selimsef / dfdc_deepfake_challenge Goto Github PK

Thank you for your work. What training data do you use? All dfdc?Your model is basically unrecognizable on FOMM video. I want to add this batch of data to train your model.

I encountered this issue during validation

  File "finetune_xy.py", line 446, in <module>
    main()
  File "finetune_xy.py", line 303, in main
    summary_writer=summary_writer)
  File "finetune_xy.py", line 311, in evaluate_val
    bce, probs, targets = validate(model, data_loader=data_val)
  File "finetune_xy.py", line 366, in validate
    fake_loss = log_loss(y[fake_idx], x[fake_idx], labels=[0, 1])
  File "/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py", line 73, in inner_f
    return f(**kwargs)
  File "/opt/conda/lib/python3.7/site-packages/sklearn/metrics/_classification.py", line 2206, in log_loss
    transformed_labels = lb.transform(y_true)
  File "/opt/conda/lib/python3.7/site-packages/sklearn/preprocessing/_label.py", line 491, in transform
    sparse_output=self.sparse_output)
  File "/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py", line 73, in inner_f
    return f(**kwargs)
  File "/opt/conda/lib/python3.7/site-packages/sklearn/preprocessing/_label.py", line 680, in label_binarize
    "binarization" % y_type)
ValueError: continuous target data is not supported with label binarization
[1]+  Exit 1                  nohup python -u finetune_xy.py --config configs/b7.json > log.out

could you explain a little bit the

data_x = []
    data_y = []
    for vid, score in probs.items():
        score = np.array(score)
        lbl = targets[vid]

        score = np.mean(score)
        lbl = np.mean(lbl)
        data_x.append(score)
        data_y.append(lbl)
    y = np.array(data_y)
    x = np.array(data_x)
    fake_idx = y > 0.1
    real_idx = y < 0.1
    fake_loss = log_loss(y[fake_idx], x[fake_idx], labels=[0, 1])
    real_loss = log_loss(y[real_idx], x[real_idx], labels=[0, 1])
    print("{}fake_loss".format(prefix), fake_loss)
    print("{}real_loss".format(prefix), real_loss)

in your code? Thank you

Hi,

I just have a quick question. You mention that you resize the videos before the face detector.
Do we need to resize the videos before we run preprocess_data.sh?
Or preprocess_data.sh would also handle the resize of the videos as well when we run face detector?

I can not find the code you resize the image. This is the closest thing I find in your code.

Thank you!

./train.sh path/to/data num-of-gpus
100%██████████████████████████████████████████████████████████████████████████████████| 2765/2765 [06:11<00:00, 7.45it/s]
Traceback (most recent call last):
File "training/pipelines/train_classifier.py", line 363, in
main()
File "training/pipelines/train_classifier.py", line 227, in main
summary_writer=summary_writer)
File "training/pipelines/train_classifier.py", line 235, in evaluate_val
bce, probs, targets = validate(model, data_loader=data_val)
File "training/pipelines/train_classifier.py", line 290, in validate
real_loss = log_loss(y[real_idx], x[real_idx], labels=[0, 1])
File "/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py", line 72, in inner_f
return f(**kwargs)
File "/opt/conda/lib/python3.7/site-packages/sklearn/metrics/_classification.py", line 2186, in log_loss
y_pred = check_array(y_pred, ensure_2d=False)
File "/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py", line 72, in inner_f
return f(**kwargs)
File "/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py", line 653, in check_array
context))
ValueError: Found array with 0 sample(s) (shape=(0,)) while a minimum of 1 is required.

I have downloaded the DFDC dataset, but metadata.json wasn't found, can you share the download link of this file, thank you very much

In face_detector.py :
self.detector = MTCNN(margin=0,thresholds=[0.85, 0.95, 0.95], device=device)

but in kernel_utils.py :
self.detector = MTCNN(margin=0, thresholds=[0.7, 0.8, 0.8], device="cuda")

this is why ?

thank you

When I run generate_fold.py,I find the "fold" is alway "0"

                for k, v in metadata.items():
                    fold = None
                    for i, fold_dirs in enumerate(folds):
                        #if part in fold_dirs:
                            fold = i
                            break
                    assert fold is not None
                    video_id = k[:-4]
                    video_fold[video_id] = fold

I debug this part and find "fold” is alway "0" and then break
Since I can't understand what does "fold" mean,I don't know how to solve it and what is the correct
could you explain it in detail? thank you!

Is there a way of using the trained weights & do inference using CPU only? My GPU can't handle inference with the current settings...

I visualized this part of the code and found that it didn't work.
Can you check it out, or open source visual code ?
Thank you very much !

Hi Selim,

Thank you for sharing your great work here, tried to use your predict_submission.sh to reproduce the submission.csv by using 7 efficientnet-b7 models on test_videos folder, but the prediction scores for all 400 videos are smaller than 0.5, most of them around 0.3-0.4, guess I did something wrong here, but just cannot figure out what could be the possible reasons, can help here?

Xu

I used Win11 WSL2+Docker. Have the same problem "CUDNN_STATUS_NOT_INITIALIZED".

Is something wrong with that?

Hello! I need to ask, Is it possible to run this code on Google Colab?

Как насчёт улучшать дипфейки, а не детектить их? :D

when i run train_classifier.py,i meet the problem.

I've been stuck here for two days.
could you explain it in detail? thank you!

when I finish Data preparation， then excute "training/pipelines/train_classifier.py", console print warn log "/diffs/XXX.png” can't open/read file

https://ai.facebook.com/blog/deepfake-detection-challenge-results-an-open-initiative-to-advance-ai/

the webpage said that the top accuracy in private evaluate dataset is about 65.18%, but I evalute the model you provided, got an accuracy about 89%, what's wrong?

looking forward for any reply.

Could you talk about it in detail ?
Thank you !

Why are you multiplying the coordinates of bounding box by 2? It will shift the region that we want to crop. I think it should not be multiplied.

Thanks.

I guess this should be mid_h = h // 2.

looking forward to that

Hi, there.
I'm trying to preprocess your model, however, I had an issue running generate_folds.py

Here is the screenshot of the error.
Run it multiple times then it shows different video file names.

I looked up the metadata.json, I guess the error is from the training video have no original video.
It is means the assert videofold[video] will not equal to videofold[ori_vid].
I not should how to fix this.
Hope to hear from you soon.

Thanks,
Silion

Hi! Thank you for your work. Does this detect face filters like you would see on instagram, for example how people use for beauty purposes, like to add makeup to the face, or change the face in general? Or does it strictly detect deepfakes?

Timm library updated version causes trouble in the code.

Installing an older version resolved the issue.
pip install timm==0.1.26

When I ran python preprocessing/detect_original_faces.py --root-dir DATA_ROOT, I encountered an error:
Traceback (most recent call last):
File "preprocessing/detect_original_faces.py", line 14, in
import face_detector, VideoDataset
ModuleNotFoundError: No module named 'VideoDataset'
How can I solve this problem?

--2020-11-02 14:27:42-- https://github.com/selimsef/dfdc_deepfake_challenge/releases/download//final_999_DeepFakeClassifier_tf_efficientnet_b7_ns_0_23
Resolving github.com (github.com)... 15.164.81.167
Connecting to github.com (github.com)|15.164.81.167|:443... connected.
HTTP request sent, awaiting response... 404 Not Found
2020-11-02 14:27:42 ERROR 404: Not Found.

i encontered this issue. What can be done to solve this problem.