.. _100LogisticIRISrst:

===============================================
100 - Logistic Regression with IRIS and pytorch
===============================================


.. only:: html

    **Links:** :download:`notebook <100_Logistic_IRIS.ipynb>`, :downloadlink:`html <100_Logistic_IRIS2html.html>`, :download:`PDF <100_Logistic_IRIS.pdf>`, :download:`python <100_Logistic_IRIS.py>`, :downloadlink:`slides <100_Logistic_IRIS.slides.html>`, :githublink:`GitHub|_doc/notebooks/101/100_Logistic_IRIS.ipynb|*`


First steps with `pytorch <http://pytorch.org/>`__ and
`Iris <http://scikit-learn.org/stable/auto_examples/datasets/plot_iris_dataset.html>`__
dataset.

**Note:** install `tqdm <https://pypi.python.org/pypi/tqdm>`__ if not
installed: ``!pip install tqdm``

.. code:: ipython3

    import time
    import numpy as np
    import pandas as pd
    import matplotlib.pyplot as plt
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    import torch.optim as optim
    from torch.autograd import Variable
    print("torch", torch.__version__)
    from torchvision import datasets, transforms
    from tqdm import tqdm
    from sklearn.datasets import load_iris
    from sklearn.preprocessing import LabelBinarizer


.. parsed-literal::
    torch 1.5.0+cpu


.. code:: ipython3

    %matplotlib inline

.. code:: ipython3

    X, Y = load_iris(return_X_y=True)
    X = X.astype("float32")
    X.shape, Y.shape




.. parsed-literal::
    ((150, 4), (150,))



.. code:: ipython3

    ftrain = np.arange(X.shape[0]) % 4 != 0
    Xtrain, Ytrain = X[ftrain, :], Y[ftrain]
    Xtest, Ytest = X[~ftrain, :], Y[~ftrain]
    Xtrain.shape, Ytrain.shape, Xtest.shape, Ytest.shape




.. parsed-literal::
    ((112, 4), (112,), (38, 4), (38,))



.. code:: ipython3

    BATCH_SIZE = 64
    TEST_BATCH_SIZE = 64
    N_EPOCHS = 2000

.. code:: ipython3

    class Net(nn.Module):
        def __init__(self):
            super(Net, self).__init__()
            self.fc1 = nn.Linear(4, 3)
        
        def forward(self, x):
            x = self.fc1(x)
            return F.log_softmax(x, dim=-1)

.. code:: ipython3

    model = Net()
    optimizer = optim.Adam(model.parameters())
    loss_fn = nn.NLLLoss()

.. code:: ipython3

    Xtrain_ = Variable(torch.from_numpy(Xtrain))
    Xtest_ = Variable(torch.from_numpy(Xtest))
    Ytrain_ = Variable(torch.from_numpy(Ytrain.astype(np.int64)))
    Ytest_ = Variable(torch.from_numpy(Ytest.astype(np.int64)))
    perfs = []
    for t in range(1, N_EPOCHS + 1):
        # Before the backward pass, use the optimizer object to zero all of the
        # gradients for the variables it will update (which are the learnable weights
        # of the model)
        optimizer.zero_grad()
        
        # Forward pass: compute predicted y by passing x to the model.
        Ypred = model(Xtrain_)
    
        # Compute and print loss.
        loss = loss_fn(Ypred , Ytrain_)
    
        # Backward pass: compute gradient of the loss with respect to model
        # parameters
        loss.backward()
    
        # Calling the step function on an Optimizer makes an update to its
        # parameters
        optimizer.step()
        
        Ypred_test = model(Xtest_)
        loss_test = loss_fn(Ypred_test, Ytest_)
        pred = Ypred_test.data.max(1, keepdim=True)[1] # get the index of the max log-probability
        accuracy = pred.eq(Ytest_.data.view_as(pred)).cpu().sum().item() / Ytest.size
        perfs.append([t, loss.data.item(), loss_test.data.item(), accuracy])

.. code:: ipython3

    df_perfs = pd.DataFrame(perfs, columns=["epoch", "train_loss", "test_loss", "accuracy"]).set_index("epoch")
    f, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))
    print("Last accuracy %.3f" % df_perfs.accuracy.iloc[-1])
    df_perfs[["train_loss", "test_loss"]].plot(ax=ax1);
    df_perfs[["accuracy"]].plot(ax=ax2);
    plt.ylim(ymin=0.7);


.. parsed-literal::
    Last accuracy 0.947



.. image:: 100_Logistic_IRIS_10_1.png