def test_multitask_dbn_mnist(learning_rate=0.01, training_epochs=10, batch_size=20,
finetune_lr=0.05, finetune_epoch=10, output_folder=None, dropout=0.2,
model_structure=[500, 100], isPCD=0, k=1, plot_weight=False):
"""
test_dbn_mnist(output_folder='/home/eric/Desktop/dbn_plots',
training_epochs=10,
model_structure=[500, 100],
finetune_epoch=10)
"""
assert output_folder is not None
from dbn_variants import MultiTask_DBN as DBN
#################################
# Data Constructing #
#################################
from sklearn.datasets import fetch_mldata
from xylearn.utils.data_util import get_train_test
from xylearn.utils.data_normalization import rescale
dataset = fetch_mldata('MNIST original')
data = get_train_test(rescale(dataset.data), dataset.target, useGPU=1, shuffle=True)
train_x, train_y = data['train']
test_x, test_y = data['test']
num_of_classes = max(train_y.eval()) - min(train_y.eval()) + 1
# man-made task label for testing
n_outs = [num_of_classes, num_of_classes, num_of_classes]
n_vis = train_x.get_value(borrow=True).shape[1]
n_train_batches = train_x.get_value(borrow=True).shape[0] / batch_size
print '... building the model'
dbn = DBN(n_ins=n_vis, n_outs=n_outs, dropout=dropout,
hidden_layers_sizes=model_structure, isPCD=isPCD)
print '... getting the pretraining functions'
pretrain_fns = dbn.get_pretrain_fns(train_set_x=train_x,
batch_size=batch_size,
k=k)
#################################
# Training the DBN #
#################################
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
plotting_time = 0.
start_time = time.clock()
import PIL.Image
from visualizer import tile_raster_images
# go through training epochs
for epoch in xrange(training_epochs):
# go through the training set
mean_cost = []
for i in xrange(dbn.n_layers):
# go through pretraining epochs
for batch_index in xrange(n_train_batches):
# for each batch, we extract the gibbs chain
new_cost = pretrain_fns[i](index=batch_index, lr=learning_rate)
mean_cost += [new_cost]
if plot_weight:
# plot all weights
weight = dbn.project_back(dbn.sigmoid_layers[i].W.T,
from_layer_i=i)
# Construct image from the weight matrix
image = PIL.Image.fromarray(tile_raster_images(
X=weight,
img_shape=(dbn.input_length, dbn.input_height), tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('pretrain_epoch_%i_layer_%i.png' % (epoch, i))
print 'Training epoch %d, cost is ' % epoch, numpy.mean(mean_cost)
end_time = time.clock()
pretraining_time = end_time - start_time
print ('Training took %f minutes' % (pretraining_time / 60.))
########################
# FINETUNING THE MODEL #
########################
# get the training, validation and testing function for the model
print '... getting the finetuning functions'
# man-made labels, list of labels
train_ys = [train_y, train_y, train_y]
finetune_fn = dbn.get_finetune_fn(train_x, train_ys, batch_size)
print '... finetunning the model'
test_score = 0.
start_time = time.clock()
# go through the each batch
for epoch in xrange(finetune_epoch):
## Pre-train layer-wise
c = []
for batch_index in xrange(n_train_batches):
c.append(finetune_fn(index=batch_index, lr=finetune_lr))
if plot_weight:
# plot all weights
weight = dbn.project_back(dbn.sigmoid_layers[-1].W.T,
from_layer_i=dbn.n_layers - 1)
# Construct image from the weight matrix
image = PIL.Image.fromarray(tile_raster_images(
X=weight,
img_shape=(dbn.input_length, dbn.input_height), tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('finetune_epoch_%i.png' % (epoch))
print 'Fine-tuning epoch %i, model cost ' % epoch,
print numpy.mean(c)
test_score = numpy.mean(c)
end_time = time.clock()
print(('Optimization complete with best test performance %f %%') %
(100. - test_score * 100.))
print >> sys.stderr, ('The fine tuning code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time)
/ 60.))
########################
# Test THE MODEL #
########################
task_id = 0
print '\nTest complete with error rate ' + str(dbn.get_error_rate(test_x, test_y, task_id=task_id)) + \
' for task_' + str(task_id)
