def channel_scaling_checker(num_examples, mode, num_batches, batch_size):
num_features = 2
monitor = Monitor(DummyModel(num_features))
dataset = DummyDataset(num_examples, num_features)
monitor.add_dataset(dataset=dataset, mode=mode,
num_batches=num_batches, batch_size=batch_size)
vis_batch = T.matrix()
mean = vis_batch.mean()
data_specs = (monitor.model.get_input_space(),
monitor.model.get_input_source())
monitor.add_channel(name='mean', ipt=vis_batch, val=mean, dataset=dataset,
data_specs=data_specs)
monitor()
assert 'mean' in monitor.channels
mean = monitor.channels['mean']
assert len(mean.val_record) == 1
actual = mean.val_record[0]
X = dataset.get_design_matrix()
if batch_size is not None and num_batches is not None:
total = min(num_examples, num_batches * batch_size)
else:
total = num_examples
expected = X[:total].mean()
if not np.allclose(expected, actual):
raise AssertionError("Expected monitor to contain %f but it has "
"%f" % (expected, actual))
def channel_scaling_checker(num_examples, mode, num_batches, batch_size):
num_features = 2
monitor = Monitor(DummyModel(num_features))
dataset = DummyDataset(num_examples, num_features)
try:
monitor.add_dataset(dataset=dataset, mode=mode,
num_batches=num_batches, batch_size=batch_size)
except NotImplementedError:
# make sure this was due to the unimplemented batch_size case
if num_batches is None:
assert num_examples % batch_size != 0
else:
assert num_examples % num_batches != 0
raise SkipTest()
vis_batch = T.matrix()
mean = vis_batch.mean()
monitor.add_channel(name='mean', ipt=vis_batch, val=mean, dataset=dataset)
monitor()
assert 'mean' in monitor.channels
mean = monitor.channels['mean']
assert len(mean.val_record) == 1
actual = mean.val_record[0]
X = dataset.get_design_matrix()
if batch_size is not None and num_batches is not None:
total = min(num_examples, num_batches * batch_size)
else:
total = num_examples
expected = X[:total].mean()
if not np.allclose(expected, actual):
raise AssertionError("Expected monitor to contain %f but it has "
"%f" % (expected, actual))
def test_tagging():
"""Test the tagging functionality of this extension."""
try:
# TODO: serial.save should be able to take an open file-like object so
# we can direct its output to a StringIO or something and not need to
# screw around like this in tests that don't actually need to touch
# the filesystem. /dev/null would work but the test would fail on
# Windows.
fd, fn = tempfile.mkstemp(suffix='.pkl')
os.close(fd)
# Test that the default key gets created.
def_model = MockModel()
def_model.monitor = Monitor(def_model)
def_ext = MonitorBasedSaveBest(channel_name='foobar', save_path=fn)
def_ext.setup(def_model, None, None)
assert 'MonitorBasedSaveBest' in def_model.tag
# Test with a custom key.
model = MockModel()
model.monitor = Monitor(model)
model.monitor.channels['foobar'] = MockChannel()
ext = MonitorBasedSaveBest(channel_name='foobar', tag_key='test123',
save_path=fn)
# Best cost is initially infinity.
ext.setup(model, None, None)
assert model.tag['test123']['best_cost'] == float("inf")
# Best cost after one iteration.
model.monitor.channels['foobar'].val_record.append(5.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test123']['best_cost'] == 5.0
# Best cost after a second, worse iteration.
model.monitor.channels['foobar'].val_record.append(7.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test123']['best_cost'] == 5.0
# Best cost after a third iteration better than 2 but worse than 1.
model.monitor.channels['foobar'].val_record.append(6.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test123']['best_cost'] == 5.0
# Best cost after a fourth, better iteration.
model.monitor.channels['foobar'].val_record.append(3.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test123']['best_cost'] == 3.0
# setting the starting epoch of saving models.
model = MockModel()
model.monitor = Monitor(model)
model.monitor.channels['foobar'] = MockChannel()
ext = MonitorBasedSaveBest(channel_name='foobar', tag_key='test12',
start_epoch=4, save_path=fn)
ext.setup(model, None, None)
assert model.tag['test12']['best_cost'] == float("inf")
# Best cost after one iteration.
model.monitor.channels['foobar'].val_record.append(5.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test12']['best_cost'] == float("inf")
# Best cost after a second, better iteration.
model.monitor.channels['foobar'].val_record.append(3.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test12']['best_cost'] == float("inf")
# Best cost after a third, worse iteration.
model.monitor.channels['foobar'].val_record.append(7.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test12']['best_cost'] == float("inf")
# Best cost after a fourth, worse iteration.
model.monitor.channels['foobar'].val_record.append(7.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test12']['best_cost'] == 7.0
# Best cost after a fifth, worse iteration.
model.monitor.channels['foobar'].val_record.append(10.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test12']['best_cost'] == 7.0
# Best cost after a fifth, better iteration.
model.monitor.channels['foobar'].val_record.append(1.0)
model.monitor.report_epoch()
ext.on_monitor(model, None, None)
assert model.tag['test12']['best_cost'] == 1.0
finally:
os.remove(fn)
return output
if __name__ == '__main__':
trainset, testset = GalaxyZoo.gzdeepdata.get_data()
# build and train classifiers for submodels
if SUBMODEL == 1:
iters = 100
if os.path.exists(DATA_DIR + 'model1saved_conv.pkl'):
model = serial.load(DATA_DIR + 'model1saved_conv.pkl')
iters = 1
# reset monitor, can't be re-used
model.monitor = Monitor(model)
else:
model = get_conv2([64, 64, 3])
get_trainer1(model, trainset, iters).main_loop()
elif SUBMODEL == 2:
iters = 50
if os.path.exists(DATA_DIR + 'model2saved_conv.pkl'):
model = serial.load(DATA_DIR + 'model2saved_conv.pkl')
iters = 1
# reset monitor, can't be re-used
model.monitor = Monitor(model)
else:
model = get_conv2([64, 64, 3])
get_trainer2(model, trainset, iters).main_loop()
outtrainset = get_output(model, trainset)
# pretrainer = get_ae_pretrainer(layer, utraindata, bsize, epochs=30)
# pretrainer.main_loop()
# serial.save(DATA_DIR+'daex_pretrained.pkl', stack)
stack = serial.load(DATA_DIR + 'daex_pretrained.pkl')
# construct DBN
dbn = construct_dbn_from_stack(stack)
# finetune softmax layer a bit
finetuner = get_finetuner(dbn, trainset, bsize, epochs=15)
finetuner.main_loop()
# now finetune layer-by-layer, boost earlier layers
lrs = [8., 6., 4., 2., 1.]
for ii, lr in zip(range(len(structure) - 1), lrs):
dbn.monitor = Monitor(dbn)
# set lr to boosted value for current layer
dbn.layers[ii].W_lr_scale = lr
finetuner = get_finetuner(dbn, trainset, bsize, epochs=30)
finetuner.main_loop()
# return to original lr
dbn.layers[ii].W_lr_scale = 1.
# total finetuner
# dbn = serial.load(DATA_DIR + 'model' + str(SUBMODEL) + 'saved_daex.pkl')
dbn.monitor = Monitor(dbn)
finetuner = get_finetuner(dbn, trainset, bsize, epochs=150)
finetuner.main_loop()
