diff options
| -rw-r--r-- | config.py | 12 | ||||
| -rw-r--r-- | models/model.py | 3 | ||||
| -rw-r--r-- | models/rgb_part_net.py | 5 |
3 files changed, 11 insertions, 9 deletions
@@ -35,7 +35,7 @@ config = { # Batch size (pr, k) # `pr` denotes number of persons # `k` denotes number of sequences per person - 'batch_size': (4, 8), + 'batch_size': (4, 6), # Number of workers of Dataloader 'num_workers': 4, # Faster data transfer from RAM to GPU if enabled @@ -62,7 +62,7 @@ config = { # Embedding dimension for each part 'embedding_dims': 256, # Triplet loss margins for HPM and PartNet - 'triplet_margins': (0.2, 0.2), + 'triplet_margins': (1.5, 1.5), }, 'optimizer': { # Global parameters @@ -81,15 +81,15 @@ 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': 0.9, + 'gamma': 1, } }, # Model metadata diff --git a/models/model.py b/models/model.py index d179361..1e22483 100644 --- a/models/model.py +++ b/models/model.py @@ -311,7 +311,8 @@ class Model: ) # Init models - model_hp = self.hp.get('model', {}) + model_hp: dict = self.hp.get('model', {}).copy() + 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) diff --git a/models/rgb_part_net.py b/models/rgb_part_net.py index fc1406c..4d7ba7f 100644 --- a/models/rgb_part_net.py +++ b/models/rgb_part_net.py @@ -76,8 +76,8 @@ class RGBPartNet(nn.Module): def _disentangle(self, x_c1_t2, x_c2_t2=None): n, t, c, h, w = x_c1_t2.size() device = x_c1_t2.device - x_c1_t1 = x_c1_t2[:, torch.randperm(t), :, :, :] if self.training: + x_c1_t1 = x_c1_t2[:, torch.randperm(t), :, :, :] ((f_a_, f_c_, f_p_), losses) = self.ae(x_c1_t2, x_c1_t1, x_c2_t2) # Decode features x_c = self._decode_cano_feature(f_c_, n, t, device) @@ -100,7 +100,8 @@ class RGBPartNet(nn.Module): else: # evaluating f_c_, f_p_ = self.ae(x_c1_t2) x_c = self._decode_cano_feature(f_c_, n, t, device) - x_p = self._decode_pose_feature(f_p_, n, t, c, h, w, device) + x_p_ = self._decode_pose_feature(f_p_, n, t, c, h, w, device) + x_p = x_p_.view(n, t, self.pn_in_channels, self.h // 4, self.w // 4) return (x_c, x_p), None, None def _decode_appr_feature(self, f_a_, n, t, device): |