TensorFlow-Slimmer image classification model library

A simple and extensible project based on TensorFlow-Slim image classification model library for

• reimplementation of CNN architecture engineering papers, e.g., MobileNetV2, ShuffleNetV2.
• example code for the NeurIPS 2019 paper: Convolution with even-sized kernels and symmetric padding
• exercises for Tsinghua lecture: Neuromorphic Computing Theory and System
  

Prerequisites

• A computer with NVIDIA GPUs
• CUDA + CuDNN
• TensorFlow (GPU version) tested: 1.14, 1.13, 1.12
• python3
• tqdm

Preparing datasets

Download and generate mnist, fashion and cifar datasets by:

cd data/xxx/
python3 download_and_convert_xxx.py

For imagenet dataset, please follow the instructions in an automated script, and change the destinations in source/datasets/dataset_factory to the converted tf.record file locations

Configuring training hyperparameters

Change or new an option file in source/options/, there are some templates for beginners.

Training devices

Change the configuration in option file:

• Single GPU trainig with GPU-2: gpu_list = [2]
• Multi-GPU training with GPU-1 and GPU-3: gpu_list = [1,3]
• CPU-only training (deprecated): gpu_list = []

Network models

Change or new a model file in source/nets/, and register it in nets_factory.py

You can also modify and add other components, such as preprocessing, dataset. Don't forget to register them in individual factory files.

Start training:

In case the option file name in source/options/ is your_options.py

cd source/
python3 top.py -o your_options

The training log files will be printed and saved in log/time_your_title.txt

You can check the training details in log files, and derive statistics for drawing curves.

And the model will be saved in model/time_your_title if you have configured the saving parameters.

Exercise for Tsinghua lecture: Neuromorphic Computing Theory and System

完成下列实验，提交实验报告，内容包括：

• 实验整个流程关键信息的记录，分析，总结
• 运行、结果等截图，收敛曲线等等
• 个人感想

Experiment 1:

使用MLP训练mnist，/options/mlp_mnist.py 记录：

• 收敛曲线（train正确率，test正确率，train loss）
• 尝试添加几层全连接，对比效果
• 尝试更改loss function 和 optimizer， 对比效果

Experiment 2:

使用LeNet训练fashion，/options/lenet_fashion.py 记录:

• 收敛曲线（train正确率，test正确率，train loss）
• 尝试添加几层卷积，以及修改preprocess，对比效果
• 尝试使用MLP训练fashion，对比和CNN的差别

Experiment 3:

• 参考ResNet论文
• 补全/options/resnet_cifar.py中的设置
• 复现其ResNet-56的实验结果