def inference(self, sess, data, multi_crop=False):
if multi_crop:
raw_data = data
(crops, flip_crops), idxs, sizes = net.multi_crop(data, size=0.75)
data = np.concatenate( crops + flip_crops, axis=0 )
# print(raw_data.shape, data.shape)
else:
idxs = sizes = None
pred, score, W, F = sess.run([self.pred_op, self.score_op, self.W, self.F], feed_dict={
self.inputs: data,
self.is_training: False
})
if multi_crop:
# print(score.shape)
num_n, num_classes = score.shape[0]/10, score.shape[1]
shape = (10, num_n, num_classes)
score = score.reshape(shape)
score = np.mean(score, axis=0)
pred = np.argmax(score, axis=1)
return pred, score, W, F, idxs, sizes
def multi_crop_test():
from scipy import misc
import matplotlib.pyplot as plt
from net_utils import multi_crop
import numpy as np
img = misc.imread('../temp/images.png', mode='RGB')
crops, mask, flip_res, flip_mask = multi_crop(np.expand_dims(img, axis=0),
size=0.7)
# [top_left, top_right, bottom_left, bottom_right, center]
f, ((ax1, ax2), (ax3, ax4), (ax5, ax6), (ax7, ax8)) = plt.subplots(4, 2)
### ax1: raw image
ax1.imshow(img[:, ::-1, :])
ax1.set_title('Raw')
## ax2: top_left
ax2.imshow(crops[0][0])
ax2.set_title('Top Left')
## ax3: top_right
ax3.imshow(crops[1][0])
ax3.set_title('Top Right')
## ax4: bottom_left
ax4.imshow(crops[2][0])
ax4.set_title('Bottom Left')
## ax5: bottom_right
ax5.imshow(crops[3][0])
ax5.set_title('Bottom Right')
## ax6: center
ax6.imshow(crops[4][0])
ax6.set_title('Center')
# ## ax7: mask
# ax7.imshow(mask[0]*20, cmap=plt.cm.jet)
# ax7.set_title('Mask')
# ## ax8
# ax8.imshow(img)
# ax8.set_title('Raw')
plt.tight_layout()
plt.show()