将tensorflow-hub模型转换为tensorflow lite(tflite)

Question

我试图使用以下代码将tensorflow-hub(.pb)中的.pb模型转换为TensorFlow Lite文件(.tflite)：

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
import numpy as np
import tensorflow_hub as hub

module_path = 'https://tfhub.dev/deepmind/biggan-deep-256/1'
tf.compat.v1.reset_default_graph()
print('Loading BigGAN module from:', module_path)
module = hub.Module(module_path)
dummy_inputs = {
    "y": tf.compat.v1.placeholder(tf.float32, [1, 1000], 'y'),
    "z": tf.compat.v1.placeholder(tf.float32, [1, 128], 'z'),
    "truncation": tf.compat.v1.placeholder(tf.float32, [], 'truncation'),
}
dummy_output = module(dummy_inputs)
print('dummy_Inputs:\n', '\n'.join(
    '  {}: {}'.format(*kv) for kv in dummy_inputs.items()))
print('dummy_Output:', dummy_output)

initializer = tf.global_variables_initializer()
sess = tf.Session()
sess.run(initializer)

save_path = "./big_gan.tflite"
_input = [dummy_inputs[k] for k in dummy_inputs]
converter = tf.lite.TFLiteConverter.from_session(sess, _input, [dummy_output])
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.experimental_new_converter = True

# errors here
tflite_model = converter.convert()

open(save_path, "wb").write(tflite_model)

返回的信息：

2022-01-28 17:10:30.240145: E tensorflow/core/grappler/grappler_item_builder.cc:670] Init node AssignVariableOp_1003 doesn't exist in graph
2022-01-28 17:10:33.853842: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:363] Ignored output_format.
2022-01-28 17:10:33.853899: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:366] Ignored drop_control_dependency.

抛出错误：

ConverterError: Input 0 of node module_apply_default/cond/AssignVariableOp was passed float from module_apply_default/cond/AssignVariableOp/Switch:0 incompatible with expected resource.

该模块按照“示例使用”的要求工作，可以正常生成图像。

有人能帮助我理解Init node AssignVariableOp_1003 doesn't exist in graph的含义以及如何修复错误吗？

Answer 1

使用TF 2.x将TF 1.x模型转换为TensorFlow Lite文件是很棘手的。我建议在Google上运行您的代码示例并切换到TF 1.x

%tensorflow_version 1.x

好像很管用。