Takekazuchi/Caracam_2

Caracam (gen 2)

This model is a fine-tuned version of google/vit-base-patch16-224 on the imagefolder dataset. It achieves the following results on the evaluation set:

• Loss: 0.7005
• Accuracy: 0.8139

Model description

This is the model the Caracam mobile app will be is using upon release
Second Generation of Caracam! with a tested accuracy of 92% and a reported accuracy of 81% this makes this model of Caracam at least 1.5x more accurate than gen 1!
If you wish to support this project please head over to my Patreon or my Cashapp
even if you don't want to support us financially thank you for even just coming to this page in the first place as the traffic helps more people find this project!

Intended uses & limitations

NOT FOR COMMERCIAL USE OUTSIDE OF OFFICIAL CARACAM MOBILE APP
Limitations
This model assumes that its input image contains a well-cropped car.
If a non-car image is given or if the car is not well-cropped, the output of the model may be meaningless.

Expected Updates

EXPECTED RELEASE DATE OF CARACAM: 12/01 (Date pushed back due to complications with tflite support for ViT model architecture)
Future versions of Caracam mobile app to have price-prediction with links via web-integration as well as more general information on the predicted car model.
User-Polls and comments with like/dislike system to be addded for feedback on new cars and anything else users want added to the app/model.

Non-Suitable Usecases

Do NOT use this model to determine whether an object is a car or not.

Suitable Usecases

DO use this model to determine the make/model/year of a car
DO take pictures of your animals to see what car they look like and send the results to my Twitter so i can retweet them!

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.9008	1.0	21451	2.0266	0.5235
1.747	2.0	42902	1.2724	0.6764
1.4453	3.0	64353	1.0682	0.7189
1.1809	4.0	85804	0.9663	0.7445
1.2318	5.0	107255	0.8934	0.7596
0.8664	6.0	128706	0.8309	0.7782
1.0645	7.0	150157	0.7863	0.7890
1.1092	8.0	171608	0.7503	0.7974
0.8655	9.0	193059	0.7204	0.8076
0.6357	10.0	214510	0.7005	0.8139

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.1+cu121
• Datasets 2.17.1
• Tokenizers 0.15.2