feat: add onnx model
Browse files- .gitignore +1 -0
- src/demo.py +93 -27
- src/init_model.py +24 -11
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
./custom_model
|
src/demo.py
CHANGED
@@ -1,58 +1,124 @@
|
|
1 |
-
|
2 |
-
from transformers import AutoConfig, PretrainedConfig, PreTrainedModel, AutoModel
|
3 |
from transformers.pipelines import PIPELINE_REGISTRY
|
|
|
4 |
|
5 |
-
|
6 |
-
|
|
|
|
|
7 |
|
8 |
-
#
|
9 |
-
|
|
|
10 |
|
11 |
-
|
12 |
-
class ONNXBaseModel(AutoModel):
|
13 |
-
config_class = ONNXBaseConfig
|
14 |
|
|
|
15 |
class ONNXBaseModel(PreTrainedModel):
|
16 |
config_class = ONNXBaseConfig
|
17 |
-
|
18 |
-
def __init__(self, config):
|
19 |
super().__init__(config)
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
AutoModel.register(ONNXBaseConfig, ONNXBaseModel)
|
26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
from transformers.pipelines import Pipeline
|
28 |
|
29 |
class ONNXBasePipeline(Pipeline):
|
30 |
def __init__(self, model, **kwargs):
|
|
|
31 |
super().__init__(model=model, **kwargs)
|
32 |
|
33 |
def _sanitize_parameters(self, **kwargs):
|
34 |
return {}, {}, {}
|
35 |
|
36 |
-
def preprocess(self,
|
37 |
-
return
|
38 |
|
39 |
-
def _forward(self,
|
40 |
-
|
|
|
|
|
41 |
|
42 |
def postprocess(self, model_outputs):
|
43 |
return model_outputs
|
44 |
|
45 |
PIPELINE_REGISTRY.register_pipeline(
|
46 |
-
task=
|
47 |
-
pipeline_class=ONNXBasePipeline
|
|
|
48 |
)
|
49 |
|
50 |
-
|
51 |
from transformers import pipeline
|
52 |
|
53 |
-
|
54 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
|
56 |
-
# 使用 pipeline
|
57 |
-
result = onnx_pipeline("Your input data here")
|
58 |
-
print(result)
|
|
|
1 |
+
from transformers import PretrainedConfig, PreTrainedModel, AutoConfig, AutoModel, modeling_utils
|
|
|
2 |
from transformers.pipelines import PIPELINE_REGISTRY
|
3 |
+
from huggingface_hub import hf_hub_download
|
4 |
|
5 |
+
import onnxruntime as ort
|
6 |
+
import torch
|
7 |
+
import os
|
8 |
+
import torch.nn as nn
|
9 |
|
10 |
+
# 1. register AutoConfig
|
11 |
+
class ONNXBaseConfig(PretrainedConfig):
|
12 |
+
model_type = 'onnx-base'
|
13 |
|
14 |
+
AutoConfig.register('onnx-base', ONNXBaseConfig)
|
|
|
|
|
15 |
|
16 |
+
# 2. register AutoModel
|
17 |
class ONNXBaseModel(PreTrainedModel):
|
18 |
config_class = ONNXBaseConfig
|
19 |
+
def __init__(self, config, base_path=None):
|
|
|
20 |
super().__init__(config)
|
21 |
+
if base_path:
|
22 |
+
model_path = base_path + '/' + config.model_path
|
23 |
+
if os.path.exists(model_path):
|
24 |
+
self.session = ort.InferenceSession(model_path)
|
25 |
+
|
26 |
+
def forward(self, input=None, **kwargs):
|
27 |
+
outs = self.session.run(None, {'input': input})
|
28 |
+
return outs
|
29 |
+
|
30 |
+
@classmethod
|
31 |
+
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
|
32 |
+
config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
|
33 |
+
is_local = os.path.isdir(pretrained_model_name_or_path)
|
34 |
+
if is_local:
|
35 |
+
base_path = pretrained_model_name_or_path
|
36 |
+
else:
|
37 |
+
config_path = hf_hub_download(repo_id=pretrained_model_name_or_path, filename='config.json')
|
38 |
+
base_path = os.path.dirname(config_path)
|
39 |
+
hf_hub_download(repo_id=pretrained_model_name_or_path, filename=config.model_path)
|
40 |
+
return cls(config, base_path=base_path)
|
41 |
+
|
42 |
+
@property
|
43 |
+
def device(self):
|
44 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
45 |
+
return torch.device(device)
|
46 |
|
47 |
AutoModel.register(ONNXBaseConfig, ONNXBaseModel)
|
48 |
|
49 |
+
# option: save config to path
|
50 |
+
local_model_path = './custom_model'
|
51 |
+
config = ONNXBaseConfig(model_path='model.onnx',
|
52 |
+
id2label={0: 'label_0', 1: 'label_1'},
|
53 |
+
label2id={0: 'label_1', 1: 'label_0'})
|
54 |
+
model = ONNXBaseModel(config, base_path='./custom_mode')
|
55 |
+
config.save_pretrained(local_model_path)
|
56 |
+
# make sure have model_type
|
57 |
+
import json
|
58 |
+
config_path = local_model_path + '/config.json'
|
59 |
+
with open(config_path, 'r') as f:
|
60 |
+
config_data = json.load(f)
|
61 |
+
config_data['model_type'] = 'onnx-base'
|
62 |
+
del config_data['transformers_version']
|
63 |
+
with open(config_path, 'w') as f:
|
64 |
+
json.dump(config_data, f, indent=2)
|
65 |
+
|
66 |
+
# save onnx
|
67 |
+
dummy_input = torch.tensor([[[[1, 2, 3], [4, 5, 6], [7, 8, 9]]]], dtype=torch.float32)
|
68 |
+
onnx_file_path = './custom_model' + '/' + 'model.onnx'
|
69 |
+
class ZeroModel(nn.Module):
|
70 |
+
def __init__(self):
|
71 |
+
super(ZeroModel, self).__init__()
|
72 |
+
def forward(self, x):
|
73 |
+
return torch.zeros_like(x)
|
74 |
+
zero_model = ZeroModel()
|
75 |
+
torch.onnx.export(zero_model, dummy_input, onnx_file_path,
|
76 |
+
input_names=['input'], output_names=['output'],
|
77 |
+
dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}})
|
78 |
+
|
79 |
+
|
80 |
+
# 2. register Pipeline
|
81 |
from transformers.pipelines import Pipeline
|
82 |
|
83 |
class ONNXBasePipeline(Pipeline):
|
84 |
def __init__(self, model, **kwargs):
|
85 |
+
self.device_id = kwargs['device']
|
86 |
super().__init__(model=model, **kwargs)
|
87 |
|
88 |
def _sanitize_parameters(self, **kwargs):
|
89 |
return {}, {}, {}
|
90 |
|
91 |
+
def preprocess(self, input):
|
92 |
+
return {'input': input}
|
93 |
|
94 |
+
def _forward(self, model_input):
|
95 |
+
with torch.no_grad():
|
96 |
+
outputs = self.model(**model_input)
|
97 |
+
return outputs
|
98 |
|
99 |
def postprocess(self, model_outputs):
|
100 |
return model_outputs
|
101 |
|
102 |
PIPELINE_REGISTRY.register_pipeline(
|
103 |
+
task='onnx-base',
|
104 |
+
pipeline_class=ONNXBasePipeline,
|
105 |
+
pt_model=ONNXBaseModel
|
106 |
)
|
107 |
|
108 |
+
# 4. show how to use
|
109 |
from transformers import pipeline
|
110 |
|
111 |
+
pipe = pipeline(
|
112 |
+
task='onnx-base',
|
113 |
+
model='m3/onnx-base',
|
114 |
+
batch_size=10,
|
115 |
+
device='cuda',
|
116 |
+
)
|
117 |
+
|
118 |
+
dummy_input = torch.tensor([[[[1, 2, 3], [4, 5, 6], [7, 8, 9]]]], dtype=torch.float32)
|
119 |
+
input_data = dummy_input.numpy()
|
120 |
+
result = pipe(
|
121 |
+
inputs=input_data, device='cuda',
|
122 |
+
)
|
123 |
+
print(result)
|
124 |
|
|
|
|
|
|
src/init_model.py
CHANGED
@@ -1,18 +1,31 @@
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
import torch.onnx
|
4 |
-
|
5 |
-
# Define a simple model
|
6 |
-
class SimpleModel(nn.Module):
|
7 |
def __init__(self):
|
8 |
-
super(
|
9 |
-
self.fc = nn.Linear(1, 1)
|
10 |
|
11 |
def forward(self, x):
|
12 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
-
# Instantiate and export the model
|
15 |
-
model = SimpleModel()
|
16 |
-
dummy_input = torch.randn(1, 10)
|
17 |
-
onnx_path = "../model.onnx"
|
18 |
-
torch.onnx.export(model, dummy_input, onnx_path, input_names=['input'], output_names=['output'])
|
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
import torch.onnx
|
4 |
+
class BaseModel(nn.Module):
|
|
|
|
|
5 |
def __init__(self):
|
6 |
+
super(BaseModel, self).__init__()
|
|
|
7 |
|
8 |
def forward(self, x):
|
9 |
+
return torch.zeros_like(x)
|
10 |
+
|
11 |
+
# create a model
|
12 |
+
model = BaseModel()
|
13 |
+
|
14 |
+
dummy_input = torch.tensor([[[[1, 2, 3], [4, 5, 6], [7, 8, 9]]]], dtype=torch.float32)
|
15 |
+
|
16 |
+
onnx_file_path = "model.onnx"
|
17 |
+
torch.onnx.export(model, dummy_input, onnx_file_path,
|
18 |
+
input_names=['input'], output_names=['output'],
|
19 |
+
dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}})
|
20 |
+
|
21 |
+
print(f"Model has been exported to {onnx_file_path}")
|
22 |
+
|
23 |
+
import onnx
|
24 |
+
import onnxruntime as ort
|
25 |
+
onnx_model = onnx.load(onnx_file_path)
|
26 |
+
onnx.checker.check_model(onnx_model)
|
27 |
+
ort_session = ort.InferenceSession(onnx_file_path)
|
28 |
+
input_data = dummy_input.numpy()
|
29 |
+
outputs = ort_session.run(None, {'input': input_data})
|
30 |
+
print("Model output:", outputs)
|
31 |
|
|
|
|
|
|
|
|
|
|