diff options
author | Jordan Gong <jordan.gong@protonmail.com> | 2021-02-28 23:12:58 +0800 |
---|---|---|
committer | Jordan Gong <jordan.gong@protonmail.com> | 2021-02-28 23:12:58 +0800 |
commit | 0d2b643d7e04eba872e8b1fc9b04478a026bb3b0 (patch) | |
tree | a53d8aa69e050397d3443a602a42ed700a8e8bb5 /models | |
parent | 10b1299eb46343e119bc413ba8774dcb1a68595e (diff) | |
parent | e04f54d0bfc8fc711e53561065d772dae1926b64 (diff) |
Merge branch 'python3.8' into python3.7
# Conflicts:
# utils/configuration.py
Diffstat (limited to 'models')
-rw-r--r-- | models/layers.py | 4 | ||||
-rw-r--r-- | models/model.py | 74 |
2 files changed, 62 insertions, 16 deletions
diff --git a/models/layers.py b/models/layers.py index ae61583..e30d0c4 100644 --- a/models/layers.py +++ b/models/layers.py @@ -80,7 +80,9 @@ class DCGANConvTranspose2d(BasicConvTranspose2d): if self.is_last_layer: return self.trans_conv(x) else: - return super().forward(x) + x = self.trans_conv(x) + x = self.bn(x) + return F.leaky_relu(x, 0.2, inplace=True) class BasicLinear(nn.Module): diff --git a/models/model.py b/models/model.py index f4654f3..aef5302 100644 --- a/models/model.py +++ b/models/model.py @@ -56,6 +56,8 @@ 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 @@ -141,6 +143,7 @@ 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() start_iter = optim_hp.pop('start_iter', 0) @@ -160,10 +163,13 @@ class Model: ) else: # Different margins self.triplet_loss = JointBatchTripletLoss( - self.rgb_pn.hpm_num_parts, triplet_is_hard, triplet_margins + self.rgb_pn.hpm_num_parts, + triplet_is_hard, triplet_is_mean, triplet_margins ) else: # Soft margins - self.triplet_loss = BatchTripletLoss(triplet_is_hard, None) + self.triplet_loss = BatchTripletLoss( + triplet_is_hard, triplet_is_mean, None + ) # Try to accelerate computation using CUDA or others self.rgb_pn = self.rgb_pn.to(self.device) @@ -213,7 +219,7 @@ class Model: y = batch_c1['label'].to(self.device) # Duplicate labels for each part y = y.repeat(self.rgb_pn.num_total_parts, 1) - trip_loss, dist, non_zero_counts = self.triplet_loss(embedding, y) + trip_loss, dist, num_non_zero = self.triplet_loss(embedding, y) losses = torch.cat(( ae_losses, torch.stack(( @@ -237,18 +243,36 @@ class Model: 'HPM': losses[3], 'PartNet': losses[4] }, self.curr_iter) - self.writer.add_scalars('Loss/non-zero counts', { - 'HPM': non_zero_counts[:self.rgb_pn.hpm_num_parts].mean(), - 'PartNet': non_zero_counts[self.rgb_pn.hpm_num_parts:].mean() - }, self.curr_iter) - self.writer.add_scalars('Embedding/distance', { - 'HPM': dist[:self.rgb_pn.hpm_num_parts].mean(), - 'PartNet': dist[self.rgb_pn.hpm_num_parts].mean() - }, self.curr_iter) - self.writer.add_scalars('Embedding/2-norm', { - 'HPM': embedding[:self.rgb_pn.hpm_num_parts].norm(), - 'PartNet': embedding[self.rgb_pn.hpm_num_parts].norm() - }, self.curr_iter) + # None-zero losses in batch + if num_non_zero is not None: + self.writer.add_scalars('Loss/non-zero counts', { + 'HPM': num_non_zero[:self.rgb_pn.hpm_num_parts].mean(), + 'PartNet': num_non_zero[self.rgb_pn.hpm_num_parts:].mean() + }, self.curr_iter) + # Embedding distance + mean_hpm_dist = dist[:self.rgb_pn.hpm_num_parts].mean(0) + self._add_ranked_scalars( + 'Embedding/HPM distance', mean_hpm_dist, + self.num_pos_pairs, self.num_pairs, self.curr_iter + ) + mean_pa_dist = dist[self.rgb_pn.hpm_num_parts:].mean(0) + self._add_ranked_scalars( + 'Embedding/ParNet distance', mean_pa_dist, + self.num_pos_pairs, self.num_pairs, self.curr_iter + ) + # Embedding norm + mean_hpm_embedding = embedding[:self.rgb_pn.hpm_num_parts].mean(0) + mean_hpm_norm = mean_hpm_embedding.norm(dim=-1) + self._add_ranked_scalars( + 'Embedding/HPM norm', mean_hpm_norm, + self.k, self.pr * self.k, self.curr_iter + ) + mean_pa_embedding = embedding[self.rgb_pn.hpm_num_parts:].mean(0) + mean_pa_norm = mean_pa_embedding.norm(dim=-1) + self._add_ranked_scalars( + 'Embedding/PartNet norm', mean_pa_norm, + self.k, self.pr * self.k, self.curr_iter + ) if self.curr_iter % 100 == 0: lrs = self.scheduler.get_last_lr() @@ -300,6 +324,24 @@ class Model: self.writer.close() break + 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], @@ -521,6 +563,8 @@ class Model: ) -> DataLoader: config: Dict = dataloader_config.copy() (self.pr, self.k) = config.pop('batch_size', (8, 16)) + 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 if self.is_train: triplet_sampler = TripletSampler(dataset, (self.pr, self.k)) return DataLoader(dataset, |