Upload convert_bert4keras_tf_to_pytorch.py

convert_bert4keras_tf_to_pytorch.py

@@ -0,0 +1,28 @@
+import os
+import tensorflow as tf
+import torch
+from collections import OrderedDict
+
+tf_checkpoint_path = "chinese_GAU-alpha-char_L-24_H-768-tf/bert_model.ckpt"
+tf_path = os.path.abspath(tf_checkpoint_path)
+init_vars = tf.train.list_variables(tf_path)
+arrays = []
+
+pytorch_state_dict = OrderedDict()
+for name, shape in init_vars:
+    array = tf.train.load_variable(tf_path, name)
+    new_name = (
+        name.replace("GAU_alpha", "gau_alpha")
+        .replace("bert", "gau_alpha")
+        .replace("/", ".")
+        .replace("layer_", "layer.")
+        .replace("kernel", "weight")
+        .replace("gamma", "weight")
+    )
+    if "embeddings" in new_name:
+        new_name = new_name + ".weight"
+    if "_dense" in new_name:
+        array = array.T
+    pytorch_state_dict[new_name] = torch.from_numpy(array)
+    
+torch.save(pytorch_state_dict, "pytorch_model.bin")