test_mask_np = ((test_mask+1)*0.5*255).detach().cpu().numpy().transpose(1,2,0)
#print( "test_mask_np.shape : ", test_mask_np.shape )
test_mask_pillow = Image.fromarray( np.uint8(test_mask_np[:,:,0]) )
test_mask_pillow.save( "_debug/test_mask.png" )
test_mask_splits, class_ids = split_masks( mask_np = test_mask_np, n_classes = args.n_classes )
for j, label in enumerate(class_ids):
test_mask_pillow_label = Image.fromarray( np.uint8(test_mask_splits[j]) )
test_mask_pillow_label.save( "_debug/test_mask_{}.png".format(label) )
"""
# 画像ファイル名
image_names.append(name)
# RLE の計算
encoded_pixels.append(convert_rle(np.round(y_pred_test_org)))
# クラスラベル
class_ids.append(0) # dummy
df_submission = pd.DataFrame(
{
'EncodedPixels': encoded_pixels,
'ClassId': class_ids
},
index=image_names)
df_submission.index.names = ['ImageId']
df_submission.to_csv(
os.path.join(args.results_dir, args.exper_name, args.submit_file))
if (args.submit):
# Kaggle API での submit
#================================
# RLE [Run Length Encoding] 形式で提出のため生成画像を元の画像サイズに変換
y_pred_test_org = np.zeros(
(len(y_pred_test), args.image_height_org, args.image_width_org),
dtype=np.float32)
for i in range(len(y_pred_test)):
y_pred_test_org[i] = cv2.resize(
y_pred_test[i, 0, :, :].squeeze(),
(args.image_height_org, args.image_width_org),
interpolation=cv2.INTER_NEAREST)
#y_pred_test_org[i] = resize( y_pred_test[i,0,:,:].squeeze(), (args.image_height_org, args.image_width_org), mode='constant', preserve_range=True )
# 提出用データに値を設定
y_sub = {
name.split(".png")[0]: convert_rle(np.round(y_pred_test_org[i] > 0.0))
for i, name in enumerate(test_image_names)
}
df_submission = pd.DataFrame.from_dict(y_sub, orient='index')
df_submission.index.names = ['id']
df_submission.columns = ['rle_mask']
df_submission.to_csv(
os.path.join(args.results_dir, args.exper_name, args.submit_file))
if (args.submit):
# Kaggle-API で submit
api = KaggleApi()
api.authenticate()
api.competition_submit(
os.path.join(args.results_dir, args.exper_name, args.submit_file),
args.exper_name, args.competition_id)
# IoU
pass
#================================
# 可視化処理
#================================
pass
#================================
# Kaggle API での submit
#================================
# 提出用データに値を設定
y_sub = []
for i in range(len(y_pred_test)):
rle = convert_rle(y_pred_test[i])
print("rle : ", rle)
y_sub.append(rle)
df_submission['rle_mask'] = y_sub
df_submission.to_csv(os.path.join(args.results_dir, args.exper_name,
args.submit_file),
index=False)
if (args.submit):
# Kaggle-API で submit
api = KaggleApi()
api.authenticate()
api.competition_submit(
os.path.join(args.results_dir, args.exper_name, args.submit_file),
args.exper_name, args.competition_id)
os.system('kaggle competitions submissions -c {}'.format(
# RLE [Run Length Encoding] 形式で提出のため生成画像を元の画像サイズに変換
y_pred_test_org = np.zeros(
(len(y_pred_test), args.image_height_org, args.image_width_org),
dtype=np.float32)
for i in range(len(y_pred_test)):
y_pred_test_org[i] = cv2.resize(
y_pred_test[i, 0, :, :].squeeze(),
(args.image_height_org, args.image_width_org),
interpolation=cv2.INTER_NEAREST)
#y_pred_test_org[i] = resize( y_pred_test[i,0,:,:].squeeze(), (args.image_height_org, args.image_width_org), mode='constant', preserve_range=True )
# 提出用データに値を設定
#threshold_best = 0.0
y_sub = {
name.split(".png")[0]:
convert_rle(np.round(y_pred_test_org[i] > threshold_best))
for i, name in enumerate(test_image_names)
}
df_submission = pd.DataFrame.from_dict(y_sub, orient='index')
df_submission.index.names = ['id']
df_submission.columns = ['rle_mask']
df_submission.to_csv(
os.path.join(args.results_dir, args.exper_name, args.submit_file))
if (args.submit):
# Kaggle-API で submit
api = KaggleApi()
api.authenticate()
api.competition_submit(
os.path.join(args.results_dir, args.exper_name, args.submit_file),
args.exper_name, args.competition_id)