iframe-proxy | Sunbelt Computer Software

210 lines (182 loc) · 7.22 KB
    Testing code for different neural network configurations.
    Adapted for Python 3.5.2
    Usage in shell:
        python3.5 test.py
    Network (network.py and network2.py) parameters:
        2nd param is epochs count
        3rd param is batch size
        4th param is learning rate (eta)
    Author:
        Michał Dobrzański, 2016
        dobrzanski.michal.daniel@gmail.com
# ----------------------
# - read the input data:
import mnist_loader
training_data, validation_data, test_data = mnist_loader.load_data_wrapper()
training_data = list(training_data)
# ---------------------
# - network.py example:
#import network
net = network.Network([784, 30, 10])
net.SGD(training_data, 30, 10, 3.0, test_data=test_data)
# ----------------------
# - network2.py example:
#import network2
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
#net.large_weight_initializer()
net.SGD(training_data, 30, 10, 0.1, lmbda = 5.0,evaluation_data=validation_data,
    monitor_evaluation_accuracy=True)
# chapter 3 - Overfitting example - too many epochs of learning applied on small (1k samples) amount od data.
# Overfitting is treating noise as a signal.
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
net.large_weight_initializer()
net.SGD(training_data[:1000], 400, 10, 0.5, evaluation_data=test_data,
    monitor_evaluation_accuracy=True,
    monitor_training_cost=True)
# chapter 3 - Regularization (weight decay) example 1 (only 1000 of training data and 30 hidden neurons)
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
net.large_weight_initializer()
net.SGD(training_data[:1000], 400, 10, 0.5,
    evaluation_data=test_data,
    lmbda = 0.1, # this is a regularization parameter
    monitor_evaluation_cost=True,
    monitor_evaluation_accuracy=True,
    monitor_training_cost=True,
    monitor_training_accuracy=True)
# chapter 3 - Early stopping implemented
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
net.SGD(training_data[:1000], 30, 10, 0.5,
    lmbda=5.0,
    evaluation_data=validation_data,
    monitor_evaluation_accuracy=True,
    monitor_training_cost=True,
    early_stopping_n=10)
# chapter 4 - The vanishing gradient problem - deep networks are hard to train with simple SGD algorithm
# this network learns much slower than a shallow one.
net = network2.Network([784, 30, 30, 30, 30, 10], cost=network2.CrossEntropyCost)
net.SGD(training_data, 30, 10, 0.1,
    lmbda=5.0,
    evaluation_data=validation_data,
    monitor_evaluation_accuracy=True)
# ----------------------
# Theano and CUDA
# ----------------------
    This deep network uses Theano with GPU acceleration support.
    I am using Ubuntu 16.04 with CUDA 7.5.
    Tutorial:
    http://deeplearning.net/software/theano/install_ubuntu.html#install-ubuntu
    The following command will update only Theano:
        sudo pip install --upgrade --no-deps theano
    The following command will update Theano and Numpy/Scipy (warning bellow):
        sudo pip install --upgrade theano
    Below, there is a testing function to check whether your computations have been made on CPU or GPU.
    If the result is 'Used the cpu' and you want to have it in gpu,     do the following:
    1) install theano:
        sudo python3.5 -m pip install Theano
    2) download and install the latest cuda:
        https://developer.nvidia.com/cuda-downloads
        I had some issues with that, so I followed this idea (better option is to download the 1,1GB package as .run file):
        http://askubuntu.com/questions/760242/how-can-i-force-16-04-to-add-a-repository-even-if-it-isnt-considered-secure-eno
        You may also want to grab the proper NVidia driver, choose it form there:
        System Settings > Software & Updates > Additional Drivers.
    3) should work, run it with:
        THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python3.5 test.py
        http://deeplearning.net/software/theano/tutorial/using_gpu.html
    4) Optionally, you can add cuDNN support from:
        https://developer.nvidia.com/cudnn
def testTheano():
    from theano import function, config, shared, sandbox
    import theano.tensor as T
    import numpy
    import time
    print("Testing Theano library...")
    vlen = 10 * 30 * 768  # 10 x #cores x # threads per core
    iters = 1000
    rng = numpy.random.RandomState(22)
    x = shared(numpy.asarray(rng.rand(vlen), config.floatX))
    f = function([], T.exp(x))
    print(f.maker.fgraph.toposort())
    t0 = time.time()
    for i in range(iters):
        r = f()
    t1 = time.time()
    print("Looping %d times took %f seconds" % (iters, t1 - t0))
    print("Result is %s" % (r,))
    if numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]):
        print('Used the cpu')
        print('Used the gpu')
# Perform check:
#testTheano()
# ----------------------
# - network3.py example:
import network3
from network3 import Network, ConvPoolLayer, FullyConnectedLayer, SoftmaxLayer # softmax plus log-likelihood cost is more common in modern image classification networks.
# read data:
training_data, validation_data, test_data = network3.load_data_shared()
# mini-batch size:
mini_batch_size = 10
# chapter 6 - shallow architecture using just a single hidden layer, containing 100 hidden neurons.
net = Network([
    FullyConnectedLayer(n_in=784, n_out=100),
    SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.1, validation_data, test_data)
# chapter 6 - 5x5 local receptive fields, 20 feature maps, max-pooling layer 2x2
net = Network([
    ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),
                  filter_shape=(20, 1, 5, 5),
                  poolsize=(2, 2)),
    FullyConnectedLayer(n_in=20*12*12, n_out=100),
    SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.1, validation_data, test_data)
# chapter 6 - inserting a second convolutional-pooling layer to the previous example => better accuracy
net = Network([
    ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),
                  filter_shape=(20, 1, 5, 5),
                  poolsize=(2, 2)),
    ConvPoolLayer(image_shape=(mini_batch_size, 20, 12, 12),
                  filter_shape=(40, 20, 5, 5),
                  poolsize=(2, 2)),
    FullyConnectedLayer(n_in=40*4*4, n_out=100),
    SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.1, validation_data, test_data)
# chapter 6 -  rectified linear units and some l2 regularization (lmbda=0.1) => even better accuracy
from network3 import ReLU
net = Network([
    ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),
                  filter_shape=(20, 1, 5, 5),
                  poolsize=(2, 2),
                  activation_fn=ReLU),
    ConvPoolLayer(image_shape=(mini_batch_size, 20, 12, 12),
                  filter_shape=(40, 20, 5, 5),
                  poolsize=(2, 2),
                  activation_fn=ReLU),
    FullyConnectedLayer(n_in=40*4*4, n_out=100, activation_fn=ReLU),
    SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.03, validation_data, test_data, lmbda=0.1)
Sunbelt Computer Software

PL/B Language Development and Support

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Sunbelt Computer Software

PL/B Language Development and Support

FilesExpand file tree

test.py

Latest commit

History

test.py

File metadata and controls
