Edit model card

finetuned-ViT-human-action-recognition-v1

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the Human_Action_Recognition dataset. It achieves the following results on the evaluation set:

  • Loss: 3.1427
  • Accuracy: 0.0791

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0002
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • num_epochs: 4
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Accuracy
1.4986 0.13 100 3.1427 0.0791
1.1929 0.25 200 3.4083 0.0726
1.2673 0.38 300 3.4615 0.0769
0.9805 0.51 400 3.9192 0.0824
1.158 0.63 500 4.2648 0.0698
1.2544 0.76 600 4.5536 0.0574
1.0073 0.89 700 4.0310 0.0819
0.9315 1.02 800 4.5154 0.0702
0.9063 1.14 900 4.7162 0.0633
0.6756 1.27 1000 4.6482 0.0626
1.0239 1.4 1100 4.6437 0.0635
0.7634 1.52 1200 4.5625 0.0752
0.8365 1.65 1300 4.9912 0.0561
0.8979 1.78 1400 5.1739 0.0356
0.9448 1.9 1500 4.8946 0.0541
0.697 2.03 1600 4.9516 0.0741
0.7861 2.16 1700 5.0090 0.0776
0.6404 2.28 1800 5.3905 0.0643
0.7939 2.41 1900 4.9159 0.1015
0.6331 2.54 2000 5.3083 0.0589
0.6082 2.66 2100 4.8538 0.0857
0.6229 2.79 2200 5.3086 0.0689
0.6964 2.92 2300 5.3745 0.0713
0.5246 3.05 2400 5.0369 0.0796
0.6097 3.17 2500 5.2935 0.0743
0.5778 3.3 2600 5.5431 0.0709
0.4196 3.43 2700 5.5508 0.0759
0.5495 3.55 2800 5.5728 0.0813
0.5932 3.68 2900 5.7992 0.0663
0.4382 3.81 3000 5.8010 0.0643
0.4827 3.93 3100 5.7529 0.0680

Framework versions

  • Transformers 4.21.2
  • Pytorch 1.12.1+cu113
  • Datasets 2.4.0
  • Tokenizers 0.12.1
Downloads last month
84
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.