python - 有没有办法在 tf.data 管道中使用 tf.keras.model.predict？

Question

我有一个训练有素的模型，我想将其用于第二个模型的 tf.data 管道中。当我尝试执行此操作时，我收到一个 ValueError: Unknown graph。正在中止。我不太清楚如何理解此错误消息。

我的代码看起来像这样:

def load_data(..., model):
    # code to load an image
    files = tf.data.Dataset.from_tensor_slices(file_list)
    images = files.map(load_image_from_file) 

    def pass_image_through_model(img):
        return model.predict(img, steps=1)

    dataset = images.map(pass_image_through_model)
    return dataset

这有什么问题吗？我得到的错误是:

    /home/.../code/dataloader.py:236 pass_image_through_model  *
        return model.predict(img, steps=1)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py:1013 predict
        use_multiprocessing=use_multiprocessing)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_arrays.py:728 predict
        callbacks=callbacks)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_arrays.py:189 model_iteration
        f = _make_execution_function(model, mode)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_arrays.py:571 _make_execution_function
        return model._make_execution_function(mode)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py:2131 _make_execution_function
        self._make_predict_function()
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py:2121 _make_predict_function
        **kwargs)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/backend.py:3760 function
        return EagerExecutionFunction(inputs, outputs, updates=updates, name=name)
    /home/.../anaconda3/envs/masters/lib/python3.7/site-packages/tensorflow_core/python/keras/backend.py:3644 __init__
        raise ValueError('Unknown graph. Aborting.')

    ValueError: Unknown graph. Aborting.

Answer 1

解决此问题的最简单方法之一是将输入直接传递给模型，而不是使用 model.preedit 方法。原因是 model.predict 返回一个 numpy.ndarray。这会导致错误，因为 tf.data 使用图执行，这意味着最好在该图中让任何操作输入并输出张量。

下面是一个快速的工作示例。

import tensorflow as tf

# Create example model
inputs = tf.keras.Input((1,))
out = tf.keras.layers.Dense(1)(inputs)
model = tf.keras.Model(inputs, out)

def map_fn(row):
    return model(row)


# Create some input data 
a = tf.constant([1, 2])

# Create the dataset
ds = tf.data.Dataset.from_tensor_slices(a).batch(1)
model_mapped_ds = ds.map(lambda x: map_fn(x))

for el in model_mapped_ds:
    print(el)

最后，下面是它在您的使用中的样子。

def pass_image_through_model(img):
    return model(img) # this returns a tensor 

@tf.function
def load_data(..., model):
    # code to load an image
    files = tf.data.Dataset.from_tensor_slices(file_list).batch(1) # Don't forget batch size!
    images = files.map(load_image_from_file) 

    dataset = images.map(pass_image_through_model)
    return dataset