def _save_preds(
model_path='/data/pneumo_log/val_1/2019_0805_0344/best_weights.hdf5',
data_path='/data/pneumo/fold/1/'):
data_path_list = glob(data_path + '*.npy')
#'/data/pneumo_log/val_1/'
base_save_path = '/'.join(model_path.split('/')[:-2])
#2019_0805_0344
folder_name = model_path.split('/')[-2]
# '/data/pneumo_log/val_1/val_predictions/2019_0805_0344/'
save_path = base_save_path + '/val_predictions/' + folder_name + '/'
# loading model
model = _load_model(model_path=model_path)
print('saving dirs: {}'.format(save_path))
data_prep._make_dir(save_path)
for data_path in tqdm(data_path_list):
# this includes .npy
file_name = data_path.split('/')[-1]
data = np.load(data_path)[()]
img = data['img']
data['pred'] = _pred_img(img, model)
np.save(save_path + file_name, data)
def make_submission(model_path,
thresh,
small_thresh=2048,
test_base_path='/data/pneumo/dicom-images-test/',
save=False):
'''
save submission file under a folder that model file is saved.
Set save=True if you want to save binary pred numpy.
all binary (small removed) npy and submission csv will be saved in /data/pneumo_log/val_1/2019_0815_1742/submission/best_weights/
'''
# with snapshot, there is a case where there are multiple model files in the dir so make `submission` folder
# best_weights
model_file_name = model_path.split('/')[-1].split('.')[0]
# 2019_0815_1742
dir_name = model_path.split('/')[-2]
# /data/pneumo_log/val_1/2019_0815_1742/submission/best_weights
save_dir = '/'.join(
model_path.split('/')[:-1]) + '/submission/' + model_file_name
data_prep._make_dir(save_dir)
# 'submission_'
file_name = 'submission_' + dir_name + '_' + model_file_name + '.csv'
test_data_path_list = glob(test_base_path + '/*/*/*.dcm')
model = pred_util._load_model(model_path=model_path)
rles = []
im_ids = []
for path in tqdm(test_data_path_list):
im_id = path.split('/')[-1].split('.dcm')[0]
im_ids.append(im_id)
im = pydicom.dcmread(path).pixel_array
# no need preprocess
pred = pred_util._pred_img(im, model)
# pred is already 1024 * 1024 but no binary. its values are 0-1
binary_pred = np.where(pred > thresh, 1, 0)
# zero out the smaller regions
if binary_pred.sum() < small_thresh:
binary_pred[:] = 0
# binary -> 0, 255 and transpose for submission format
if save:
# save numpy. This is usually needed for ensemble. In order to use ensemble_util._ensemble_preds on the data later,
# save it as dictionary like {'pred':}
np.save(save_dir + '/' + im_id, {
'pred': binary_pred,
'pred_row': pred
})
binary_pred = (binary_pred.T * 255).astype(np.uint8)
rles.append(mask2rle(binary_pred, 1024, 1024))
sub_df = pd.DataFrame({'ImageId': im_ids, 'EncodedPixels': rles})
sub_df.loc[sub_df.EncodedPixels == '', 'EncodedPixels'] = '-1'
sub_df.to_csv(save_dir + '/' + file_name, index=False)
if save:
# for later ensemble
return save_dir
def make_ensemble_submission(model_path_list,
thresh_list=None,
small_thresh=2048,
test_base_path='/data/pneumo/dicom-images-test',
column_name='score',
save=False,
save_path=None):
'''
ensemble: This function call make_submission() each and save each submission csv.
1. save binary pred based on each thresh as numpy (w/ remove small masks)
2. ensemble (max vote)
3. remove small masks
4. mask2rle
5. save submission file
'''
if save_path is None:
# make save path based on first model path
save_path = '/'.join(
model_path_list[0].split('/')[:-1]) + '/ensemble_submission/'
print('start making ensemble submission under {}'.format(save_path))
data_prep._make_dir(save_path)
# get best thresholds if thresh_list is None
if thresh_list is None:
thresh_list = []
for model_path in model_path_list:
thresh = _get_best_threshold(model_path, column_name=column_name)
thresh_list.append(thresh)
binary_mask_path_list = []
for i, model_path in enumerate(model_path_list):
print('start pred test data by {}'.format(model_path))
binary_mask_path = make_submission(model_path,
thresh_list[i],
small_thresh=small_thresh,
test_base_path=test_base_path,
save=True)
# save path should be like /data/pneumo_log/val_1/2019_0815_1742/submission/best_weights/
binary_mask_path_list.append(binary_mask_path)
# save ensembled data
ensemble_util.ensemble_dirs(binary_mask_path_list=binary_mask_path_list,
cpu_num=16,
save_path=save_path,
data_key='pred')
# assume the pred is already binary. don't need to set thresh
sub_df = _make_submission_from_predictions(save_path,
small_thresh=small_thresh)
file_name = 'ensemble_submission.csv'
sub_df.to_csv(save_path + '/' + file_name, index=False)
print('saved submission file at {}'.format(save_path + '/' + file_name))
print(
'$kaggle competitions submit siim-acr-pneumothorax-segmentation -f {} -m "snapshot ensembles"'
.format(save_path + '/' + file_name))
def ensemble_dirs(binary_mask_path_list=[], cpu_num=16, save_path='', data_key='pred'):
'''
save ensembled numpy at save_path. ensemble all of files under each folder in binary_mask_path_list
if TTA, change data_key to 'mean_pred'
'''
image_id_list = [path.split('/')[-1].split('.npy')[0] for path in glob(binary_mask_path_list[0] + '/*.npy')]
p = Pool(processes=cpu_num)
data_prep._make_dir(save_path)
job_args = [(image_id, save_path, binary_mask_path_list, data_key) for image_id in image_id_list]
list(tqdm(p.imap(_wrap_ensemble_preds, job_args), total=len(job_args)))
def _load_eval_w_thresh_list(data_path, thresh_list=[0.5], save_binary=False):
'''
set save_path if save_binary is True.
'''
data = np.load(data_path)[()]
mask = data['mask']
pred = data['pred']
image_id = data_path.split('/')[-1].split('.npy')[0]
if 'aug_pred' in data.keys():
# TTA
aug_pred = data['aug_pred']
mean_pred = data['mean_pred']
result_list = []
for thresh in thresh_list:
result = {}
dice, binary_pred = evaluation(mask, pred, thresh)
# for save. need mask here to use this functin to the saved data to evaluate
data = {'pred': binary_pred, 'mask': mask}
result['score'] = dice
if 'aug_pred' in data.keys():
aug_dice, aug_binary_pred = evaluation(mask, aug_pred, thresh)
mean_dice, mean_binary_pred = evaluation(mask, mean_pred, thresh)
result['aug_score'] = aug_dice
result['mean_score'] = mean_dice
data['aug_pred'] = aug_binary_pred
data['mean_pred'] = mean_binary_pred
result['image_id'] = image_id
result['thresh'] = thresh
if save_binary:
save_dir = '/'.join(data_path.split('/')
[:-1]) + '/binary_thresh_' + str(thresh) + '/'
data_prep._make_dir(save_dir)
file_name = save_dir + '/' + str(image_id)
np.save(file_name, data)
result_list.append(result)
return result_list
def _save_preds(
model_path='/data/pneumo_log/val_1/2019_0805_0344/best_weights.hdf5',
data_path='/data/pneumo/fold/1/',
tta=True):
'''
save prediction as numpy. if tta=True, save horizontal flip too. The saved 'aug_pred' mask has been fliped again.
'''
print('start pred with {}'.format(model_path))
data_path_list = glob(data_path + '*.npy')
#'/data/pneumo_log/val_1/'
base_save_path = '/'.join(model_path.split('/')[:-2])
#2019_0805_0344
folder_name = model_path.split('/')[-2]
# '/data/pneumo_log/val_1/val_predictions/2019_0805_0344/best_weights/'
# extended for snapshots
save_path = base_save_path + '/val_predictions/' + folder_name + '/' + model_path.split(
'/')[-1].split('.')[0] + '/'
# loading model
model = _load_model(model_path=model_path)
print('saving dirs: {}'.format(save_path))
data_prep._make_dir(save_path)
for data_path in tqdm(data_path_list):
# this includes .npy
file_name = data_path.split('/')[-1]
data = np.load(data_path)[()]
img = data['img']
data['pred'] = _pred_img(img, model)
if tta:
flip = np.flip(img, axis=-1)
flip_pred = _pred_img(flip, model)
# save after flip
data['aug_pred'] = flip_pred
data['mean_pred'] = 0.5 * data['pred'] + 0.5 * data['aug_pred']
np.save(save_path + file_name, data)
def save_make_pseudo_data(
pred_data_dir='/data/pneumo_log/val_1/2019_0815_1742/submission/snapshot_model_2/',
zero_max=0.005,
one_min=0.8,
cpu_num=16,
test_base_path='/data/pneumo/dicom-images-test/',
test_data=True):
'''
save pseudo label as dictionary {'img':, 'mask'} under pred_data_dir+'/pseudo/'
This can be applied to train data (fold) too. set test_data=False
'''
if test_data:
save_path = pred_data_dir + '/pseudo/'
else:
save_path = pred_data_dir + '/pseudo_train_fold/'
data_prep._make_dir(save_path)
print('start to make pseudo label under {}'.format(save_path))
pred_data_path_list = glob(pred_data_dir + '/*.npy')
p = Pool(processes=cpu_num)
job_args = [(pred_data_path, save_path, zero_max, one_min, test_base_path,
test_data) for pred_data_path in pred_data_path_list]
list(tqdm(p.imap(_wrap_save_pseudo_label, job_args), total=len(job_args)))