diff options
Diffstat (limited to 'models/model.py')
| -rw-r--r-- | models/model.py | 58 |
1 files changed, 44 insertions, 14 deletions
diff --git a/models/model.py b/models/model.py index d11617b..912d0b9 100644 --- a/models/model.py +++ b/models/model.py @@ -51,7 +51,6 @@ class Model: self.total_iters = self.meta.get('total_iters', (80000, 80000, 80000)) self.is_train: bool = True - self.train_size: int = 74 self.in_channels: int = 3 self.pr: Optional[int] = None self.k: Optional[int] = None @@ -67,6 +66,7 @@ class Model: self.optimizer: Optional[optim.Adam] = None self.scheduler: Optional[optim.lr_scheduler.StepLR] = None self.writer: Optional[SummaryWriter] = None + self.image_log_on = system_config.get('image_log_on', False) self.CASIAB_GALLERY_SELECTOR = { 'selector': {'conditions': ClipConditions({r'nm-0[1-4]'})} @@ -138,19 +138,18 @@ class Model: # Prepare for model, optimizer and scheduler model_hp = self.hp.get('model', {}) optim_hp: Dict = self.hp.get('optimizer', {}).copy() + start_iter = optim_hp.pop('start_iter', 0) ae_optim_hp = optim_hp.pop('auto_encoder', {}) pn_optim_hp = optim_hp.pop('part_net', {}) hpm_optim_hp = optim_hp.pop('hpm', {}) fc_optim_hp = optim_hp.pop('fc', {}) sched_hp = self.hp.get('scheduler', {}) - self.rgb_pn = RGBPartNet(self.train_size, self.in_channels, **model_hp) + self.rgb_pn = RGBPartNet(self.in_channels, **model_hp, + image_log_on=self.image_log_on) # Try to accelerate computation using CUDA or others self.rgb_pn = self.rgb_pn.to(self.device) self.optimizer = optim.Adam([ {'params': self.rgb_pn.ae.parameters(), **ae_optim_hp}, - {'params': self.rgb_pn.pn.parameters(), **pn_optim_hp}, - {'params': self.rgb_pn.hpm.parameters(), **hpm_optim_hp}, - {'params': self.rgb_pn.fc_mat, **fc_optim_hp}, ], **optim_hp) self.scheduler = optim.lr_scheduler.StepLR(self.optimizer, **sched_hp) self.writer = SummaryWriter(self._log_name) @@ -168,10 +167,20 @@ class Model: # Training start start_time = datetime.now() - running_loss = torch.zeros(4).to(self.device) + running_loss = torch.zeros(5, device=self.device) print(f"{'Iter':^5} {'Loss':^6} {'Xrecon':^8} {'PoseSim':^8}", - f"{'CanoCons':^8} {'BATrip':^8} {'LR':^9}") + f"{'CanoCons':^8} {'BATripH':^8} {'BATripP':^8} LR(s)") for (batch_c1, batch_c2) in dataloader: + if self.curr_iter == start_iter: + self.optimizer.add_param_group( + {'params': self.rgb_pn.pn.parameters(), **pn_optim_hp} + ) + self.optimizer.add_param_group( + {'params': self.rgb_pn.hpm.parameters(), **hpm_optim_hp} + ) + self.optimizer.add_param_group( + {'params': self.rgb_pn.fc_mat, **fc_optim_hp} + ) self.curr_iter += 1 # Zero the parameter gradients self.optimizer.zero_grad() @@ -179,12 +188,10 @@ class Model: x_c1 = batch_c1['clip'].to(self.device) x_c2 = batch_c2['clip'].to(self.device) y = batch_c1['label'].to(self.device) - losses = self.rgb_pn(x_c1, x_c2, y) + losses, images = self.rgb_pn(x_c1, x_c2, y) loss = losses.sum() loss.backward() self.optimizer.step() - # Step scheduler - self.scheduler.step() # Statistics and checkpoint running_loss += losses.detach() @@ -192,15 +199,39 @@ class Model: 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' + 'Canonical consistency loss', 'Batch All triplet loss (HPM)', + 'Batch All triplet loss (PartNet)' ], losses)), self.curr_iter) + if self.image_log_on: + (appearance_image, canonical_image, pose_image) = images + self.writer.add_images( + 'Canonical image', canonical_image, self.curr_iter + ) + for i in range(self.pr * self.k): + self.writer.add_images( + f'Original image/batch {i}', x_c1[i], self.curr_iter + ) + self.writer.add_images( + f'Appearance image/batch {i}', + appearance_image[:, i, :, :, :], + self.curr_iter + ) + self.writer.add_images( + f'Pose image/batch {i}', + pose_image[:, i, :, :, :], + self.curr_iter + ) if self.curr_iter % 100 == 0: + lrs = self.scheduler.get_last_lr() print(f'{self.curr_iter:5d} {running_loss.sum() / 100:6.3f}', - '{:f} {:f} {:f} {:f}'.format(*running_loss / 100), - f'{self.scheduler.get_last_lr()[0]:.3e}') + '{:f} {:f} {:f} {:f} {:f}'.format(*running_loss / 100), + ' '.join(('{:.3e}'.format(lr) for lr in lrs))) running_loss.zero_() + # Step scheduler + self.scheduler.step() + if self.curr_iter % 1000 == 0: torch.save({ 'iter': self.curr_iter, @@ -396,7 +427,6 @@ class Model: self, dataset_config: Dict ) -> Union[CASIAB]: - self.train_size = dataset_config.get('train_size', 74) self.in_channels = dataset_config.get('num_input_channels', 3) self._dataset_sig = self._make_signature( dataset_config, |