diff options
| -rw-r--r-- | config.py | 18 | ||||
| -rw-r--r-- | models/model.py | 27 |
2 files changed, 22 insertions, 23 deletions
@@ -70,8 +70,6 @@ config = { }, 'optimizer': { # Global parameters - # Iteration start to optimize non-disentangling parts - # 'start_iter': 0, # Initial learning rate of Adam Optimizer 'lr': 1e-4, # Coefficients used for computing running averages of @@ -85,15 +83,15 @@ config = { # 'amsgrad': False, # Local parameters (override global ones) - # 'auto_encoder': { - # 'weight_decay': 0.001 - # }, + 'auto_encoder': { + 'weight_decay': 0.001 + }, }, 'scheduler': { - # Period of learning rate decay - 'step_size': 500, - # Multiplicative factor of decay - 'gamma': 1, + # Step start to decay + 'start_step': 15_000, + # Multiplicative factor of decay in the end + 'final_gamma': 0.001, } }, # Model metadata @@ -107,6 +105,6 @@ config = { # Restoration iteration (multiple models, e.g. nm, bg and cl) 'restore_iters': (0, 0, 0), # Total iteration for training (multiple models) - 'total_iters': (80_000, 80_000, 80_000), + 'total_iters': (25_000, 25_000, 25_000), }, } diff --git a/models/model.py b/models/model.py index 61470d9..2fb2b39 100644 --- a/models/model.py +++ b/models/model.py @@ -144,7 +144,6 @@ class Model: triplet_is_mean = model_hp.pop('triplet_is_mean', True) triplet_margins = model_hp.pop('triplet_margins', None) 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', {}) @@ -181,14 +180,17 @@ class Model: {'params': self.rgb_pn.hpm.parameters(), **hpm_optim_hp}, {'params': self.rgb_pn.fc_mat, **fc_optim_hp} ], **optim_hp) - sched_gamma = sched_hp.get('gamma', 0.9) - sched_step_size = sched_hp.get('step_size', 500) + sched_final_gamma = sched_hp.get('final_gamma', 0.001) + sched_start_step = sched_hp.get('start_step', 15_000) + + def lr_lambda(epoch): + passed_step = epoch - sched_start_step + all_step = self.total_iter - sched_start_step + return sched_final_gamma ** (passed_step / all_step) self.scheduler = optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda=[ - lambda epoch: sched_gamma ** (epoch // sched_step_size), - lambda epoch: 0 if epoch < start_iter else 1, - lambda epoch: 0 if epoch < start_iter else 1, - lambda epoch: 0 if epoch < start_iter else 1, + lr_lambda, lr_lambda, lr_lambda, lr_lambda ]) + self.writer = SummaryWriter(self._log_name) self.rgb_pn.train() @@ -208,7 +210,7 @@ class Model: running_loss = torch.zeros(5, device=self.device) print(f"{'Time':^8} {'Iter':^5} {'Loss':^6}", f"{'Xrecon':^8} {'CanoCons':^8} {'PoseSim':^8}", - f"{'BATripH':^8} {'BATripP':^8} {'LRs':^19}") + f"{'BATripH':^8} {'BATripP':^8} {'LR':^9}") for (batch_c1, batch_c2) in dataloader: self.curr_iter += 1 # Zero the parameter gradients @@ -279,10 +281,7 @@ class Model: lrs = self.scheduler.get_last_lr() # Write learning rates self.writer.add_scalar( - 'Learning rate/Auto-encoder', lrs[0], self.curr_iter - ) - self.writer.add_scalar( - 'Learning rate/Others', lrs[1], self.curr_iter + 'Learning rate', lrs[0], self.curr_iter ) # Write disentangled images if self.image_log_on: @@ -306,7 +305,7 @@ class Model: 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])) + f'{lrs[0]:.3e}') running_loss.zero_() # Step scheduler @@ -382,6 +381,8 @@ class Model: # Init models model_hp: dict = self.hp.get('model', {}).copy() + model_hp.pop('triplet_is_hard', True) + model_hp.pop('triplet_is_mean', True) model_hp.pop('triplet_margins', None) self.rgb_pn = RGBPartNet(self.in_channels, self.in_size, **model_hp) # Try to accelerate computation using CUDA or others |