else:
mask_pred_val = model.predict(val)
# save my val pred pickle
model_name = model_filename.split('.')[1].split('/')[2]
my_val_dir = './results/my_val_pkl'
#if test_aug:
# pkl.dump(mask_pred_val, open(os.path.join(my_val_dir, model_name + '_testaug.pkl'), 'wb'))
#else:
# pkl.dump(mask_pred_val, open(os.path.join(my_val_dir, model_name + '.pkl'), 'wb'))
for pixel_threshold in [0.5]: #np.arange(0.3,1,0.05):
mask_pred_val = np.where(mask_pred_val>=pixel_threshold, 1, 0)
mask_pred_val = mask_pred_val * 255
mask_pred_val = mask_pred_val.astype(np.uint8)
dice, jacc, jacc_th = dice_jacc_mean(val_mask, mask_pred_val)
print model_filename
print "Resized val dice coef : {:.4f}".format(dice)
print "Resized val jacc coef : {:.4f}".format(jacc)
print "Resized val jacc_th coef : {:.4f}".format(jacc_th)
if test_aug:
pkl.dump((dice, jacc, jacc_th), open(os.path.join(my_val_dir, model_name + '_testaug_score.pkl'), 'wb'))
else:
pkl.dump((dice, jacc, jacc_th), open(os.path.join(my_val_dir, model_name + '_score.pkl'), 'wb'))
def predict_challenge(challenge_folder, challenge_predicted_folder, plot=False):
challenge_list = ISIC.list_from_folder(challenge_folder)
challenge_resized_folder = challenge_folder + "_{}_{}".format(height, width)
if not os.path.exists(challenge_resized_folder):
train = None
train_mask = None # clear memory
print("Load best checkpoint")
model.load_weights(model_filename) # load best saved checkpoint
# evaluate model
mask_pred_val = model.predict(val)
mask_pred_test = model.predict(test)
for pixel_threshold in [0.5]: #np.arange(0.3,1,0.05):
mask_pred_val = np.where(mask_pred_val >= pixel_threshold, 1, 0)
mask_pred_val = mask_pred_val * 255
mask_pred_val = mask_pred_val.astype(np.uint8)
print("Validation Predictions Max: {}, Min: {}".format(
np.max(mask_pred_val), np.min(mask_pred_val)))
print(model.evaluate(val, val_mask, batch_size=batch_size, verbose=1))
dice, jacc = dice_jacc_mean(val_mask, mask_pred_val, smooth=0)
print(model_filename)
print("Resized val dice coef : {:.4f}".format(dice))
print("Resized val jacc coef : {:.4f}".format(jacc))
mask_pred_test = np.where(mask_pred_test >= pixel_threshold, 1, 0)
mask_pred_test = mask_pred_test * 255
mask_pred_test = mask_pred_test.astype(np.uint8)
print(model.evaluate(test, test_mask, batch_size=batch_size,
verbose=1))
dice, jacc = dice_jacc_mean(test_mask, mask_pred_test, smooth=0)
print("Resized test dice coef : {:.4f}".format(dice))
print("Resized test jacc coef : {:.4f}".format(jacc))
else:
print('Load model')
model.load_weights(model_filename)
# evaluate model
if test_aug:
mask_pred_val = np.array([my_predict(model, x) for x in tqdm(val)])
else:
mask_pred_val = model.predict(val, batch_size=batch_size)[:, 0, :, :]
hist, bin_edges = np.histogram(mask_pred_val.flatten(), bins=5)
print hist
print bin_edges
for pixel_threshold in [0.5]: #, 0.5, 0.4]:
print "pixel_threshold is {}".format(pixel_threshold)
mask_pred_val_256 = (np.where(mask_pred_val >= pixel_threshold, 1, 0) *
255).astype(np.uint8)
dice, jacc, jacc_th = dice_jacc_mean(val_mask, mask_pred_val_256)
print "Resized val dice coef : {:.4f}".format(dice)
print "Resized val jacc coef : {:.4f}".format(jacc)
if do_filter:
print 'Use Classifier model to filter the masks'
size_1_2 = 256
base_folder = './datasets/isic2018_crop_{}_{}_{}'.format(
attribute, size_1_2, size_1_2)
image_folder = os.path.join(base_folder, 'image')
val = ISIC.load_images(val_list, size_1_2, size_1_2, image_folder)
# filter by pre_cls model
pre_cls_model_filename = './weights2018_task2/{}_{}_crop_pre_cls.h5'.format(
'vgg', size_1_2)
pre_cls_model = models.VGG16(size_1_2,
size_1_2,
if do_evaluate:
model.load_weights(model_filename)
val_pred_mask = model.predict(val_image)
save_pred_images(val_pred_mask, val_list, pred_val_mask_folder)
for pixel_threshold in [0.5]: # np.arange(0.3, 1, 0.05):
val_pred_mask = np.where(val_pred_mask >= pixel_threshold, 1, 0)
val_pred_mask = val_pred_mask * 255
val_pred_mask = val_pred_mask.astype(np.uint8)
print "Validation Prediction Max:{}, Min:{}".format(
np.max(val_pred_mask), np.min(val_pred_mask))
print model.evaluate(val_image,
val_mask,
batch_size=batch_size,
verbose=1)
dice, jacc = dice_jacc_mean(val_mask, val_pred_mask, smooth=0)
print "Resized val dice coef: {:.4f}".format(dice)
print "Resized val jacc coef: {:.4f}".format(jacc)
if do_test:
[te_image,
te_mask] = ISIC.load_image(base_te_image_folder, base_te_mask_folder,
te_list, height, width, remove_mean_imagenet,
remove_mean_dataset, rescale_mask)
te_image = te_image / 127.5 - 1
model.load_weights(model_filename)
te_pred_mask = model.predict(te_image)
save_pred_images(te_pred_mask, te_list, pred_te_mask_folder)
for pixel_threshold in [0.5]: # np.arange(0.3, 1, 0.05):
te_pred_mask = np.where(te_pred_mask >= pixel_threshold, 1, 0)
te_pred_mask = te_pred_mask * 255