rahular.github.io

If you are looking for a SVM based approach, see here.

MNIST, by far is one of the most used toy datasets in machine learning. It contains 70,000 images of handwritten digits. The standard train-test split is 60,000-10,000. More details can be found here.

In this notebook, I show how applying very basic algorithms to this dataset gives surprisingly good results. It is always recommended to start off simple even on more complex datasets. Though simple models may not learn well if the data is complex, having a notebook like this as a starting point helps a lot in identifying the areas which need improvement. Also, the underlying process of experimentation remains the same.

Libraries used:

Here is a brief outline about the next sections:

  1. Pre-processing of data: normalization and one-hot encoding the outputs
  2. Augmenting the dataset with artificially generated data
  3. Training a linear model: single layer fully connected neural network
  4. Training a small CNN: A network similar to VGG16 in terms of architecture
  5. Adding regularization: Dropout and Batch Normalization

As always, what follows is mostly code. I have tried to add some explanations along the way :)

Pre-processing

Get MNIST test and train sets and see their dimensions

from keras.datasets import mnist
(train_x, train_y), (test_x, test_y) = mnist.load_data()

print train_x.shape, train_y.shape
print test_x.shape, test_y.shape
Using TensorFlow backend.


(60000, 28, 28) (60000,)
(10000, 28, 28) (10000,)

Add a channel dimension to this data as Keras expects it later when we reach CNNs. Since I use a tf backend, the channel is my 3rd dimension

import numpy as np

train_x = np.expand_dims(train_x, 3)
test_x = np.expand_dims(test_x, 3)

print train_x.shape
print test_x.shape
(60000, 28, 28, 1)
(10000, 28, 28, 1)

We also need to convert the labels into one hot encodings

from keras.utils.np_utils import to_categorical

print 'Before one hot encoding: ', train_y[5]

train_y = to_categorical(train_y)
test_y = to_categorical(test_y)

print 'After one hot encoding: ', train_y[5]
Before one hot encoding:  2
After one hot encoding:  [ 0.  0.  1.  0.  0.  0.  0.  0.  0.  0.]

Create a method which takes a data point as input and returns it’s normalized value

mean = train_x.mean().astype(np.float32)
std = train_x.std().astype(np.float32)

def norm_input(x):
    return (x - mean) / std

Create train and test batches with and without data augmentation. Note that we can perform normalization easily with the ImageGenerator itself, rather than doing it manually (as above)

from keras.preprocessing import image

# without augmentation
generator = image.ImageDataGenerator()
train_batches = generator.flow(train_x, train_y, batch_size=64)
test_batches = generator.flow(test_x, test_y, batch_size=64)

# with augmentation
da_generator = image.ImageDataGenerator(rotation_range=8, width_shift_range=0.08, shear_range=0.3, 
                           height_shift_range=0.08, zoom_range=0.08)
da_train_batches = da_generator.flow(train_x, train_y, batch_size=64)
da_test_batches = da_generator.flow(test_x, test_y, batch_size=64)

Let us write a simple training method which does the following:

def train(model, train, test):
    model.fit_generator(generator=train, samples_per_epoch=train.N, nb_epoch=1, 
                           validation_data=test, nb_val_samples=test.N)
    model.optimizer.lr = 0.1
    model.fit_generator(generator=train, samples_per_epoch=train.N, nb_epoch=5, 
                           validation_data=test, nb_val_samples=test.N)
    model.optimizer.lr = 0.01
    model.fit_generator(generator=train, samples_per_epoch=train.N, nb_epoch=5, 
                           validation_data=test, nb_val_samples=test.N) 

Linear model

Create a simple fully connected network with a single hidden layer. The hidden layer has 10 neurons. We use standard SGD for optimization and categorical crossentropy loss function.

Lamda layer allows us to pass an arbitrary function which is applied on each data point.

Flatten layer just squashes the data into a single dimension.

from keras.models import Sequential
from keras.layers import Dense, Lambda, Flatten
from keras.optimizers import sgd

def get_linear_model():
    model = Sequential()
    model.add(Lambda(norm_input, input_shape=(28, 28, 1)))
    model.add(Flatten())
    model.add(Dense(10, activation='softmax'))
    model.compile(sgd(), loss='categorical_crossentropy', metrics=['accuracy'])

    return model

Linear model without augmentation

linear_model = get_linear_model()
train(linear_model, train_batches, test_batches)
Epoch 1/1
60000/60000 [==============================] - 7s - loss: 0.5115 - acc: 0.8513 - val_loss: 0.3399 - val_acc: 0.9037
Epoch 1/5
60000/60000 [==============================] - 7s - loss: 0.3418 - acc: 0.9015 - val_loss: 0.3137 - val_acc: 0.9093
Epoch 2/5
60000/60000 [==============================] - 7s - loss: 0.3198 - acc: 0.9081 - val_loss: 0.2976 - val_acc: 0.9147
Epoch 3/5
60000/60000 [==============================] - 7s - loss: 0.3055 - acc: 0.9133 - val_loss: 0.2921 - val_acc: 0.9152
Epoch 4/5
60000/60000 [==============================] - 7s - loss: 0.2975 - acc: 0.9155 - val_loss: 0.2873 - val_acc: 0.9175
Epoch 5/5
60000/60000 [==============================] - 7s - loss: 0.2922 - acc: 0.9166 - val_loss: 0.2885 - val_acc: 0.9186
Epoch 1/5
60000/60000 [==============================] - 7s - loss: 0.2876 - acc: 0.9186 - val_loss: 0.2875 - val_acc: 0.9174
Epoch 2/5
60000/60000 [==============================] - 7s - loss: 0.2849 - acc: 0.9194 - val_loss: 0.2825 - val_acc: 0.9229
Epoch 3/5
60000/60000 [==============================] - 7s - loss: 0.2814 - acc: 0.9214 - val_loss: 0.2737 - val_acc: 0.9216
Epoch 4/5
60000/60000 [==============================] - 7s - loss: 0.2782 - acc: 0.9227 - val_loss: 0.2780 - val_acc: 0.9202
Epoch 5/5
60000/60000 [==============================] - 7s - loss: 0.2731 - acc: 0.9231 - val_loss: 0.2732 - val_acc: 0.9251

Linear model with augmentation

da_linear_model = get_linear_model()
train(da_linear_model, da_train_batches, da_test_batches)
Epoch 1/1
60000/60000 [==============================] - 7s - loss: 1.1433 - acc: 0.6333 - val_loss: 0.9586 - val_acc: 0.7045
Epoch 1/5
60000/60000 [==============================] - 8s - loss: 0.9645 - acc: 0.7021 - val_loss: 0.9101 - val_acc: 0.7168
Epoch 2/5
60000/60000 [==============================] - 7s - loss: 0.9418 - acc: 0.7100 - val_loss: 0.9046 - val_acc: 0.7242
Epoch 3/5
60000/60000 [==============================] - 7s - loss: 0.9307 - acc: 0.7167 - val_loss: 0.8987 - val_acc: 0.7314
Epoch 4/5
60000/60000 [==============================] - 7s - loss: 0.9301 - acc: 0.7180 - val_loss: 0.8909 - val_acc: 0.7285
Epoch 5/5
60000/60000 [==============================] - 7s - loss: 0.9183 - acc: 0.7222 - val_loss: 0.8600 - val_acc: 0.7375
Epoch 1/5
60000/60000 [==============================] - 7s - loss: 0.9138 - acc: 0.7227 - val_loss: 0.8701 - val_acc: 0.7423
Epoch 2/5
60000/60000 [==============================] - 7s - loss: 0.9080 - acc: 0.7246 - val_loss: 0.8778 - val_acc: 0.7394
Epoch 3/5
60000/60000 [==============================] - 7s - loss: 0.9091 - acc: 0.7271 - val_loss: 0.8683 - val_acc: 0.7401
Epoch 4/5
60000/60000 [==============================] - 7s - loss: 0.9122 - acc: 0.7251 - val_loss: 0.8657 - val_acc: 0.7360
Epoch 5/5
60000/60000 [==============================] - 7s - loss: 0.9075 - acc: 0.7256 - val_loss: 0.8810 - val_acc: 0.7361

VGG-like CNN model

CNN, which specialize in image data, is expected to out-perform the previous model. Let’s see by how much!

from keras.models import Sequential
from keras.layers import Dense, Lambda, Flatten, Convolution2D, MaxPooling2D
from keras.optimizers import sgd

def get_vgg_model():
    model = Sequential()
    model.add(Lambda(norm_input, input_shape=(28, 28, 1)))
    model.add(Convolution2D(32, 3, 3, activation='relu'))
    model.add(Convolution2D(32, 3, 3, activation='relu'))
    model.add(MaxPooling2D())
    model.add(Convolution2D(64, 3, 3, activation='relu'))
    model.add(Convolution2D(64, 3, 3, activation='relu'))
    model.add(MaxPooling2D())
    model.add(Flatten())
    model.add(Dense(512, activation='relu'))
    model.add(Dense(10, activation='softmax'))
    model.compile(sgd(), loss='categorical_crossentropy', metrics=['accuracy'])

    return model

CNN without augmentation

vgg_model = get_vgg_model()
train(vgg_model, train_batches, test_batches)
Epoch 1/1
60000/60000 [==============================] - 84s - loss: 1.3413 - acc: 0.5329 - val_loss: 0.1907 - val_acc: 0.9452
Epoch 1/5
60000/60000 [==============================] - 85s - loss: 0.1524 - acc: 0.9533 - val_loss: 0.1001 - val_acc: 0.9685
Epoch 2/5
60000/60000 [==============================] - 85s - loss: 0.0948 - acc: 0.9714 - val_loss: 0.0678 - val_acc: 0.9776
Epoch 3/5
60000/60000 [==============================] - 86s - loss: 0.0725 - acc: 0.9774 - val_loss: 0.0665 - val_acc: 0.9790
Epoch 4/5
60000/60000 [==============================] - 87s - loss: 0.0575 - acc: 0.9819 - val_loss: 0.0563 - val_acc: 0.9819
Epoch 5/5
60000/60000 [==============================] - 84s - loss: 0.0486 - acc: 0.9849 - val_loss: 0.0470 - val_acc: 0.9841
Epoch 1/5
60000/60000 [==============================] - 86s - loss: 0.0435 - acc: 0.9864 - val_loss: 0.0442 - val_acc: 0.9853
Epoch 2/5
60000/60000 [==============================] - 85s - loss: 0.0361 - acc: 0.9887 - val_loss: 0.0441 - val_acc: 0.9858
Epoch 3/5
60000/60000 [==============================] - 85s - loss: 0.0317 - acc: 0.9899 - val_loss: 0.0399 - val_acc: 0.9857
Epoch 4/5
60000/60000 [==============================] - 83s - loss: 0.0283 - acc: 0.9911 - val_loss: 0.0438 - val_acc: 0.9856
Epoch 5/5
60000/60000 [==============================] - 85s - loss: 0.0248 - acc: 0.9919 - val_loss: 0.0464 - val_acc: 0.9860

By now, one should have observed that the training accuracy is slightly higher than the validation accuracy which is an indication of over fitting. We will address that in the next section

CNN with augmentation

da_vgg_model = get_vgg_model()
train(da_vgg_model, da_train_batches, da_test_batches)
Epoch 1/1
60000/60000 [==============================] - 84s - loss: 1.9180 - acc: 0.3261 - val_loss: 0.6366 - val_acc: 0.7977
Epoch 1/5
60000/60000 [==============================] - 85s - loss: 0.3559 - acc: 0.8900 - val_loss: 0.2062 - val_acc: 0.9375
Epoch 2/5
60000/60000 [==============================] - 84s - loss: 0.2029 - acc: 0.9371 - val_loss: 0.1989 - val_acc: 0.9349
Epoch 3/5
60000/60000 [==============================] - 82s - loss: 0.1558 - acc: 0.9524 - val_loss: 0.1327 - val_acc: 0.9571
Epoch 4/5
60000/60000 [==============================] - 82s - loss: 0.1317 - acc: 0.9584 - val_loss: 0.1093 - val_acc: 0.9648
Epoch 5/5
60000/60000 [==============================] - 84s - loss: 0.1167 - acc: 0.9645 - val_loss: 0.1002 - val_acc: 0.9672
Epoch 1/5
60000/60000 [==============================] - 86s - loss: 0.1051 - acc: 0.9671 - val_loss: 0.0934 - val_acc: 0.9708
Epoch 2/5
60000/60000 [==============================] - 87s - loss: 0.0964 - acc: 0.9698 - val_loss: 0.0999 - val_acc: 0.9692
Epoch 3/5
60000/60000 [==============================] - 86s - loss: 0.0885 - acc: 0.9728 - val_loss: 0.0906 - val_acc: 0.9716
Epoch 4/5
60000/60000 [==============================] - 86s - loss: 0.0824 - acc: 0.9738 - val_loss: 0.0717 - val_acc: 0.9771
Epoch 5/5
60000/60000 [==============================] - 87s - loss: 0.0789 - acc: 0.9745 - val_loss: 0.0738 - val_acc: 0.9756

VGG-like CNN model with regularization

The purpose of Regularization is to avoid over-fitting. Here we apply:

For sake of demonstration, we will only consider un-augmented data from here on.

Dropout

from keras.models import Sequential
from keras.layers import Dense, Lambda, Flatten, Convolution2D, MaxPooling2D, Dropout
from keras.optimizers import sgd

def get_dropout_vgg_model():
    model = Sequential()
    model.add(Lambda(norm_input, input_shape=(28, 28, 1)))
    model.add(Convolution2D(32, 3, 3, activation='relu'))
    model.add(Convolution2D(32, 3, 3, activation='relu'))
    model.add(MaxPooling2D())
    model.add(Dropout(0.5))
    model.add(Convolution2D(64, 3, 3, activation='relu'))
    model.add(Convolution2D(64, 3, 3, activation='relu'))
    model.add(MaxPooling2D())
    model.add(Dropout(0.5))
    model.add(Flatten())
    model.add(Dense(512, activation='relu'))
    model.add(Dense(10, activation='softmax'))
    model.compile(sgd(), loss='categorical_crossentropy', metrics=['accuracy'])

    return model
dropout_vgg_model = get_dropout_vgg_model()
train(dropout_vgg_model, train_batches, test_batches)
Epoch 1/1
60000/60000 [==============================] - 93s - loss: 2.2000 - acc: 0.1931 - val_loss: 0.8338 - val_acc: 0.7178
Epoch 1/5
60000/60000 [==============================] - 89s - loss: 0.4403 - acc: 0.8606 - val_loss: 0.1492 - val_acc: 0.9556
Epoch 2/5
60000/60000 [==============================] - 89s - loss: 0.2202 - acc: 0.9334 - val_loss: 0.1008 - val_acc: 0.9692
Epoch 3/5
60000/60000 [==============================] - 88s - loss: 0.1682 - acc: 0.9483 - val_loss: 0.0831 - val_acc: 0.9746
Epoch 4/5
60000/60000 [==============================] - 87s - loss: 0.1443 - acc: 0.9555 - val_loss: 0.0641 - val_acc: 0.9791
Epoch 5/5
60000/60000 [==============================] - 87s - loss: 0.1250 - acc: 0.9607 - val_loss: 0.0564 - val_acc: 0.9815
Epoch 1/5
60000/60000 [==============================] - 86s - loss: 0.1126 - acc: 0.9652 - val_loss: 0.0519 - val_acc: 0.9824
Epoch 2/5
60000/60000 [==============================] - 88s - loss: 0.1040 - acc: 0.9677 - val_loss: 0.0436 - val_acc: 0.9855
Epoch 3/5
60000/60000 [==============================] - 90s - loss: 0.0956 - acc: 0.9699 - val_loss: 0.0451 - val_acc: 0.9859
Epoch 4/5
60000/60000 [==============================] - 87s - loss: 0.0893 - acc: 0.9706 - val_loss: 0.0386 - val_acc: 0.9871
Epoch 5/5
60000/60000 [==============================] - 87s - loss: 0.0832 - acc: 0.9738 - val_loss: 0.0390 - val_acc: 0.9878

Dropout + Batch Normalization

from keras.models import Sequential
from keras.layers import Dense, Lambda, Flatten, Convolution2D, MaxPooling2D, Dropout, BatchNormalization
from keras.optimizers import sgd

def get_bn_dropout_vgg_model():
    model = Sequential()
    model.add(Lambda(norm_input, input_shape=(28, 28, 1)))
    model.add(Convolution2D(32, 3, 3, activation='relu'))
    model.add(BatchNormalization(axis=3))
    model.add(Convolution2D(32, 3, 3, activation='relu'))
    model.add(BatchNormalization(axis=3))
    model.add(MaxPooling2D())
    model.add(Convolution2D(64, 3, 3, activation='relu'))
    model.add(BatchNormalization(axis=3))
    model.add(Convolution2D(64, 3, 3, activation='relu'))
    model.add(BatchNormalization(axis=3))
    model.add(MaxPooling2D())
    model.add(Flatten())
    model.add(Dense(512, activation='relu'))
    model.add(BatchNormalization())
    model.add(Dropout(0.5))
    model.add(Dense(10, activation='softmax'))
    model.compile(sgd(), loss='categorical_crossentropy', metrics=['accuracy'])

    return model
bn_dropout_vgg_model = get_bn_dropout_vgg_model()
train(bn_dropout_vgg_model, train_batches, test_batches)
Epoch 1/1
60000/60000 [==============================] - 177s - loss: 0.1690 - acc: 0.9485 - val_loss: 0.0366 - val_acc: 0.9871
Epoch 1/5
60000/60000 [==============================] - 180s - loss: 0.0581 - acc: 0.9820 - val_loss: 0.0292 - val_acc: 0.9897
Epoch 2/5
60000/60000 [==============================] - 185s - loss: 0.0424 - acc: 0.9867 - val_loss: 0.0220 - val_acc: 0.9926
Epoch 3/5
60000/60000 [==============================] - 184s - loss: 0.0365 - acc: 0.9886 - val_loss: 0.0206 - val_acc: 0.9935
Epoch 4/5
60000/60000 [==============================] - 182s - loss: 0.0281 - acc: 0.9910 - val_loss: 0.0201 - val_acc: 0.9932
Epoch 5/5
60000/60000 [==============================] - 177s - loss: 0.0258 - acc: 0.9919 - val_loss: 0.0189 - val_acc: 0.9938
Epoch 1/5
60000/60000 [==============================] - 177s - loss: 0.0222 - acc: 0.9932 - val_loss: 0.0181 - val_acc: 0.9937
Epoch 2/5
60000/60000 [==============================] - 177s - loss: 0.0197 - acc: 0.9939 - val_loss: 0.0162 - val_acc: 0.9944
Epoch 3/5
60000/60000 [==============================] - 177s - loss: 0.0165 - acc: 0.9949 - val_loss: 0.0195 - val_acc: 0.9929
Epoch 4/5
60000/60000 [==============================] - 178s - loss: 0.0154 - acc: 0.9955 - val_loss: 0.0182 - val_acc: 0.9939
Epoch 5/5
60000/60000 [==============================] - 178s - loss: 0.0150 - acc: 0.9952 - val_loss: 0.0216 - val_acc: 0.9928

We can see the 4th epoch went up to 0.9939 and came back down in the 5th. Therefore, it is always a good idea to save the weights after every epoch and use the one which gave the best validation accuracy.

In the hope of pushing it back up just a little bit more, let’s decay the learning rate some more and train it for 5 additional epochs

bn_dropout_vgg_model.optimizer.lr = 0.001
bn_dropout_vgg_model.fit_generator(generator=train_batches, samples_per_epoch=train_batches.N, nb_epoch=5, 
                       validation_data=test_batches, nb_val_samples=test_batches.N) 
Epoch 1/5
60000/60000 [==============================] - 175s - loss: 0.0125 - acc: 0.9963 - val_loss: 0.0187 - val_acc: 0.9931
Epoch 2/5
60000/60000 [==============================] - 175s - loss: 0.0116 - acc: 0.9962 - val_loss: 0.0136 - val_acc: 0.9955
Epoch 3/5
60000/60000 [==============================] - 175s - loss: 0.0104 - acc: 0.9968 - val_loss: 0.0190 - val_acc: 0.9938
Epoch 4/5
60000/60000 [==============================] - 180s - loss: 0.0090 - acc: 0.9976 - val_loss: 0.0173 - val_acc: 0.9947
Epoch 5/5
60000/60000 [==============================] - 177s - loss: 0.0087 - acc: 0.9972 - val_loss: 0.0188 - val_acc: 0.9943

<keras.callbacks.History at 0x7fd21ecc1e10>

So with 20 minutes of coding, we were able to reach an accuracy of 0.9955 (epoch 2) which means an error of 0.45%! This puts the model among the best models in the world! Don’t believe me? Have a look here.