Convolutional Net With Lasange - image resizing

Question

I've been training some neural and convolutional nets with Lasagne and was doing most of my data/image preprocessing using Python. However I would like to incorporate some of this into my Lasagne layers to make my code more flexible.

Is there a Lasagne layer which can resize the input images?

score 0 · Answer 1 · answered Sep 17 '15 at 11:13

Rather than doing that in a layer, you can use a nolearn.lasagne.BatchIterator; in the following snippet I resample my original 1D signal to a 1000 points signal:

from nolearn.lasagne import BatchIterator
from scipy.signal import resample
import numpy as np

class ResampleIterator(BatchIterator):

    def __init__(self, batch_size, newSize):
        super(ResampleIterator, self).__init__(batch_size)
        self.newSize = newSize

    def transform(self, Xb, yb):
        X_new = resample(Xb, self.newSize, axis=2).astype(np.float32)    
        return X_new, yb


 myNet = NeuralNet(
 # define your usual other parameters (layers, etc) here

    # and these are the lines you are interested in:
    batch_iterator_train=CropIterator(batch_size=128, newSize=1000),
    batch_iterator_test=CropIterator(batch_size=128, newSize=1000),
    )

I don't know if you've used nolearn yet, you can read more about it (installation, examples) here

Convolutional Net With Lasange - image resizing

1 Answers1