Configuration changes: scores limit 5->10, min_prob 0.05->0.1, cur_grad_scale more aggressive increase

egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py

@@ -649,7 +649,7 @@ class AttentionDownsample(torch.nn.Module):
         scores = (src * self.query).sum(dim=-1, keepdim=True)
 
         scores =  penalize_abs_values_gt(scores,
-                                         limit=5.0,
+                                         limit=10.0,
                                          penalty=1.0e-04)
 
         weights = scores.softmax(dim=1)

egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py

@@ -482,7 +482,7 @@ class ActivationBalancer(torch.nn.Module):
             scale_gain_factor: float = 0.02,
             min_abs: float = 0.2,
             max_abs: float = 100.0,
-            min_prob: float = 0.05,
+            min_prob: float = 0.1,
     ):
         super(ActivationBalancer, self).__init__()
         self.num_channels = num_channels

egs/librispeech/ASR/pruned_transducer_stateless7/train.py

@@ -850,7 +850,7 @@ def train_one_epoch(
             # of the grad scaler is configurable, but we can't configure it to have different
             # behavior depending on the current grad scale.
             cur_grad_scale = scaler._scale.item()
-            if cur_grad_scale < 1.0:
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
                 scaler.update(cur_grad_scale * 2.0)
             if cur_grad_scale < 0.01:
                 logging.warn("Grad scale is small: {cur_grad_scale}")