hca97 / mosquito-classifiction Goto Github PK

In the challenge, there are two main species (GENUS, SPECIES).

The mosquito classifier consistently identifies mosquitos within the same species, which is reasonable given that mosquitos belonging to the same species often share similar characteristics.

Using a step-by-step classifier might help. First, we figure out if the mosquito is a GENUS or SPECIES. Then, we use another classifier to find out the exact kind. This way, we might get better results. @fkemeth what do you think?

Here are possible implementations:

Idea 1

Idea 2

Idea 3

Idea 4

all the models performed very badly in round 2 investigate why.

Assumptions:

• #8
• #9
• #10

So far our model uses BN for normalization but BN might not be the best one.

Things to be tested:

• LayerNorm
• No normalization

Hi @fkemeth, I want to make a LinkedIn post about our results for some PR! 😊 Since this was a group effort, I'd like your permission to share the post. If you're not okay with it, I won't post it.

See here:
https://arxiv.org/abs/2109.01903

Since mosquitoes look like maybe we can use ArcFace Loss rather than Softmax. The ArcFace Loss is designed to make each class more separable than Softmax. I am not sure how well it works on our use case though, what do you think @fkemeth?

Ref: https://arxiv.org/pdf/1801.07698.pdf

Helps us understand when losses diverge.

https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.callbacks.TQDMProgressBar.html#lightning.pytorch.callbacks.TQDMProgressBar

Hi @fkemeth,

We need to choose our top submissions: https://www.aicrowd.com/challenges/mosquitoalert-challenge-2023/dynamic_contents.

We need to pick three submissions. I suggest one submission should be our best solution, one you want to test, and one that I want to test. Does that sound okay?

We have to submit them before 11 PM tomorrow. Maybe we can do the submission tomorrow evening?

model is trained on GPU with mixed precision do you think this can affect on CPU performance? I know in GPU and CPU the matrix multiplications are done differently but there shouldn't be too much of a difference.

Evaluate how good the model is when trained on both datasets.

• https://neuralmagic.com/blog/yolov8-detection-10x-faster-with-deepsparse-500-fps-on-a-cpu/
• https://docs.ultralytics.com/modes/export/
• https://docs.neuralmagic.com/get-started/transfer-a-sparsified-model/cv-object-detection
• https://sparsezoo.neuralmagic.com/models/yolov8-m-voc_coco-pruned75?hardware=deepsparse-c6i.12xlarge&tab=3

See here for the pl implementation

https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.callbacks.EarlyStopping.html#lightning.pytorch.callbacks.EarlyStopping

https://discourse.aicrowd.com/t/external-datasets-used-by-participants/9217

As the competition ends, it would be good to recreate our best solution. This way, we can be sure we know how to get the same results again.

• Reproduce Best YOLO model
• Reproduce Best CLIP model
• #33
• #32

• use GANs to create artificial data (FastGAN, StyleGAN-ADA)
• https://discourse.aicrowd.com/t/external-dataset-notice-on-usage-declaration/8999
• aegypti - https://www.kaggle.com/datasets/pradeepisawasan/aedes-mosquitos
• albopictus, aegypti - https://data.mendeley.com/datasets/zw4p9kj6nt/2
• #28
• #26

One advantage of CovNext over ViT models is that we can vary the image size (use larger images and keep the original aspect ratio) which might help us improve our score. But first, we need to see if the CovNext can perform as well as the ViT models.

The training loss curves of YOLO don't look good, I think the reason might we are doing fine tuning and the initial lr might be too much or too little.

We need more extensive testing maybe we can do cross-validation? Since the clip classifier is fast to train (takes 30-40 min to train)