An open API service indexing awesome lists of open source software.

https://github.com/luocfprime/dtd_experiment


https://github.com/luocfprime/dtd_experiment

Last synced: 3 months ago
JSON representation

Awesome Lists containing this project

README

          

# Experiment on DTD Dataset

## Experiment Setting

| config | use pretrained | freeze backbone | freeze until epoch | gradient clip | augmentation | use scheduler |
|:------:|:--------------:|:---------------:|:------------------:|:-------------:|:------------:|:-------------:|
| exp1 | √ | √ | 15 | √ | √ | √ |
| exp2 | √ | √ | 15 | √ | x | √ |
| exp3 | √ | √ | 15 | √ | √ | x |
| exp4 | √ | √ | 15 | x | √ | √ |
| exp5 | √ | √ | - | √ | √ | √ |
| exp6 | √ | x | - | √ | √ | √ |
| exp7 | x | - | - | √ | √ | √ |
| exp8 | x | - | - | x | √ | √ |
| exp9 | x | - | - | x | x | √ |
| exp10 | x | - | - | x | x | x |

> **freeze until epoch:** freeze backbone until this epoch, then unfreeze
> **gradient clip:** clip gradient norm to 10
> **augmentation:** random rotation, random horizontal flip, color jitter

- exp1: full
- exp2: w/o augmentation
- exp3: w/o scheduler
- exp4: w/o gradient clip
- exp5: w/o unfreeze (keep backbone frozen during entire training)
- exp6: w/o freeze backbone
- exp7: w/o pretrained
- exp8: w/o pretrained, w/o gradient clip
- exp9: w/o pretrained, w/o gradient clip, w/o augmentation
- exp10: w/o pretrained, w/o gradient clip, w/o augmentation, w/o scheduler

**Other hyperparameters settings:**

- batch size: 64
- optimizer: SGD
- learning rate: 0.01
- momentum: 0.9
- weight decay: 0.0001
- scheduler: StepLR (if used)
- lr_scheduler_step_size: 10
- lr_scheduler_gamma: 0.1

## Experiment Result

See the full wandb log [here](https://wandb.ai/lcf235/ResNet18_DTD_full_experiment).

## To Reproduce

Go to the root directory of this project, then run the following command to reproduce full experiment:

```bash
ls -l config | awk '{print $9}' | grep .yml | xargs -n 1 -I {} python train.py config/{}
```