def generate_predictions_with_augmented_images(net, X, combinations):
t0 = time()
n = X.shape[0]
num_features = 121
num_combination = len(combinations)
predictions = np.empty((n, num_features))
# TODO optimize by vectorizing the for loop
# (although X_aug will be very very big and probably won't fit in memory?)
for i, x in enumerate(X):
if i % 1000 == 0:
print('Finished predicting %i images. Took %i seconds so far' % (i, time() - t0))
# Image only has one channel (grayscale)
# im_affine_transform only works for images with shape of (height, width)
x = x[0]
x_aug = np.array([
im_affine_transform(x, scale=scale, rotation=rotation, translate_y=translate_y, translate_x=translate_x)
for scale, rotation, translate_y, translate_x in combinations
])
x_aug = x_aug.reshape(num_combination, 1, args.hw, args.hw).astype(np.float32)
# TODO do we need to normalize this back to 0?
pred = net.predict_proba(x_aug).mean(axis=0)
predictions[i] = pred
return predictions
t0 = time()
# For each image generate a number of variation of the same images using
# affine transform and average out prediction
predictions = np.empty((X.shape[0], num_combination, 121))
for i, x in enumerate(X):
if i % 2500 == 0:
print('Finished predicting %i images. Took %i seconds so far' %
(i, time() - t0))
# Image only has one channel (grayscale)
# im_affine_transform only works for images with shape of (height, width)
x = x[0]
x_aug = np.array([
im_affine_transform(x,
scale=scale,
rotation=rotation,
translate_y=translate_y,
translate_x=translate_x)
for scale, rotation, translate_y, translate_x in combinations
])
x_aug = x_aug.reshape(num_combination, 1, hw, hw).astype(np.float32)
pred = net.predict_proba(x_aug)
predictions[i] = pred
print('Finished predicting all %i images. Took %i seconds in total' %
(X.shape[0], time() - t0))
np.save(args.pred_aug_fname, predictions)
print('Predictions saved to %s' % args.pred_aug_fname)
print(predictions.shape)