Saving train state of step 1000

config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "./distil-large-v3-init",
+  "_name_or_path": "openai/whisper-large-v3",
   "activation_dropout": 0.0,
   "activation_function": "gelu",
   "apply_spec_augment": false,

distil-whisper/events.out.tfevents.1713341751.mycena-3090.144385.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca47056266d431f63008e943b3644392c7d27bbc0ee4e0e662135e39b3914d04
+size 12458

run_distillation.py

@@ -77,6 +77,9 @@ def chinese_wer(ref, hyp):
     返回:
     float: 計算出的 WER
     """
+    if type(ref) == list and type(hyp) == list:
+        ref = "".join(ref)
+        hyp = "".join(hyp)
     # 將字符串分割成字符列表
     ref_chars = list(ref.replace(" ", ""))
     hyp_chars = list(hyp.replace(" ", ""))
@@ -1297,10 +1300,8 @@ def main():
         # we do not want to group tokens when computing the metrics
         label_str = tokenizer.batch_decode(labels, skip_special_tokens=True)
         # !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!1
-        print("pred_str: ", pred_str)
-        print("label_str: ", label_str)
-        print("!!!!!!!!!!!!!!!!!!!!!!!!!")
-        wer_ortho = 100 * metric.compute(predictions=pred_str, references=label_str)
+
+        # wer_ortho = 100 * metric.compute(predictions=pred_str, references=label_str)
         wer_ortho = 100 * chinese_wer(pred_str, label_str)
 
         # normalize everything and re-compute the WER