VedantDave/TVSSL
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Update 'train_mm_moco.py'

Browse Source
This commit is contained in:
Vedant Dave 2023-09-13 08:39:36 +00:00
parent ee5ce8cb82
commit 9fd405c42e
1 changed files with 18 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
train_mm_moco.py
View File

@ -100,6 +100,12 @@ class MultiModalMoCo(nn.Module):
q_vv = self.phi_vision_q(vision_base_q)
q_vt = self.phi_tactile_q(vision_base_q)
with torch.no_grad(): # no gradient
# Update key encoders
self._momentum_update_key_encoder(self.vision_base_q, self.vision_base_k)
self._momentum_update_key_encoder(self.tactile_base_q, self.tactile_base_k)
self._momentum_update_key_encoder(self.phi_vision_q, self.phi_vision_k)
# Compute key features
vision_base_k = self.vision_base_k(x_vision_k)
k_vv = self.phi_vision_k(vision_base_k)
k_tv = self.phi_tactile_k(vision_base_k)
@ -108,18 +114,15 @@ class MultiModalMoCo(nn.Module):
q_tv = self.Phi_vision_q(tactile_base_q)
q_tt = self.Phi_tactile_q(tactile_base_q)
tactile_base_k = self.tactile_base_k(x_tactile_k)
k_vt = self.Phi_vision_k(tactile_base_k)
k_tt = self.Phi_tactile_k(tactile_base_k)
with torch.no_grad(): # no gradient
# Update key encoders
self._momentum_update_key_encoder(self.vision_base_q, self.vision_base_k)
self._momentum_update_key_encoder(self.tactile_base_q, self.tactile_base_k)
self._momentum_update_key_encoder(self.phi_vision_q, self.phi_vision_k)
self._momentum_update_key_encoder(self.phi_tactile_q, self.phi_tactile_k)
self._momentum_update_key_encoder(self.Phi_vision_q, self.Phi_vision_k)
self._momentum_update_key_encoder(self.Phi_tactile_q, self.Phi_tactile_k)
# Compute key features
tactile_base_k = self.tactile_base_k(x_tactile_k)
k_vt = self.Phi_vision_k(tactile_base_k)
k_tt = self.Phi_tactile_k(tactile_base_k)
# Compute the contrastive loss for each pair of queries and keys
vision_vision_intra = self.moco_contrastive_loss(q_vv, k_vv)