The mistakes in the dataset

Hi @anna_mrukwa

Thank you for your keen observation and for letting us know!

We will try our best to reduce these noise in future Blitz puzzles.

However, since the number of bad samples here is low and because the ongoing AI Blitz 13 is about to end, we think it would not be wise to update the dataset at this point.
Thank you again for reporting.

Regards,
Shubhamai

View post 1

Extract images from the main image

Hi @jinoooooooooo

Yes! We have a function added in the Baseline notebook of Face Recognition Puzzle named get_target_face which simply takes the target face_no ( which will be between “00” and “99” [ in string ] ) and target_image .

def get_target_face(face_no, target_image):


  # Top-Left x, y corrdinates of the specific face 
  x, y = (int(face_no[0]))*216, (int(face_no[1]))*216

  target_face = target_image[x:x+216, y:y+216]

  return target_face

# Showing a sample face from a sample target image 

sample_target_face = get_target_face("96", sample_target)

You can find more about it in the baseline. I hope this helps!

Shubhamai

View post 1

How to fine tune models

Hi @diego26dav

For these image-to-image translation tasks, I would really suggest you to check out this TensorFlow notebook on Image-to-image translation using GANs, they have some really good implementation on this for getting started. You can also check out this blog on the topic.

Shubhamai

View post 0

🚨 AI Blitz 13 ⚡ Faces [New Challenge Launch]

Hi @notnanton

Glad to hear that you liked the challenge And yes, you are allowed to use pre-trained networks, including the one you mentioned. You are also allowed to fine-tune its weights. The things that the rule doesn’t allow is using models that are not free to use and publically available. I hope this helps. Let me know if you have any more questions.

Shubhamai

View post 0

Getting error while submitting the submission.csv

Hi, is this possible if you can share the notebook you used to make submissions in dms ? It will help us a lot to figure out the issue. In the meantime, you can also try looking at our Getting Started notebook for sentiment classification and see if there’s any major difference in terms of what commands you are using to make a submission. I hope it helps.

Shubhamai

View post 0

Getting error while submitting the submission.csv

Hi. Can you please share with us the steps to reproduce the error so that we can figure out the exact cause of this issue ?

Regards
Shubhamai

View post 0

Getting error while submitting the submission.csv

Hi @mohd_saqib

Thanks for letting us know. We are looking into the issue, will get back to you asap.

Shubhamai

View post 0

Request for the original images

Hi @jinoooooooooo

Welcome back to blitz . I can understand but sadly we cannot provide the images because the challenge is based on using the embeddings to classify the sentiment. Adding images will change a whole lot to the original challenge.

Other than that I looked at your submissions and found out you were having some issues. So the exact submission format is that you first put your submission.csv file and original_notebook.ipynb ( the notebook you generated the predictions from ) in a zip file. This is what the final zip file should look like -

── submission.zip
   ├── submission.csv
   └── original_notebook.ipynb

And then you can submit your submission.zip in on the challenge page. I hope this help Let me know if you have any more questions.

Shubhamai

View post 1

How embeddings were gotten?

The exact layer depends on the model you are using and what layer you want to extract the features from. Commonly, for transfer learning, it is the layer just before the fully connected layer. So for ex. in vgg16, it’s the last max-pooling layer. Same with resnet18, the features are extracted from the layer before the fully connected layer.

vgg16 architecture1774×1130 74.7 KB

Here’s another blog you can read if you want to learn more about vgg16 feature extraction. I hope this helps

Shubhamai

View post 0

How embeddings were gotten?

Hi @vad13irt

So, to generate the embeddings, we first took all of the images which had their corresponding labels and put them into ( for ex. resnet18 ) model, and then we extracted the features of the images from a certain layer of the model. This blog is also quite good if you want to understand the whole process.

Let me know if you have any questions. Enjoy Blitz

Shubhamai

View post 1

What's the definition of external data?

Hi youssef_nader3

Apologies for the late reply here. You can freely use publicly available pre-trained models & embeddings for submission in Blitz XII. But use of external raw data is not allowed. The challenge rule generally prohibits the use of other public/private datasets which you can use to train your model or improve the scores. I hope this helps. Let me know if you still have any doubts

Best
Shubhamai

View post 1

Two questions to the " [Getting Started Notebook] Object Detection"

Hi, it’s fixed now!

For the -

1. Why score = 0 & Why " Generating Predictions" has the variable ‘n’ in two for-cycles ?

I looked into the issue and turns out there were few mistakes we made in our baselines such as -

• Not loading the model weights after training it.
  cfg.MODEL.WEIGHTS = os.path.join(cfg.OUTPUT_DIR, "model_final.pth")

• Bounding boxes were in the wrong format
  The right format for bounding boxes is (x, y, w, h) instead of (x1, y1, x2, y2) ( which detectron2 outputted )

As for the -

2. and “image_id”:n+1?

The image_id helps the evaluator identify predictions for each image. So there are 1000 unique images ids because there are 1000 testing images.

To know what is the image_id of each image, it is simply the file number ( for ex. 345 in 345.jpg ) + 1 because the image_id starts from 1 in our ground truth file.

I hope this helps, the getting started notebook is also updated! let me know if you still had any doubts

Shubhamai

View post 1

Two questions to the " [Getting Started Notebook] Object Detection"

Thanks for letting us know I am looking into the notebook and will give you updates asap

View post 1

505 Error everytime i open leaderboards

Hi, it should be working now

View post 1

505 Error everytime i open leaderboards

Thanks for letting us know! We are looking into the issue

View post 1

Submission Guidelines

Yes, you can make submissions from your own/local machine, but the submission has to be a .ipynb notebook. In case you want to import a custom python file, you can add that python file in the assets directory and then import it in the .ipynb notebook. I hope this helps :), let me know if you have any doubts!

Cheers
Shubhamai

View post 1

Submission issue on google cloud platform

Hi, can you please tell which specific puzzle you are having issues with ? We are looking into the issue however an another way you can make the submission is by wrapping up all necessary files in a zip file such such that the zip file contents tree looks something like this -
Using Docking ISS puzzle as an example

submission.zip
├── assets
│   └── submission.csv
└── original_notebook.ipynb

• The original_notebook.ipynb is the notebook you used to train & make the predictions.
• The submisison.csv contains the predictions of the puzzle ( the format can be different for different puzzles ).

After that you can simply submit the zip file by running
aicrowd submission create -c docking-iss -f YOUR_ZIP_FILE_PATH

View post 0

0 score but submission evaluated

I tried looking into the issue and also re-runned the submission but I couldn’t able to find anything wrong from our side, can you please check and see if the bounding boxes were in the right format (x, y, w, h) or the ImageID corresponding to bounding box is correct.

Shubhamai

View post 1

ISS Docking leaderboard

The same question was also replied in discord too but i will also send it here in case another participant also had the same question

Yep, so the Distance Loss is mainly taken into account, but in case, there was a same score in Distance Loss, the second Docking Port Loss is taken into account.

View post 1

Submission failed

Sorry for a little late reply but I looked into the notebook & the output .mp4 predictions, and found out that the videos the gen_predictions function is outputting contains only single frame.

And it’s because the predictions variable, from the model.nn_module.predict(image_batch) in gen_predictions function, has shape torch.Size([1, 1, 128, 128]) which represents a single image, the image_batch variable has also the same shape.

So i find out the main reason for this error is because you changed the batch size of test_loader from 23 ( in getting started code ) to 1. The main reason of using batch size of 23 in getting started code was because that each batch will represent each video. So, you will change to change your batch size to 23 and hopefully it will all work.

Let me surely know if you have any more doubts.

Cheers
Shubhamai

View post 1