appian42 / kaggle-rsna-intracranial-hemorrhage Goto Github PK

total 16
-rw-r--r-- 1 root root 2934 Oct 19 02:37 README.md
drwxr-xr-x 2 root root 4096 Oct 19 02:37 bin
drwxr-xr-x 2 root root 4096 Oct 19 02:37 conf
drwxr-xr-x 6 root root 4096 Oct 19 02:37 src
['/kaggle/working/RSNA666/src/preprocess/dicom_to_dataframe.py', '--input', '/kaggle/input/rsna-intracranial-hemorrhage-detection/stage_1_train.csv', '--output', '/kaggle/input/train_raw.pkl', '--imgdir', '/kaggle/input/rsna-intracranial-hemorrhage-detection/stage_1_train_images']
read /kaggle/input/rsna-intracranial-hemorrhage-detection/stage_1_train.csv (4045572 records)
100%|█████████████████████████████| 4045572/4045572 [00:14<00:00, 278177.31it/s]
removed ID_6431af929
making records...
args.n_pool= 4
------1--------
0%| | 0/674257 [00:00<?, ?it/s]

I test your code on Kagggle.
wait for your help,Thanks

Thanks a lot for your sharing. When I read the code, I find that in src/preprocess/create_dataset.py, the add_adjacent_labels perhaps works weird.

            if j == 0:
                left = labels[j-1]
            else:
                left = ''

Doesn't it should be

            if j == 0:
                left = ''
            else:
                left = labels[j-1]

def add_adjacent_labels(df):
    df = df.sort_values('PositionOrd')

    records = []
    print('making adjacent labels...')
    for index,group in tqdm(df.groupby('StudyInstanceUID')):

        labels = list(group.labels)
        for j,id in enumerate(group.ID):
            if j == 0:
                left = labels[j-1]
            else:
                left = ''
            if j+1 == len(labels):
                right = ''
            else:
                right = labels[j+1]

            records.append({
                'LeftLabel': left,
                'RightLabel': right,
                'ID': id,
            })
    return pd.merge(df, pd.DataFrame(records), on='ID')

Looking forward for your response!
Thanks again for your sharing!!

Hi, Appian. Thanks for sharing the code. I have a question about model.avg_pool. You use nn.AdaptiveAvgPool2d(1) to make the feature map become 1x1, is that too small for the training process? I think that it may lose some information or details which are important to the classification task. Have you ever tried other size, for example 5x5 or 7x7? Thanks for answering.

Hi, Is this a baseline?