Create japanese_bert.py

text/japanese_bert.py

@@ -0,0 +1,26 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
+model = AutoModelForMaskedLM.from_pretrained("./bert/bert-base-japanese-v3").to(device)
+
+def get_bert_feature(text, word2ph):
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors='pt')
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = model(**inputs, output_hidden_states=True)
+        res = torch.cat(res['hidden_states'][-3:-2], -1)[0].cpu()
+    assert inputs['input_ids'].shape[-1] == len(word2ph)
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+
+    return phone_level_feature.T