File size: 7,012 Bytes
89cf463
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
# -- coding: utf-8 --
# @Time : 2021/11/29
# @Author : ykk648
# @Project : https://github.com/ykk648/AI_power

"""
todo: io_binding https://onnxruntime.ai/docs/api/python/api_summary.html
"""
import os

import onnxruntime
import numpy as np
from cv2box import MyFpsCounter
import re
def get_output_info(onnx_session):
    output_name = []
    output_shape = []
    for node in onnx_session.get_outputs():
        output_name.append(node.name)
        output_shape.append(node.shape)
    return output_name, output_shape


def get_input_info(onnx_session):
    input_name = []
    input_shape = []
    for node in onnx_session.get_inputs():
        input_name.append(node.name)
        input_shape.append(node.shape)
    return input_name, input_shape


def get_input_feed(input_name, image_tensor):
    """
    Args:
        input_name:
        image_tensor: [image tensor, ...]
    Returns:
    """
    input_feed = {}
    for index, name in enumerate(input_name):
        input_feed[name] = image_tensor[index]
    return input_feed


class ONNXModel:
    def __init__(self, onnx_path, provider='gpu', debug=False, input_dynamic_shape=None, model_name=''):
        self.provider = provider
        trt_cache_path = './cache/' + str(self.provider) + '/' + str(model_name)
        if self.provider == 'gpu':
            self.providers = (
                "CUDAExecutionProvider",
                {'device_id': 0, }
            )
        elif self.provider == 'trt':
            os.makedirs(trt_cache_path, exist_ok=True)
            self.providers = (
                'TensorrtExecutionProvider',
                {'trt_engine_cache_enable': True, 'trt_engine_cache_path': trt_cache_path, 'trt_fp16_enable': False, }
            )
        elif self.provider == 'trt16':
            os.makedirs(trt_cache_path, exist_ok=True)
            self.providers = (
                'TensorrtExecutionProvider',
                {'trt_engine_cache_enable': True, 'trt_engine_cache_path': trt_cache_path, 'trt_fp16_enable': True,
                 'trt_dla_enable': False}
            )
        elif self.provider == 'trt8':
            os.makedirs(trt_cache_path, exist_ok=True)
            self.providers = (
                'TensorrtExecutionProvider',
                {'trt_engine_cache_enable': True, 'trt_int8_enable': 1, }
            )
        else:
            self.providers = "CPUExecutionProvider"
        #onnxruntime.set_default_logger_severity(2)
        session_options = onnxruntime.SessionOptions()
        session_options.log_severity_level = 3
        try:
            self.onnx_session = onnxruntime.InferenceSession(onnx_path, session_options, providers=[self.providers])
        except Exception as e:
            if type(e.args[0])==str and 'TensorRT EP could not deserialize engine from cache' in e.args[0]:
                res = re.match('.*TensorRT EP could not deserialize engine from cache: (.*)', e.args[0])
                os.remove(res.group(1))
                print('waiting generate new model...')
                self.onnx_session = onnxruntime.InferenceSession(onnx_path, session_options, providers=[self.providers])
            else:
                raise e
        print(model_name,self.onnx_session.get_providers())
        if 'trt' in self.provider:
            assert 'Tensorrt' in self.onnx_session.get_providers()[0], 'Tensorrt start failure'
        # sessionOptions.intra_op_num_threads = 3
        self.input_name, self.input_shape = get_input_info(self.onnx_session)
        self.output_name, self.output_shape = get_output_info(self.onnx_session)

        self.input_dynamic_shape = input_dynamic_shape

        if self.input_dynamic_shape is not None:
            self.input_dynamic_shape = self.input_dynamic_shape if isinstance(self.input_dynamic_shape, list) else [
                self.input_dynamic_shape]

        if debug:
            print('onnx version: {}'.format(onnxruntime.__version__))
            print("input_name:{}, shape:{}".format(self.input_name, self.input_shape))
            print("output_name:{}, shape:{}".format(self.output_name, self.output_shape))

        self.warm_up()
        self.speed_test()
        self.speed_test()
    def warm_up(self):
        if not self.input_dynamic_shape:
            try:
                self.forward([np.random.rand(*([1]+self.input_shape[i][1:])).astype(np.float32)
                              for i in range(len(self.input_shape))])
            except TypeError:
                print('Model may be dynamic, plz name the \'input_dynamic_shape\' !')
        else:
            self.forward([np.random.rand(*self.input_dynamic_shape[i]).astype(np.float32)
                          for i in range(len(self.input_shape))])
        print('Model warm up done !')

    def speed_test(self):
        if not self.input_dynamic_shape:
            input_tensor = [np.random.rand(*([1]+self.input_shape[i][1:])).astype(np.float32)
                            for i in range(len(self.input_shape))]
        else:
            input_tensor = [np.random.rand(*self.input_dynamic_shape[i]).astype(np.float32)
                            for i in range(len(self.input_shape))]

        with MyFpsCounter('[{}] onnx 10 times'.format(self.provider)) as mfc:
            for i in range(10):
                _ = self.forward(input_tensor)

    def forward(self, image_tensor_in, trans=False):
        """
        Args:
            image_tensor_in: image_tensor [image_tensor] [image_tensor_1, image_tensor_2]
            trans: apply trans for image_tensor or first image_tensor(list)
        Returns:
            model output
        """
        if not isinstance(image_tensor_in, list) or len(image_tensor_in) == 1:
            image_tensor_in = image_tensor_in[0] if isinstance(image_tensor_in, list) else image_tensor_in
            if trans:
                image_tensor_in = image_tensor_in.transpose(2, 0, 1)[np.newaxis, :]
            image_tensor_in = [np.ascontiguousarray(image_tensor_in)]
        else:
            # for multi input, only trans first tensor
            if trans:
                image_tensor_in[0] = image_tensor_in[0].transpose(2, 0, 1)[np.newaxis, :]
            image_tensor_in = [np.ascontiguousarray(image_tensor) for image_tensor in image_tensor_in]

        input_feed = get_input_feed(self.input_name, image_tensor_in)
        temp_result = self.onnx_session.run(self.output_name, input_feed=input_feed)
        if len(temp_result)==1:
            return temp_result
        else:
            while np.any(np.isnan(temp_result[0])) or np.any(np.isnan(temp_result[1])):
                temp_result = self.onnx_session.run(self.output_name, input_feed=input_feed)
            return temp_result

    def batch_forward(self, bach_image_tensor, trans=False):
        if trans:
            bach_image_tensor = bach_image_tensor.transpose(0, 3, 1, 2)
        input_feed = get_input_feed(self.input_name, bach_image_tensor)
        return self.onnx_session.run(self.output_name, input_feed=input_feed)