CustomDLCoder/extract_fc.py at main · zhoumingyi/CustomDLCoder · GitHub

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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
import numpy as np
from utils.utils import *

def extract_fc(op, kwargs, interpreter, unknown_config):
    try:
        activation = op['builtin_options']['fused_activation_function']
    except:
        kwargs['activation='] = 'activation=kTfLiteActNone'
        print("Warning: no activation function found")
    else:
        kwargs['activation='] = 'activation=' + conv_activation_parser(activation)
    try:
        weights_format = op['builtin_options']['weights_format']
    except:
        kwargs['weights_format='] = 'weights_format=kTfLiteFullyConnectedWeightsFormatDefault'
        print("Warning: no weights_format function found")
    else:
        kwargs['weights_format='] = 'weights_format=' + weights_format_parser(weights_format)
    try:
        asymmetric_quantize_inputs = op['builtin_options']['asymmetric_quantize_inputs']
    except:
        kwargs['asymmetric_quantize_inputs='] = 'asymmetric_quantize_inputs=false'
        print("Warning: no asymmetric_quantize_inputs function found")
    else:
        kwargs['asymmetric_quantize_inputs='] = 'asymmetric_quantize_inputs=' + str.lower(str(asymmetric_quantize_inputs))
    try:
        keep_num_dims = op['builtin_options']['keep_num_dims']
    except:
        kwargs['keep_num_dims='] = 'keep_num_dims=false'
        print("Warning: no keep_num_dims function found")
    else:
        kwargs['keep_num_dims='] = 'keep_num_dims=' + str.lower(str(keep_num_dims))
    if len(op['inputs']) > 2:
        kwargs['has_conv_bias='] = 'has_conv_bias=true'
    else:
        kwargs['has_conv_bias='] = 'has_conv_bias=false'
        kwargs['const TfLiteType bias_type=;'] = ''
        kwargs['const int bias_dims_size=;'] = ''
        kwargs['const int32_t bias_dims_raw=;'] = ''
        kwargs['bias_raw=;'] = ''
        kwargs['const float scale_bias=;'] = ''
        kwargs['const int32_t zero_point_bias=;'] = ''
        kwargs['bias_tensor_data=bias_raw;'] = ''
        kwargs['RuntimeShape(bias_dims_size, bias_dims_raw), bias_tensor_data,'] = 'GetTensorShape(bias),GetTensorData<float>(bias),'

    unknown_status = {}
    for tensor_details in interpreter.get_tensor_details():
        if tensor_details['index'] == op['inputs'][1]:

            add_aug_num = 0
            input_arg = []
            try:
                filter_tensor = interpreter.get_tensor(tensor_details["index"])
            except:
                kwargs['filter_raw=;'] = ''
                kwargs['filter_tensor_data=filter_raw;'] = ''
                add_aug_num += 1
                input_arg.append('float* input_v_' + str(i))
                kwargs['input_2_placeholder,'] = ''
            else:
                if np.all(filter_tensor == 0):
                    kwargs['filter_raw=;'] = ''
                    kwargs['filter_tensor_data=filter_raw;'] = ''
                    add_aug_num += 1
                    input_arg.append('float* input_v_' + str(i))
                    kwargs['input_2_placeholder'] = 'float* filter_tensor_data'
                else:
                    filter_tensor = interpreter.get_tensor(tensor_details["index"])
                    filter_item_num = 1
                    for i in range(0, len(tensor_details['shape'])):
                        filter_item_num = filter_item_num * tensor_details['shape'][i]
                    tflite_type, type_str = conv_data_type_parser(tensor_details['dtype'])
                    kwargs['filter_raw='] = type_str + ' filter_raw[' + str(filter_item_num) + ']=' + '{' + str(filter_tensor.flatten('C').tolist()).strip('[').strip(']') + '}'
                    kwargs['filter_tensor_data=filter_raw'] = type_str + '* filter_tensor_data=filter_raw'
                    kwargs['input_2_placeholder,'] = ''

            # if np.size(filter_tensor) > 4000000:
            #     rows_per_split = np.size(filter_tensor) // 100
            #     print("rows_per_split: ", rows_per_split)
            #     # Create and initialize the split arrays
            #     split_arrays = []
            #     for i in range(100):
            #         start_row = i * rows_per_split
            #         end_row = (i + 1) * rows_per_split if i < 99 else None
            #         split_arrays.append(filter_tensor[start_row:end_row, :])
            #     # split_arrays = np.array_split(filter_tensor, 10)
            #     kwargs['filter_raw='] = ''

            #     for i, split_array in enumerate(split_arrays):
            #         header_file_name = f"array_part_{i + 1}.h"

            #         with open(header_file_name, 'w') as header_file:
            #             header_file.write(f'# ifndef ARRAY_PART_{i + 1}_H\n')
            #             header_file.write(f'# define ARRAY_PART_{i + 1}_H\n\n')
            #             header_file.write('extern float myLargeArrayPart{}[{}];\n\n'.format(i + 1, split_array.shape[0] * split_array.shape[1]))
            #             header_file.write('float myLargeArrayPart' + str(i + 1) + '[' + str(split_array.shape[0] * split_array.shape[1]) + ']=' + '{' + str(split_array.flatten('C').tolist()).strip('[').strip(']'))

            #             # # Write array initialization values
            #             # for row in split_array:
            #             #     header_file.write('    {')
            #             #     header_file.write(', '.join(map(str, row)))
            #             #     header_file.write('},\n')

            #             header_file.write('};\n\n')
            #             header_file.write('# endif')
            # filter_dims_raw = '{' + ','.join(str(e) for e in list(filter_tensor.shape)) +'}'
            filter_dims_raw = '{' + ','.join(str(e) for e in list(tuple(tensor_details['shape']))) +'}'
            filter_dims_size = len(tensor_details['shape'])
            quantization_filter = tensor_details['quantization']
            kwargs['filter_dims_size='] = 'filter_dims_size=' + str(filter_dims_size)
            kwargs['filter_dims_raw='] = 'filter_dims_raw[' + str(filter_dims_size) + ']=' + filter_dims_raw
            kwargs['filter_type='] = 'filter_type=' + tflite_type
            kwargs['scale_filter='] = 'scale_filter=' + str(quantization_filter[0])
            kwargs['zero_point_filter='] = 'zero_point_filter=' + str(quantization_filter[1])

        elif tensor_details['index'] == op['inputs'][2]:
            bias_tensor = interpreter.get_tensor(tensor_details["index"])
            # if np.all(bias_tensor) == 0:
            #     kwargs['has_conv_bias='] = 'has_conv_bias=false'
            #     kwargs['const TfLiteType bias_type=;'] = ''
            #     kwargs['const int bias_dims_size=;'] = ''
            #     kwargs['const int32_t bias_dims_raw=;'] = ''
            #     kwargs['bias_raw=;'] = ''
            #     kwargs['const float scale_bias=;'] = ''
            #     kwargs['const int32_t zero_point_bias=;'] = ''
            #     kwargs['bias_tensor_data=bias_raw;'] = ''
            #     kwargs['RuntimeShape(bias_dims_size, bias_dims_raw), bias_tensor_data,'] = 'GetTensorShape(bias), GetTensorData<float>(bias),'
            # elif kwargs['has_conv_bias='] == 'has_conv_bias=true':
            bias_item_num = bias_tensor.size
            # bias_channel = bias_tensor.shape[0]
            bias_dims_raw = '{' + ','.join(str(e) for e in list(bias_tensor.shape)) + '}'
            bias_dims_size = len(bias_tensor.shape)
            tflite_type, type_str = conv_data_type_parser(bias_tensor.dtype)
            quantization_bias = tensor_details['quantization']
            kwargs['bias_type='] = 'bias_type=' + tflite_type
            kwargs['bias_dims_size='] = 'bias_dims_size=' + str(bias_dims_size)
            kwargs['bias_dims_raw='] = 'bias_dims_raw[' + str(bias_dims_size) + ']=' + bias_dims_raw
            kwargs['bias_raw='] = type_str + ' bias_raw[' + str(bias_item_num) + ']=' + '{' + str(bias_tensor.tolist()).strip('[').strip(']') + '}'
            kwargs['scale_bias='] = 'scale_bias=' + str(quantization_bias[0])
            kwargs['zero_point_bias='] = 'zero_point_bias=' + str(quantization_bias[1])
            kwargs['bias_tensor_data=bias_raw'] = type_str + '* bias_tensor_data=bias_raw'
            kwargs['  const TfLiteTensor* bias = nullptr;'] = ''
        elif tensor_details['index'] == op['outputs'][0]:
            # output_tensor = interpreter.get_tensor(tensor_details["index"])
            # output_channel = tensor_details['shape'][1]
            output_num = 1
            for i in range(0, len(tensor_details['shape'])):
                output_num = output_num * tensor_details['shape'][i]
            output_dims_raw = '{' + ','.join(str(e) for e in list(tensor_details['shape'])) + '}'
            output_dims_size = len(tensor_details['shape'])
            tflite_type, type_str = conv_data_type_parser(tensor_details['dtype'])
            quantization_output = tensor_details['quantization']
            kwargs['output_dims_size='] = 'output_dims_size=' + str(output_dims_size)
            kwargs['output_dims_raw='] = 'output_dims_raw[' + str(output_dims_size) + ']=' + output_dims_raw
            kwargs['output_num='] = 'output_num=' + str(output_num)
            kwargs['output_type='] = 'output_type=' + tflite_type
            kwargs['scale_output='] = 'scale_output=' + str(quantization_output[0])
            kwargs['zero_point_output='] = 'zero_point_output=' + str(quantization_output[1])
        elif tensor_details['index'] == op['inputs'][0]:
            # input_tensor = interpreter.get_tensor(tensor_details["index"])
            # if len(tensor_details['shape']) == 4:
            #     input_channel = tensor_details['shape'][3]
            #     input_height = tensor_details['shape'][1]
            #     input_width = tensor_details['shape'][2]
            #     input_dims_raw = '{' + '1,' + str(input_height) + ',' + str(input_width)  + ',' + str(input_channel) + '}'
            # else:
            #     input_channel = tensor_details['shape'][1]
            #     input_dims_raw = '{' + '1,' + str(input_channel) + '}'
            input_dims_raw = '{' + ','.join(str(e) for e in list(tensor_details['shape'])) + '}'
            input_dims_size = len(tensor_details['shape'])
            tflite_type, type_str = conv_data_type_parser(tensor_details['dtype'])
            quantization_input = tensor_details['quantization']
            kwargs['input_dims_size='] = 'input_dims_size=' + str(input_dims_size)
            kwargs['input_dims_raw='] = 'input_dims_raw[' + str(input_dims_size) + ']=' + input_dims_raw
            kwargs['input_type='] = 'input_type=' + tflite_type
            kwargs['scale_input='] = 'scale_input=' + str(quantization_input[0])
            kwargs['zero_point_input='] = 'zero_point_input=' + str(quantization_input[1])
            # kwargs['input_tensor_data=input_raw'] = type_str + '* input_tensor_data=input_raw'
    unknown_status['lhs_cacheable=true'] = ["lhs_cacheable=true", "lhs_cacheable=false"]
    unknown_status['rhs_cacheable=true'] = ["rhs_cacheable=false", "rhs_cacheable=true"]
    unknown_status["opname"] = "FullyConnectedOptions"
    unknown_config.append(unknown_status)

    return kwargs, unknown_config