Add Deformable ConvNets

deformable_convnets/mnist/__init__.py

@@ -0,0 +1 @@
+

deformable_convnets/mnist/models.py

@@ -0,0 +1,45 @@
+import chainer
+import chainer.functions as F
+import chainer.links as L
+
+
+class Convnet(chainer.Chain):
+
+    def __init__(self, n_out):
+        super(Convnet, self).__init__(
+            # the size of the inputs to each layer will be inferred
+            l1=L.Convolution2D(None, 32, 3, 1, 1),
+            l2=L.Convolution2D(None, 32, 3, 1, 1),
+            l3=L.Convolution2D(None, 32, 3, 1, 1),
+            fc=L.Linear(None, n_out),
+        )
+
+    def __call__(self, x):
+        B = x.shape[0]
+        h1 = F.relu(self.l1(x))  # 28
+        h2 = F.relu(self.l2(h1))
+        h2 = F.max_pooling_2d(h2, ksize=4, stride=2, pad=1)  # 14
+        h3 = F.relu(self.l3(h2))
+        self.feat = h3
+        return self.fc(h3.reshape(B, -1))
+
+
+class DeformableConvnet(chainer.Chain):
+
+    def __init__(self, n_out):
+        super(DeformableConvnet, self).__init__(
+            # the size of the inputs to each layer will be inferred
+            l1=L.Convolution2D(None, 32, 3, 1, 1),
+            l2=L.DeformableConvolution2D(None, 32, 3, 1, 1),
+            l3=L.DeformableConvolution2D(None, 32, 3, 1, 1),
+            fc=L.Linear(None, n_out),
+        )
+
+    def __call__(self, x):
+        B = x.shape[0]
+        h1 = F.relu(self.l1(x))  # 28
+        h2 = F.relu(self.l2(h1))
+        h2 = F.max_pooling_2d(h2, ksize=4, stride=2, pad=1)  # 14
+        h3 = F.relu(self.l3(h2))
+        self.feat = h3
+        return self.fc(h3.reshape(B, -1))

deformable_convnets/mnist/scale_transform.py

@@ -0,0 +1,8 @@
+from chainercv import transforms
+
+
+def transform(in_data):
+    img, label = in_data
+    img = transforms.random_expand(img, max_ratio=3)
+    img = transforms.resize(img, (28, 28))
+    return img, label

deformable_convnets/mnist/test_mnist.py

@@ -0,0 +1,67 @@
+import argparse
+import matplotlib.pyplot as plt
+import numpy as np
+
+import chainer
+import chainer.functions as F
+import chainer.links as L
+from chainer import training
+from chainer.training import extensions
+from chainer.dataset.convert import concat_examples
+
+from chainercv.datasets import TransformDataset
+from chainercv import transforms
+
+from models import Convnet
+from models import DeformableConvnet
+from scale_transform import transform
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Chainer example: MNIST')
+    parser.add_argument('--resume', '-r', default='result/model_iter_9',
+                        help='Resume the training from snapshot')
+    parser.add_argument('--deformable', '-d', type=int, default=1,
+                        help='use deformable convolutions')
+    args = parser.parse_args()
+
+    if args.deformable == 1:
+        model = DeformableConvnet(10)
+    else:
+        model = Convnet(10)
+    chainer.serializers.load_npz(args.resume, model)
+
+    train, test = chainer.datasets.get_mnist(ndim=3)
+    test = TransformDataset(test, transform)
+
+    test_iter = chainer.iterators.SerialIterator(test, batch_size=1,
+                                                 repeat=False, shuffle=False)
+
+    threshold = 1
+    for i in range(1):
+        batch = test_iter.next()
+        in_arrays = concat_examples(batch, device=None)
+        in_vars = tuple(chainer.Variable(x) for x in in_arrays)
+        img, label = in_vars
+        model(img)
+        feat = model.feat
+        H, W = feat.shape[2:]
+        center = F.sum(feat[:, :, H / 2, W / 2])
+        center.grad= np.ones_like(center.data)
+        model.zerograds()
+        img.zerograd()
+        center.backward(retain_grad=True)
+
+        img_grad = img.grad[0]  # (1, 28, 28)
+
+        img_grad_abs = (np.abs(img_grad) / np.max(np.abs(img_grad)) * 255)[0]  # 28, 28
+        img_grad_abs[np.isnan(img_grad_abs)] = 0
+        y_indices, x_indices = np.where(img_grad_abs > threshold)
+        plt.scatter(x_indices, y_indices, c='red')
+        vis_img = transforms.chw_to_pil_image(255 * img.data[0])[:, :, 0]
+        plt.imshow(vis_img, interpolation='nearest', cmap='gray')
+        plt.show()
+
+
+if __name__ == '__main__':
+    main()

deformable_convnets/mnist/train_mnist.py

@@ -0,0 +1,109 @@
+import argparse
+
+import chainer
+import chainer.links as L
+from chainer import training
+from chainer.training import extensions
+
+from chainercv.datasets import TransformDataset
+from chainercv import transforms
+
+from models import Convnet
+from models import DeformableConvnet
+from scale_transform import transform
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Chainer example: MNIST')
+    parser.add_argument('--batchsize', '-b', type=int, default=100,
+                        help='Number of images in each mini-batch')
+    parser.add_argument('--epoch', '-e', type=int, default=20,
+                        help='Number of sweeps over the dataset to train')
+    parser.add_argument('--gpu', '-g', type=int, default=-1,
+                        help='GPU ID (negative value indicates CPU)')
+    parser.add_argument('--deformable', '-d', type=int, default=1,
+                        help='use deformable convolutions')
+    args = parser.parse_args()
+
+    print('GPU: {}'.format(args.gpu))
+    print('# Minibatch-size: {}'.format(args.batchsize))
+    print('# epoch: {}'.format(args.epoch))
+    print('deformable: {}'.format(args.deformable))
+    print('')
+
+    # Set up a neural network to train
+    # Classifier reports softmax cross entropy loss and accuracy at every
+    # iteration, which will be used by the PrintReport extension below.
+    # model = L.Classifier(Convnet(10))
+    if args.deformable == 1:
+        feat_extractor = DeformableConvnet(10)
+    else:
+        feat_extractor = Convnet(10)
+    model = L.Classifier(feat_extractor)
+    if args.gpu >= 0:
+        chainer.cuda.get_device(args.gpu).use()  # Make a specified GPU current
+        model.to_gpu()  # Copy the model to the GPU
+
+    # Setup an optimizer
+    optimizer = chainer.optimizers.Adam()
+    optimizer.setup(model)
+
+    # Load the MNIST dataset
+    train, test = chainer.datasets.get_mnist(ndim=3)
+    train = TransformDataset(train, transform)
+    test = TransformDataset(test, transform)
+
+    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
+    test_iter = chainer.iterators.SerialIterator(test, args.batchsize,
+                                                 repeat=False, shuffle=False)
+
+    # Set up a trainer
+    updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
+    trainer = training.Trainer(updater, (args.epoch, 'epoch'))
+
+    # Evaluate the model with the test dataset for each epoch
+    trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu))
+
+    # Dump a computational graph from 'loss' variable at the first iteration
+    # The "main" refers to the target link of the "main" optimizer.
+    trainer.extend(extensions.dump_graph('main/loss'))
+
+    # Take a snapshot for each specified epoch
+    if args.deformable:
+        snapshot_fn = 'model_iter_{.updater.epoch}'
+    else:
+        snapshot_fn = 'non_deformable_{.updater.epoch}'
+    trainer.extend(extensions.snapshot_object(feat_extractor, snapshot_fn),
+                   trigger=(1, 'epoch'))
+
+    # Write a log of evaluation statistics for each epoch
+    trainer.extend(extensions.LogReport())
+
+    # Save two plot images to the result dir
+    if extensions.PlotReport.available():
+        trainer.extend(
+            extensions.PlotReport(['main/loss', 'validation/main/loss'],
+                                  'epoch', file_name='loss.png'))
+        trainer.extend(
+            extensions.PlotReport(
+                ['main/accuracy', 'validation/main/accuracy'],
+                'epoch', file_name='accuracy.png'))
+
+    # Print selected entries of the log to stdout
+    # Here "main" refers to the target link of the "main" optimizer again, and
+    # "validation" refers to the default name of the Evaluator extension.
+    # Entries other than 'epoch' are reported by the Classifier link, called by
+    # either the updater or the evaluator.
+    trainer.extend(extensions.PrintReport(
+        ['epoch', 'main/loss', 'validation/main/loss',
+         'main/accuracy', 'validation/main/accuracy', 'elapsed_time']))
+
+    # Print a progress bar to stdout
+    trainer.extend(extensions.ProgressBar())
+
+    # Run the training
+    trainer.run()
+
+
+if __name__ == '__main__':
+    main()