() split_size = h // self.num_part x = x.split(split_size, dim=2) x = [self.avg_pool(x_) + self.max_pool(x_) for x_ in x] - x = [x_.view(t, n, c) for x_ in x] + x = [x_.view(n, t, c) for x_ in x] x = torch.stack(x) - # p, t, n, c + # p, n, t, c x = self.tfa(x) return x diff --git a/models/rgb_part_net.py b/models/rgb_part_net.py index 0e7d8b3..8ebcfd3 100644 --- a/models/rgb_part_net.py +++ b/models/rgb_part_net.py @@ -1,8 +1,5 @@ -import random - import torch import torch.nn as nn -import torch.nn.functional as F from models.auto_encoder import AutoEncoder from models.hpm import HorizontalPyramidMatching @@ -59,24 +56,18 @@ class RGBPartNet(nn.Module): return x @ self.fc_mat def forward(self, x_c1, x_c2=None, y=None): - # Step 0: Swap batch_size and time dimensions for next step - # n, t, c, h, w - x_c1 = x_c1.transpose(0, 1) - if self.training: - x_c2 = x_c2.transpose(0, 1) - # Step 1: Disentanglement - # t, n, c, h, w - ((x_c_c1, x_p_c1), images, losses) = self._disentangle(x_c1, x_c2) + # n, t, c, h, w + ((x_c, x_p), losses, images) = self._disentangle(x_c1, x_c2) # Step 2.a: Static Gait Feature Aggregation & HPM # n, c, h, w - x_c = self.hpm(x_c_c1) + x_c = self.hpm(x_c) # p, n, c # Step 2.b: FPFE & TFA (Dynamic Gait Feature Aggregation) - # t, n, c, h, w - x_p = self.pn(x_p_c1) + # n, t, c, h, w + x_p = self.pn(x_p) # p, n, c # Step 3: Cat feature map together and fc @@ -91,113 +82,60 @@ class RGBPartNet(nn.Module): else: return x.unsqueeze(1).view(-1) - def _disentangle(self, x_c1, x_c2=None): - t, n, c, h, w = x_c1.size() - device = x_c1.device + def _disentangle(self, x_c1_t2, x_c2_t2=None): + n, t, c, h, w = x_c1_t2.size() + device = x_c1_t2.device + x_c1_t1 = x_c1_t2[:, torch.randperm(t), :, :, :] if self.training: - # Encoded appearance, canonical and pose features - f_a_c1, f_c_c1, f_p_c1 = [], [], [] - # Features required to calculate losses - f_p_c2 = [] - xrecon_loss, cano_cons_loss = [], [] - for t2 in range(t): - t1 = random.randrange(t) - output = self.ae(x_c1[t2], x_c1[t1], x_c2[t2]) - (f_c1_t2, f_p_t2, losses) = output - - (f_a_c1_t2, f_c_c1_t2, f_p_c1_t2) = f_c1_t2 - if self.image_log_on: - f_a_c1.append(f_a_c1_t2) - # Save canonical features and pose features - f_c_c1.append(f_c_c1_t2) - f_p_c1.append(f_p_c1_t2) - - # Losses per time step - # Used in pose similarity loss - (_, f_p_c2_t2) = f_p_t2 - f_p_c2.append(f_p_c2_t2) - - # Cross reconstruction loss and canonical loss - (xrecon_loss_t2, cano_cons_loss_t2) = losses - xrecon_loss.append(xrecon_loss_t2) - cano_cons_loss.append(cano_cons_loss_t2) - if self.image_log_on: - f_a_c1 = torch.stack(f_a_c1) - f_c_c1_mean = torch.stack(f_c_c1).mean(0) - f_p_c1 = torch.stack(f_p_c1) - f_p_c2 = torch.stack(f_p_c2) - + ((f_a_, f_c_, f_p_), losses) = self.ae(x_c1_t2, x_c1_t1, x_c2_t2) # Decode features - appearance_image, canonical_image, pose_image = None, None, None with torch.no_grad(): - # Decode average canonical features to higher dimension - x_c_c1 = self.ae.decoder( - torch.zeros((n, self.f_a_dim), device=device), - f_c_c1_mean, - torch.zeros((n, self.f_p_dim), device=device), - cano_only=True - ) - # Decode pose features to images - f_p_c1_ = f_p_c1.view(t * n, -1) - x_p_c1_ = self.ae.decoder( - torch.zeros((t * n, self.f_a_dim), device=device), - torch.zeros((t * n, self.f_c_dim), device=device), - f_p_c1_ - ) - x_p_c1 = x_p_c1_.view(t, n, c, h, w) + x_c = self._decode_cano_feature(f_c_, n, t, device) + x_p = self._decode_pose_feature(f_p_, n, t, c, h, w, device) + i_a, i_c, i_p = None, None, None if self.image_log_on: - # Decode appearance features - f_a_c1_ = f_a_c1.view(t * n, -1) - appearance_image_ = self.ae.decoder( - f_a_c1_, - torch.zeros((t * n, self.f_c_dim), device=device), - torch.zeros((t * n, self.f_p_dim), device=device) - ) - appearance_image = appearance_image_.view(t, n, c, h, w) + i_a = self._decode_appr_feature(f_a_, n, t, c, h, w, device) # Continue decoding canonical features - canonical_image = self.ae.decoder.trans_conv3(x_c_c1) - canonical_image = torch.sigmoid( - self.ae.decoder.trans_conv4(canonical_image) - ) - pose_image = x_p_c1 - - # Losses - xrecon_loss = torch.sum(torch.stack(xrecon_loss)) - pose_sim_loss = self._pose_sim_loss(f_p_c1, f_p_c2) * 10 - cano_cons_loss = torch.mean(torch.stack(cano_cons_loss)) + i_c = self.ae.decoder.trans_conv3(x_c) + i_c = torch.sigmoid(self.ae.decoder.trans_conv4(i_c)) + i_p = x_p - return ((x_c_c1, x_p_c1), - (appearance_image, canonical_image, pose_image), - (xrecon_loss, pose_sim_loss, cano_cons_loss)) + return (x_c, x_p), losses, (i_a, i_c, i_p) else: # evaluating - x_c1_ = x_c1.view(t * n, c, h, w) - (f_c_c1_, f_p_c1_) = self.ae(x_c1_) - - # Canonical features - f_c_c1 = f_c_c1_.view(t, n, -1) - f_c_c1_mean = f_c_c1.mean(0) - x_c_c1 = self.ae.decoder( - torch.zeros((n, self.f_a_dim)), - f_c_c1_mean, - torch.zeros((n, self.f_p_dim)), - cano_only=True - ) - - # Pose features - x_p_c1_ = self.ae.decoder( - torch.zeros((t * n, self.f_a_dim)), - torch.zeros((t * n, self.f_c_dim)), - f_p_c1_ - ) - x_p_c1 = x_p_c1_.view(t, n, c, h, w) - - return (x_c_c1, x_p_c1), None, None - - @staticmethod - def _pose_sim_loss(f_p_c1: torch.Tensor, - f_p_c2: torch.Tensor) -> torch.Tensor: - f_p_c1_mean = f_p_c1.mean(dim=0) - f_p_c2_mean = f_p_c2.mean(dim=0) - return F.mse_loss(f_p_c1_mean, f_p_c2_mean) + f_c_, f_p_ = self.ae(x_c1_t2) + x_c = self._decode_cano_feature(f_c_, n, t, device) + x_p = self._decode_pose_feature(f_p_, n, t, c, h, w, device) + return (x_c, x_p), None, None + + def _decode_appr_feature(self, f_a_, n, t, c, h, w, device): + # Decode appearance features + x_a_ = self.ae.decoder( + f_a_, + torch.zeros((n * t, self.f_c_dim), device=device), + torch.zeros((n * t, self.f_p_dim), device=device) + ) + x_a = x_a_.view(n, t, c, h, w) + return x_a + + def _decode_cano_feature(self, f_c_, n, t, device): + # Decode average canonical features to higher dimension + f_c = f_c_.view(n, t, -1) + x_c = self.ae.decoder( + torch.zeros((n, self.f_a_dim), device=device), + f_c.mean(1), + torch.zeros((n, self.f_p_dim), device=device), + cano_only=True + ) + return x_c + + def _decode_pose_feature(self, f_p_, n, t, c, h, w, device): + # Decode pose features to images + x_p_ = self.ae.decoder( + torch.zeros((n * t, self.f_a_dim), device=device), + torch.zeros((n * t, self.f_c_dim), device=device), + f_p_ + ) + x_p = x_p_.view(n, t, c, h, w) + return x_p -- cgit v1.2.3