train_model_a.py

Question

有没有办法在另一个模型B中使用tf.estimator训练的模型A？

这是情况，假设我有一个训练有素的模型A＆＃39;使用model_a_fn（）。＆＃39;模型A＆＃39;获取图像作为输入，并输出一些类似于MNIST分类器的向量浮动值。还有另一个＆＃39;模型B＆＃39;在model_b_fn（）中定义。它还将图像作为输入，并且需要矢量输出＆＃39;模型A＆＃39;在培训＆＃39;模型B＆＃39;。

基本上我想训练＆＃39;模型B＆＃39;需要输入作为图像＆amp;预测输出＆＃39;模型A＆＃39;。（不再需要训练＆＃39;模型A＆＃39;只是为了在训练时获得预测输出＆＃39;模型B＆＃39;）

我已经尝试过三种情况：

在model_b_fn（）
已导出＆＃39;模型A＆＃39;使用tf.estimator.export_savedmodel（），并创建预测函数。用params dict将它传递给model_b_fn（）。
与2相同，但恢复＆＃39;模型A＆＃39;在model_b_fn（）

但所有案例都显示错误：

...必须与...相同的图表。
TypeError：无法挑选_thread.RLock对象
TypeError：Feed的值不能是tf.Tensor对象。

这是我用过的代码......只附上重要的部分

train_model_a.py

def model_a_fn(features, labels, mode, params):
    # ...
    # ...
    # ...
    return

def main():
    # model checkpoint location
    model_a_dir = './model_a'

    # create estimator for Model A
    model_a = tf.estimator.Estimator(model_fn=model_a_fn, model_dir=model_a_dir)

    # train Model A
    model_a.train(input_fn=lambda : input_fn_a)
    # ...
    # ...
    # ...

    # export model a
    model_a.export_savedmodel(model_a_dir, serving_input_receiver_fn=serving_input_receiver_fn)
    # exported to ./model_a/123456789
    return

if __name__ == '__main__':
    main()

train_model_b_case_1.py

# follows model_a's input format
def bypass_input_fn(x):
    features = {
        'x': x,
    }
    return features

def model_b_fn(features, labels, mode, params):
    # parse input
    inputs = tf.reshape(features['x'], shape=[-1, 28, 28, 1])

    # get Model A's response
    model_a = params['model_a']
    predictions = model_a.predict(
        input_fn=lambda: bypass_input_fn(inputs)
    )
    for results in predictions:
        # Error occurs!!!
        model_a_output = results['class_id']

    # build Model B
    layer1 = tf.layers.conv2d(inputs, 32, 5, same, activation=tf.nn.relu)
    layer1 = tf.layers.max_pooling2d(layer1, pool_size=[2, 2], strides=2)

    # ...
    # some layers added...
    # ...

    flatten = tf.layers.flatten(prev_layer)
    layern = tf.layers.dense(10)

    # let say layern's output shape and model_a_output's output shape is same
    add_layer = tf.add(flatten, model_a_output)

    # ...
    # do more... stuff
    # ...
    return

def main():
    # load pretrained model A
    model_a_dir = './model_a'
    model_a = tf.estimator.Estimator(model_fn=model_a_fn, model_dir=model_a_dir)

    # model checkpoint location
    model_b_dir = './model_b/'

    # create estimator for Model A
    model_b = tf.estimator.Estimator(
        model_fn=model_b_fn,
        model_dir=model_b_dir,
        params={
            'model_a': model_a,
        }
    )

    # train Model B
    model_b.train(input_fn=lambda : input_fn_b)
    return

if __name__ == '__main__':
    main()

train_model_b_case_2.py

def model_b_fn(features, labels, mode, params):
    # parse input
    inputs = tf.reshape(features['x'], shape=[-1, 28, 28, 1])

    # get Model A's response
    model_a_predict_fn = params['model_a_predict_fn']
    model_a_prediction = model_a_predict_fn(
        {
            'x': inputs
        }
    )
    model_a_output = model_a_prediction['output']

    # build Model B
    layer1 = tf.layers.conv2d(inputs, 32, 5, same, activation=tf.nn.relu)
    layer1 = tf.layers.max_pooling2d(layer1, pool_size=[2, 2], strides=2)

    # ...
    # some layers added...
    # ...

    flatten = tf.layers.flatten(prev_layer)
    layern = tf.layers.dense(10)

    # let say layern's output shape and model_a_output's output shape is same
    add_layer = tf.add(flatten, model_a_output)

    # ...
    # do more... stuff
    # ...
    return

def main():
    # load pretrained model A
    model_a_dir = './model_a/123456789'
    model_a_predict_fn = tf.contrib.predictor.from_saved_model(export_dir=model_a_dir)

    # model checkpoint location
    model_b_dir = './model_b/'

    # create estimator for Model A
    # Error occurs!!!
    model_b = tf.estimator.Estimator(
        model_fn=model_b_fn,
        model_dir=model_b_dir,
        params={
            'model_a_predict_fn': model_a_predict_fn,
        }
    )

    # train Model B
    model_b.train(input_fn=lambda : input_fn_b)
    return

if __name__ == '__main__':
    main()

train_model_b_case_3.py

def model_b_fn(features, labels, mode, params):
    # parse input
    inputs = tf.reshape(features['x'], shape=[-1, 28, 28, 1])

    # get Model A's response
    model_a_predict_fn = tf.contrib.predictor.from_saved_model(export_dir=params['model_a_dir'])
    # Error occurs!!!
    model_a_prediction = model_a_predict_fn(
        {
            'x': inputs
        }
    )
    model_a_output = model_a_prediction['output']

    # build Model B
    layer1 = tf.layers.conv2d(inputs, 32, 5, same, activation=tf.nn.relu)
    layer1 = tf.layers.max_pooling2d(layer1, pool_size=[2, 2], strides=2)

    # ...
    # some layers added...
    # ...

    flatten = tf.layers.flatten(prev_layer)
    layern = tf.layers.dense(10)

    # let say layern's output shape and model_a_output's output shape is same
    add_layer = tf.add(flatten, model_a_output)

    # ...
    # do more... stuff
    # ...
    return

def main():
    # load pretrained model A
    model_a_dir = './model_a/123456789'

    # model checkpoint location
    model_b_dir = './model_b/'

    # create estimator for Model A
    # Error occurs!!!
    model_b = tf.estimator.Estimator(
        model_fn=model_b_fn,
        model_dir=model_b_dir,
        params={
            'model_a_dir': model_a_dir,
        }
    )

    # train Model B
    model_b.train(input_fn=lambda : input_fn_b)
    return

if __name__ == '__main__':
    main()

所以在另一个tf.estimator中使用经过培训的自定义tf.estimator的任何想法都可以吗？

Answer 1

我已经找到解决这个问题的一种方法。

如果遇到相同的问题，可以使用此方法。

创建一个运行tensorflow.contrib.predictor.from_saved_model（）的函数->将其称为'pretrained_predictor（）'
在模型B的model_fn（）中，在预定义的'pretrained_predictor（）'上方调用tensorflow.py_func（）

例如，请参见https://github.com/moono/tf-cnn-mnist/blob/master/4_3_estimator_within_estimator.py以获取简单的用例。

Tensorflow，在另一个tf.estimator model_fn中使用tf.estimator训练的模型

train_model_a.py

train_model_b_case_1.py

train_model_b_case_2.py

train_model_b_case_3.py

1 个答案: