diff options
Diffstat (limited to 'models/model.py')
| -rw-r--r-- | models/model.py | 282 |
1 files changed, 33 insertions, 249 deletions
diff --git a/models/model.py b/models/model.py index ceadb92..25c8a4f 100644 --- a/models/model.py +++ b/models/model.py @@ -6,22 +6,18 @@ from typing import Union, Optional import numpy as np import torch import torch.nn as nn -import torch.nn.functional as F import torch.optim as optim from torch.utils.data import DataLoader from torch.utils.data.dataloader import default_collate from torch.utils.tensorboard import SummaryWriter from tqdm import tqdm -from models.hpm import HorizontalPyramidMatching -from models.part_net import PartNet from models.rgb_part_net import RGBPartNet from utils.configuration import DataloaderConfiguration, \ HyperparameterConfiguration, DatasetConfiguration, ModelConfiguration, \ SystemConfiguration from utils.dataset import CASIAB, ClipConditions, ClipViews, ClipClasses from utils.sampler import TripletSampler -from utils.triplet_loss import BatchTripletLoss class Model: @@ -62,8 +58,6 @@ class Model: self.in_size: tuple[int, int] = (64, 48) self.pr: Optional[int] = None self.k: Optional[int] = None - self.num_pairs: Optional[int] = None - self.num_pos_pairs: Optional[int] = None self._gallery_dataset_meta: Optional[dict[str, list]] = None self._probe_datasets_meta: Optional[dict[str, dict[str, list]]] = None @@ -73,8 +67,6 @@ class Model: self._dataset_sig: str = 'undefined' self.rgb_pn: Optional[RGBPartNet] = None - self.triplet_loss_hpm: Optional[BatchTripletLoss] = None - self.triplet_loss_pn: Optional[BatchTripletLoss] = None self.optimizer: Optional[optim.Adam] = None self.scheduler: Optional[optim.lr_scheduler.StepLR] = None self.writer: Optional[SummaryWriter] = None @@ -166,71 +158,23 @@ class Model: )) # Prepare for model, optimizer and scheduler model_hp: dict = self.hp.get('model', {}).copy() - triplet_is_hard = model_hp.pop('triplet_is_hard', True) - 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() - ae_optim_hp = optim_hp.pop('auto_encoder', {}) - hpm_optim_hp = optim_hp.pop('hpm', {}) - pn_optim_hp = optim_hp.pop('part_net', {}) sched_hp = self.hp.get('scheduler', {}) - ae_sched_hp = sched_hp.get('auto_encoder', {}) - hpm_sched_hp = sched_hp.get('hpm', {}) - pn_sched_hp = sched_hp.get('part_net', {}) self.rgb_pn = RGBPartNet(self.in_channels, self.in_size, **model_hp, image_log_on=self.image_log_on) - # Hard margins - if triplet_margins: - self.triplet_loss_hpm = BatchTripletLoss( - triplet_is_hard, triplet_is_mean, triplet_margins[0] - ) - self.triplet_loss_pn = BatchTripletLoss( - triplet_is_hard, triplet_is_mean, triplet_margins[1] - ) - else: # Soft margins - self.triplet_loss_hpm = BatchTripletLoss( - triplet_is_hard, triplet_is_mean, None - ) - self.triplet_loss_pn = BatchTripletLoss( - 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 # Try to accelerate computation using CUDA or others self.rgb_pn = self.rgb_pn.to(self.device) - self.triplet_loss_hpm = self.triplet_loss_hpm.to(self.device) - self.triplet_loss_pn = self.triplet_loss_pn.to(self.device) - - self.optimizer = optim.Adam([ - {'params': self.rgb_pn.ae.parameters(), **ae_optim_hp}, - {'params': self.rgb_pn.hpm.parameters(), **hpm_optim_hp}, - {'params': self.rgb_pn.pn.parameters(), **pn_optim_hp}, - ], **optim_hp) - - # Scheduler + self.optimizer = optim.Adam(self.rgb_pn.parameters(), **optim_hp) start_step = sched_hp.get('start_step', 15_000) final_gamma = sched_hp.get('final_gamma', 0.001) - ae_start_step = ae_sched_hp.get('start_step', start_step) - ae_final_gamma = ae_sched_hp.get('final_gamma', final_gamma) - ae_all_step = self.total_iter - ae_start_step - hpm_start_step = hpm_sched_hp.get('start_step', start_step) - hpm_final_gamma = hpm_sched_hp.get('final_gamma', final_gamma) - hpm_all_step = self.total_iter - hpm_start_step - pn_start_step = pn_sched_hp.get('start_step', start_step) - pn_final_gamma = pn_sched_hp.get('final_gamma', final_gamma) - pn_all_step = self.total_iter - 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, - ]) - + all_step = self.total_iter - start_step + self.scheduler = optim.lr_scheduler.LambdaLR( + self.optimizer, + lambda t: final_gamma ** ((t - start_step) / all_step) + if t > start_step else 1, + ) self.writer = SummaryWriter(self._log_name) # Set seeds for reproducibility @@ -259,24 +203,18 @@ class Model: # forward + backward + optimize x_c1 = batch_c1['clip'].to(self.device) x_c2 = batch_c2['clip'].to(self.device) - embed_c, embed_p, ae_losses, images = self.rgb_pn(x_c1, x_c2) - y = batch_c1['label'].to(self.device) - losses, hpm_result, pn_result = self._classification_loss( - embed_c, embed_p, ae_losses, y - ) + losses, features, images = self.rgb_pn(x_c1, x_c2) loss = losses.sum() loss.backward() self.optimizer.step() self.scheduler.step() # Learning rate - self.writer.add_scalars('Learning rate', dict(zip(( - 'Auto-encoder', 'HPM', 'PartNet' - ), self.scheduler.get_last_lr())), self.curr_iter) - # Other stats - self._write_stat( - 'Train', embed_c, embed_p, hpm_result, pn_result, loss, losses + self.writer.add_scalar( + 'Learning rate', self.scheduler.get_last_lr()[0], self.curr_iter ) + # Other stats + self._write_stat('Train', loss, losses) if self.curr_iter % 100 == 99: # Write disentangled images @@ -295,32 +233,33 @@ class Model: self.writer.add_images( f'Pose image/batch {i}', p, self.curr_iter ) - - # Validation - embed_c = self._flatten_embedding(embed_c) - embed_p = self._flatten_embedding(embed_p) - self._write_embedding('HPM Train', embed_c, x_c1, y) - self._write_embedding('PartNet Train', embed_p, x_c1, y) + f_a, f_c, f_p = features + for i, (f_a_i, f_c_i, f_p_i) in enumerate( + zip(f_a, f_c, f_p) + ): + self.writer.add_images( + f'Appearance features/Layer {i}', + f_a_i[:, :3, :, :], self.curr_iter + ) + self.writer.add_images( + f'Canonical features/Layer {i}', + f_c_i[:, :3, :, :], self.curr_iter + ) + for j, p in enumerate(f_p_i): + self.writer.add_images( + f'Pose features/Layer {i}/batch{j}', + p[:, :3, :, :], self.curr_iter + ) # Calculate losses on testing batch batch_c1, batch_c2 = next(val_dataloader) x_c1 = batch_c1['clip'].to(self.device) x_c2 = batch_c2['clip'].to(self.device) with torch.no_grad(): - embed_c, embed_p, ae_losses, _ = self.rgb_pn(x_c1, x_c2) - y = batch_c1['label'].to(self.device) - losses, hpm_result, pn_result = self._classification_loss( - embed_c, embed_p, ae_losses, y - ) + losses, _, _ = self.rgb_pn(x_c1, x_c2) loss = losses.sum() - self._write_stat( - 'Val', embed_c, embed_p, hpm_result, pn_result, loss, losses - ) - embed_c = self._flatten_embedding(embed_c) - embed_p = self._flatten_embedding(embed_p) - self._write_embedding('HPM Val', embed_c, x_c1, y) - self._write_embedding('PartNet Val', embed_p, x_c1, y) + self._write_stat('Val', loss, losses) # Checkpoint if self.curr_iter % 1000 == 999: @@ -333,117 +272,15 @@ class Model: self.writer.close() - def _classification_loss(self, embed_c, embed_p, ae_losses, y): - # Duplicate labels for each part - y_triplet = y.repeat(self.rgb_pn.num_parts, 1) - hpm_result = self.triplet_loss_hpm( - embed_c, y_triplet[:self.rgb_pn.hpm.num_parts] - ) - pn_result = self.triplet_loss_pn( - embed_p, y_triplet[self.rgb_pn.hpm.num_parts:] - ) - losses = torch.stack(( - *ae_losses, - hpm_result.pop('loss').mean(), - pn_result.pop('loss').mean() - )) - return losses, hpm_result, pn_result - - 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) - padded_frame = torch.cat((padding, frame, padding), dim=-1) - self.writer.add_embedding( - embed, - metadata=y.cpu().tolist(), - label_img=padded_frame, - global_step=self.curr_iter, - tag=tag - ) - - def _flatten_embedding(self, embed): - return embed.detach().transpose(0, 1).reshape(self.k * self.pr, -1) - def _write_stat( - self, postfix, embed_c, embed_p, hpm_result, pn_result, loss, losses + self, postfix, loss, losses ): # Write losses to TensorBoard self.writer.add_scalar(f'Loss/all {postfix}', loss, self.curr_iter) self.writer.add_scalars(f'Loss/disentanglement {postfix}', dict(zip(( 'Cross reconstruction loss', 'Canonical consistency loss', 'Pose similarity loss' - ), losses[:3])), self.curr_iter) - self.writer.add_scalars(f'Loss/triplet loss {postfix}', { - 'HPM': losses[3], - 'PartNet': losses[4] - }, self.curr_iter) - # None-zero losses in batch - if hpm_result['counts'] is not None and pn_result['counts'] is not None: - self.writer.add_scalars(f'Loss/non-zero counts {postfix}', { - 'HPM': hpm_result['counts'].mean(), - 'PartNet': pn_result['counts'].mean() - }, self.curr_iter) - # Embedding distance - mean_hpm_dist = hpm_result['dist'].mean(0) - self._add_ranked_scalars( - f'Embedding/HPM distance {postfix}', mean_hpm_dist, - self.num_pos_pairs, self.num_pairs, self.curr_iter - ) - mean_pn_dist = pn_result['dist'].mean(0) - self._add_ranked_scalars( - f'Embedding/ParNet distance {postfix}', mean_pn_dist, - self.num_pos_pairs, self.num_pairs, self.curr_iter - ) - # Embedding norm - mean_hpm_embedding = embed_c.mean(0) - mean_hpm_norm = mean_hpm_embedding.norm(dim=-1) - self._add_ranked_scalars( - f'Embedding/HPM norm {postfix}', mean_hpm_norm, - self.k, self.pr * self.k, self.curr_iter - ) - mean_pa_embedding = embed_p.mean(0) - mean_pa_norm = mean_pa_embedding.norm(dim=-1) - self._add_ranked_scalars( - f'Embedding/PartNet norm {postfix}', mean_pa_norm, - self.k, self.pr * self.k, self.curr_iter - ) - - def _add_ranked_scalars( - self, - main_tag: str, - metric: torch.Tensor, - num_pos: int, - num_all: int, - global_step: int - ): - rank = metric.argsort() - pos_ile = 100 - (num_pos - 1) * 100 // num_all - self.writer.add_scalars(main_tag, { - '0%-ile': metric[rank[-1]], - f'{100 - pos_ile}%-ile': metric[rank[-num_pos]], - '50%-ile': metric[rank[num_all // 2 - 1]], - f'{pos_ile}%-ile': metric[rank[num_pos - 1]], - '100%-ile': metric[rank[0]] - }, global_step) - - def predict_all( - self, - iters: tuple[int], - dataset_config: DatasetConfiguration, - dataset_selectors: dict[ - str, dict[str, Union[ClipClasses, ClipConditions, ClipViews]] - ], - dataloader_config: DataloaderConfiguration, - ) -> dict[str, torch.Tensor]: - # Transform data to features - gallery_samples, probe_samples = self.transform( - iters, dataset_config, dataset_selectors, dataloader_config - ) - # Evaluate features - accuracy = self.evaluate(gallery_samples, probe_samples) - - return accuracy + ), losses)), self.curr_iter) def transform( self, @@ -466,9 +303,6 @@ 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 self.rgb_pn = self.rgb_pn.to(self.device) @@ -509,54 +343,6 @@ class Model: 'feature': torch.cat((feature_c, feature_p)).view(-1) } - @staticmethod - def evaluate( - gallery_samples: dict[str, dict[str, Union[list, torch.Tensor]]], - probe_samples: dict[str, dict[str, Union[list, torch.Tensor]]], - num_ranks: int = 5 - ) -> dict[str, torch.Tensor]: - conditions = list(probe_samples.keys()) - gallery_views_meta = gallery_samples['meta']['views'] - probe_views_meta = probe_samples[conditions[0]]['meta']['views'] - accuracy = { - condition: torch.empty( - len(gallery_views_meta), len(probe_views_meta), num_ranks - ) - for condition in conditions - } - - for condition in conditions: - gallery_samples_c = gallery_samples[condition] - (labels_g, _, views_g, features_g) = gallery_samples_c.values() - views_g = np.asarray(views_g) - probe_samples_c = probe_samples[condition] - (labels_p, _, views_p, features_p, _) = probe_samples_c.values() - views_p = np.asarray(views_p) - accuracy_c = accuracy[condition] - for (v_g_i, view_g) in enumerate(gallery_views_meta): - gallery_view_mask = (views_g == view_g) - f_g = features_g[gallery_view_mask] - y_g = labels_g[gallery_view_mask] - for (v_p_i, view_p) in enumerate(probe_views_meta): - probe_view_mask = (views_p == view_p) - f_p = features_p[probe_view_mask] - y_p = labels_p[probe_view_mask] - # Euclidean distance - f_p_squared_sum = torch.sum(f_p ** 2, dim=1).unsqueeze(1) - f_g_squared_sum = torch.sum(f_g ** 2, dim=1).unsqueeze(0) - f_p_times_f_g_sum = f_p @ f_g.T - dist = torch.sqrt(F.relu( - f_p_squared_sum - 2*f_p_times_f_g_sum + f_g_squared_sum - )) - # Ranked accuracy - rank_mask = dist.argsort(1)[:, :num_ranks] - positive_mat = torch.eq(y_p.unsqueeze(1), - y_g[rank_mask]).cumsum(1).gt(0) - positive_counts = positive_mat.sum(0) - total_counts, _ = dist.size() - accuracy_c[v_g_i, v_p_i, :] = positive_counts / total_counts - return accuracy - def _load_pretrained( self, iters: tuple[int], @@ -629,8 +415,6 @@ class Model: nn.init.zeros_(m.bias) elif isinstance(m, nn.Linear): nn.init.xavier_uniform_(m.weight) - elif isinstance(m, (HorizontalPyramidMatching, PartNet)): - nn.init.xavier_uniform_(m.fc_mat) def _parse_dataset_config( self, |