H5模型转pb模型
这个实际上是个伪需求，直接h5转tflite就好了，但是就是没办法，总有些东西不支持新的方法。下面记录一下怎样把tf2.0生成的h5模型转成tf1.10的pb模型。
加载模型
首先用老版本的tensorflow.keras加载模型：
import tensorflow as tf
import numpy as np
from functools import reduce, wraps
from typing import List
import os
from tensorflow.contrib.lite.python import lite
kl = tf.keras.layers
k = tf.keras
kr = tf.keras.regularizers
K = tf.keras.backend
h5_model = k.models.load_model('infer.h5')
当然会碰到许多不兼容的地方：
ValueError: ('Unrecognized keyword arguments:', dict_keys(['ragged']))
这个是因为老的k.Input不支持ragged参数，修改如下：
def __init__(self,
input_shape=None,
batch_size=None,
dtype=None,
input_tensor=None,
sparse=False,
name=None,
**kwargs):
if 'batch_input_shape' in kwargs:
batch_input_shape = kwargs.pop('batch_input_shape')
if input_shape and batch_input_shape:
raise ValueError('Only provide the input_shape OR '
'batch_input_shape argument to '
'InputLayer, not both at the same time.')
batch_size = batch_input_shape[0]
input_shape = batch_input_shape[1:]
# NOTE 注释这里：
# if kwargs:
# raise ValueError('Unrecognized keyword arguments:', kwargs.keys())
ValueError: Unknown initializer: GlorotUniform
老版本的tf.keras没有这个初始化类，从新的tensorflow里面拷贝一个来就好了。
class GlorotNormal(k.initializers.VarianceScaling): |
TypeError: ('Keyword argument not understood:', 'threshold')
这个是因为Relu层的实现不一样。
先把tf.keras.backend.relu替换为如下：
def relu(x, alpha=0., max_value=None, threshold=0):
if alpha != 0.:
if max_value is None and threshold == 0:
return nn.leaky_relu(x, alpha=alpha)
if threshold != 0:
negative_part = nn.relu(-x + threshold)
else:
negative_part = nn.relu(-x)
clip_max = max_value is not None
if threshold != 0:
# computes x for x > threshold else 0
x = x * math_ops.cast(math_ops.greater(x, threshold), floatx())
elif max_value == 6:
# if no threshold, then can use nn.relu6 native TF op for performance
x = nn.relu6(x)
clip_max = False
else:
x = nn.relu(x)
if clip_max:
max_value = _to_tensor(max_value, x.dtype.base_dtype)
zero = _to_tensor(0., x.dtype.base_dtype)
x = clip_ops.clip_by_value(x, zero, max_value)
if alpha != 0.:
alpha = _to_tensor(alpha, x.dtype.base_dtype)
x -= alpha * negative_part
return x
再把tf.keras.layers.advanced_activations里面的Relu替换为如下：
class ReLU(Layer):
def __init__(self, max_value=None, negative_slope=0, threshold=0, **kwargs):
super(ReLU, self).__init__(**kwargs)
if max_value is not None and max_value < 0.:
raise ValueError('max_value of Relu layer '
'cannot be negative value: ' + str(max_value))
if negative_slope < 0.:
raise ValueError('negative_slope of Relu layer '
'cannot be negative value: ' + str(negative_slope))
self.support_masking = True
if max_value is not None:
max_value = K.cast_to_floatx(max_value)
self.max_value = max_value
self.negative_slope = K.cast_to_floatx(negative_slope)
self.threshold = K.cast_to_floatx(threshold)
def call(self, inputs):
# alpha is used for leaky relu slope in activations instead of
# negative_slope.
return K.relu(inputs,
alpha=self.negative_slope,
max_value=self.max_value,
threshold=self.threshold)
def get_config(self):
config = {
'max_value': self.max_value,
'negative_slope': self.negative_slope,
'threshold': self.threshold
}
base_config = super(ReLU, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def compute_output_shape(self, input_shape):
return input_shape
TypeError: ('Keyword argument not understood:', 'interpolation')
这个是因为上采样层不同，替换tensorflow/python/keras/layers/convolutional.py里面的UpSampling2D：
|
再替换tensorflow/python/keras/backend.py里面的resize_images：
|
转换模型
加载模型好了其实就简单了：
def test_convert_pb_2(): |
得到如下： (vim3l) ➜ retinaface python ./test_model.py
2020-03-18 00:02:18.150198: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
WARNING:tensorflow:No training configuration found in save file: the model was *not* compiled. Compile it manually.
['input_1']
['concatenate_3/concat', 'concatenate_4/concat', 'concatenate_5/concat']
然后终于可以用他的**转换工具转换了
