怎么将densenet+CTC在keras下的h5模型转为TensorFlow下的pb模型呀？亲们！

[JentMan]

我用DenseNet + CTC算法做中文识别，我重新训练出了keras下的h5模型(里面保存了权重、网络结构和配置信息等)，现在将其转为TensorFlow下的pb模型时，提示“ValueError: Unknown loss function: ”，我该做怎么做啊？load_model('weights_densenet***.h5', custom_objects={})中custom_objects该怎么写呢？

[wqt2019]

-- coding:utf-8 --

from keras import backend as K import os import shutil import tensorflow as tf import keys import densenet from keras.layers import Input from keras.models import Model

from keras.utils import multi_gpu_model def preprocess_image(im):

等比例将图像高度缩放到32

# im=tf.image.rgb_to_grayscale(im)
im_shape = tf.shape(im)
h=im_shape[1]
w=im_shape[2]
height=tf.constant(32,tf.int32)
scale = tf.divide(tf.cast(h,tf.float32),32)
width = tf.divide(tf.cast(w,tf.float32),scale)
width =tf.cast(width,tf.int32)
resize_image = tf.image.resize_images(im, [height,width], method=tf.image.ResizeMethod.BILINEAR, align_corners=True)
resize_image = tf.cast(resize_image, tf.float32) / 255 - 0.5
width = tf.reshape(width, [1])
height = tf.reshape(height, [1])
im_info = tf.concat([height, width], 0)
im_info = tf.concat([im_info, [1]], 0)
im_info = tf.reshape(im_info, [1, 3])
im_info = tf.cast(im_info, tf.float32)
return resize_image,im_info

reload(densenet)

characters = keys.alphabet[:] characters = characters[1:] + u'卍' nclass = len(characters) input = Input(shape=(32, None, 1), name='the_input')

input = Input(tensor=resize_image)

y_pred= densenet.dense_cnn(input, nclass) basemodel = Model(inputs=input, outputs=y_pred) modelPath = os.path.join(os.getcwd(), './densenet/weights_densenet.h5')

basemodel = multi_gpu_model(basemodel, gpus=8)

basemodel.load_weights(modelPath)

model model_mulit_gpu

export_path = './model/1/' if os.path.exists(export_path): shutil.rmtree(export_path) version = 1 path='model' K.set_learning_phase(0) if not os.path.exists(path): os.mkdir(path) export_path = os.path.join(tf.compat.as_bytes(path),tf.compat.as_bytes(str(version))) builder = tf.saved_model.builder.SavedModelBuilder(export_path) print('input info:', basemodel.input_names, '---', basemodel.input) print('output info:', basemodel.output_names, '---', basemodel.output) input_tensor = basemodel.input output_tensor=basemodel.output

tensor_info_input = tf.saved_model.utils.build_tensor_info(raw_image)

model_input = tf.saved_model.utils.build_tensor_info(input_tensor) model_output = tf.saved_model.utils.build_tensor_info(output_tensor)

im_info_output = tf.saved_model.utils.build_tensor_info(im_info)

print('im_info_input tensor shape', im_info.shape)

prediction_signature = ( tf.saved_model.signature_def_utils.build_signature_def( inputs={'images': model_input}, outputs={'prediction': model_output}, method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))

prediction_signature = (

tf.saved_model.signature_def_utils.build_signature_def(

inputs={'images': tensor_info_input},

outputs={'prediction': model_output,

'im_info': im_info_output},

method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))

with K.get_session() as sess: builder.add_meta_graph_and_variables( sess=sess, tags=[tf.saved_model.tag_constants.SERVING], signature_def_map={'predict_images':prediction_signature,} ) builder.save()

[yjfncu]

识别模型中，原来使用h5预测的时候输入图片宽度不定，但是转换后的pb模型输入要求图片宽度必须为280，怎么转换后的pb才能让输入宽度不定呢

[cuimiao187561]

大佬们我把识别模型h5转pb识别效果不一样，加载pb模型测试效果很差，这是什么原因导致的？@yjfncu @wqt2019 @JentMan

[yjfncu]

可以啊，你把保存模型这行代码中save_weights=False,改成True，然后转pb的时候使用load_model加载h5并转换就行，具体看我博客：https://blog.csdn.net/yjfncu

在 2019-11-13 10:42:05，"cuimiao187561" notifications@github.com 写道：

大佬们我把识别模型h5转pb识别效果不一样，加载pb模型测试效果很差，这是什么原因导致的？@yjfncu@wqt2019@JentMan

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.