![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1630583%2F33870c3bcd56543625c2009f2e27af43%2Faugument100000.gif?generation=1665905790619453&alt=media\")
Thank you for organizing very exciting competition.
\nAnd also the following codes helped me a lot, thank you very much! @its7171
\nhttps://www.kaggle.com/code/its7171/dfl-benchmark-training
\nhttps://www.kaggle.com/code/its7171/dfl-benchmark-inference
def hoge(src_imgs):\n # len(src_imgs)==11+1 frames\n\n # sampling parametors\n angle = random.uniform(-0.1, 0.1)\n translate_x = random.uniform(-1, 1)\n translate_y = random.uniform(-1, 1)\n scale = random.uniform(-0.0005, 0.0005)\n\n dst_imgs = torch.zeros_like(src_imgs)\n for i,t in enumerate(-6,6):\n dst_imgs[i] = torchvision.transforms.functional.affine(src_imgs[i],\n angle=float(t*angle),\n translate=(t*translate_y, t*translate_x),\n scale=1.0 + t*scale,\n interpolation = torchvision.transforms.InterpolationMode.BILINEAR)\n return dst_imgs\nThe left column in the figure below shows 11 frames of video before augment,
\n and the center and right columns are examples of the augment applied.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1630583%2Fb292f8832e026f08f0eb4d655baef581%2Faugument175000.gif?generation=1665906158387381&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1630583%2F3b9a6d71b4729efab357078adb0471ff%2Faugument075000.gif?generation=1665906759837057&alt=media\")
For the labels, I used a Gaussian distribution (sigma=20 frames) as shown below.
\nThese soft label designs are same for the three events.
\nBinary cross-entropy was used for the loss function.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1630583%2F38f26f2c7f32dff2713f658fdfdd189a%2Flabel_design.png?generation=1665906218833070&alt=media\")
Ensemble of 5 efficientnet_b1 models with different CVs.
\nIt was the last four days and I was frazzled.
\nAs a last ditch effort, I tried to annotate all the unannotated time.
\nAfter annotating 5 videos (1606b0e6_0, 407c5a9e_1, 4ffd5986_0, 9a97dae4_1, ecf251d4_0) I ran out of steam and started training the model.
I spent quite a bit of time trying to parallelize cv2.VideoCapture with Dataloader, but it did not work.
\nI also tried decoders such as decord, dali, pyav, howevwer these did not work either.
\nFinally, I used cv2.VideoCapture for video encoding,
\nand grayscaling and resizing of the image were applied on CPU (np.array), and then difference, absolute , creating input-batch (rolling & copy) were done on the GPU (nn.sensor).
This speed up allowed us to infer on 6 models (efficientnet_b1 with 11x960x512 tensor) within the time limit.
\nDue to make margins, we submitted an ensemble of five models, finally.
| submit | \nLB | \n
|---|---|
| ①： my baseline ( 512x288 ) | \n0.625 | \n
| ②： ① + larger iamge ( 1024x576 ) | \n0.717 | \n
| ③： ② + affine augument | \n0.748 | \n
| ④： ③ + 5-model ensemble | \n0.768 | \n
| ⑤： ④ + extra annotation | \n0.792 | \n
I modified the model in Scores ② above and added experimental results on inter-frame differences.
\nI compared two models.
\nThe first model takes as input an absolute value image of the difference between current and previous.
\nThe second model takes as input a 2-channel image that concatenates the current and previous images.
\nThey are denoted as with_absdiff and without_absdiff, respectively.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1630583%2F4e8f391116cea04ec030ba6f6a4aade9%2Fextra_exp.png?generation=1666051814016820&alt=media\")
The metrics for each epoch are shown in the figure below.
\nFor most metrics, with_absdiff outperforms without_absdiff.
\nGenerally, CNNs are said to automatically extract features from images and videos, but that is the case when there is a large amount of data.
\nWhen data is limited, as in this case, it loses out to simple feature extraction, such as absolute value of differences.
What is interesting is the relationship between train_loss and valid_loss.
\nSince epoch 3, without_absdiff has been smaller than with_absdiff on train_loss.
\nOn the other hand, with regard to valid_loss, with_absdiff is always smaller than without_absdiff.
\nThis is because without_absdiff extracts poor features for the two channels of input (current and previous frames).
Is it an experiment where only 2 adjacent frames are input instead of the above input of 11 channels?
Yes.
\nTo simplify problem, additional experiments are limited to only two frames.
How many steps were used for each epoch (step axis of the graph)?
\nOr is it really an 8-step fine-tuning?
The x-axis indicates the number of epochs.
\nEach epoch has 13540 steps and mini-batch size is 16.
Hi @chack3 , how to find ur original notebook 4th place training and inference, it is public then share link with me, i am building cutoom solution for my action classification use case?
", "votes": 0, "replies": [] }, { "id": 1993251, "author_name": "Rahul Pandey", "author_url": "", "post_date": "2022-10-18T08:11:53.960000", "content": "What a wonderful achievement👍. I have learn something new from your post.
", "votes": 0, "replies": [] } ], "comment_count": 8, "messages": [ { "id": 1989867, "postDate": "2022-10-16T07:55:32.657Z", "content": "Thank you for organizing very exciting competition.
\nAnd also the following codes helped me a lot, thank you very much! @its7171
\nhttps://www.kaggle.com/code/its7171/dfl-benchmark-training
\nhttps://www.kaggle.com/code/its7171/dfl-benchmark-inference
def hoge(src_imgs):\n # len(src_imgs)==11+1 frames\n\n # sampling parametors\n angle = random.uniform(-0.1, 0.1)\n translate_x = random.uniform(-1, 1)\n translate_y = random.uniform(-1, 1)\n scale = random.uniform(-0.0005, 0.0005)\n\n dst_imgs = torch.zeros_like(src_imgs)\n for i,t in enumerate(-6,6):\n dst_imgs[i] = torchvision.transforms.functional.affine(src_imgs[i],\n angle=float(t*angle),\n translate=(t*translate_y, t*translate_x),\n scale=1.0 + t*scale,\n interpolation = torchvision.transforms.InterpolationMode.BILINEAR)\n return dst_imgs\nThe left column in the figure below shows 11 frames of video before augment,
\n and the center and right columns are examples of the augment applied.
For the labels, I used a Gaussian distribution (sigma=20 frames) as shown below.
\nThese soft label designs are same for the three events.
\nBinary cross-entropy was used for the loss function.
Ensemble of 5 efficientnet_b1 models with different CVs.
\nIt was the last four days and I was frazzled.
\nAs a last ditch effort, I tried to annotate all the unannotated time.
\nAfter annotating 5 videos (1606b0e6_0, 407c5a9e_1, 4ffd5986_0, 9a97dae4_1, ecf251d4_0) I ran out of steam and started training the model.
I spent quite a bit of time trying to parallelize cv2.VideoCapture with Dataloader, but it did not work.
\nI also tried decoders such as decord, dali, pyav, howevwer these did not work either.
\nFinally, I used cv2.VideoCapture for video encoding,
\nand grayscaling and resizing of the image were applied on CPU (np.array), and then difference, absolute , creating input-batch (rolling & copy) were done on the GPU (nn.sensor).
This speed up allowed us to infer on 6 models (efficientnet_b1 with 11x960x512 tensor) within the time limit.
\nDue to make margins, we submitted an ensemble of five models, finally.
| submit | \nLB | \n
|---|---|
| ①： my baseline ( 512x288 ) | \n0.625 | \n
| ②： ① + larger iamge ( 1024x576 ) | \n0.717 | \n
| ③： ② + affine augument | \n0.748 | \n
| ④： ③ + 5-model ensemble | \n0.768 | \n
| ⑤： ④ + extra annotation | \n0.792 | \n
I modified the model in Scores ② above and added experimental results on inter-frame differences.
\nI compared two models.
\nThe first model takes as input an absolute value image of the difference between current and previous.
\nThe second model takes as input a 2-channel image that concatenates the current and previous images.
\nThey are denoted as with_absdiff and without_absdiff, respectively.
The metrics for each epoch are shown in the figure below.
\nFor most metrics, with_absdiff outperforms without_absdiff.
\nGenerally, CNNs are said to automatically extract features from images and videos, but that is the case when there is a large amount of data.
\nWhen data is limited, as in this case, it loses out to simple feature extraction, such as absolute value of differences.
What is interesting is the relationship between train_loss and valid_loss.
\nSince epoch 3, without_absdiff has been smaller than with_absdiff on train_loss.
\nOn the other hand, with regard to valid_loss, with_absdiff is always smaller than without_absdiff.
\nThis is because without_absdiff extracts poor features for the two channels of input (current and previous frames).
Is it an experiment where only 2 adjacent frames are input instead of the above input of 11 channels?
Yes.
\nTo simplify problem, additional experiments are limited to only two frames.
How many steps were used for each epoch (step axis of the graph)?
\nOr is it really an 8-step fine-tuning?
The x-axis indicates the number of epochs.
\nEach epoch has 13540 steps and mini-batch size is 16.
Hi @chack3 , how to find ur original notebook 4th place training and inference, it is public then share link with me, i am building cutoom solution for my action classification use case?
", "rawMarkdown": "Hi @chack3 , how to find ur original notebook 4th place training and inference, it is public then share link with me, i am building cutoom solution for my action classification use case?" }, { "id": 1993251, "postDate": "2022-10-18T08:11:53.960Z", "content": "What a wonderful achievement👍. I have learn something new from your post.
", "rawMarkdown": "What a wonderful achievement👍. I have learn something new from your post." } ], "complete": true, "source": "live", "fetched_at": "2026-10-22T08:53:48Z", "competition_id": 37244, "competition_slug": "dfl-bundesliga-data-shootout", "team_id": 8957311, "solution_url": "/competitions/dfl-bundesliga-data-shootout/writeups/ohkawa3-public-5th-place-solution" }