{
  "cells": [
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "%matplotlib inline"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "\n# Deep CCA\n\nThis example demonstrates how to easily train Deep CCA models and variants\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "import numpy as np\nimport pytorch_lightning as pl\nfrom matplotlib import pyplot as plt\nfrom torch.utils.data import Subset"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "from cca_zoo.data import Split_MNIST_Dataset\nfrom cca_zoo.deepmodels import (\n    DCCA,\n    CCALightning,\n    get_dataloaders,\n    architectures,\n    DCCA_NOI,\n    DCCA_SDL,\n    BarlowTwins,\n)\n\n\ndef plot_latent_label(model, dataloader, num_batches=100):\n    fig, ax = plt.subplots(ncols=model.latent_dims)\n    for j in range(model.latent_dims):\n        ax[j].set_title(f\"Dimension {j}\")\n        ax[j].set_xlabel(\"View 1\")\n        ax[j].set_ylabel(\"View 2\")\n    for i, (data, label) in enumerate(dataloader):\n        z = model(*data)\n        zx, zy = z\n        zx = zx.to(\"cpu\").detach().numpy()\n        zy = zy.to(\"cpu\").detach().numpy()\n        for j in range(model.latent_dims):\n            ax[j].scatter(zx[:, j], zy[:, j], c=label.numpy(), cmap=\"tab10\")\n        if i > num_batches:\n            plt.colorbar()\n            break\n\n\nn_train = 500\nn_val = 100\ntrain_dataset = Split_MNIST_Dataset(mnist_type=\"MNIST\", train=True)\nval_dataset = Subset(train_dataset, np.arange(n_train, n_train + n_val))\ntrain_dataset = Subset(train_dataset, np.arange(n_train))\ntrain_loader, val_loader = get_dataloaders(train_dataset, val_dataset, batch_size=128)\n\n# The number of latent dimensions across models\nlatent_dims = 2\n# number of epochs for deep models\nepochs = 20\n\nencoder_1 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)\nencoder_2 = architectures.Encoder(latent_dims=latent_dims, feature_size=392)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Deep CCA\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "dcca = DCCA(latent_dims=latent_dims, encoders=[encoder_1, encoder_2])\ndcca = CCALightning(dcca)\ntrainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)\ntrainer.fit(dcca, train_loader, val_loader)\nplot_latent_label(dcca.model, train_loader)\nplt.suptitle(\"DCCA\")\nplt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Deep CCA by Non-Linear Orthogonal Iterations\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "dcca_noi = DCCA_NOI(\n    latent_dims=latent_dims, N=len(train_dataset), encoders=[encoder_1, encoder_2]\n)\ndcca_noi = CCALightning(dcca_noi)\ntrainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)\ntrainer.fit(dcca_noi, train_loader, val_loader)\nplot_latent_label(dcca_noi.model, train_loader)\nplt.title(\"DCCA by Non-Linear Orthogonal Iterations\")\nplt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Deep CCA by Stochastic Decorrelation Loss\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "dcca_sdl = DCCA_SDL(\n    latent_dims=latent_dims, N=len(train_dataset), encoders=[encoder_1, encoder_2]\n)\ndcca_sdl = CCALightning(dcca_sdl)\ntrainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)\ntrainer.fit(dcca_sdl, train_loader, val_loader)\nplot_latent_label(dcca_sdl.model, train_loader)\nplt.title(\"DCCA by Stochastic Decorrelation\")\nplt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Deep CCA by Barlow Twins\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "barlowtwins = BarlowTwins(latent_dims=latent_dims, encoders=[encoder_1, encoder_2])\nbarlowtwins = CCALightning(barlowtwins)\ntrainer = pl.Trainer(max_epochs=epochs, enable_checkpointing=False)\ntrainer.fit(dcca, train_loader, val_loader)\nplot_latent_label(dcca_sdl.model, train_loader)\nplt.title(\"DCCA by Barlow Twins\")\nplt.show()"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.7.9"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}