By using model Resnet-18 which was built by myself, I used Pytorch to classify image of 10 animal types.
Here is my pytorch implementation of the model described in the RESNET paper.
Note: Because this is the model I built, instead of this is trained by COCO dataset, I trained by dataset of 10 animal classes, which I present later.
Statistics of datasets I used for experiments. These datasets could be download from link
| Classes | Train samples | Test samples |
|---|---|---|
| butterfly | 1902 | 210 |
| cat | 1508 | 160 |
| chicken | 2790 | 308 |
| cow | 1684 | 182 |
| dog | 4373 | 490 |
| elephant | 1306 | 140 |
| horse | 2357 | 266 |
| sheep | 1638 | 182 |
| spider | 4345 | 476 |
| squirre | 1680 | 182 |
For optimizer and learning rate, I use:
- SGD optimizer with different learning rates (0.01 in most cases).
Additionally, in the my model, I will set up 100 epochs (using early stopping if after 5 epochs, if there is not greater score, it will stop train proccess) ,which is seen as a loop over batch_size: 16
If you want to train a model with default parameters, you could run:
python train_animal.py If you want to adjust your preference parameters, here is some option you can choose:
| Parameters | Abbreviation | Default | Description |
|---|---|---|---|
| --batch-size | -b | 16 | Select suitable batch size |
| --data-path | -p | '../../' | directory contains dataset |
| --lr | 1e-2 | modify learning rate | |
| --epochs | -e | 100 | modify epoch number |
| --log-path | -l | tensorboard | directory contains metrics visualization |
| --checkpoint | -sc | tensorboard/animals/epoch_last.pt | directory which saves the train model |
For example: python train.py -p dataset_location --log-patch directory-name
How to view tensorboard:
tensorboard --logdir directory/contain/tensorboard/
You could preview my evaluating process throughout Tenserboard. After 22 epochs, I reached:
-
The highest accuracy score: 0.7 in epoch 19
-
Confusion matrix:
-
Cross entropy loss in 2 processes:
After buiding models, I started to implement the testing process by file test_cnn.py
Regarding to the above confusion matrix, this model had good prediction in 4 classes: spider, chicken, horse and butterfly.
In contrast, it scored the worst prediction in class: Cat.