add more debugging, shuffle dataset max train

join_datasets_asr_ctc.py

@@ -479,6 +479,11 @@ def load_raw_datasets(training_args, data_args):
                 f"{', '.join(raw_datasets['train'].column_names)}."
             )
 
+        # If dataset_seed is set, shuffle train
+        if data_args.dataset_seed is not None:
+            raw_datasets["train"] = raw_datasets["train"].shuffle(seed=data_args.dataset_seed)
+
+
         if data_args.max_train_samples is not None:
             raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
         other_columns_train = [col for col in raw_datasets["train"].column_names if col not in min_columns_train]
@@ -771,10 +776,6 @@ def preprocess_audio_datasets(raw_datasets, tokenizer, feature_extractor, traini
             input_columns=["input_length"],
         )
 
-    # If dataset_seed is set, shuffle train
-    if data_args.dataset_seed is not None:
-        vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(seed=data_args.dataset_seed)
-
     # TODO: Log sample of datasets in the right way (see wandb docs)
     pd_train = vectorized_datasets["train"].select(range(10)).to_pandas()
     pd_eval = vectorized_datasets["eval"].select(range(10)).to_pandas()
@@ -872,6 +873,21 @@ def main():
         data_args=data_args
     )
 
+    # Inspect datasets
+    logger.info("Inspect datasets")
+    avg = []
+    std = []
+    import numpy as np
+    for input_ in vectorized_datasets["train"][:10]["input_values"]:
+        avg.append(np.average(input_))
+        std.append(np.std(input_))
+    for input_ in vectorized_datasets["eval"][:10]["input_values"]:
+        avg.append(np.average(input_))
+        std.append(np.std(input_))
+    
+    logger.info(f"Average values: {avg}")
+    logger.info(f"Std values: {std}")
+
     # 7. Next, we can prepare the training.
     # Let's use word error rate (WER) as our evaluation metric,
     # instantiate a data collator and the trainer

join_datasets_asr_ctc_run.sh

@@ -30,6 +30,7 @@ python join_datasets_asr_ctc.py \
 	--mask_time_length="10" \
 	--mask_feature_prob="0.25" \
 	--mask_feature_length="64" \
+	--dataset_seed="42" \
 	--gradient_checkpointing \
 	--use_auth_token \
 	--preprocessing_only \

vocab.json

@@ -1 +1 @@
-{"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "ä": 27, "å": 28, "ö": 29, "|": 0, "[UNK]": 30, "[PAD]": 31}
+{"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "\u00e4": 27, "\u00e5": 28, "\u00f6": 29, "|": 0, "[UNK]": 30, "[PAD]": 31}