Training in progress, step 1000

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e2570949865d0774ba0c804ba877390aaaf5741c0eedac5bcb3a4da58c87943e
+oid sha256:66c2866840783f9227a6fe5132558e2be3d9e0497ec1651e9cce3a9a6bf89696
 size 3086784957

run_speech_recognition_seq2seq_streaming.py

@@ -484,6 +484,33 @@ def main():
         batch["labels"] = tokenizer(input_str).input_ids
         return batch
 
+    def prepare_dataset_bayar(batch):
+        MAX_LENGTH = 225
+        MAX_AUDIO_DURATION = 30
+        sample = batch[audio_column_name][0]
+        DEFAULT_SAMPLING_RATE = sample['sampling_rate']
+
+        bs = len(batch[text_column_name])
+        result = {'input_features': [], 'labels': []}
+        list_arr, list_text, total = [], [], 0
+        for i in range(bs + 1):
+            if i == bs or total + batch[audio_column_name][i]['array'].shape[0] / DEFAULT_SAMPLING_RATE > MAX_AUDIO_DURATION:
+                if total == 0: continue # because it could be evenly distributed when i == bs
+                tokens = tokenizer((' '.join(list_text))).input_ids
+                if len(tokens) > MAX_LENGTH: continue # too long -> might mislead to not-aligning problem
+
+                result['input_features'].append(feature_extractor(np.concatenate(list_arr), sampling_rate=DEFAULT_SAMPLING_RATE).input_features[0])
+                result['labels'].append(tokens)
+#                result['input_length'].append(len(np.concatenate(list_arr)))
+                list_arr, list_text, total = [], [], 0
+            if i < bs:
+                duration = batch[audio_column_name][i]['array'].shape[0] / DEFAULT_SAMPLING_RATE
+                if duration > MAX_AUDIO_DURATION: continue
+                total += duration
+                list_arr.append(batch[audio_column_name][i]['array'])
+                list_text.append(batch[text_column_name][i])
+        return result
+
     def prepare_dataset_concatenated(batch):
         """
         - ignoring do_lower_case, do_remove_punctuation
@@ -541,7 +568,9 @@ def main():
 
     with training_args.main_process_first(desc="dataset map pre-processing"):
         vectorized_datasets = raw_datasets.map(
-            prepare_dataset,
+            prepare_dataset_bayar,
+            batched=True,
+            batch_size=64,
             remove_columns=raw_datasets_features,
         ).with_format("torch")
 
@@ -557,11 +586,6 @@ def main():
     def is_audio_in_length_range(length):
         return min_input_length < length < max_input_length
 
-    if training_args.do_train:
-        vectorized_datasets["train"] = vectorized_datasets["train"].filter(
-            is_audio_in_length_range,
-            input_columns=["input_length"],
-        )
 
     # 8. Load Metric
     metric = evaluate.load("wer")

runs/Dec20_04-28-50_129-146-123-136/events.out.tfevents.1671514074.129-146-123-136.2405293.0

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c537e2c595696a280619d99cf29d56f02069cb421afdadd8c2f364030881728e
-size 10865
+oid sha256:e53df992c9891d4248761877f6cbbdc67edcff1e69a48ce384178b8bed3cdef5
+size 11179

runs/Dec20_10-12-02_129-146-123-136/1671531744.4050705/events.out.tfevents.1671531744.129-146-123-136.3808886.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1ffebe9ce297b1b99ecd6dab7bbe24d7684379aea436f80a19a062561005daa
+size 5881

runs/Dec20_10-12-02_129-146-123-136/events.out.tfevents.1671531744.129-146-123-136.3808886.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e271be54ec7c2ee8d4f866c1eb16c070ac31894fc81fc91f4ef9cfc7e61b5c99
+size 10865

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:352b94eb50530c27c29e9ab71118514bfe55c7b1ed12d6de27277d35c58388ae
+oid sha256:b4e85ca367fa64bd161858b432c71b1ebc8ebc010d9d91fc0d176f9c5e1e9c0c
 size 4731