训练时使用固定大小, 方便编程实现和速度优化

部署时使用任意大小, 提高体验

但是采用固定大小输入做训练, 部署时采用任意大小, 可能效果有点差别吧....

pb文件的网络结构

import tensorflow as tfoutput_graph_def = tf.GraphDef()PB_PATH = r"./pb/feathers.pb"with open(PB_PATH, "rb") as f:    output_graph_def.ParseFromString(f.read())    tf.import_graph_def(        output_graph_def,        name='',  # 默认name为import,类似scope    )with tf.Session() as sess:    sess.run(tf.global_variables_initializer())    tf.summary.FileWriter('./log/', sess.graph)

tensorboard.exe --logdir .

将固定输入大小的pb文件,转化为任意输入大小的tfjs格式 

import tensorflow as tffrom tensorflow.python.framework import graph_utilimport tensorflowjs as tfjssess = tf.Session()output_graph_def = tf.GraphDef()#   feathers starry candyPB_PATH = r"./pb/candy.pb"TFJS_PATH = r'./tfjs/candy'in_image = tf.placeholder(tf.float32, (None, None, None, 3), name='in_x')with open(PB_PATH, "rb") as f:    output_graph_def.ParseFromString(f.read())    tf.import_graph_def(        output_graph_def,        input_map={            'in_x:0': in_image        },        name='',  # 默认name为import,类似scope        # return_elements=['generator/mul:0']    )sess.run(tf.global_variables_initializer())output = sess.graph.get_tensor_by_name("generator/output:0")with tf.Session() as sess:    sess.run(tf.global_variables_initializer())    constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ['generator/output'])    with tf.gfile.FastGFile("./pb/tmp.pb", mode='wb') as f:        f.write(constant_graph.SerializeToString())tfjs.converters.tf_saved_model_conversion_pb.convert_tf_frozen_model(    "./pb/tmp.pb",    'generator/output',    TFJS_PATH)