# SqueezeNet

**Repository Path**: Iranb/squeezenet

## Basic Information

- **Project Name**: SqueezeNet
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-12-22
- **Last Updated**: 2024-05-31

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# SqueezeNet
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size: [https://arxiv.org/abs/1602.07360](https://arxiv.org/abs/1602.07360)
Iandola F N, Han S, Moskewicz M W, et al.

## Overview

Here we provide the implementation of Squeezenet in TensorFlow. The repository is organised as follows:

- `src/data/` contains the necessary dataset tfrecord genetator and tfrecord reader;
    * an implementation of build tf_record dataset (`build_tfrecord.py`);
    * an implementation of read tf_record dataset (`read_tfrecord.py`);
- `src/pre_trained/` contains a pre-trained SqueezeNet model;
- `src/models.py` contains the implementation of the SqueezeNet network ;
- `src/config.py` contains the config of the SqueezeNet network ;



Finally, `train_npu.py` puts all of the above together and may be used to execute a full training run on Tint image net.
The reported result from paper is TOP-1 Accuracy 57.5% and TOP-5 Accuracy 80.3%. The model we reproduce can achieve TOP-1 Accuracy 70.0% and TOP-5 Accuracy 90.2% about 20H train on tiny-ImageNet


## Dependencies

The script has been tested running under Python 3.7 Ascend 910 environment, with the following packages installed (along with their dependencies):

- `tensorflow - 1.15`
- `easydict`

dataset TFRECORD  obs path: `obs://ma-iranb/data/squeezenet/`

## Usage

```
python3 train_npu.py
```


## License
MIT