def test_weightsHistogram():
"""The histogram of the weights of a RBM is
an array of pairs."""
net = RBM(16, 23)
hist = net.weightsHistogram()
for h in hist:
assert len(h) == 2
def test_load():
"""A DBN can be loaded from a Pickle file."""
net = DBN([RBM(4, 5), RBM(5, 6)], 'Test')
net.save()
net = DBN.load('Test')
assert len(net) == 2
os.remove(nets.full('Test.pkl'))
def test_generationLength():
"""A DBN can generate a sample whose length
is equal to the number of visible units of
its first RBM."""
net = DBN([RBM(4, 7), RBM(7, 3)])
sample = net.evaluate([0, 1, 1, 0.4])
assert len(sample) == 4
def test_saveAndLoad():
"""A DBN serialised to a file is equivalent to
a DBN loaded from the same file."""
net_1 = DBN([RBM(4, 5), RBM(5, 6)], 'Test')
net_1.save()
net_2 = DBN.load('Test')
assert [np.equal(net_1[i].W, net_2[i].W) for i in range(2)]
os.remove(nets.full('Test.pkl'))
def test_receptiveField():
"""The receptive fields of a DBN have a shape whose
area is equal to the number of visible units."""
net = DBN([RBM(49, 15), RBM(15, 26)])
rc = net.receptiveField(1, 10)
assert rc.shape == (49, )
rc = net.receptiveField(2, 10)
assert rc.shape == (49, )
def test_DBN_error():
"""The reconstruction error of a DBN is less than 1."""
net = DBN([RBM(16, 15), RBM(15, 10)], 'Test')
trainset = DataSet.fromWhatever('top_left')
for train_info in net.learn(trainset):
pass
mean_err = np.sqrt(((trainset[0] - net.evaluate(trainset[0]))**2).mean())
assert mean_err <= 1
def test_RBM_error():
"""The reconstruction error of a RBM is less than 1."""
net = RBM(16, 15)
trainset = DataSet.fromWhatever('top_left')
trainer = CDTrainer(net)
for err in trainer.run(trainset):
pass
mean_err = np.sqrt(((trainset[0] - net.evaluate(trainset[0]))**2).mean())
assert mean_err <= 1
def test_RBM_learn():
"""A RBM can learn to generate samples from a dataset."""
net = RBM(16, 15)
trainset = DataSet.fromWhatever('top_left')
trainer = CDTrainer(net)
for err in trainer.run(trainset):
pass
def test_DBN_learn():
"""A DBN can learn to generate samples from a dataset."""
net = DBN([RBM(16, 15), RBM(15, 10)], 'Test')
trainset = DataSet.fromWhatever('top_left')
for train_info in net.learn(trainset):
pass
def test_init():
"""A DBN can be constructed from a list of RBMs."""
net = DBN([RBM(4, 7), RBM(7, 3)], 'Test')
assert len(net) == 2
def test_append():
"""A RBM can be appended to a DBN."""
net = DBN()
net.append(RBM(3, 4))
net.append(RBM(4, 7))
assert len(net) == 2
def train(request):
"""Set up a network to be trained according to
the parameters in the HTTP request."""
print(request.POST['pass'])
if not authorized(request.POST['pass']):
return HttpResponse(status = 401)
trainset_name = request.POST['dataset']
trainset = DataSet.fromWhatever(trainset_name)
try:
num_layers = 1 + int(request.POST['num_hid_layers'])
except ValueError:
num_layers = 1
# return HttpResponse({'error': 'you haven\'t specified [...]'})
std_dev = float(request.POST['std_dev'])
net = DBN(name = trainset_name)
vis_size = int(request.POST['vis_sz'])
for layer in range(1, num_layers):
hid_size = int(request.POST['hid_sz_' + str(layer)])
print('creating a', vis_size, 'x', hid_size, 'RBM...')
net.append(RBM(vis_size, hid_size, std_dev = std_dev))
vis_size = hid_size # for constructing the next RBM
epochs = request.POST['epochs']
config = {
'max_epochs' : int(epochs if (epochs != 'inf') else maxsize),
'batch_size' : int(request.POST['batch_size']),
'learn_rate' : float(request.POST['learn_rate']),
'momentum' : float(request.POST['momentum']),
'std_dev' : std_dev,
'spars_target' : float(request.POST['spars_target'])
}
# sanity check for batch size:
if len(trainset) % config['batch_size'] != 0:
print('encountered batch size', config['batch_size'], 'for dataset with', len(trainset), 'examples: adjusting batch size to', end = ' ')
while len(trainset) % config['batch_size'] != 0:
config['batch_size'] -= 1
print(config['batch_size'])
random_id = ''.join(random.choice(string.ascii_uppercase + string.digits) for i in range(10))
training_jobs[random_id] = {
'birthday': time(),
'network': net,
'generator': net.learn(trainset, Configuration(**config))
}
# delete the old client job that is being replaced (if any):
last_job = request.POST['last_job_id']
if last_job in training_jobs:
del training_jobs[last_job]
# delete a random pending job older than five hours:
random_old_job = random.choice(list(training_jobs.keys()))
if time() - training_jobs[random_old_job]['birthday'] > 18000:
print('deleting old job n.', random_old_job)
del training_jobs[random_old_job] # risky...
return HttpResponse(random_id)