bubing
diff --git a/‎recipes/LibriSpeech/ASR/seq2seq/experiment.py‎
Lines changed: 1 addition & 1 deletion b/‎recipes/LibriSpeech/ASR/seq2seq/experiment.py‎
Lines changed: 1 addition & 1 deletion
diff --git a/‎speechbrain/alignment/aligner.py‎
Lines changed: 15 additions & 9 deletions b/‎speechbrain/alignment/aligner.py‎
Lines changed: 15 additions & 9 deletions
diff --git a/‎speechbrain/decoders/decoders.py‎
Lines changed: 0 additions & 255 deletions b/‎speechbrain/decoders/decoders.py‎
Lines changed: 0 additions & 255 deletions
@@ -71,7 +71,7 @@ def compute_forward(self, x, y, stage):
         # Forward pass
         feats = self.hparams.compute_features(wavs)
         feats = self.hparams.normalize(feats, wav_lens)
-        x = self.hparams.enc(feats)
+        x = self.hparams.enc(feats.detach())
         e_in = self.hparams.emb(y_in)
         h, _ = self.hparams.dec(e_in, x, wav_lens)
 
 
@@ -518,15 +518,19 @@ def _make_trans_prob(self, phn_lens_abs):
         )
 
         # clear probabilities for too-long sequences
-        mask_a = torch.arange(U_max).to(device)[None, :] < phn_lens_abs[:, None]
+        mask_a = (
+            torch.arange(U_max, device=device)[None, :] < phn_lens_abs[:, None]
+        )
         mask_a = mask_a.unsqueeze(2)
         mask_a = mask_a.expand(-1, -1, U_max)
         mask_b = mask_a.permute(0, 2, 1)
         trans_prob = trans_prob * (mask_a & mask_b).float()
 
         ## put -infs in place of zeros:
         trans_prob = torch.where(
-            trans_prob == 1, trans_prob, torch.tensor(-float("Inf")).to(device)
+            trans_prob == 1,
+            trans_prob,
+            torch.tensor(-float("Inf"), device=device),
         )
 
         ## normalize
@@ -575,7 +579,7 @@ def _make_emiss_pred_useful(
         emiss_pred_acc_lens = torch.where(
             mask_lens[:, :, None],
             emission_pred,
-            torch.tensor([0.0]).to(device),
+            torch.tensor([0.0], device=device),
         )
 
         # manipulate phn tensor, and then 'torch.gather'
@@ -590,7 +594,7 @@ def _make_emiss_pred_useful(
         emiss_pred_useful = torch.where(
             mask_phn_lens[:, None, :],
             emiss_pred_useful,
-            torch.tensor([self.neg_inf]).to(device),
+            torch.tensor([self.neg_inf], device=device),
         )
 
         emiss_pred_useful = emiss_pred_useful.permute(0, 2, 1)
@@ -647,8 +651,8 @@ def _dp_forward(
 
         # initialise
         alpha_matrix = self.neg_inf * torch.ones(
-            [batch_size, U_max, fb_max_length]
-        ).to(device)
+            [batch_size, U_max, fb_max_length], device=device
+        )
         alpha_matrix[:, :, 0] = pi_prob + emiss_pred_useful[:, :, 0]
 
         for t in range(1, fb_max_length):
@@ -733,9 +737,11 @@ def _dp_viterbi(
         trans_prob = trans_prob.to(device)
 
         v_matrix = self.neg_inf * torch.ones(
-            [batch_size, U_max, fb_max_length]
-        ).to(device)
-        backpointers = -99 * torch.ones([batch_size, U_max, fb_max_length])
+            [batch_size, U_max, fb_max_length], device=device
+        )
+        backpointers = -99 * torch.ones(
+            [batch_size, U_max, fb_max_length], device=device
+        )
 
         # initialise
         v_matrix[:, :, 0] = pi_prob + emiss_pred_useful[:, :, 0]