# alexnet-tiny-imagenet

**Repository Path**: walkeriii/alexnet-tiny-imagenet

## Basic Information

- **Project Name**: alexnet-tiny-imagenet
- **Description**: Trains AlexNet on the Tiny Imagenet dataset from Stanford.
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-06-05
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

This took me a while to do mostly because of how long it took to unzip the dataset (tiny-imagenet-200) and how large the network is (for my measly Intel iGPU). Even just training the last layer took my laptop half an hour to get through one-third of an epoch.

After ending training there, I saved the model and tested it against the validation set. I’m seeing about 40% accuracy which is extremely good especially for only getting through a third of an epoch and shows how powerful transfer learning can be.

I have already provided a trained model in the folder `saved_models`. To train/save another model run `train.py` using the following structure:

`python3 train.py --data /tiny/imagenet/dir/ --save /dir/to/save/model/`

To test the created model from the above, run `test.py` like so:

`python3 test.py --model /dir/containing/model/--data /tiny/imagenet/dir/`

If you don’t specify the data directory, it defaults to `./data/tiny-imagenet-200`and if you don’t specify the model directory, it defaults to `./saved_models`.

We can see the results of the model in real life when I pass my webcam output to it:
![net classifying soda bottle](./webcam_pics/soda.PNG "soda bottle")
![net classifying phone](./webcam_pics/phone.PNG "phone")


It is much faster and more reliable than LeNet-5 which I trained previously on CIFAR-100 instead. It would be interesting to test Alexnet(current net) vs LeNet-5 when they’re both trained on tiny-imagenet-200.