simnets-tf
SimNets is a generalization of Convolutional Neural Networks that was first proposed by Cohen et al. (at NIPS 2014 DL Workshop and CVPR 2016), in whish they showed it to be well-suited for classification under limited computional resources. It was later discovered that under some settings SimNets realize Convolutional Arithmetic Circuits, where all computation is done in log-space instead of linear-space. We have previously released an implementation of SimNets in Caffe, and now we release our implementaiton in Tensorflow, with wrapper layers for Keras. You can read the documentation at https://huji-deep.github.io/simnets-tf/.
SimNets implementation in TensorFlow
Binary installation
Binary installation requires a cuda toolkit installation >= 7.5.
Download the .whl file from the GitHub release tab, then type:
python -m pip install <whl file>
all requirements should be installed automatically.
Building from Source
Building from source requires:
- A working c++ compiler with c++11 support (gcc >= 4.7)
- Cuda toolkit installed (for nvcc)
- CMake >= 3.0 (
apt install cmake) - TensorFlow installed for the Python interpreter you intend to use
Important: The following command should run without error:
python -c 'import tensorflow as tf'
To build the project type the following commands:
Python 2.7:
git clone --recursive https://github.com/HUJI-Deep/simnets-tf.git
cd simnets-tf
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DSIMNETS_PYTHON_VERSION=2.7 -DCMAKE_INSTALL_PREFIX=install
make -j simnet_ops
Python 3.5:
git clone --recursive https://github.com/HUJI-Deep/simnets-tf.git
cd simnets-tf
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DSIMNETS_PYTHON_VERSION=3.5 -DCMAKE_INSTALL_PREFIX=install
make -j simnet_ops
To test the code you can now type:
make test_simnet_ops
This should run for about two minutes and return without any errors.
Now you can create a .whl file:
make create_wheel
Finally, to install the simnets-tf package type (remember to use the right interpreter):
cd install/dist
python -m pip install <whl file>
The installation is successful if the following runs (again, remember to use the right interpreter):
python -c 'import simnets'
Usage example
Keras
import simnets.keras as sk
import keras
from keras.datasets import mnist
from keras.models import Model
from keras.layers import Input, Flatten, AveragePooling2D, Lambda
from keras import backend as K
import numpy as np
batch_size = 64
num_classes = 10
sim_kernel = 2
sim_channels = 32
mex_channels = sim_channels
epochs = 3
# input image dimensions
img_rows, img_cols = 28, 28
# the data, shuffled and split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
input_shape = (1, img_rows, img_cols)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train = x_train / 255.0 - 0.5
x_test = x_test / 255.0 - 0.5
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
def sum_pooling_layer(x, pool_size):
x = AveragePooling2D(pool_size=pool_size, padding='valid')(x)
x = Lambda(lambda x: x * pool_size[0] * pool_size[1])(x)
return x
a = Input(shape=(1, img_rows, img_cols))
b = sk.Similarity(sim_channels,
ksize=[2, 2], strides=[2, 2], similarity_function='L2',
normalization_term=True, padding=[2, 2], out_of_bounds_value=np.nan, ignore_nan_input=True)(a)
while b.shape[-2:] != (1, 1):
mex_channels *= 2
b = sk.Mex(mex_channels,
blocks=[int(b.shape[-3]), 1, 1], strides=[int(b.shape[-3]), 1, 1],
softmax_mode=True, normalize_offsets=True,
use_unshared_regions=True, unshared_offset_region=[2])(b)
b = sum_pooling_layer(b, pool_size=(2, 2))
b = sk.Mex(num_classes,
blocks=[mex_channels, 1, 1], strides=[mex_channels, 1, 1],
softmax_mode=True, normalize_offsets=True,
use_unshared_regions=True, shared_offset_region=[1])(b)
b = Flatten()(b)
model = Model(inputs=[a], outputs=[b])
print(model.summary())
def softmax_loss(y_true, y_pred):
return K.categorical_crossentropy(y_pred, y_true, True)
model.compile(loss=softmax_loss,
optimizer=keras.optimizers.nadam(lr=1e-2, epsilon=1e-6),
metrics=['accuracy'])
sk.perform_unsupervised_init(model, 'kmeans', layers=None, data=x_train, batch_size=100)
model.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(x_test, y_test))
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
Low level
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from simnets import similarity
from simnets.unsupervised import similarity_unsupervised_init
import matplotlib.pyplot as plt
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
sess = tf.InteractiveSession()
x = tf.placeholder(tf.float32, shape=[None, 784])
xr = tf.reshape(x, [-1, 1, 28, 28])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
shape = [10, 1, 28, 28]
templates = tf.Variable(tf.truncated_normal(shape, stddev=0.1))
weights_var = tf.Variable(tf.truncated_normal(shape, stddev=0.1))
weights = tf.abs(weights_var)
sim = similarity(xr, templates, weights, similarity_function='L2', ksize=[28,28], strides=[28,28], padding=[0,0])
y = tf.reshape(sim, [-1, 10])
kmo_init, kmo = similarity_unsupervised_init('kmeans', sim, templates, weights_var)
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
sess.run(tf.global_variables_initializer())
for idx in range(300):
batch = mnist.train.next_batch(100)
if idx == 0:
kmo_init.run(feed_dict={x: batch[0], y_: batch[1]})
kmo.run(feed_dict={x: batch[0], y_: batch[1]})
if (idx + 1) % 100 == 0:
print('kmeans', idx+1, '/', 1000)
def normalize(img):
return (img - img.min()) / (img.max() - img.min())
templates_np = tf.get_default_session().run(templates)
plt.figure(1)
for i in range(10):
plt.subplot(4,3, i+1)
plt.imshow(normalize(templates_np[i,0,...]))
plt.show()
for idx in range(1000):
batch = mnist.train.next_batch(100)
train_step.run(feed_dict={x: batch[0], y_: batch[1]})
if (idx + 1) % 100 == 0:
print(idx+1, '/', 1000)
correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print('Accuracy:', accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels}))
