[PYTHON] I wanted to challenge the classification of CIFAR-10 using Chainer's trainer

Introduction

The other day here I learned that Chainer can be written very concisely, so I challenged the image classification of CIFAR-10 that I wanted to try from before. I tried I wanted to write ..., but I have only a poor CPU environment, so I have not been able to confirm execution. I moved it all day and advanced about 2 epoch, so it's probably correct ... ^^; Regarding the implementation, I referred to the blog of here.

Implementation

Load the image of CIFAR-10

Download and load the CIFAR-10 data from here. Since it is like pickle, it is read by the following function.

def unpickle(file):
    fp = open(file, 'rb')
    if sys.version_info.major == 2:
        data = pickle.load(fp)
    elif sys.version_info.major == 3:
        data = pickle.load(fp, encoding='latin-1')                                                                                    
    fp.close()

    return data

neural network

I referred to the blog I introduced earlier. I'm still not sure how to design this area ...

class Cifar10Model(chainer.Chain):

    def __init__(self):
        super(Cifar10Model,self).__init__(
                conv1 = F.Convolution2D(3, 32, 3, pad=1),
                conv2 = F.Convolution2D(32, 32, 3, pad=1),
                conv3 = F.Convolution2D(32, 32, 3, pad=1),
                conv4 = F.Convolution2D(32, 32, 3, pad=1),
                conv5 = F.Convolution2D(32, 32, 3, pad=1),
                conv6 = F.Convolution2D(32, 32, 3, pad=1),
                l1 = L.Linear(512, 512),
                l2 = L.Linear(512,10))

    def __call__(self, x, train=True):
        h = F.relu(self.conv1(x))
        h = F.max_pooling_2d(F.relu(self.conv2(h)), 2)
        h = F.relu(self.conv3(h))
        h = F.max_pooling_2d(F.relu(self.conv4(h)), 2)
        h = F.relu(self.conv5(h))
        h = F.max_pooling_2d(F.relu(self.conv6(h)), 2)
        h = F.dropout(F.relu(self.l1(h)), train=train)
        return self.l2(h)

Data read

I got a little clogged here. When using Chainer's new function trainer, I pass the data I want to learn to the iterator, but in tutorials etc.

train_iter = chainer.iterators.SerialIterator(train, 100)
test_iter = chainer.iterators.SerialIterator(test, 100,repeat=False, shuffle=False)

I didn't know how to pass the label etc. After a lot of research, I found that Tuple_dataset should be used.

train = chainer.tuple_dataset.TupleDataset(train_data, train_label)

It seems to be good to do like this.

Below is the entire code for the read part.

x_train = None
y_train = []
for i in range(1,6):
    data_dic = unpickle("cifar-10-batches-py/data_batch_{}".format(i))
    if i == 1:
        x_train = data_dic['data']
    else:
        x_train = np.vstack((x_train, data_dic['data']))
    y_train += data_dic['labels']

test_data_dic = unpickle("cifar-10-batches-py/test_batch")
x_test = test_data_dic['data']
x_test = x_test.reshape(len(x_test),3,32,32)
y_test = np.array(test_data_dic['labels'])
x_train = x_train.reshape((len(x_train),3, 32, 32))
y_train = np.array(y_train)
x_train = x_train.astype(np.float32)
x_test = x_test.astype(np.float32)
x_train /= 255
x_test/=255                                                                                                                     
y_train = y_train.astype(np.int32)
y_test = y_test.astype(np.int32)

train = tuple_dataset.TupleDataset(x_train, y_train)
test = tuple_dataset.TupleDataset(x_test, y_test)

Learning part

I am learning with the neural network defined earlier. The code is just a little tweak to the tutorial MNIST. I am surprised to be able to write it insanely concisely.

model = L.Classifier(Cifar10Model())
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)

train_iter = chainer.iterators.SerialIterator(train, 100)
test_iter = chainer.iterators.SerialIterator(test, 100,repeat=False, shuffle=False)

updater = training.StandardUpdater(train_iter, optimizer, device=-1)
trainer = training.Trainer(updater, (40, 'epoch'), out="logs")
trainer.extend(extensions.Evaluator(test_iter, model, device=-1))
trainer.extend(extensions.LogReport())
trainer.extend(extensions.PrintReport( ['epoch', 'main/loss', 'validation/main/loss', 'main/accuracy', 'validation/main/accuracy']))
trainer.extend(extensions.ProgressBar())                                                                                          
trainer.run()

result

When you run it, a progress bar will appear to tell you how much you are learning.

Estimated time to finish: 6 days

~~ I gave up ~~ (Fixed on 2016.08.15) I did my best

I read the output log into dictionary and graphed it with matplotlib

in conclusion

~~ I couldn't confirm that the result was correct, but I learned how to use ~~ trainer. After all GPU is indispensable to study Deep Learning