建议在 tensorflow 2.0 中调试 `tf.data.Dataset` 操作

Question

What is the equivalent of Panda's df.head() for tf datasets?

Following the documentation here I've constructed the following toy examples:

dset = tf.data.Dataset.from_tensor_slices((tf.constant([1.,2.,3.]), tf.constant([4.,4.,4.]), tf.constant([5.,6.,7.])))
print(dset)

outputs

<TensorSliceDataset shapes: ((), (), ()), types: (tf.float32, tf.float32, tf.float32)>

I would prefer to get back something resembling a tensor , so to get some values I'll make an iterator.

dset_iter = dset.__iter__()
print(dset_iter.next())

outputs

(<tf.Tensor: id=122, shape=(), dtype=float32, numpy=1.0>,
 <tf.Tensor: id=123, shape=(), dtype=float32, numpy=4.0>,
 <tf.Tensor: id=124, shape=(), dtype=float32, numpy=5.0>)

So far so good. Let's try some windowing...

windowed = dset.window(2)
print(windowed)

outputs

<WindowDataset shapes: (<tensorflow.python.data.ops.dataset_ops.DatasetStructure object at 0x1349b25c0>, <tensorflow.python.data.ops.dataset_ops.DatasetStructure object at 0x1349b27b8>, <tensorflow.python.data.ops.dataset_ops.DatasetStructure object at 0x1349b29b0>), types: (<tensorflow.python.data.ops.dataset_ops.DatasetStructure object at 0x1349b25c0>, <tensorflow.python.data.ops.dataset_ops.DatasetStructure object at 0x1349b27b8>, <tensorflow.python.data.ops.dataset_ops.DatasetStructure object at 0x1349b29b0>)>

Ok, use the iterator trick again:

windowed_iter = windowed.__iter__()
windowed_iter.next()

outputs

(<_VariantDataset shapes: (), types: tf.float32>,
 <_VariantDataset shapes: (), types: tf.float32>,
 <_VariantDataset shapes: (), types: tf.float32>)

What? A WindowDataset 's iterator gives back a tuple of other dataset objects?
I would expect the first item in this WindowDataset to be the tensor with values [[1.,4.,5.],[2.,4.,6.]]. Maybe this is still true, but it isn't readily apparent to me from this 3-tuple of datasets. Ok. Let's get their iterators...

vd = windowed_iter.get_next()
vd0, vd1, vd2 = vd[0], vd[1], vd[2]
vd0i, vd1i, vd2i = vd0.__iter__(), vd1.__iter__(), vd2.__iter__()
print(vd0i.next(), vd1i.next(), vd2i.next())

outputs

(<tf.Tensor: id=357, shape=(), dtype=float32, numpy=1.0>,
 <tf.Tensor: id=358, shape=(), dtype=float32, numpy=4.0>,
 <tf.Tensor: id=359, shape=(), dtype=float32, numpy=5.0>)

As you can see, this workflow is quickly becoming a mess. I like how Tf2.0 is attempting to make the framework more interactive and pythonic. Are there good examples of the datasets api conforming to this vision too?

Answer 1

I was in a similar situation. I eventually ended up using zip

train_dataset = train_dataset.window(10, shift=5)
for step_dataset in train_dataset:
    for (images, labels, paths) in zip(*step_dataset):
        train_step(images, labels)