https://github.com/ternaus/boost_imagenet
Example on how to fine tune models to boost the performance of models on ImageNet
https://github.com/ternaus/boost_imagenet
Last synced: about 2 months ago
JSON representation
Example on how to fine tune models to boost the performance of models on ImageNet
- Host: GitHub
- URL: https://github.com/ternaus/boost_imagenet
- Owner: ternaus
- Created: 2022-09-26T19:09:18.000Z (almost 4 years ago)
- Default Branch: main
- Last Pushed: 2022-09-28T21:55:36.000Z (almost 4 years ago)
- Last Synced: 2025-03-20T21:43:57.143Z (over 1 year ago)
- Language: Python
- Size: 15.6 KB
- Stars: 1
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Boost Imagenet
An example code for using [Ternaus.com](https://ternaus.com) for improving the performance of a neural network on
the ImageNet dataset.
### Save the model
```bash
import timm
import torch
model = timm.create_model('resnet34', pretrained=True)
torch.jit.save(m, MODEL_SCHECKPOINT_FILE)
```
### Evaluate on the validations set
```bash
python -m src.evaluate -i PATH_TO_IMAGENET_VAL \
-r PATH_TO_REAL_JSON \
-c PATH_TO_CHECKPOINT \
-o vanila.json
```
and outputs file like:
```bash
{
"accuracy": {
"1": 81.82206375301578,
"5": 95.0765420500886
},
"losses": [
{
"filename": "ILSVRC2012_val_00000002.JPEG",
"loss": 0.08855907618999481
},
{
"filename": "ILSVRC2012_val_00000003.JPEG",
"loss": 0.00010000000000000009
}
]
}
```
### Get images that are similar to the hard ones
Get token at [Ternaus.com/account](https://ternaus.com/account)
#### Get urls to similar images:
```bash
python -m src.get_similar_images -i PATH_TO_JSON_WITH_LOSSES \
-o OUTPUT_PATH_TO_JSON_FOR_SIMILAR_IMAGES \
-im PATH_TO_VAL_IMAGES \
--num_images NUM_SIMILAR_IMAGES_PER_IMAGE
```
#### Download images:
```bash
python -m src.download_images \
-a PATH_TOIMAGENET_VAL_ANNOTATIONS \
-u OUTPUT_PATH_TO_JSON_FOR_SIMILAR_IMAGES \
-o OUTPUT_PATH \
-j NUM_JOBS
```
It creates a folder structure like:
```bash
n12985857
n12985857.txt
n12998815
n12998815.txt
```
where `n12985857` is a folder with images and `n12985857.txt` is a file with urls to images.