使用tensorflow估计器创建多输入 [英] what use tensorflow estimator create multi-input

问题描述

对不起，我的英语不好=.=

Sorry my English is poor =。=

我创建了一个keras模型并使用tf.keras.estimator.model_to_estimator转换为估算器，但是该模型是多输入的，我该如何创建数据集来提供数据?

I create a keras model and use tf.keras.estimator.model_to_estimator convert to estimator but the model is multi-input, what can I create Dataset feed the data?

这是我的型号代码:

model = VGG19(include_top=False, input_shape=(182, 182 , 3))
y = model.output
y = keras.layers.Flatten()(y)
y = keras.layers.Dense(512, activation='relu')(y)
y = keras.layers.Dense(256, activation='relu')(y)
y = keras.layers.Dense(128, activation='relu')(y)
model = keras.Model(inputs=model.input, outputs=y)

input_image = keras.layers.Input(shape=(182, 182, 3))
input_anchor = keras.layers.Input(shape=(182, 182, 3))
out_image = model(input_image)
out_anchor = model(input_anchor)

out = keras.layers.concatenate([out_image, out_anchor])
out = keras.layers.Dense(1, activation='sigmoid')(out)
img_model = keras.Model([input_image, input_anchor], out)

face_model.compile(optimizer=tf.train.AdamOptimizer(1e-4, loss='binary_crossentropy', metrics=['accuracy'])

distribution = tf.contrib.distribute.CollectiveAllReduceStrategy(num_gpus_per_worker=0)
config = tf.estimator.RunConfig(model_dir='/home/zjq/test/image_model.h5', train_distribute=distribution)

est_model = tf.keras.estimator.model_to_estimator(keras_model=image_model, config=config)

现在，我有一个输入列表，形状为[(100000，182，182，3)，(100000，182，182，3)，(100000，1)]，如何定义输入函数return tf .data.DataSet?

Now, I have a input list, the shape is [(100000, 182, 182, 3), (100000, 182, 182, 3), (100000, 1)],how to define the input function return tf.data.DataSet?

推荐答案

首先，输入输入占位符:

First, name input placeholders:

input_image = keras.layers.Input(shape=(182, 182, 3),name='image')
input_anchor = keras.layers.Input(shape=(182, 182, 3),name='anchor')

如果输入数据为train_data并且形状为[(100000, 182, 182, 3), (100000, 182, 182, 3), (100000, 1)]，请执行以下操作:

If your input data are train_data and shape is [(100000, 182, 182, 3), (100000, 182, 182, 3), (100000, 1)], then do it:

BATCH_SIZE = 512
EPOCHS = 4

def input_fn(data, epochs, batch_size):
    # Convert the inputs to a Dataset.
    dataset = tf.data.Dataset.from_tensor_slices(({'image':data[0],'anchor':data[1]}, data[2]))
    # Shuffle, repeat, and batch the examples.
    SHUFFLE_SIZE = 1000
    dataset = dataset.shuffle(SHUFFLE_SIZE).repeat(epochs).batch(batch_size)
    dataset = dataset.prefetch(2)
    # Return the dataset.
    return dataset
est_model.train(lambda :input_fn(train_data,EPOCHS,BATCH_SIZE))

可以根据需要调整参数BATCH_SIZE，EPOCHS和SHUFFLE_SIZE.

The parameters BATCH_SIZE,EPOCHS and SHUFFLE_SIZE can be adjusted according to your needs.

这篇关于使用tensorflow估计器创建多输入的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！