you need to get below files first.
Last active
September 5, 2024 06:20
-
-
Save tworuler/bd7bd4c6cd9a8fbbeb060e7b64cfa008 to your computer and use it in GitHub Desktop.
Convert BlazeFace .tflite to .pb
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| #!/usr/bin/env python | |
| # coding: utf-8 | |
| import os | |
| import numpy as np | |
| import json | |
| import tensorflow as tf | |
| os.environ['CUDA_VISIBLE_DEVICES'] = '0' | |
| schema = "schema.fbs" | |
| binary = "flatc" | |
| model_path = "mediapipe/models/facedetector_front.tflite" | |
| output_pb_path = "facedetector_front.pb" | |
| output_savedmodel_path = "saved_model" | |
| model_json_path = "/tmp/facedetector_front.json" | |
| num_tensors = 176 | |
| output_node_names = ['classificators', 'regressors'] | |
| def gen_model_json(): | |
| if not os.path.exists(model_json_path): | |
| cmd = (binary + | |
| " -t --strict-json --defaults-json -o /tmp {schema} -- {input}". | |
| format(input=model_path, schema=schema)) | |
| os.system(cmd) | |
| def parse_json(): | |
| j = json.load(open(model_json_path)) | |
| op_types = [v['builtin_code'] for v in j['operator_codes']] | |
| # print('op types:', op_types) | |
| ops = j['subgraphs'][0]['operators'] | |
| # print('num of ops:', len(ops)) | |
| return ops, op_types | |
| def make_graph(ops, op_types, interpreter): | |
| tensors = {} | |
| input_details = interpreter.get_input_details() | |
| output_details = interpreter.get_output_details() | |
| # print(input_details) | |
| for input_detail in input_details: | |
| tensors[input_detail['index']] = tf.placeholder( | |
| dtype=input_detail['dtype'], | |
| shape=input_detail['shape'], | |
| name=input_detail['name']) | |
| for index, op in enumerate(ops): | |
| print('op: ', op) | |
| op_type = op_types[op['opcode_index']] | |
| if op_type == 'CONV_2D': | |
| input_tensor = tensors[op['inputs'][0]] | |
| weights_detail = interpreter._get_tensor_details(op['inputs'][1]) | |
| bias_detail = interpreter._get_tensor_details(op['inputs'][2]) | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| # print('weights_detail: ', weights_detail) | |
| # print('bias_detail: ', bias_detail) | |
| # print('output_detail: ', output_detail) | |
| weights_array = interpreter.get_tensor(weights_detail['index']) | |
| weights_array = np.transpose(weights_array, (1, 2, 3, 0)) | |
| bias_array = interpreter.get_tensor(bias_detail['index']) | |
| weights = tf.Variable(weights_array, name=weights_detail['name']) | |
| bias = tf.Variable(bias_array, name=bias_detail['name']) | |
| options = op['builtin_options'] | |
| output_tensor = tf.nn.conv2d( | |
| input_tensor, | |
| weights, | |
| strides=[1, options['stride_h'], options['stride_w'], 1], | |
| padding=options['padding'], | |
| dilations=[ | |
| 1, options['dilation_h_factor'], | |
| options['dilation_w_factor'], 1 | |
| ], | |
| name=output_detail['name'] + '/conv2d') | |
| output_tensor = tf.add( | |
| output_tensor, bias, name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'DEPTHWISE_CONV_2D': | |
| input_tensor = tensors[op['inputs'][0]] | |
| weights_detail = interpreter._get_tensor_details(op['inputs'][1]) | |
| bias_detail = interpreter._get_tensor_details(op['inputs'][2]) | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| # print('weights_detail: ', weights_detail) | |
| # print('bias_detail: ', bias_detail) | |
| # print('output_detail: ', output_detail) | |
| weights_array = interpreter.get_tensor(weights_detail['index']) | |
| weights_array = np.transpose(weights_array, (1, 2, 3, 0)) | |
| bias_array = interpreter.get_tensor(bias_detail['index']) | |
| weights = tf.Variable(weights_array, name=weights_detail['name']) | |
| bias = tf.Variable(bias_array, name=bias_detail['name']) | |
| options = op['builtin_options'] | |
| output_tensor = tf.nn.depthwise_conv2d( | |
| input_tensor, | |
| weights, | |
| strides=[1, options['stride_h'], options['stride_w'], 1], | |
| padding=options['padding'], | |
| # dilations=[ | |
| # 1, options['dilation_h_factor'], | |
| # options['dilation_w_factor'], 1 | |
| # ], | |
| name=output_detail['name'] + '/depthwise_conv2d') | |
| output_tensor = tf.add( | |
| output_tensor, bias, name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'MAX_POOL_2D': | |
| input_tensor = tensors[op['inputs'][0]] | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| options = op['builtin_options'] | |
| output_tensor = tf.nn.max_pool( | |
| input_tensor, | |
| ksize=[ | |
| 1, options['filter_height'], options['filter_width'], 1 | |
| ], | |
| strides=[1, options['stride_h'], options['stride_w'], 1], | |
| padding=options['padding'], | |
| name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'PAD': | |
| input_tensor = tensors[op['inputs'][0]] | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| paddings_detail = interpreter._get_tensor_details(op['inputs'][1]) | |
| # print('output_detail:', output_detail) | |
| # print('paddings_detail:', paddings_detail) | |
| paddings_array = interpreter.get_tensor(paddings_detail['index']) | |
| paddings = tf.Variable( | |
| paddings_array, name=paddings_detail['name']) | |
| output_tensor = tf.pad( | |
| input_tensor, paddings, name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'RELU': | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| input_tensor = tensors[op['inputs'][0]] | |
| output_tensor = tf.nn.relu( | |
| input_tensor, name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'RESHAPE': | |
| input_tensor = tensors[op['inputs'][0]] | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| options = op['builtin_options'] | |
| output_tensor = tf.reshape( | |
| input_tensor, options['new_shape'], name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'ADD': | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| input_tensor_0 = tensors[op['inputs'][0]] | |
| input_tensor_1 = tensors[op['inputs'][1]] | |
| output_tensor = tf.add( | |
| input_tensor_0, input_tensor_1, name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| elif op_type == 'CONCATENATION': | |
| output_detail = interpreter._get_tensor_details(op['outputs'][0]) | |
| input_tensor_0 = tensors[op['inputs'][0]] | |
| input_tensor_1 = tensors[op['inputs'][1]] | |
| options = op['builtin_options'] | |
| output_tensor = tf.concat([input_tensor_0, input_tensor_1], | |
| options['axis'], | |
| name=output_detail['name']) | |
| tensors[output_detail['index']] = output_tensor | |
| else: | |
| raise ValueError(op_type) | |
| def main(): | |
| gen_model_json() | |
| ops, op_types = parse_json() | |
| interpreter = tf.lite.Interpreter(model_path) | |
| interpreter.allocate_tensors() | |
| # input_details = interpreter.get_input_details() | |
| # output_details = interpreter.get_output_details() | |
| # print(input_details) | |
| # print(output_details) | |
| # for i in range(num_tensors): | |
| # detail = interpreter._get_tensor_details(i) | |
| # print(detail) | |
| make_graph(ops, op_types, interpreter) | |
| config = tf.ConfigProto() | |
| config.gpu_options.allow_growth = True | |
| graph = tf.get_default_graph() | |
| # writer = tf.summary.FileWriter(os.path.splitext(output_pb_path)[0]) | |
| # writer.add_graph(graph) | |
| # writer.flush() | |
| # writer.close() | |
| with tf.Session(config=config, graph=graph) as sess: | |
| sess.run(tf.global_variables_initializer()) | |
| graph_def = tf.graph_util.convert_variables_to_constants( | |
| sess=sess, | |
| input_graph_def=graph.as_graph_def(), | |
| output_node_names=output_node_names) | |
| with tf.gfile.GFile(output_pb_path, 'wb') as f: | |
| f.write(graph_def.SerializeToString()) | |
| tf.saved_model.simple_save( | |
| sess, | |
| output_savedmodel_path, | |
| inputs={'input': graph.get_tensor_by_name('input:0')}, | |
| outputs={ | |
| 'classificators': graph.get_tensor_by_name('classificators:0'), | |
| 'regressors': graph.get_tensor_by_name('regressors:0') | |
| }) | |
| if __name__ == '__main__': | |
| main() |
I'm running into issues parsing the json. There are missing opcode_index attributes:
('op: ', {u'inputs': [0, 1, 2], u'builtin_options': {u'stride_w': 2, u'stride_h': 2}, u'builtin_options_type': u'Conv2DOptions', u'outputs': [3]})
Traceback (most recent call last):
File "blazeface_tflite_to_pb.py", line 203, in <module>
main()
File "blazeface_tflite_to_pb.py", line 173, in main
make_graph(ops, op_types, interpreter)
File "blazeface_tflite_to_pb.py", line 46, in make_graph
op_type = op_types[op['opcode_index']]
KeyError: 'opcode_index'
How should i handle these?
flatc -t --strict-json --defaults-json -o schema.fbs -- face_detection_front.tflite
flatc: error: current schema has no file_identifier: cannot test if "face_detection_front.tflite" matches the schema, use --raw-binary to read this file anyway.
I was able to use this great code example to successfully convert the BlazeFace, FaceMesh and 3D Hand Pose into a saved_model. If it's useful to you, I'm happy. I used Tensorflow-GPU v1.15.2.
・My repository
https://github.com/PINTO0309/PINTO_model_zoo
・Example
I was able to use this great code example to successfully convert the BlazeFace, FaceMesh and 3D Hand Pose into a saved_model. If it's useful to you, I'm happy. I used
Tensorflow-GPU v1.15.2.・My repository
https://github.com/PINTO0309/PINTO_model_zoo・Example
@PINTO0309 the links are broken. Can you fix them?
@SocialKitLtd
I fixed it.
@PINTO0309 Thanks! But I can't seems to find the .pb model. Is it there? Or I'm miss understanding what you wrote
@SocialKitLtd Run download.sh or download_new.sh to download all the materials you need.
@PINTO0309 On which of the folders? Sorry, I don't really have deep understand ML yet:
01_float32
@PINTO0309 Works!! :) Katsuya, I sent you an email from [email protected]. Did you get it by chance?
These locations are now not valid-
flatc: https://github.com/tensorflow/tensorflow/blob/release_1.14.0/tensorflow/lite/tools/visualize.py#L38
shema: https://github.com/tensorflow/tensorflow/blob/release_1.14.0/tensorflow/lite/tools/visualize.py#L37
Author
These locations are now not valid-
flatc: https://github.com/tensorflow/tensorflow/blob/release_1.14.0/tensorflow/lite/tools/visualize.py#L38
shema: https://github.com/tensorflow/tensorflow/blob/release_1.14.0/tensorflow/lite/tools/visualize.py#L37
They are still valid at:
flatc: https://github.com/tensorflow/tensorflow/blob/v1.14.0/tensorflow/lite/tools/visualize.py#L38
schema: https://github.com/tensorflow/tensorflow/blob/v1.14.0/tensorflow/lite/tools/visualize.py#L37
RuntimeError: Sorry, schema file cannot be found at '../schema/schema.fbs' . where do i get this file?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
how to replace PReLU?