[FIX] clip overflowing probs (#1335)

doctr/models/recognition/master/pytorch.py

@@ -259,7 +259,7 @@ def __call__(
             for encoded_seq in out_idxs.cpu().numpy()
         ]
 
-        return list(zip(word_values, probs.numpy().tolist()))
+        return list(zip(word_values, probs.numpy().clip(0, 1).tolist()))
 
 
 def _master(

doctr/models/recognition/master/tensorflow.py

@@ -260,7 +260,7 @@ def __call__(
         decoded_strings_pred = tf.sparse.to_dense(decoded_strings_pred.to_sparse(), default_value="not valid")[:, 0]
         word_values = [word.decode() for word in decoded_strings_pred.numpy().tolist()]
 
-        return list(zip(word_values, probs.numpy().tolist()))
+        return list(zip(word_values, probs.numpy().clip(0, 1).tolist()))
 
 
 def _master(arch: str, pretrained: bool, backbone_fn, pretrained_backbone: bool = True, **kwargs: Any) -> MASTER:

doctr/models/recognition/parseq/pytorch.py

@@ -401,7 +401,7 @@ def __call__(
             for encoded_seq in out_idxs.cpu().numpy()
         ]
         # compute probabilties for each word up to the EOS token
-        probs = [preds_prob[i, : len(word)].mean().item() for i, word in enumerate(word_values)]
+        probs = [preds_prob[i, : len(word)].clip(0, 1).mean().item() for i, word in enumerate(word_values)]
 
         return list(zip(word_values, probs))
 

doctr/models/recognition/parseq/tensorflow.py

@@ -432,7 +432,7 @@ def __call__(
         word_values = [word.decode() for word in decoded_strings_pred.numpy().tolist()]
 
         # compute probabilties for each word up to the EOS token
-        probs = [preds_prob[i, : len(word)].numpy().mean().item() for i, word in enumerate(word_values)]
+        probs = [preds_prob[i, : len(word)].numpy().clip(0, 1).mean().item() for i, word in enumerate(word_values)]
 
         return list(zip(word_values, probs))
 

doctr/models/recognition/sar/pytorch.py

@@ -323,7 +323,7 @@ def __call__(
             for encoded_seq in out_idxs.detach().cpu().numpy()
         ]
 
-        return list(zip(word_values, probs.numpy().tolist()))
+        return list(zip(word_values, probs.numpy().clip(0, 1).tolist()))
 
 
 def _sar(

doctr/models/recognition/sar/tensorflow.py

@@ -362,7 +362,7 @@ def __call__(
         decoded_strings_pred = tf.sparse.to_dense(decoded_strings_pred.to_sparse(), default_value="not valid")[:, 0]
         word_values = [word.decode() for word in decoded_strings_pred.numpy().tolist()]
 
-        return list(zip(word_values, probs.numpy().tolist()))
+        return list(zip(word_values, probs.numpy().clip(0, 1).tolist()))
 
 
 def _sar(

doctr/models/recognition/vitstr/pytorch.py

@@ -167,7 +167,7 @@ def __call__(
             for encoded_seq in out_idxs.cpu().numpy()
         ]
         # compute probabilties for each word up to the EOS token
-        probs = [preds_prob[i, : len(word)].mean().item() for i, word in enumerate(word_values)]
+        probs = [preds_prob[i, : len(word)].clip(0, 1).mean().item() for i, word in enumerate(word_values)]
 
         return list(zip(word_values, probs))
 

doctr/models/recognition/vitstr/tensorflow.py

@@ -175,7 +175,7 @@ def __call__(
         word_values = [word.decode() for word in decoded_strings_pred.numpy().tolist()]
 
         # compute probabilties for each word up to the EOS token
-        probs = [preds_prob[i, : len(word)].numpy().mean().item() for i, word in enumerate(word_values)]
+        probs = [preds_prob[i, : len(word)].numpy().clip(0, 1).mean().item() for i, word in enumerate(word_values)]
 
         return list(zip(word_values, probs))