Felix Marty
commited on
Commit
•
be527a9
1
Parent(s):
f75daf5
working version?
Browse files- onnx_export.py +58 -54
onnx_export.py
CHANGED
@@ -4,9 +4,7 @@ from optimum.exporters.onnx import OnnxConfigWithPast, export, validate_model_ou
|
|
4 |
|
5 |
from tempfile import TemporaryDirectory
|
6 |
|
7 |
-
from transformers import AutoConfig, is_torch_available
|
8 |
-
|
9 |
-
from transformers import AutoConfig
|
10 |
|
11 |
from pathlib import Path
|
12 |
|
@@ -29,55 +27,54 @@ def previous_pr(api: "HfApi", model_id: str, pr_title: str) -> Optional["Discuss
|
|
29 |
return discussion
|
30 |
|
31 |
def convert_onnx(model_id: str, task: str, folder: str):
|
32 |
-
model_class = TasksManager.get_model_class_for_task(task)
|
33 |
-
config = AutoConfig.from_pretrained(model_id)
|
34 |
-
model = model_class.from_config(config)
|
35 |
-
|
36 |
-
device = "cpu" # ?
|
37 |
-
|
38 |
-
# Dynamic axes aren't supported for YOLO-like models. This means they cannot be exported to ONNX on CUDA devices.
|
39 |
-
# See: https://github.com/ultralytics/yolov5/pull/8378
|
40 |
-
if model.__class__.__name__.startswith("Yolos") and device != "cpu":
|
41 |
-
return
|
42 |
-
|
43 |
-
onnx_config_class_constructor = TasksManager.get_exporter_config_constructor(model_type=config.model_type, exporter="onnx", task=task, model_name=model_id)
|
44 |
-
onnx_config = onnx_config_class_constructor(model.config)
|
45 |
-
|
46 |
-
# We need to set this to some value to be able to test the outputs values for batch size > 1.
|
47 |
-
if (
|
48 |
-
isinstance(onnx_config, OnnxConfigWithPast)
|
49 |
-
and getattr(model.config, "pad_token_id", None) is None
|
50 |
-
and task == "sequence-classification"
|
51 |
-
):
|
52 |
-
model.config.pad_token_id = 0
|
53 |
-
|
54 |
-
if is_torch_available():
|
55 |
-
from optimum.exporters.onnx.utils import TORCH_VERSION
|
56 |
-
|
57 |
-
if not onnx_config.is_torch_support_available:
|
58 |
-
print(
|
59 |
-
"Skipping due to incompatible PyTorch version. Minimum required is"
|
60 |
-
f" {onnx_config.MIN_TORCH_VERSION}, got: {TORCH_VERSION}"
|
61 |
-
)
|
62 |
-
|
63 |
-
onnx_inputs, onnx_outputs = export(
|
64 |
-
model, onnx_config, onnx_config.DEFAULT_ONNX_OPSET, Path(folder), device=device
|
65 |
-
)
|
66 |
-
atol = onnx_config.ATOL_FOR_VALIDATION
|
67 |
-
if isinstance(atol, dict):
|
68 |
-
atol = atol[task.replace("-with-past", "")]
|
69 |
-
validate_model_outputs(
|
70 |
-
onnx_config,
|
71 |
-
model,
|
72 |
-
Path(folder),
|
73 |
-
onnx_outputs,
|
74 |
-
atol,
|
75 |
-
)
|
76 |
-
|
77 |
-
# TODO: iterate in folder and add all
|
78 |
-
operations = [CommitOperationAdd(path_in_repo=local.split("/")[-1], path_or_fileobj=local) for local in local_filenames]
|
79 |
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
|
83 |
def convert(api: "HfApi", model_id: str, task:str, force: bool=False) -> Optional["CommitInfo"]:
|
@@ -98,7 +95,14 @@ def convert(api: "HfApi", model_id: str, task:str, force: bool=False) -> Optiona
|
|
98 |
new_pr = pr
|
99 |
raise Exception(f"Model {model_id} already has an open PR check out {url}")
|
100 |
else:
|
101 |
-
convert_onnx(model_id, task, folder)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
102 |
finally:
|
103 |
shutil.rmtree(folder)
|
104 |
return new_pr
|
@@ -113,12 +117,12 @@ if __name__ == "__main__":
|
|
113 |
"""
|
114 |
parser = argparse.ArgumentParser(description=DESCRIPTION)
|
115 |
parser.add_argument(
|
116 |
-
"model_id",
|
117 |
type=str,
|
118 |
help="The name of the model on the hub to convert. E.g. `gpt2` or `facebook/wav2vec2-base-960h`",
|
119 |
)
|
120 |
parser.add_argument(
|
121 |
-
"task",
|
122 |
type=str,
|
123 |
help="The task the model is performing",
|
124 |
)
|
|
|
4 |
|
5 |
from tempfile import TemporaryDirectory
|
6 |
|
7 |
+
from transformers import AutoConfig, AutoTokenizer, is_torch_available
|
|
|
|
|
8 |
|
9 |
from pathlib import Path
|
10 |
|
|
|
27 |
return discussion
|
28 |
|
29 |
def convert_onnx(model_id: str, task: str, folder: str):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
+
# Allocate the model
|
32 |
+
model = TasksManager.get_model_from_task(task, model_id, framework="pt")
|
33 |
+
model_type = model.config.model_type.replace("_", "-")
|
34 |
+
model_name = getattr(model, "name", None)
|
35 |
+
|
36 |
+
onnx_config_constructor = TasksManager.get_exporter_config_constructor(
|
37 |
+
model_type, "onnx", task=task, model_name=model_name
|
38 |
+
)
|
39 |
+
onnx_config = onnx_config_constructor(model.config)
|
40 |
+
|
41 |
+
needs_pad_token_id = (
|
42 |
+
isinstance(onnx_config, OnnxConfigWithPast)
|
43 |
+
and getattr(model.config, "pad_token_id", None) is None
|
44 |
+
and task in ["sequence_classification"]
|
45 |
+
)
|
46 |
+
if needs_pad_token_id:
|
47 |
+
#if args.pad_token_id is not None:
|
48 |
+
# model.config.pad_token_id = args.pad_token_id
|
49 |
+
try:
|
50 |
+
tok = AutoTokenizer.from_pretrained(model_id)
|
51 |
+
model.config.pad_token_id = tok.pad_token_id
|
52 |
+
except Exception:
|
53 |
+
raise ValueError(
|
54 |
+
"Could not infer the pad token id, which is needed in this case, please provide it with the --pad_token_id argument"
|
55 |
+
)
|
56 |
+
|
57 |
+
# Ensure the requested opset is sufficient
|
58 |
+
opset = onnx_config.DEFAULT_ONNX_OPSET
|
59 |
+
|
60 |
+
output = Path(folder).joinpath("model.onnx")
|
61 |
+
onnx_inputs, onnx_outputs = export(
|
62 |
+
model,
|
63 |
+
onnx_config,
|
64 |
+
opset,
|
65 |
+
output,
|
66 |
+
)
|
67 |
+
|
68 |
+
atol = onnx_config.ATOL_FOR_VALIDATION
|
69 |
+
if isinstance(atol, dict):
|
70 |
+
atol = atol[task.replace("-with-past", "")]
|
71 |
+
|
72 |
+
validate_model_outputs(onnx_config, model, output, onnx_outputs, atol)
|
73 |
+
print(f"All good, model saved at: {output}")
|
74 |
+
|
75 |
+
operations = [CommitOperationAdd(path_in_repo=file_name, path_or_fileobj=os.path.join(folder, file_name)) for file_name in os.listdir(folder)]
|
76 |
+
|
77 |
+
return operations
|
78 |
|
79 |
|
80 |
def convert(api: "HfApi", model_id: str, task:str, force: bool=False) -> Optional["CommitInfo"]:
|
|
|
95 |
new_pr = pr
|
96 |
raise Exception(f"Model {model_id} already has an open PR check out {url}")
|
97 |
else:
|
98 |
+
operations = convert_onnx(model_id, task, folder)
|
99 |
+
|
100 |
+
new_pr = api.create_commit(
|
101 |
+
repo_id=model_id,
|
102 |
+
operations=operations,
|
103 |
+
commit_message=pr_title,
|
104 |
+
create_pr=True,
|
105 |
+
)
|
106 |
finally:
|
107 |
shutil.rmtree(folder)
|
108 |
return new_pr
|
|
|
117 |
"""
|
118 |
parser = argparse.ArgumentParser(description=DESCRIPTION)
|
119 |
parser.add_argument(
|
120 |
+
"--model_id",
|
121 |
type=str,
|
122 |
help="The name of the model on the hub to convert. E.g. `gpt2` or `facebook/wav2vec2-base-960h`",
|
123 |
)
|
124 |
parser.add_argument(
|
125 |
+
"--task",
|
126 |
type=str,
|
127 |
help="The task the model is performing",
|
128 |
)
|