def vgg19():
"""
VGG19 network architecture with random parameters. Parameters
can be loaded using ``neupy.storage`` module.
Originally VGG19 was built in order to solve image classification
problem. It was used in the ImageNet competition. The goal of the
competition is to build a model that classifies image into one of
the 1,000 categories. Categories include animals, objects, transports
and so on.
VGG19 has roughly 143 million parameters.
Examples
--------
>>> from neupy import architectures
>>> vgg19 = architectures.vgg19()
>>> vgg19
(?, 224, 224, 3) -> [... 47 layers ...] -> (?, 1000)
>>>
>>> from neupy import algorithms
>>> optimizer = algorithms.Momentum(vgg19)
See Also
--------
:architecture:`vgg16` : VGG16 network
:architecture:`squeezenet` : SqueezeNet network
:architecture:`resnet50` : ResNet50 network
References
----------
Very Deep Convolutional Networks for Large-Scale Image Recognition.
https://arxiv.org/abs/1409.1556
"""
HalfPadConvolution = partial(layers.Convolution, padding='SAME')
return layers.join(
layers.Input((224, 224, 3)),
HalfPadConvolution((3, 3, 64), name='conv1_1') > layers.Relu(),
HalfPadConvolution((3, 3, 64), name='conv1_2') > layers.Relu(),
layers.MaxPooling((2, 2)),
HalfPadConvolution((3, 3, 128), name='conv2_1') > layers.Relu(),
HalfPadConvolution((3, 3, 128), name='conv2_2') > layers.Relu(),
layers.MaxPooling((2, 2)),
HalfPadConvolution((3, 3, 256), name='conv3_1') > layers.Relu(),
HalfPadConvolution((3, 3, 256), name='conv3_2') > layers.Relu(),
HalfPadConvolution((3, 3, 256), name='conv3_3') > layers.Relu(),
HalfPadConvolution((3, 3, 256), name='conv3_4') > layers.Relu(),
layers.MaxPooling((2, 2)),
HalfPadConvolution((3, 3, 512), name='conv4_1') > layers.Relu(),
HalfPadConvolution((3, 3, 512), name='conv4_2') > layers.Relu(),
HalfPadConvolution((3, 3, 512), name='conv4_3') > layers.Relu(),
HalfPadConvolution((3, 3, 512), name='conv4_4') > layers.Relu(),
layers.MaxPooling((2, 2)),
HalfPadConvolution((3, 3, 512), name='conv5_1') > layers.Relu(),
HalfPadConvolution((3, 3, 512), name='conv5_2') > layers.Relu(),
HalfPadConvolution((3, 3, 512), name='conv5_3') > layers.Relu(),
HalfPadConvolution((3, 3, 512), name='conv5_4') > layers.Relu(),
layers.MaxPooling((2, 2)),
layers.Reshape(),
layers.Linear(4096, name='dense_1') > layers.Relu(),
layers.Dropout(0.5),
layers.Linear(4096, name='dense_2') > layers.Relu(),
layers.Dropout(0.5),
layers.Linear(1000, name='dense_3') > layers.Softmax(),
)
x_labeled, x_unlabeled, y_labeled, y_unlabeled = train_test_split(
data.astype(np.float32),
target.astype(np.float32),
test_size=(1 - n_labeled / n_samples))
x_labeled_4d = x_labeled.reshape((n_labeled, 1, 28, 28))
x_unlabeled_4d = x_unlabeled.reshape((n_unlabeled, 1, 28, 28))
encoder = layers.join(
layers.Input((1, 28, 28)),
layers.Convolution((16, 3, 3)) > layers.Relu(),
layers.Convolution((16, 3, 3)) > layers.Relu(),
layers.MaxPooling((2, 2)),
layers.Convolution((32, 3, 3)) > layers.Relu(),
layers.MaxPooling((2, 2)),
layers.Reshape(),
layers.Relu(256),
layers.Relu(128),
)
decoder = layers.join(
layers.Relu(256),
layers.Relu(32 * 5 * 5),
layers.Reshape((32, 5, 5)),
layers.Upscale((2, 2)),
layers.Convolution((16, 3, 3), padding='full') > layers.Relu(),
layers.Upscale((2, 2)),
layers.Convolution((16, 3, 3), padding='full') > layers.Relu(),
layers.Convolution((1, 3, 3), padding='full') > layers.Sigmoid(),
layers.Reshape(),
)
def test_gru_connection_exceptions(self):
network = layers.join(layers.GRU(10), layers.Reshape())
with self.assertRaises(LayerConnectionError):
layers.join(layers.Input(1), network)
network = algorithms.Momentum(
[
[
[
# 3 categorical inputs
layers.Input(3),
# Train embedding matrix for categorical inputs.
# It has 18 different unique categories (6 categories
# per each of the 3 columns). Next layer projects each
# category into 4 dimensional space. Output shape from
# the layer should be: (batch_size, 3, 4)
layers.Embedding(n_unique_categories, 4),
# Reshape (batch_size, 3, 4) to (batch_size, 12)
layers.Reshape(),
],
[
# 17 numerical inputs
layers.Input(17),
]
],
# Concatenate (batch_size, 12) and (batch_size, 17)
# into one matrix with shape (batch_size, 29)
layers.Concatenate(),
layers.Relu(128),
layers.Relu(32) > layers.Dropout(0.5),
layers.Sigmoid(1)
],
step=0.2,
test_size=(1 - n_labeled / n_samples))
x_labeled_4d = x_labeled.reshape((n_labeled, 28, 28, 1))
x_unlabeled_4d = x_unlabeled.reshape((n_unlabeled, 28, 28, 1))
# We will features trained in the encoder and the first part for the future
# classifier. At first we pre-train them with unlabeled data, since we have
# a lot of it and we hope to learn some common features from it.
encoder = layers.join(
layers.Input((28, 28, 1)),
layers.Convolution((3, 3, 16)) > layers.Relu(),
layers.Convolution((3, 3, 16)) > layers.Relu(),
layers.MaxPooling((2, 2)),
layers.Convolution((3, 3, 32)) > layers.Relu(),
layers.MaxPooling((2, 2)),
layers.Reshape(),
layers.Relu(256),
layers.Relu(128),
)
# Notice that in the decoder every operation reverts back changes from the
# encoder layer. Upscale replaces MaxPooling and Convolutional layer
# without padding replaced with large padding that increase size of the image.
decoder = layers.join(
layers.Relu(256),
layers.Relu(32 * 5 * 5),
layers.Reshape((5, 5, 32)),
layers.Upscale((2, 2)),
layers.Convolution((3, 3, 16), padding=2) > layers.Relu(),
layers.Upscale((2, 2)),
layers.Convolution((3, 3, 16), padding=2) > layers.Relu(),