Merge pull request #7 from megagonlabs/feature/setup_model

add setup_model.py

Author: hiroshi-matsuda-rit

ginza_transformers/layers/hf_shim_custom.py

@@ -1,5 +1,4 @@
 import sys
-from typing import Any
 from io import BytesIO
 from pathlib import Path
 import srsly
@@ -15,6 +14,17 @@
 from transformers import AutoModel, AutoConfig, AutoTokenizer
 
 
+def override_hf_shims_to_bytes():
+    assert hf_shim.HFShim.to_bytes is not HFShimCustom.to_bytes
+    origin = hf_shim.HFShim.to_bytes
+    hf_shim.HFShim.to_bytes = HFShimCustom.to_bytes
+    return origin
+
+def recover_hf_shims_to_bytes(origin):
+    assert hf_shim.HFShim.to_bytes is HFShimCustom.to_bytes
+    hf_shim.HFShim.to_bytes = origin
+
+
 def override_hf_shims_from_bytes():
     assert hf_shim.HFShim.from_bytes is not HFShimCustom.from_bytes
     origin = hf_shim.HFShim.from_bytes
@@ -28,6 +38,44 @@ def recover_hf_shims_from_bytes(origin):
 
 class HFShimCustom(HFShim):
 
+    def to_bytes(self):
+        config = {}
+        tok_dict = {}
+        # weights_bytes = {}
+        tok_cfg = {}
+        trf_cfg = {}
+        hf_model = self._hfmodel
+        if hf_model.transformer is not None:
+            tok_dict = {}
+            config = hf_model.transformer.config.to_dict()
+            tokenizer = hf_model.tokenizer
+            with make_tempdir() as temp_dir:
+                if hasattr(tokenizer, "vocab_file"):
+                    vocab_file_name = tokenizer.vocab_files_names["vocab_file"]
+                    vocab_file_path = str((temp_dir / vocab_file_name).absolute())
+                    with open(vocab_file_path, "wb") as fileh:
+                        fileh.write(hf_model.vocab_file_contents)
+                    tokenizer.vocab_file = vocab_file_path
+                tokenizer.save_pretrained(str(temp_dir.absolute()))
+                for x in temp_dir.glob("**/*"):
+                    if x.is_file():
+                        tok_dict[x.name] = x.read_bytes()
+            filelike = BytesIO()
+            torch.save(self._model.state_dict(), filelike)
+            filelike.seek(0)
+            # weights_bytes = filelike.getvalue()
+        else:
+            tok_cfg = hf_model._init_tokenizer_config
+            trf_cfg = hf_model._init_transformer_config
+        msg = {
+            "config": config,
+            # "state": weights_bytes,
+            "tokenizer": tok_dict,
+            "_init_tokenizer_config": tok_cfg,
+            "_init_transformer_config": trf_cfg,
+        }
+        return srsly.msgpack_dumps(msg)
+
     def from_bytes(self, bytes_data):
         msg = srsly.msgpack_loads(bytes_data)
         config_dict = msg["config"]
@@ -62,34 +110,35 @@ def from_bytes(self, bytes_data):
                     with open(vocab_file_path, "rb") as fileh:
                         vocab_file_contents = fileh.read()
 
-            try:
+            ops = get_current_ops()
+            if ops.device_type == "cpu":
+                map_location = "cpu"
+            else:  # pragma: no cover
+                device_id = torch.cuda.current_device()
+                map_location = f"cuda:{device_id}"
+
+            if "state" in msg:
                 transformer = AutoModel.from_config(config)
-            except OSError as e:
+                filelike = BytesIO(msg["state"])
+                filelike.seek(0)
+                transformer.load_state_dict(torch.load(filelike, map_location=map_location))
+            else:
                 try:
-                    transformer = AutoModel.from_pretrained(config["_name_or_path"], local_files_only=True)
+                    transformer = AutoModel.from_pretrained(config._name_or_path, local_files_only=True)
                 except OSError as e2:
-                    print("trying to download model from huggingface hub:", config["_name_or_path"], "...", file=sys.stderr)
-                    transformer = AutoModel.from_pretrained(config["_name_or_path"])
+                    print("trying to download model from huggingface hub:", config._name_or_path, "...", file=sys.stderr)
+                    transformer = AutoModel.from_pretrained(config._name_or_path)
                     print("succeded", file=sys.stderr)
 
+            transformer.to(map_location)
+            self._model = transformer
             self._hfmodel = HFObjects(
                 tokenizer,
                 transformer,
                 vocab_file_contents,
                 SimpleFrozenDict(),
                 SimpleFrozenDict(),
             )
-            self._model = transformer
-            filelike = BytesIO(msg["state"])
-            filelike.seek(0)
-            ops = get_current_ops()
-            if ops.device_type == "cpu":
-                map_location = "cpu"
-            else:  # pragma: no cover
-                device_id = torch.cuda.current_device()
-                map_location = f"cuda:{device_id}"
-            self._model.load_state_dict(torch.load(filelike, map_location=map_location))
-            self._model.to(map_location)
         else:
             self._hfmodel = HFObjects(
                 None,

ginza_transformers/setup_model.py

@@ -0,0 +1,31 @@
+import sys
+
+import spacy
+
+from .layers.hf_shim_custom import override_hf_shims_to_bytes, recover_hf_shims_to_bytes
+
+
+def main():
+    org_spacy_model_path = sys.argv[1]
+    dst_spacy_model_path = sys.argv[2]
+    transformers_model_name = sys.argv[3]
+
+    nlp = spacy.load(org_spacy_model_path)
+    transformer = nlp.get_pipe("transformer")
+    for i, node in enumerate(transformer.model.walk()):
+        if node.shims:
+            break
+    else:
+        assert False
+    node.shims[0]._hfmodel.transformer.config._name_or_path = transformers_model_name
+    node.shims[0]._hfmodel.tokenizer.save_pretrained(transformers_model_name)
+    node.shims[0]._hfmodel.transformer.save_pretrained(transformers_model_name)
+    override_hf_shims_to_bytes()
+    try:
+        origin = nlp.to_disk(dst_spacy_model_path)
+    finally:
+        recover_hf_shims_to_bytes(origin)
+
+
+if __name__ == "__main__":
+    main()