def predict_mask(test_img, crop_num = 5000, model_path = './scadec_Hydra/maskCNN/CNN_mask-2000.meta'):
cropSize = (28,28)
if len(np.shape(test_img)) == 3:
test_img = test_img[:,:,0]
test_img_patches, boxes = random_crop(test_img, crop_size = cropSize, crop_num = crop_num,
save = False, saveMode='npy')
test_img_patches = np.reshape(test_img_patches,list(np.shape(test_img_patches))+[1])
test_img_patches_grad = np.reshape(get_gradients(test_img_patches[:,:,:,0]),(len(boxes), 28, 28, 1))
test_data = np.zeros((len(boxes), 28,28,2))
for idx in range(len(boxes)):
test_data[idx, :, :, :] = np.concatenate([test_img_patches[idx,:,:], test_img_patches_grad[idx,:,:]], axis = 2)
with tf.Session() as sess:
new_saver = tf.train.import_meta_graph(model_path)
new_saver.restore(sess, '.'.join(model_path.split('.')[:-1]))
graph = tf.get_default_graph()
xs = graph.get_operation_by_name('xs').outputs[0]
pred = graph.get_operation_by_name('pred').outputs[0]
y_pre_raw = sess.run(pred, feed_dict={xs:test_data})
# print(y_pre_raw[:50,0])
y_pre = y_pre_raw[:,0]>0.9
mask = np.zeros(np.shape(test_img))
for idx, box in enumerate(boxes):
if y_pre[idx] == True:
mask[box[0]:box[1],box[2]:box[3]] += 0.01
return mask
def predict_masks_old2(test_imgs, crop_num = 5000, model_path = './scadec_Hydra/maskCNN/CNN_mask-2000.meta'):
print('Computing masks...')
startTime = time.time()
# 1. Get croped patches
img_n, img_h, img_w, img_c = np.shape(test_imgs)
masks = np.zeros(np.shape(test_imgs))
cropSize = (28,28)
with tf.Session() as sess:
# Get feed data(img+gradient)
new_saver = tf.train.import_meta_graph(model_path)
new_saver.restore(sess, '.'.join(model_path.split('.')[:-1]))
graph = tf.get_default_graph()
xs = graph.get_operation_by_name('xs').outputs[0]
pred = graph.get_operation_by_name('pred').outputs[0]
# Get patches
print('Getting patches...')
data = []
for imgIdx in range(img_n):
img = np.squeeze(test_imgs[imgIdx])
img_patches, boxes = random_crop(img, crop_size = cropSize, crop_num = crop_num,
save = False, saveMode='npy')
img_patches = np.reshape(img_patches,list(np.shape(img_patches))+[1])
img_patches_grad = np.reshape(get_gradients(img_patches[:,:,:,0]),(len(boxes), 28, 28, 1))
datum = np.zeros((len(boxes), 28,28,2))
for idx in range(len(boxes)):
datum[idx, :, :, :] = np.concatenate([img_patches[idx,:,:], img_patches_grad[idx,:,:]], axis = 2)
data.append(datum)
# Get Predictions
print('Predicting...')
y_pre_raws = []
for imgIdx in range(img_n):
y_pre_raws.append(sess.run(pred, feed_dict={xs:data[imgIdx]}))
# Get masks
# y_pres = []
print('Computing masks...')
for imgIdx in range(img_n):
y_pre = y_pre_raws[imgIdx][:,0]>0.9
mask = np.zeros(np.shape(img))
for idx, box in enumerate(boxes):
if y_pre[idx] == True:
mask[box[0]:box[1],box[2]:box[3]] += 0.01
masks[imgIdx] = np.reshape(mask, [img_h, img_w, img_c])
endTime = time.time()
print('Finished computing masks within {%.2f} mins for {} images!'.format((endTime - startTime)/60, img_n))
return masks
np.arange(99, 245, 5), :, :, :]
return data, vdata, truths, vtruths
#%% 1. Getting patches
crop_num = 5000
cropSize = (28, 28)
imgs = get_data(para_dict_use_train, DEBUG_MODE=False)
img_n, img_h, img_w, img_c = np.shape(imgs)
data = np.zeros([img_n * crop_num, img_h, img_w, img_c])
for imgIdx in range(img_n):
img = np.squeeze(imgs[imgIdx])
img_patches, boxes = random_crop(img,
crop_size=cropSize,
crop_num=crop_num,
save=False,
saveMode='npy')
img_patches = np.reshape(img_patches, list(np.shape(img_patches)) + [1])
img_patches_grad = np.reshape(get_gradients(img_patches[:, :, :, 0]),
(len(boxes), 28, 28, 1))
datum = np.zeros((len(boxes), 28, 28, 2))
for idx in range(len(boxes)):
datum[idx, :, :, :] = np.concatenate(
[img_patches[idx, :, :], img_patches_grad[idx, :, :]], axis=2)
data[img_n]
model_path = './scadec_Hydra/maskCNN/CNN_mask-2000.meta'
def predict_masks(test_imgs,
crop_num=5000,
model_path='./scadec_Hydra/maskCNN/CNN_mask-2000.meta',
save_path='..data/masks/',
saveName='patches',
loadName='patches',
savePatches=False,
loadPatches=False,
saveYPre=False,
loadYPre=False,
saveMasks=False,
loadMaskes=False): #, savePath = None):
# import os
# print(os.getcwd())
print('Computing masks...')
startTime = time.time()
# 1. Get croped patches
img_n, img_h, img_w, img_c = np.shape(test_imgs)
# print(np.shape(masks))
cropSize = (28, 28)
savefile_img_num = 2000
with tf.Session() as sess:
# Get feed data(img+gradient)
new_saver = tf.train.import_meta_graph(model_path)
new_saver.restore(sess, '.'.join(model_path.split('.')[:-1]))
graph = tf.get_default_graph()
xs = graph.get_operation_by_name('xs').outputs[0]
pred = graph.get_operation_by_name('pred').outputs[0]
# Get patches
print('Getting patches...')
if loadPatches:
patches, boxeses = np.load(save_path +
'{}_{}_{}_patches_boxeses.npy'.format(
loadName, cropSize[0], cropSize[1]))
# with open('./scadec_Hydra/maskCNN/pre_computed/{}.pkl'.format(loadName), 'rb') as f: # Python 3: open(..., 'wb')
# data = pickle.load(f)
elif (not loadYPre) and (not loadMaskes):
patches = np.zeros([img_n * crop_num, cropSize[0], cropSize[1], 2])
boxeses = []
# for preTraIdx in tqdm(range(500), ncols=75):
for imgIdx in tqdm(range(img_n), ncols=75):
img = np.squeeze(test_imgs[imgIdx])
img_patches, boxes = random_crop(img,
crop_size=cropSize,
crop_num=crop_num,
save=False,
saveMode='npy')
img_patches = np.reshape(img_patches,
list(np.shape(img_patches)) + [1])
img_patches_grad = np.reshape(
get_gradients(img_patches[:, :, :, 0]),
(len(boxes), 28, 28, 1))
datum = np.zeros((len(boxes), 28, 28, 2))
for idx in range(len(boxes)):
datum[idx, :, :, :] = np.concatenate(
[img_patches[idx, :, :], img_patches_grad[idx, :, :]],
axis=2)
# data.append(datum)
patches[imgIdx * crop_num:(imgIdx + 1) *
crop_num, :, :, :] = datum
boxeses.append(boxes)
if savePatches:
print('Saving...')
# np.save('./scadec_Hydra/maskCNN/pre_computed/{}_{}_{}_patches_boxeses'.format(saveName, cropSize[0], cropSize[1]), [patches, boxeses])
np.save(
save_path +
'{}_{}_{}_patches'.format(saveName, cropSize[0], cropSize[1]),
patches)
# Get Predictions
print('Predicting...')
if loadYPre:
y_pre_raws = np.load(save_path + '{}_{}_{}_ypre_raws.npy'.format(
saveName, cropSize[0], cropSize[1]))
boxeses = np.load(save_path + '{}_{}_{}_boxeses.npy'.format(
saveName, cropSize[0], cropSize[1]))
elif not loadMaskes:
y_pre_raws = []
for imgIdx in tqdm(range(img_n), ncols=75):
y_pre_raws.append(
sess.run(pred,
feed_dict={
xs:
patches[imgIdx * crop_num:(imgIdx + 1) *
crop_num]
}))
# a = 1
if saveYPre:
np.save(
save_path + '{}_{}_{}_ypre_raws.npy'.format(
saveName, cropSize[0], cropSize[1]), y_pre_raws)
np.save(
save_path + '{}_{}_{}_boxeses.npy'.format(
saveName, cropSize[0], cropSize[1]), boxeses)
# Get masks
# y_pres = []
print('Computing masks...')
fMean = np.ones((5, 5)) / 25
masks = np.zeros(np.shape(test_imgs))
for imgIdx in tqdm(range(img_n), ncols=75):
# y_pre = y_pre_raws[imgIdx][:,0]>0.9
boxes = boxeses[imgIdx]
mask = np.zeros(np.shape(test_imgs[0, :, :, 0]))
for idx, box in enumerate(boxes):
mask[box[0]:box[1],
box[2]:box[3]] += 0.01 * y_pre_raws[imgIdx][idx, 0]
# if y_pre[idx] == True:
# mask[box[0]:box[1],box[2]:box[3]] += 0.01
mask = convolve2d(mask, fMean, mode='same')
masks[imgIdx] = np.reshape(mask, [img_h, img_w, img_c])
if saveMasks:
np.save(
save_path +
'{}_{}_{}_masks'.format(saveName, cropSize[0], cropSize[1]),
masks)
endTime = time.time()
print('Finished computing masks for {} within {} mins for {} images!'.
format(saveName, (endTime - startTime) / 60, img_n))
return masks