An exploration of the trade-off between image quality and accuracy.
The dataset is downloaded from Kaggle (https://www.kaggle.com/c/dogs-vs-cats).
After getting the data, move the train.zip file to data/ and decompress the file. The structure of data/ should be like as follows:
├── __init__.py
├── prepare.py
├── quantize.py
├── test.zip (not used)
├── train.zip
└── train
├── cat.m.jpg [12500 images]
└── dog.n.jpg [12500 images]
Then the images of different qualities and correspondingly prepared h5 files (e.g. quality_0\data.h5) could be generated using quantize.py and prepare.py, respecitively.
Note: For this dataset, h5 file is not a very good choice because generated data.h5 is too large (25GB) and it also takes a lot of time to read the data into the memory (> 4GB). I will search for more efficient ways of utilizing data.
From this time, the model is a simple multi-layer CNNs in order to shrink the parameters easily. The architecure of model is given in model.py. Or you can obatin the architecure and parameters in parameters/models/.
To train the model on images with quality scale of 5, just type the following command in your bash.
python train.py --quality 5
python train_paras.py --setting 2 (for parameter minimizing)
Besides image quality, other hyper-parameters such as learning rate and training epochs rate can also be specified. You can get more information by adding -h while runing the script:
python train.py -h
python train_paras.py -h (for parameter minimizing)
Having finished trainiing, you could collect the result using logs/collect.py script like:
cd logs/
python collect.py --quality 15 --setting 1
It is worthy to notice that the relationship between settings and models are list in model_paras.py.
After having trained the models, you obatin the results using logs/collect.py script. Then the collected results should be set in plot.py for plotting.
python plot.py
The detailed information about different models for parameter minimizing is given in parameters/models/.
To reproduce the results (architecure figure & parameters txt), run the parameters/model.py scirpt.
cd parameters
python model.py --all True
