def convert_mask(mask):
y_bin = binarize_sample((mask, mask_thresh, min_size_thresh, dilation))
if np.sum(y_bin) > 0:
rle = mask_to_rle(y_bin.T, 1024, 1024)
else:
rle = " -1"
return rle
def convert_one(sample_id):
if "+" in model_name:
model_lst = model_name.split("+")
else:
model_lst = [model_name]
y_pred = []
for model in model_lst:
for fold in range(n_fold):
filename = osp.join(dumps_dir, f"{fold}_{model}_test",
f"{sample_id}.png")
y_pred.append(cv2.imread(filename, 0) / 255.0)
y_pred = np.mean(np.array(y_pred), axis=0)
cv2.imwrite(
f"/mnt/ssd2/dataset/pneumo/predictions/uint8/mean/{sample_id}.png",
np.uint8(255 * y_pred))
y_bin = binarize_sample((y_pred, mask_thresh, min_size_thresh, dilation))
if np.sum(y_bin) > 0:
rle = mask_to_rle(y_bin.T, 1024, 1024)
else:
rle = " -1"
return rle
def convert_one(sample_id):
path_pattern = osp.join(predict_dir, "tmp", f"{model_name}*{sample_id}*")
fns = glob(path_pattern)
assert len(fns) == 1, path_pattern
y_pred = np.load(fns[0])
agreement = float(osp.basename(fns[0]).split("|")[1])
if agreement < 0.65 and correction:
y_bin = binarize_sample((y_pred, mask_thresh, 1500, 2))
else:
y_bin = binarize_sample(
(y_pred, mask_thresh, min_size_thresh, dilation))
if np.sum(y_bin) > 0:
rle = mask_to_rle(y_bin.T, 1024, 1024)
else:
rle = " -1"
return rle
def main():
os.makedirs("outs/", exist_ok=True)
model_name = "sx50_sx101_se154"
sample_df = pd.read_csv("tables/stage_2_sample_submission.csv")
sample_df = sample_df.drop_duplicates("ImageId",
keep="last").reset_index(drop=True)
pred_dict = {}
for subm_path in sorted(glob("subm/st2*csv")):
print(subm_path)
df = pd.read_csv(subm_path)
for index, row in tqdm.tqdm(df.iterrows(), total=len(df)):
image_id = row["ImageId"]
annot = row["EncodedPixels"].strip()
if annot != "-1":
mask = rle2mask(annot, 1024, 1024)
else:
mask = np.zeros((1024, 1024))
if image_id in pred_dict:
pred_dict[image_id] += mask
else:
pred_dict[image_id] = mask
del df
print(len(pred_dict))
#################################################################
threshold_list = [
0
] # 0 - if union, 1 - if with certainty of 2 for many models
for threshold in threshold_list:
sublist = []
for index, row in tqdm.tqdm(sample_df.iterrows(),
total=len(sample_df)):
image_id = row["ImageId"]
pred = pred_dict[image_id]
if pred.sum() > 0:
out_cut = np.copy(pred)
out_cut[np.nonzero(out_cut <= threshold)] = 0.0
out_cut[np.nonzero(out_cut > threshold)] = 1.0
out_cut = ndimage.binary_fill_holes(out_cut).astype(
out_cut.dtype)
# imsave(model_name+'/'+image_id + '.png', out_cut)
rle = mask_to_rle(out_cut, 1024, 1024)
sublist.append([image_id, rle])
else:
rle = " -1"
sublist.append([image_id, rle])
submission_df = pd.DataFrame(sublist,
columns=sample_df.columns.values)
submission_df.to_csv(
f"subm/st2_union_submission_corr1_{model_name}_{threshold}.csv",
index=False)