Pass target_length to TBMediaLogger so predicted + synthesized mels match ground truth dimensions

scripts/train_first.py

@@ -462,6 +462,12 @@ def _run_standalone(cfg, resume_path: str | None,
             # Get the original audio for this row
             row_audio = smoke_rows[i].audio
 
+            # Compute the ground truth mel's time dimension — the predicted mel
+            # will be upsampled to this length so both mels have the same
+            # aspect ratio in TensorBoard's IMAGES tab
+            gt_mel = mels[i]   # (80, T_gt) — precomputed earlier
+            gt_T = gt_mel.shape[1]
+
             # 1. Log ground truth mel + original audio
             media_logger.log_ground_truth(
                 step=step,
@@ -471,6 +477,7 @@ def _run_standalone(cfg, resume_path: str | None,
             )
 
             # 2. Log predicted mel + Griffin-Lim reconstructed audio
+            #    UPSAMPLE to match ground truth's time dimension
             # Convert pred_slice (1, T, D) → (n_mels, T) mel
             pred_mel = pred_slice.squeeze(0).detach().cpu().numpy().T  # (D, T)
             # Pad to (n_mels=80, T) if D < 80
@@ -481,10 +488,15 @@ def _run_standalone(cfg, resume_path: str | None,
                 pred_mel = padded
             else:
                 pred_mel = pred_mel[:media_logger.n_mels, :]
-            media_logger.log_reconstructed(step=step, mel=pred_mel)
+            # Upsample predicted mel to match ground truth's T dimension
+            # so both images have the same width in TensorBoard
+            media_logger.log_reconstructed(
+                step=step, mel=pred_mel, target_length=gt_T,
+            )
 
             # 3. Log synthesized TTS sample (text → tokens → model → audio)
             #    every 2 steps so we don't bloat the TB logs
+            #    Also upsample to the ground truth's T for consistent display
             if step % 2 == 0 or step == 1:
                 tts_text = tts_samples[tts_idx % len(tts_samples)]
                 tts_idx += 1
@@ -493,6 +505,7 @@ def _run_standalone(cfg, resume_path: str | None,
                     text=tts_text,
                     phonemizer=phonemizer,
                     model=model,
+                    target_length=gt_T,
                 )
 
             # Log to LossLogger (also pushes losses.json to HF on a cadence)