diff options
| -rw-r--r-- | config.py | 14 | ||||
| -rw-r--r-- | models/auto_encoder.py | 2 | ||||
| -rw-r--r-- | models/model.py | 61 |
3 files changed, 46 insertions, 31 deletions
@@ -8,6 +8,8 @@ config = { 'save_dir': 'runs', # Recorde disentangled image or not 'image_log_on': False, + # Image log per n steps + 'image_log_steps': 100, # The number of subjects for validating (Part of testing set) 'val_size': 20, }, @@ -37,7 +39,7 @@ config = { # Batch size (pr, k) # `pr` denotes number of persons # `k` denotes number of sequences per person - 'batch_size': (4, 6), + 'batch_size': (4, 5), # Number of workers of Dataloader 'num_workers': 4, # Faster data transfer from RAM to GPU if enabled @@ -84,18 +86,22 @@ config = { # Local parameters (override global ones) # 'auto_encoder': { - # 'weight_decay': 0.001 + # 'lr': 1e-3 # }, }, 'scheduler': { # Step start to decay 'start_step': 500, + # Step stop decaying + # 'stop_step': 30_000, # Multiplicative factor of decay in the end 'final_gamma': 0.01, # Local parameters (override global ones) - # 'hpm': { - # 'final_gamma': 0.001 + # 'auto_encoder': { + # 'start_step': 0, + # 'stop_step': 500, + # 'final_gamma': 0.5 # } } }, diff --git a/models/auto_encoder.py b/models/auto_encoder.py index 1028767..d11ec99 100644 --- a/models/auto_encoder.py +++ b/models/auto_encoder.py @@ -154,7 +154,7 @@ class AutoEncoder(nn.Module): return ( (f_a_c1_t2, f_c_c1_t2, f_p_c1_t2), - (xrecon_loss / 10, cano_cons_loss, pose_sim_loss * 10) + (xrecon_loss, cano_cons_loss, pose_sim_loss * 10) ) else: # evaluating return f_a_c1_t2, f_c_c1_t2, f_p_c1_t2 diff --git a/models/model.py b/models/model.py index 78a9c0f..896d8d3 100644 --- a/models/model.py +++ b/models/model.py @@ -76,6 +76,7 @@ class Model: 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.image_log_steps = system_config.get('image_log_steps', 100) self.val_size = system_config.get('val_size', 10) self.CASIAB_GALLERY_SELECTOR = { @@ -193,8 +194,8 @@ class Model: triplet_is_hard, triplet_is_mean, None ) - self.num_pairs = (self.pr*self.k-1) * (self.pr*self.k) // 2 - self.num_pos_pairs = (self.k*(self.k-1)//2) * self.pr + self.num_pairs = (self.pr * self.k - 1) * (self.pr * self.k) // 2 + self.num_pos_pairs = (self.k * (self.k - 1) // 2) * self.pr # Try to accelerate computation using CUDA or others self.rgb_pn = self.rgb_pn.to(self.device) @@ -208,24 +209,31 @@ class Model: ], **optim_hp) # Scheduler - start_step = sched_hp.get('start_step', 15_000) + start_step = sched_hp.get('start_step', 0) + stop_step = sched_hp.get('stop_step', self.total_iter) final_gamma = sched_hp.get('final_gamma', 0.001) ae_start_step = ae_sched_hp.get('start_step', start_step) + ae_stop_step = ae_sched_hp.get('stop_step', stop_step) ae_final_gamma = ae_sched_hp.get('final_gamma', final_gamma) - ae_all_step = self.total_iter - ae_start_step + ae_all_step = ae_stop_step - ae_start_step hpm_start_step = hpm_sched_hp.get('start_step', start_step) + hpm_stop_step = hpm_sched_hp.get('stop_step', stop_step) hpm_final_gamma = hpm_sched_hp.get('final_gamma', final_gamma) - hpm_all_step = self.total_iter - hpm_start_step + hpm_all_step = hpm_stop_step - hpm_start_step pn_start_step = pn_sched_hp.get('start_step', start_step) + pn_stop_step = pn_sched_hp.get('stop_step', stop_step) pn_final_gamma = pn_sched_hp.get('final_gamma', final_gamma) - pn_all_step = self.total_iter - pn_start_step + pn_all_step = pn_stop_step - pn_start_step self.scheduler = optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda=[ - lambda t: ae_final_gamma ** ((t - ae_start_step) / ae_all_step) - if t > ae_start_step else 1, - lambda t: hpm_final_gamma ** ((t - hpm_start_step) / hpm_all_step) - if t > hpm_start_step else 1, - lambda t: pn_final_gamma ** ((t - pn_start_step) / pn_all_step) - if t > pn_start_step else 1, + lambda t: 1 if t <= ae_start_step + else ae_final_gamma ** ((t - ae_start_step) / ae_all_step) + if ae_start_step < t <= ae_stop_step else ae_final_gamma, + lambda t: 1 if t <= hpm_start_step + else hpm_final_gamma ** ((t - hpm_start_step) / hpm_all_step) + if hpm_start_step < t <= hpm_stop_step else hpm_final_gamma, + lambda t: 1 if t <= pn_start_step + else pn_final_gamma ** ((t - pn_start_step) / pn_all_step) + if pn_start_step < t <= pn_stop_step else pn_final_gamma, ]) self.writer = SummaryWriter(self._log_name) @@ -275,24 +283,25 @@ class Model: 'Train', embed_c, embed_p, hpm_result, pn_result, loss, losses ) - if self.curr_iter % 100 == 99: - # Write disentangled images - if self.image_log_on: - i_a, i_c, i_p = images + # Write disentangled images + if self.image_log_on and self.curr_iter % self.image_log_steps \ + == self.image_log_steps - 1: + i_a, i_c, i_p = images + self.writer.add_images( + 'Appearance image', i_a, self.curr_iter + ) + self.writer.add_images( + 'Canonical image', i_c, self.curr_iter + ) + for i, (o, p) in enumerate(zip(x_c1, i_p)): self.writer.add_images( - 'Appearance image', i_a, self.curr_iter + f'Original image/batch {i}', o, self.curr_iter ) self.writer.add_images( - 'Canonical image', i_c, self.curr_iter + f'Pose image/batch {i}', p, self.curr_iter ) - for i, (o, p) in enumerate(zip(x_c1, i_p)): - self.writer.add_images( - f'Original image/batch {i}', o, self.curr_iter - ) - self.writer.add_images( - f'Pose image/batch {i}', p, self.curr_iter - ) + if self.curr_iter % 100 == 99: # Validation embed_c = self._flatten_embedding(embed_c) embed_p = self._flatten_embedding(embed_p) @@ -349,7 +358,7 @@ class Model: def _write_embedding(self, tag, embed, x, y): frame = x[:, 0, :, :, :].cpu() n, c, h, w = frame.size() - padding = torch.zeros(n, c, h, (h-w) // 2) + padding = torch.zeros(n, c, h, (h - w) // 2) padded_frame = torch.cat((padding, frame, padding), dim=-1) self.writer.add_embedding( embed, |