Modify convert (#395)

dbiir · Oct 19, 2023 · 4f9d551 · 4f9d551
1 parent 3e82267
commit 4f9d551
Show file tree

Hide file tree

Showing 7 changed files with 66 additions and 66 deletions.
diff --git a/scripts/convert_bert_from_huggingface_to_uer.py b/scripts/convert_bert_from_huggingface_to_uer.py
@@ -72,8 +72,8 @@ def main():
         output_model["target.sp.linear_2.bias"] = input_model["cls.seq_relationship.bias"]
     output_model["target.mlm.linear_1.weight"] = input_model["cls.predictions.transform.dense.weight"]
     output_model["target.mlm.linear_1.bias"] = input_model["cls.predictions.transform.dense.bias"]
-    output_model["target.layer_norm.gamma"] = input_model["cls.predictions.transform.LayerNorm.weight"]
-    output_model["target.layer_norm.beta"] = input_model["cls.predictions.transform.LayerNorm.bias"]
+    output_model["target.mlm.layer_norm.gamma"] = input_model["cls.predictions.transform.LayerNorm.weight"]
+    output_model["target.mlm.layer_norm.beta"] = input_model["cls.predictions.transform.LayerNorm.bias"]
     output_model["target.mlm.linear_2.weight"] = input_model["cls.predictions.decoder.weight"]
     output_model["target.mlm.linear_2.bias"] = input_model["cls.predictions.bias"]
 

diff --git a/scripts/convert_bert_from_original_tf_to_uer.py b/scripts/convert_bert_from_original_tf_to_uer.py
@@ -93,8 +93,8 @@ def main():
         output_model["target.sp.linear_2.bias"] = input_model["cls/seq_relationship/output_bias"]
     output_model["target.mlm.linear_1.weight"] = input_model["cls/predictions/transform/dense/kernel"]
     output_model["target.mlm.linear_1.bias"] = input_model["cls/predictions/transform/dense/bias"]
-    output_model["target.layer_norm.gamma"] = input_model["cls/predictions/transform/LayerNorm/gamma"]
-    output_model["target.layer_norm.beta"] = input_model["cls/predictions/transform/LayerNorm/beta"]
+    output_model["target.mlm.layer_norm.gamma"] = input_model["cls/predictions/transform/LayerNorm/gamma"]
+    output_model["target.mlm.layer_norm.beta"] = input_model["cls/predictions/transform/LayerNorm/beta"]
     output_model["target.mlm.linear_2.weight"] = input_model["bert/embeddings/word_embeddings"]
     output_model["target.mlm.linear_2.bias"] = input_model["cls/predictions/output_bias"]
 

diff --git a/scripts/convert_bert_from_uer_to_huggingface.py b/scripts/convert_bert_from_uer_to_huggingface.py
@@ -70,8 +70,8 @@ def main():
         output_model["cls.seq_relationship.bias"] = input_model["target.sp.linear_2.bias"]
     output_model["cls.predictions.transform.dense.weight"] = input_model["target.mlm.linear_1.weight"]
     output_model["cls.predictions.transform.dense.bias"] = input_model["target.mlm.linear_1.bias"]
-    output_model["cls.predictions.transform.LayerNorm.weight"] = input_model["target.layer_norm.gamma"]
-    output_model["cls.predictions.transform.LayerNorm.bias"] = input_model["target.layer_norm.beta"]
+    output_model["cls.predictions.transform.LayerNorm.weight"] = input_model["target.mlm.layer_norm.gamma"]
+    output_model["cls.predictions.transform.LayerNorm.bias"] = input_model["target.mlm.layer_norm.beta"]
     output_model["cls.predictions.decoder.weight"] = input_model["target.mlm.linear_2.weight"]
     output_model["cls.predictions.bias"] = input_model["target.mlm.linear_2.bias"]
 

diff --git a/scripts/convert_bert_from_uer_to_original_tf.py b/scripts/convert_bert_from_uer_to_original_tf.py
@@ -86,8 +86,8 @@ def main():
         output_model["cls/seq_relationship/output_bias"] = input_model["target.sp.linear_2.bias"]
     output_model["cls/predictions/transform/dense/kernel"] = input_model["target.mlm.linear_1.weight"]
     output_model["cls/predictions/transform/dense/bias"] = input_model["target.mlm.linear_1.bias"]
-    output_model["cls/predictions/transform/LayerNorm/gamma"] = input_model["target.layer_norm.gamma"]
-    output_model["cls/predictions/transform/LayerNorm/beta"] = input_model["target.layer_norm.beta"]
+    output_model["cls/predictions/transform/LayerNorm/gamma"] = input_model["target.mlm.layer_norm.gamma"]
+    output_model["cls/predictions/transform/LayerNorm/beta"] = input_model["target.mlm.layer_norm.beta"]
     output_model["cls/predictions/output_bias"] = input_model["target.mlm.linear_2.bias"]
 
     tf_vars = []

diff --git a/scripts/convert_t5_from_uer_to_huggingface.py b/scripts/convert_t5_from_uer_to_huggingface.py
@@ -34,69 +34,69 @@
     input_model["target.lm.output_layer.weight"]
 
 for i in range(args.layers_num):
-    output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.0.weight"] = \
-        input_model["encoder.block." + str(i) + ".layer.0.SelfAttention.q.weight"]
-    output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.1.weight"] = \
-        input_model["encoder.block." + str(i) + ".layer.0.SelfAttention.k.weight"]
-    output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.2.weight"] = \
-        input_model["encoder.block." + str(i) + ".layer.0.SelfAttention.v.weight"]
-    output_model["encoder.transformer." + str(i) + ".self_attn.final_linear.weight"] = \
-        input_model["encoder.block." + str(i) + ".layer.0.SelfAttention.o.weight"]
-    output_model["encoder.transformer." + str(i) + ".layer_norm_1.weight"] = \
-        input_model["encoder.block." + str(i) + ".layer.0.layer_norm.weight"]
+    output_model["encoder.block." + str(i) + ".layer.0.SelfAttention.q.weight"] = \
+        input_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.0.weight"]
+    output_model["encoder.block." + str(i) + ".layer.0.SelfAttention.k.weight"] = \
+        input_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.1.weight"]
+    output_model["encoder.block." + str(i) + ".layer.0.SelfAttention.v.weight"] = \
+        input_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.2.weight"]
+    output_model["encoder.block." + str(i) + ".layer.0.SelfAttention.o.weight"] = \
+        input_model["encoder.transformer." + str(i) + ".self_attn.final_linear.weight"]
+    output_model["encoder.block." + str(i) + ".layer.0.layer_norm.weight"] = \
+        input_model["encoder.transformer." + str(i) + ".layer_norm_1.weight"]
 
     if args.type == "t5-v1_1":
-        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_gate.weight"] = \
-            input_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wi_0.weight"]
-        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"] = \
-            input_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wi_1.weight"]
-        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"] = \
-            input_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wo.weight"]
+        output_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wi_0.weight"] = \
+            input_model["encoder.transformer." + str(i) + ".feed_forward.linear_gate.weight"]
+        output_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wi_1.weight"] = \
+            input_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"]
+        output_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wo.weight"] = \
+            input_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"]
     else:
-        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"] = \
-            input_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wi.weight"]
-        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"] = \
-            input_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wo.weight"]
-    output_model["encoder.transformer." + str(i) + ".layer_norm_2.weight"] = \
-        input_model["encoder.block." + str(i) + ".layer.1.layer_norm.weight"]
+        output_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wi.weight"] = \
+            input_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"]
+        output_model["encoder.block." + str(i) + ".layer.1.DenseReluDense.wo.weight"] = \
+            input_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"]
+    output_model["encoder.block." + str(i) + ".layer.1.layer_norm.weight"] = \
+        input_model["encoder.transformer." + str(i) + ".layer_norm_2.weight"]
 
 for i in range(args.decoder_layers_num):
-    output_model["decoder.transformer_decoder." + str(i) + ".self_attn.linear_layers.0.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.0.SelfAttention.q.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".self_attn.linear_layers.1.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.0.SelfAttention.k.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".self_attn.linear_layers.2.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.0.SelfAttention.v.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".self_attn.final_linear.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.0.SelfAttention.o.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".layer_norm_1.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.0.layer_norm.weight"]
+    output_model["decoder.block." + str(i) + ".layer.0.SelfAttention.q.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".self_attn.linear_layers.0.weight"]
+    output_model["decoder.block." + str(i) + ".layer.0.SelfAttention.k.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".self_attn.linear_layers.1.weight"]
+    output_model["decoder.block." + str(i) + ".layer.0.SelfAttention.v.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".self_attn.linear_layers.2.weight"]
+    output_model["decoder.block." + str(i) + ".layer.0.SelfAttention.o.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".self_attn.final_linear.weight"]
+    output_model["decoder.block." + str(i) + ".layer.0.layer_norm.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".layer_norm_1.weight"]
 
-    output_model["decoder.transformer_decoder." + str(i) + ".context_attn.linear_layers.0.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.q.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".context_attn.linear_layers.1.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.k.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".context_attn.linear_layers.2.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.v.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".context_attn.final_linear.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.o.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".layer_norm_2.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.1.layer_norm.weight"]
+    output_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.q.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".context_attn.linear_layers.0.weight"]
+    output_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.k.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".context_attn.linear_layers.1.weight"]
+    output_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.v.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".context_attn.linear_layers.2.weight"]
+    output_model["decoder.block." + str(i) + ".layer.1.EncDecAttention.o.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".context_attn.final_linear.weight"]
+    output_model["decoder.block." + str(i) + ".layer.1.layer_norm.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".layer_norm_2.weight"]
 
     if args.type == "t5-v1_1":
-        output_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_gate.weight"] = \
-            input_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wi_0.weight"]
-        output_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_1.weight"] = \
-            input_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wi_1.weight"]
-        output_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_2.weight"] = \
-            input_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wo.weight"]
+        output_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wi_0.weight"] = \
+            input_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_gate.weight"]
+        output_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wi_1.weight"] = \
+            input_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_1.weight"]
+        output_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wo.weight"] = \
+            input_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_2.weight"]
     else:
-        output_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_1.weight"] = \
-            input_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wi.weight"]
-        output_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_2.weight"] = \
-            input_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wo.weight"]
-    output_model["decoder.transformer_decoder." + str(i) + ".layer_norm_3.weight"] = \
-        input_model["decoder.block." + str(i) + ".layer.2.layer_norm.weight"]
+        output_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wi.weight"] = \
+            input_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_1.weight"]
+        output_model["decoder.block." + str(i) + ".layer.2.DenseReluDense.wo.weight"] = \
+            input_model["decoder.transformer_decoder." + str(i) + ".feed_forward.linear_2.weight"]
+    output_model["decoder.block." + str(i) + ".layer.2.layer_norm.weight"] = \
+        input_model["decoder.transformer_decoder." + str(i) + ".layer_norm_3.weight"]
 
 output_model["encoder.final_layer_norm.weight"] = \
     input_model["encoder.layer_norm.weight"]

diff --git a/scripts/convert_xlmroberta_from_huggingface_to_uer.py b/scripts/convert_xlmroberta_from_huggingface_to_uer.py
@@ -67,9 +67,9 @@
     input_model["lm_head.dense.weight"]
 output_model["target.mlm.linear_1.bias"] = \
     input_model["lm_head.dense.bias"]
-output_model["target.layer_norm.gamma"] = \
+output_model["target.mlm.layer_norm.gamma"] = \
     input_model["lm_head.layer_norm.weight"]
-output_model["target.layer_norm.beta"] = \
+output_model["target.mlm.layer_norm.beta"] = \
     input_model["lm_head.layer_norm.bias"]
 output_model["target.mlm.linear_2.weight"] = \
     input_model["lm_head.decoder.weight"]

diff --git a/scripts/convert_xlmroberta_from_uer_to_huggingface.py b/scripts/convert_xlmroberta_from_uer_to_huggingface.py
@@ -68,9 +68,9 @@
 output_model["lm_head.dense.bias"] = \
     input_model["target.mlm.linear_1.bias"]
 output_model["lm_head.layer_norm.weight"] = \
-    input_model["target.layer_norm.gamma"]
+    input_model["target.mlm.layer_norm.gamma"]
 output_model["lm_head.layer_norm.bias"] = \
-    input_model["target.layer_norm.beta"]
+    input_model["target.mlm.layer_norm.beta"]
 output_model["lm_head.decoder.weight"] = \
     input_model["target.mlm.linear_2.weight"]
 output_model["lm_head.decoder.bias"] = \