How to fourth layer to tensor to let convolution consume it

LazGp 0
Viewed 23

I am trying to add preprocessing layers to a tensorflow model but can't figure out how to use tf.keras.layers.Reshape correctly.

I have some previous preprocessing layers that bring a numpy image to a tensorflow tensor of shape TensorShape([50, 50, 3]).

The first layer of the model that I am trying to connect these preprocessing layers to is a convolutional layer which requires a four dimensional input. When I call the model on a tensor of dimensions Tensorshape([50,50,3]), with three dimensions, I get the error:

Input 0 of layer "Conv1" is incompatible with the layer: expected min_ndim=4, found ndim=3. 
Full shape received: (50, 50, 3)

Call arguments received by layer "model_1" (type Functional):
  • inputs=tf.Tensor(shape=(50, 50, 3), dtype=float32)
  • training=False
  • mask=None

Changing the input tensor to a numpy array and sizing it to a (1,50,50,3) array, then inputting that into the model works fine. I want to do this with a tensorflow layer however so that I can save the preprocessing and model inference into a single Saved Model format tensorflow file without having to do python preprocessing.

tf.expand_dims(input_tensor, axis=0) works, but it's not a layer object so I can't use it. It looks like tf.keras.layers.Reshape((1,50,50,3), input_shape=(50,50,3)) or tf.keras.layers.Reshape((1,50,50,3)) is the way to go then.

My silly problem is that I just can't figure out how to use tf.keras.layers.Reshape, even using the tf/keras documentation. When I pass in a tensor of dimensions TensorShape([50, 50, 3]), to tf.keras.layers.Reshape((1,50,50,3)) I get the error message:

Input to reshape is a tensor with 7500 values, but the requested shape has 375000 [Op:Reshape]

Call arguments received by layer "reshape_9" (type Reshape):
  • inputs=tf.Tensor(shape=(50, 50, 3), dtype=float32)

So, how do I get this to work? All I want is to end up with the same Tensor as inputted but with TensorShape([1,50,50,3]) instead of TensorShape([50,50,3]), and I want this to happen with a tf.keras.layers object. So really I just want to perform an identity transformation that adds a fourth dimension of size one at the beginning of a tensor so that a convolutional layer can consume it, and I want this from a tf.keras.layers object.

deep-learning machine-learning keras tensorflow python
2 Answers

It's because the images are put together into batches. Try

tf.keras.layers.Reshape((-1,50,50,3))

Or you can use a Lambda layer instead:

tf.keras.layers.Lambda(lambda x: tf.expand_dims(x, axis=-1))

The value -1 means the reshape or expand_dims function will calculate that shape dimension for you, depending on the number of values it receives.

AndrzejO0

You need to make sure to add an extra dimension for the batch size, if you are passing in a single image the batch size would be 1. You can use np.expand_dims to add the extra dimension.

Rarblack0
Related