python - 使用 tensorflow 2.0 进行模型预测会导致 python 内核死机并重新启动

Question

我通过简单地使用 tf.keras 而不是 keras 将我的模型从 tensorflow 1.15 移植到 2.0。即使使用 GPU，该模型在 tf 1.15 下也可以正常工作，但是在 tf 2.0 或 2.1 的情况下，当我调用 mode.predict() 时，它会导致 python 内核死亡并重新启动。根本没有错误消息。对于 tensorflow 2.1，现象是一样的。

模型创建工作正常，因此编译模型，从文件加载权重。问题是由调用 model.predict() 引起的。

有什么提示吗？

import numpy as np
from IPython.display import clear_output
import time
import tensorflow as tf
import os
os.environ["PATH"] += os.pathsep + "C:/Program Files (x86)/Graphviz2.38/bin/"
from tensorflow.keras.layers import Lambda
from tensorflow.keras.layers import Input, Dense, Conv2D, Conv1D, Flatten, BatchNormalization, Activation, LeakyReLU, add, Subtract
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Dense, Activation, Flatten, Input, Concatenate, Reshape
from tensorflow.keras.optimizers import Adam, Nadam
from tensorflow.keras.utils import plot_model
from tensorflow.keras.initializers import RandomNormal
from tensorflow.keras import regularizers

import tensorflow.keras.backend as K
def residual_layer( input_block, filters, kernel_size, use_bias=True, use_reg = True,
                reg_const = 0.0001):

    x = input_block
    if use_reg:
        x = Conv2D(
            filters=filters
            , kernel_size=kernel_size
            , data_format="channels_first"
            , padding='same'
            , use_bias=use_bias
            , activation='linear'
            , kernel_regularizer=regularizers.l2(reg_const)
        )(x)
    else:
        x = Conv2D(
            filters=filters
            , kernel_size=kernel_size
            , data_format="channels_first"
            , padding='same'
            , use_bias=use_bias
            , activation='linear'
        )(x)

    x = BatchNormalization()(x)
    #x = BatchNormalization(axis=1)(x)

    x = add([input_block, x])

    x = LeakyReLU()(x)

    return (x)


def conv_layer( x, filters, kernel_size, padding='same', use_bias=True, use_reg = True,
                reg_const = 0.0001, use_batchnorm = True, use_LeakyReLU = True):
    if use_reg:
        x = Conv2D(
            filters=filters
            , kernel_size=kernel_size
            , data_format="channels_first"
            , padding=padding
            , use_bias=use_bias
            , activation='relu'
            , kernel_regularizer=regularizers.l2(reg_const)
        )(x)
    else:
        x = Conv2D(
            filters=filters
            , kernel_size=kernel_size
            , data_format="channels_first"
            , padding=padding
            , use_bias=use_bias
            , activation='relu'
        )(x)
    if use_batchnorm:
        x = BatchNormalization(axis=1)(x)
    if use_LeakyReLU:
        x = LeakyReLU()(x)

    return (x)


def LeakyReLU_layer(x):
    x = LeakyReLU()(x)

    return (x)

def value_head(x, use_bias = True, use_reg = True,
                reg_const = 0.0001, use_batchnorm = True, use_LeakyReLU = True):
    if use_reg:
        x = Conv2D(
            filters=1
            , kernel_size=(1, 1)
            , data_format="channels_first"
            , padding='same'
            , use_bias=use_bias
            , activation='linear'
            , kernel_regularizer=regularizers.l2(reg_const)
        )(x)
    else:
        x = Conv2D(
            filters=1
            , kernel_size=(1, 1)
            , data_format="channels_first"
            , padding='same'
            , use_bias=use_bias
            , activation='linear'
        )(x)

    if use_batchnorm:
        x = BatchNormalization(axis=1)(x)
    if use_LeakyReLU:
        x = LeakyReLU()(x)

    x = Flatten()(x)

    x = Dense(
        29*29*2
        , use_bias=use_bias
        , activation='relu'
        , kernel_regularizer=regularizers.l2(reg_const)
    )(x)
    if use_LeakyReLU:
        x = LeakyReLU()(x)

    x = Dense(
        1
        , use_bias=use_bias
        , activation='tanh'
        , kernel_regularizer=regularizers.l2(reg_const)
        , name='value_head'
    )(x)

    return (x)


def build_model_v01_only_value(weightsFileName = None):

    main_input = Input(shape=(3,19,19,), name='main_input')


    x = conv_layer(main_input, 64, (5,5), padding='same')
    x = conv_layer(x, 64, (3, 3), padding='same')
    x = conv_layer(x, 64, (3, 3), padding='same')
    x = conv_layer(x, 64, (3, 3), padding='same')


    for h in range(5):
        x = residual_layer(x, 64, (3,3))

    vh = value_head(x)

    modelv = Model(inputs=[main_input], outputs=[vh])

    modelv.compile(loss={'value_head': 'mean_squared_error'},
                  optimizer=Adam(), metrics=['mae'])



    plot_model(modelv, 'gb_modelv_01.png', show_shapes=True)

    if weightsFileName != None:
        modelv.load_weights(weightsFileName)

    return modelv


vm = build_model_v01_only_value("memoryweights/vm_0018.hdf5")

Xin= np.zeros((noOfMirrorsForEval+1, 3, 19, 19), dtype=float)
# ... some calculation of Xin
v_predict = vm.predict(Xin).flatten()

标签： pythontensorflowkeras