[Not for Merge]: Visualize the gradient of each node in the lattice.

egs/librispeech/ASR/pruned_transducer_stateless/model.py

@@ -166,4 +166,23 @@ def forward(
             reduction="sum",
         )
 
-        return (simple_loss, pruned_loss)
+        B = px_grad.size(0)
+        S = px_grad.size(1)
+        T = px_grad.size(2) - 1
+        # px_grad's shape (B, S, T+1)
+        # py_grad's shape (B, S+1, T)
+
+        px_grad_pad = torch.zeros(
+            (B, 1, T + 1), dtype=px_grad.dtype, device=px_grad.device
+        )
+        py_grad_pad = torch.zeros(
+            (B, S + 1, 1), dtype=px_grad.dtype, device=px_grad.device
+        )
+
+        px_grad_padded = torch.cat([px_grad, px_grad_pad], dim=1)
+        py_grad_padded = torch.cat([py_grad, py_grad_pad], dim=2)
+
+        # tot_grad's shape (B, S+1, T+1)
+        tot_grad = px_grad_padded + py_grad_padded
+
+        return (simple_loss, pruned_loss, tot_grad, x_lens, y_lens)

egs/librispeech/ASR/pruned_transducer_stateless/train.py

@@ -35,7 +35,7 @@
 import logging
 from pathlib import Path
 from shutil import copyfile
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple
 
 import k2
 import sentencepiece as spm
@@ -434,14 +434,20 @@ def compute_loss(
     y = k2.RaggedTensor(y).to(device)
 
     with torch.set_grad_enabled(is_training):
-        simple_loss, pruned_loss = model(
+        simple_loss, pruned_loss, tot_grad, x_lens, y_lens, = model(
             x=feature,
             x_lens=feature_lens,
             y=y,
             prune_range=params.prune_range,
             am_scale=params.am_scale,
             lm_scale=params.lm_scale,
         )
+        cut_ids = [c.id for c in supervisions["cut"]]
+        save_and_plot_tot_grad(tot_grad, cut_ids, x_lens, y_lens)
+        import sys
+
+        sys.exit()
+
         loss = params.simple_loss_scale * simple_loss + pruned_loss
 
     assert loss.requires_grad == is_training
@@ -491,6 +497,53 @@ def compute_validation_loss(
     return tot_loss
 
 
+def save_and_plot_tot_grad(
+    tot_grad: torch.Tensor,
+    cut_ids: List[str],
+    x_lens: torch.Tensor,
+    y_lens: torch.Tensor,
+):
+    """Save and plot the tot_grad.
+
+    Args:
+      tot_grad:
+        A tensor of shape (B, U+1, T+1). It contains the gradient of
+        each node in the lattice.
+      cut_ids:
+        A list of size B, containing the cut ID of each utterance in the batch.
+      x_lens:
+        A 1-D tensor of shape (B,), specifying the number of valid acoustic
+        frames in tot_grad for each utterance in the batch.
+      y_lens:
+        A 1-D tensor of shape (B,), specifying the number of valid tokens
+        in tot_grad for each utterance in the batch.
+    """
+    import matplotlib.pyplot as plt
+
+    tot_grad = tot_grad.detach().cpu().permute(0, 2, 1)
+
+    ext = "png"  # supported types: png, ps, pdf, svg
+
+    x_lens = x_lens.tolist()
+    y_lens = y_lens.tolist()
+
+    tot_grad = tot_grad.unbind(0)
+    for i in range(len(cut_ids)):
+        cid = cut_ids[i]
+        T = x_lens[i]
+        U = y_lens[i]
+        grad = tot_grad[i][:T, :U]
+
+        filename = f"{cid}.{ext}"
+        logging.info(f"Saving to {filename}")
+        #  plt.matshow(grad.t(), origin="lower", cmap="gray")
+        plt.matshow(grad.t(), origin="lower")
+        plt.xlabel("t")
+        plt.ylabel("u")
+        plt.title(cid)
+        plt.savefig(filename)
+
+
 def train_one_epoch(
     params: AttributeDict,
     model: nn.Module,
@@ -577,6 +630,7 @@ def maybe_log_param_relative_changes():
             batch=batch,
             is_training=True,
         )
+
         # summary stats
         tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info