Art Style Classifier With CNN (VGG16-BN)

Created and maintained by Joshua Ning

Getting Started

Download data from artbench
Download pre-trained model from google drive and move it into /models
Use Conda Python 3.9.7

Dependencies:

1. matplotlib
2. torch
3. numpy
4. pandas
5. torchinfo
6. torchmetrics

File Structures

Only data_256 is required to start training
model_train_val.ipynb can train and validate, is optimized for GPU with CUDA
351.ipynb is a notebook with more documentation, essentially the same as model_train_val.ipynb
train.py is a python script that can only train(for server)
The stats and models will be saved in the respective directories

How to Train and Test

1. make sure all the dependencies are installed
2. open model_train_val.ipynb
3. run code blocks 1-4
4. For training only:
   • In code block 5, if training for the first time, change save_point variable to 0
   • In code block 5, if you want to continue training from a previous model, enter the last epoch number
   • Skip code block 6 if training for the first time
   • Change the hyper-parameter variables on top of code block 7 if needed
   • Change the output file directory in code block 7 on lines 34 and 38 if needed
5. For evaluation only:
   • Code block 8 is for evaluating multiple epochs of a trained model
     • Make sure line 13 file path is consistent with the saved model path
   • Code block 9 is for evaluating single model
     • Make sure line 6 file path is consistent with the saved model path
   • Code block 10 is for generating a decision matrix, which must be run after code block 9

Available models

1. test_v1.pth A small model from tutorial to test loading and saving data
   Trained for 1 epoch, accuracy of 10%
2. test_v2.pth mini-VGG BN
   Trained for 3 epoch, batch size = 32, accuracy of 40%
3. epoch_19_bs16 VGG-16 BN
   Trained for 19 epoch, batch size = 16, accuracy of 53.9%

Stats

1. stats.csv mini-VGG BN stats
2. /stats_saves_vgg_16 VGG 16 batch size = 16 training stats
3. /stats_saves_vgg16_32 VGG 16 batch size = 32 training stats
4. test_eval_stats_32.csv VGG 16 batch size = 32 evaluation stats
5. test_eval_stats.csv VGG 16 batch size = 16 evaluation stats
6. stats_plot.m Plot data with MATLAB

Media

Plots and graphs generated throughout the project