TensorFlow读取模型

TensorFlow保存和读取模型

加载图结构和参数：

import tensorflow as tf

ckpt = tf.train.get_checkpoint_state('./model/')   #通过检查点文件锁定最新的模型
saver = tf.train.import_meta_graph(file_path[0])   #载入图结构，存在在.meta文件中

with tf.Session() as sess:
	saver.restore(sess,file_path[1])               #载入参数，参数保存在两个文件中    

参考链接

查看TensorFlow的预训练权重.npy文件

  import numpy as np 
  pre_train = np.load("pnet.npy",allow_pickle = True,encoding = 'latin1')
  print(pre_train)

参考链接

将TensorFlow的.npy模型文件转换为caffe模型

直接附代码：

import sys
import tensorflow as tf
import caffe
import numpy as np
import cv2
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))

from tensorflow.python import pywrap_tensorflow
checkpoint_path = "./pnet.npy"   # .npy模型文件的存放地址
pre_train = np.load(checkpoint_path, allow_pickle=True, encoding="latin1")
dict = pre_train.item()
dict1 = dict.items()


cf_prototxt = "./det1.prototxt"  # caffe模型的网络结构
cf_model = "./det1.caffemodel"   # 要生成的caffe模型


def tensor4d_transform(tensor):
    return tensor.transpose((3, 2, 0, 1))
def tensor2d_transform(tensor):
    return tensor.transpose((1, 0))

def tf2caffe(checkpoint_path,cf_prototxt,cf_model):
    net = caffe.Net(cf_prototxt, caffe.TRAIN)

    for key1, value1 in dict1:
         if "PReLU" in key1:
            pass
         else:
            dict2 = value1.items()
            for key2, value2 in dict2:
                if key2 == "weights":
                    net.params[key1][0].data.flat = value2.flat
                elif key2 =='biases':
                    net.params[key1][1].data.flat = value2.flat
                else:
                    pass

    net.save(cf_model)
    print("\n- Finished.\n")

if __name__ == '__main__':
    tf2caffe(checkpoint_path, cf_prototxt, cf_model)

将caffe模型转换为TensorFlow的.npy模型文件

参考

我python3转换mtcnn的模型试了一下，各种报错，解决了一个还有一个，最后实在试不下去，放弃了，最后卡在这里：

Type                 Name                                          Param               Output
----------------------------------------------------------------------------------------------
Input                data                                             --       (1, 3, 12, 12)
Convolution          conv1                                 (10, 3, 3, 3)      (1, 10, 10, 10)
Pooling              pool1                                            --        (1, 10, 5, 5)
Convolution          conv2                                (16, 10, 3, 3)        (1, 16, 3, 3)
Convolution          conv3                                (32, 16, 3, 3)        (1, 32, 1, 1)
Convolution          conv4-1                               (2, 32, 1, 1)         (1, 2, 1, 1)
Convolution          conv4-2                               (4, 32, 1, 1)         (1, 4, 1, 1)
Softmax              prob1                                            --         (1, 2, 1, 1)
Converting data...
Saving data...
Traceback (most recent call last):
  File "convert.py", line 61, in <module>
    main()
  File "convert.py", line 57, in main
    args.phase)
  File "convert.py", line 34, in convert
    np.save(data_out, data)
  File "/home/yaspeed/.local/lib/python3.6/site-packages/numpy/lib/npyio.py", line 536, in save
    pickle_kwargs=pickle_kwargs)
  File "/home/yaspeed/.local/lib/python3.6/site-packages/numpy/lib/format.py", line 633, in write_array
    pickle.dump(array, fp, protocol=2, **pickle_kwargs)
AttributeError: Can't pickle local object 'DataInjector.load_using_caffe.<locals>.<lambda>'