Spaces:
Running
Running
File size: 5,624 Bytes
48af67d 8bf6199 5bfda01 8bf6199 5bfda01 8bf6199 5bfda01 8bf6199 5bfda01 a5da707 8bf6199 a5da707 8bf6199 5bfda01 8bf6199 5bfda01 8bf6199 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 |
---
title: YoloV3 on PASCAL VOC Dataset From Scratch (Slide for GradCam output)
emoji: 🚀
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 3.39.0
app_file: app.py
pinned: false
license: mit
---
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
# [GithubREPO](https://github.com/deepanshudashora/ERAV1/tree/master/session13)
# Training Procedure
#### [Experiment 1](https://github.com/deepanshudashora/ERAV1/tree/master/session13/lightning_version/Experiments)
1. The model is trained on Tesla T4 (15GB GPU memory)
2. The training is completed in two phases
3. The first phase contains 20 epochs and the second phase contains another 20 epochs
4. In the first training we see loss dropping correctly but in the second training it drops less
5. We run our two training loops separately and do not run any kind of validation on them, except for validation loss
#### [Experiment 2](https://github.com/deepanshudashora/ERAV1/tree/master/session13/lightning_version)
1. The model is trained on 2 Tesla t4 GPUs, with distributed training using PyTorch lightning
2. For doing the distributed training we use the strategy ```ddp_notebook_find_unused_parameters_true```
* Later we evaluate the model and get the numbers
* The lightning generally saves the model as .ckpt format, so we convert it to torch format by saving state dict as .pt format
* For doing this we use these two lines of code
```
best_model = torch.load(weights_path)
torch.save(best_model['state_dict'], f'best_model.pth')
litemodel = YOLOv3(num_classes=num_classes)
litemodel.load_state_dict(torch.load("best_model.pth",map_location='cpu'))
device = "cpu"
torch.save(litemodel.state_dict(), PATH)
```
* The model starts overfitting on the dataset after 30 epochs
* Future Improvements
1. Train the model in 1 shot instead of two different phases
2. Keep a better batch size (Basically earn more money and buy a good GPU)
3. Data transformation also plays a vital role here
4. OneCycle LR range needs to be appropriately modified for a better LR
# Data Transformation
Along with the transforms mentioned in the [config file](https://github.com/deepanshudashora/ERAV1/blob/master/session13/lightning_version/config.py), we also apply **mosaic transform** on 75% images
[Reference](https://www.kaggle.com/code/nvnnghia/awesome-augmentation/notebook)
# Accuracy Report
```
Class accuracy is: 85.015236%
No obj accuracy is: 98.522491%
Obj accuracy is: 65.760597%
MAP: 0.4661380648612976
```
# [Training Logs](https://github.com/deepanshudashora/ERAV1/blob/master/session13/lightning_version/training_logs/csv_training_logs/lightning_logs/version_0/metrics.csv)
#### For faster execution we run the validation step after 20 epochs for the first 20 epochs of training and after that after every 5 epochs till 40 epochs
```
lr-Adam step train_loss epoch val_loss
786 NaN 19499 4.653981 37.0 NaN
787 0.000160 19549 NaN NaN NaN
788 NaN 19549 4.864988 37.0 NaN
789 0.000160 19599 NaN NaN NaN
790 NaN 19599 5.241925 37.0 NaN
791 0.000160 19649 NaN NaN NaN
792 NaN 19649 5.020171 37.0 NaN
793 0.000161 19699 NaN NaN NaN
794 NaN 19699 4.245292 38.0 NaN
795 0.000161 19749 NaN NaN NaN
796 NaN 19749 4.541957 38.0 NaN
797 0.000161 19799 NaN NaN NaN
798 NaN 19799 3.837740 38.0 NaN
799 0.000161 19849 NaN NaN NaN
800 NaN 19849 4.239679 38.0 NaN
801 0.000161 19899 NaN NaN NaN
802 NaN 19899 4.034101 38.0 NaN
803 0.000161 19949 NaN NaN NaN
804 NaN 19949 5.010788 38.0 NaN
805 0.000161 19999 NaN NaN NaN
806 NaN 19999 3.980245 38.0 NaN
807 0.000161 20049 NaN NaN NaN
808 NaN 20049 4.641729 38.0 NaN
809 0.000161 20099 NaN NaN NaN
810 NaN 20099 4.563717 38.0 NaN
811 0.000161 20149 NaN NaN NaN
812 NaN 20149 4.422552 38.0 NaN
813 0.000161 20199 NaN NaN NaN
814 NaN 20199 4.925357 38.0 NaN
815 0.000161 20249 NaN NaN NaN
816 NaN 20249 4.788391 39.0 NaN
817 0.000161 20299 NaN NaN NaN
818 NaN 20299 4.478580 39.0 NaN
819 0.000161 20349 NaN NaN NaN
820 NaN 20349 4.624731 39.0 NaN
821 0.000161 20399 NaN NaN NaN
822 NaN 20399 4.425498 39.0 NaN
823 0.000161 20449 NaN NaN NaN
824 NaN 20449 4.361921 39.0 NaN
825 0.000161 20499 NaN NaN NaN
826 NaN 20499 4.318252 39.0 NaN
827 0.000161 20549 NaN NaN NaN
828 NaN 20549 4.013813 39.0 NaN
829 0.000161 20599 NaN NaN NaN
830 NaN 20599 4.476331 39.0 NaN
831 0.000161 20649 NaN NaN NaN
832 NaN 20649 4.192605 39.0 NaN
833 0.000161 20699 NaN NaN NaN
834 NaN 20699 4.065756 39.0 NaN
835 NaN 20719 NaN 39.0 4.348697
```
|