/PCANet_python_ver

CSE512 Project.

Primary LanguageJupyter NotebookMIT LicenseMIT

PCANet_python_ver

SBU CSE512 Project

A simple implementation of PCANet: A Simple Deep Learning Baseline for Image Classification? https://arxiv.org/abs/1404.3606

File list:

  • PCANet.py:

    Usage:

    • Example: See interface.py and PCANet-MNIST.ipynb

    • General use process:

    1. Load data into numpy ndarray(check data_loader.py as example)
    2. Initialize PCANet class, e.g.:

    net = PCANet(k1=7, k2=7, L1=8, L2=8, block_size=7, *other_parameters)

    Parameters:

    k1
# the mean remove patch size width, required
k2
# the mean remove patch size height, required
L1
# the number of filters in the First stage, required
L2
# the number of filters in the Second stage, required
block_size
# the block size of histogram, required
overlapping_radio=0 
# overlapping radio, 1-0, optional, default 0
linear_classifier='svm' 
# linear_classifier, a linear_classifer or 'svm', 'svm' means sklearn.svm.SVC(), optional, default 'svm'
spp_parm=None
# parameters for spp, needs to be a list, eg[4,2,1], optional, default none
dim_reduction=None
# dim reduction after spp, the number of dimension reduce to, optional, default none
    1. Train PCANet with train data, e.g.:

    net.fit(train_data)

    1. Predict test data's label with trained PCANet, e.g.:

    prediction = net.predict(test_data)

  • interface.py

    A toy example of using PCANet

  • data_loader.py

    Simple data loader for MNIST and CIFAR10

    Sample usage:

    train_images, train_labels, test_images, test_labels = load_mnist('path/to/mnist’)

    train_images, train_labels, test_images, test_labels = load_CIFAR10('path/to/CIFAR10’)

  • Experiment Notebooks:

    • PCANet_experiment_on_MNIST.ipynb

      A notebook shows the experiment results on MNIST.

    • PCANet_experiment_on_CIFAR10.ipynb

      A notebook shows the experiment results on CIFAR10.

    • VGG11_experiment_on_CIFAR10.ipynb

      For comparison, a notebook for training a simple VGG11 on CIFAR10.

Trained models

The models are named by its training data sets, which means that each model is trained by the related data set. So, you will select the proper model depending on what testing data set you want to use.

  • Usage: Use model = torch.load(PATH_to_the_model) to load the model you need, then perform prediction = model.predict(test_images) to get the predicted results.