def create_celeba_streams(training_batch_size, monitoring_batch_size,
include_targets=False):
"""Creates CelebA data streams.
Parameters
----------
training_batch_size : int
Batch size for training.
monitoring_batch_size : int
Batch size for monitoring.
include_targets : bool
If ``True``, use both features and targets. If ``False``, use
features only.
Returns
-------
rval : tuple of data streams
Data streams for the main loop, the training set monitor,
the validation set monitor and the test set monitor.
"""
sources = ('features', 'targets') if include_targets else ('features',)
train_set = CelebA('64', ('train',), sources=sources)
valid_set = CelebA('64', ('valid',), sources=sources)
test_set = CelebA('64', ('test',), sources=sources)
return create_streams(train_set, valid_set, test_set, training_batch_size,
monitoring_batch_size)
def create_celeba_data_streams(batch_size, monitoring_batch_size, rng=None):
train_set = CelebA('64', ('train', ), sources=('features', ))
valid_set = CelebA('64', ('valid', ), sources=('features', ))
main_loop_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
batch_size,
rng=rng))
train_monitor_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(5000, monitoring_batch_size, rng=rng))
valid_monitor_stream = DataStream.default_stream(
valid_set,
iteration_scheme=ShuffledScheme(5000, monitoring_batch_size, rng=rng))
return main_loop_stream, train_monitor_stream, valid_monitor_stream
def test_celeba():
data_path = config.data_path
try:
config.data_path = '.'
f = h5py.File('celeba_64.hdf5', 'w')
f['features'] = numpy.arange(10 * 3 * 64 * 64, dtype='uint8').reshape(
(10, 3, 64, 64))
f['targets'] = numpy.arange(10 * 40, dtype='uint8').reshape((10, 40))
split_dict = {
'train': {
'features': (0, 6),
'targets': (0, 6)
},
'valid': {
'features': (6, 8),
'targets': (6, 8)
},
'test': {
'features': (8, 10),
'targets': (8, 10)
}
}
f.attrs['split'] = H5PYDataset.create_split_array(split_dict)
f.close()
dataset = CelebA(which_format='64', which_sets=('train', ))
assert_equal(dataset.filename, 'celeba_64.hdf5')
finally:
config.data_path = data_path
os.remove('celeba_64.hdf5')
def main(args):
# Load Config Module from source file
config_module = imp.load_source('config', args.config_path)
# Get configuration parameters
cfg = config_module.cfg
# Define name of npz file to which the model parameters will be saved
weights_fname = str(args.config_path)[:-3]+'.npz'
model = config_module.get_model(interp=False)
print('Compiling theano functions...')
# Compile functions
tfuncs, tvars,model = make_training_functions(cfg,model)
# Test set for interpolations
test_set = CelebA('64',('test',),sources=('features',))
# Loop across epochs
offset = True
params = list(set(lasagne.layers.get_all_params(model['l_out'],trainable=True)+\
lasagne.layers.get_all_params(model['l_discrim'],trainable=True)+\
[x for x in lasagne.layers.get_all_params(model['l_out'])+\
lasagne.layers.get_all_params(model['l_discrim']) if x.name[-4:]=='mean' or x.name[-7:]=='inv_std']))
metadata = GANcheckpoints.load_weights(weights_fname, params)
epoch = args.epoch if args.epoch>0 else metadata['epoch'] if 'epoch' in metadata else 0
print('loading weights, epoch is '+str(epoch))
model['l_IAF_mu'].reset("Once")
model['l_IAF_ls'].reset("Once")
# Open Test Set
test_set.open()
np.random.seed(epoch*42+5)
# Generate Random Samples, averaging latent vectors across masks
samples = np.uint8(from_tanh(tfuncs['sample'](np.random.randn(27,cfg['num_latents']).astype(np.float32))))
np.random.seed(epoch*42+5)
# Get Reconstruction/Interpolation Endpoints
endpoints = np.uint8(test_set.get_data(request = list(np.random.choice(test_set.num_examples,6,replace=False)))[0])
# Get reconstruction latents
Ze = np.asarray(tfuncs['Zfn'](to_tanh(np.float32(endpoints))))
# Get Interpolant Latents
Z = np.asarray([Ze[2 * i, :] * (1 - j) + Ze[2 * i + 1, :] * j for i in range(3) for j in [x/6.0 for x in range(7)]],dtype=np.float32)
# Get all images
images = np.append(samples,np.concatenate([np.insert(endpoints[2*i:2*(i+1),:,:,:],1,np.uint8(from_tanh(tfuncs['sample'](Z[7*i:7*(i+1),:]))),axis=0) for i in range(3)],axis=0),axis=0)
# Plot images
plot_image_grid(images,6,9,'pics/'+str(args.config_path)[:-3]+'_sample'+str(epoch)+'.png')
# Close test set
test_set.close(state=None)
def get_celeba(split, sources, load_in_memory):
from fuel.datasets import CelebA
return CelebA('64', split, sources=sources, load_in_memory=load_in_memory)