Add a resizing layer to a keras sequential model
I think you should consider using tensorflow's resize_images layer.
https://www.tensorflow.org/api_docs/python/tf/image/resize_images
It appears keras does not include this, and perhaps because the feature does not exist in theano. I have written a custom keras layer that does the same. It's a quick hack, so it might not work well in your case.
import keras
import keras.backend as K
from keras.utils import conv_utils
from keras.engine import InputSpec
from keras.engine import Layer
from tensorflow import image as tfi
class ResizeImages(Layer):
"""Resize Images to a specified size
# Arguments
output_size: Size of output layer width and height
data_format: A string,
one of `channels_last` (default) or `channels_first`.
The ordering of the dimensions in the inputs.
`channels_last` corresponds to inputs with shape
`(batch, height, width, channels)` while `channels_first`
corresponds to inputs with shape
`(batch, channels, height, width)`.
It defaults to the `image_data_format` value found in your
Keras config file at `~/.keras/keras.json`.
If you never set it, then it will be "channels_last".
# Input shape
- If `data_format='channels_last'`:
4D tensor with shape:
`(batch_size, rows, cols, channels)`
- If `data_format='channels_first'`:
4D tensor with shape:
`(batch_size, channels, rows, cols)`
# Output shape
- If `data_format='channels_last'`:
4D tensor with shape:
`(batch_size, pooled_rows, pooled_cols, channels)`
- If `data_format='channels_first'`:
4D tensor with shape:
`(batch_size, channels, pooled_rows, pooled_cols)`
"""
def __init__(self, output_dim=(1, 1), data_format=None, **kwargs):
super(ResizeImages, self).__init__(**kwargs)
data_format = conv_utils.normalize_data_format(data_format)
self.output_dim = conv_utils.normalize_tuple(output_dim, 2, 'output_dim')
self.data_format = conv_utils.normalize_data_format(data_format)
self.input_spec = InputSpec(ndim=4)
def build(self, input_shape):
self.input_spec = [InputSpec(shape=input_shape)]
def compute_output_shape(self, input_shape):
if self.data_format == 'channels_first':
return (input_shape[0], input_shape[1], self.output_dim[0], self.output_dim[1])
elif self.data_format == 'channels_last':
return (input_shape[0], self.output_dim[0], self.output_dim[1], input_shape[3])
def _resize_fun(self, inputs, data_format):
try:
assert keras.backend.backend() == 'tensorflow'
assert self.data_format == 'channels_last'
except AssertionError:
print "Only tensorflow backend is supported for the resize layer and accordingly 'channels_last' ordering"
output = tfi.resize_images(inputs, self.output_dim)
return output
def call(self, inputs):
output = self._resize_fun(inputs=inputs, data_format=self.data_format)
return output
def get_config(self):
config = {'output_dim': self.output_dim,
'padding': self.padding,
'data_format': self.data_format}
base_config = super(ResizeImages, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
I thought I should post an updated answer, since the accepted answer is wrong and there are some major updates in the recent Keras release.
To add a resizing layer, according to documentation:
tf.keras.layers.experimental.preprocessing.Resizing(height, width, interpolation="bilinear", crop_to_aspect_ratio=False, **kwargs)
For you, it should be:
from tensorflow.keras.layers.experimental.preprocessing import Resizing
model = Sequential()
model.add(Resizing(224,224))
The accepted answer uses the Reshape layer, which works like NumPy's reshape, which can be used to reshape a 4x4 matrix into a 2x8 matrix, but that will result in the image loosing locality information:
0 0 0 0
1 1 1 1 -> 0 0 0 0 1 1 1 1
2 2 2 2 2 2 2 2 3 3 3 3
3 3 3 3
Instead, image data should be rescaled / "resized" using, e.g., Tensorflows image_resize
.
But beware about the correct usage and the bugs!
As shown in the related question, this can be used with a lambda layer:
model.add( keras.layers.Lambda(
lambda image: tf.image.resize_images(
image,
(224, 224),
method = tf.image.ResizeMethod.BICUBIC,
align_corners = True, # possibly important
preserve_aspect_ratio = True
)
))
In your case, as you have a 160x320 image, you also have to decide whether to keep the aspect ratio, or not. If you want to use a pre-trained network, then you should use the same kind of resizing that the network was trained for.