if __name__=="__main__":
model_config_dict = OrderedDict()
model_config_dict['batch_size'] = 64
model_config_dict['num_display'] = 16*16
model_config_dict['hidden_distribution'] = 1.
model_config_dict['epochs'] = 200
is_training = True
is_continue = False
last_batch_idx = 0
if is_training is True:
#################
# LOAD DATA SET #
#################
_ , data_stream = faces(batch_size=model_config_dict['batch_size'])
expert_size_list = [1024]
hidden_size_list = [100]
num_filters_list = [128]
lr_list = [1e-3]
lambda_eng_list = [1e-5]
for lr in lr_list:
for num_filters in num_filters_list:
for hidden_size in hidden_size_list:
for expert_size in expert_size_list:
for lambda_eng in lambda_eng_list:
model_config_dict['hidden_size'] = hidden_size
model_config_dict['expert_size'] = expert_size
model_config_dict['min_num_gen_filters'] = num_filters
os.makedirs(model_dir)
if not os.path.exists(samples_dir):
os.makedirs(samples_dir)
f_log = open('logs/%s.ndjson'%desc, 'wb')
log_fields = [
'n_epochs',
'n_updates',
'n_examples',
'n_seconds',
'1k_va_nnd',
'10k_va_nnd',
'100k_va_nnd',
'g_cost',
'd_cost',
]
tr_data, te_data, tr_stream, val_stream, te_stream = faces(ntrain=ntrain) # Only tr_data/tr_stream are used.
tr_handle = tr_data.open()
vaX, = tr_data.get_data(tr_handle, slice(0, 10000))
vaX = transform(vaX)
vis_idxs = py_rng.sample(np.arange(len(vaX)), nvis)
vaX_vis = inverse_transform(vaX[vis_idxs])
color_grid_vis(vaX_vis, (14, 14), 'samples/%s_etl_test.png'%desc)
sample_zmb = floatX(np_rng.uniform(-1., 1., size=(nvis, nz)))
vaX = vaX.reshape(len(vaX), -1)
# DEFINE NETWORKS.
relu = activations.Rectify()
sigmoid = activations.Sigmoid()
lrelu = activations.LeakyRectify()
tanh = activations.Tanh()
bce = T.nnet.binary_crossentropy
if batch_count % 100 == 0:
# sample data
save_as = samples_dir + "/" + model_name + "_SAMPLES{}.png".format(batch_count)
sample_data = sampling_function(fixed_hidden_data)[0]
sample_data = np.asarray(sample_data)
color_grid_vis(inverse_transform(sample_data).transpose([0, 2, 3, 1]), (16, 16), save_as)
np.save(file=samples_dir + "/" + model_name + "_MOMENT_COST", arr=np.asarray(moment_cost_list))
np.save(file=samples_dir + "/" + model_name + "_VAE_COST", arr=np.asarray(vae_cost_list))
if __name__ == "__main__":
batch_size = 128
num_epochs = 100
_, data_stream = faces(batch_size=batch_size)
num_hiddens = 1024
learning_rate = 1e-4
l2_weight = 1e-5
optimizer = Adagrad(lr=sharedX(learning_rate), regularizer=Regularizer(l2=l2_weight))
model_test_name = (
model_name
+ "_HIDDEN{}".format(int(num_hiddens))
+ "_REG{}".format(int(-np.log10(l2_weight)))
+ "_LR{}".format(int(-np.log10(learning_rate)))
)
train_model(
model_name=model_test_name,
save_as = samples_dir + "/" + model_test_name + "_MODEL.pkl"
save_model(tensor_params_list=decoder_parameters, save_to=save_as)
if __name__ == "__main__":
model_config_dict = OrderedDict()
model_config_dict["batch_size"] = 12 * 12
model_config_dict["num_display"] = 16 * 16
model_config_dict["hidden_distribution"] = 1.0
model_config_dict["epochs"] = 200
#################
# LOAD DATA SET #
#################
_, data_stream = faces(batch_size=model_config_dict["batch_size"])
hidden_size_list = [1024]
num_filters_list = [128]
lr_list = [1e-4]
dropout_list = [False]
lambda_eng_list = [1e-10]
lambda_gen_list = [1e-10]
for lr in lr_list:
for num_filters in num_filters_list:
for hidden_size in hidden_size_list:
for lambda_eng in lambda_eng_list:
for lambda_gen in lambda_gen_list:
model_config_dict["hidden_size"] = hidden_size
model_config_dict["min_num_gen_filters"] = num_filters
l2 = 1e-5 # l2 weight decay
nvis = 196 # # of samples to visualize during training
b1 = 0.5 # momentum term of adam
nc = 3 # # of channels in image
nbatch = 128 # # of examples in batch
npx = 64 # # of pixels width/height of images
nz = 100 # # of dim for Z
ngf = 128 # # of gen filters in first conv layer
ndf = 128 # # of discrim filters in first conv layer
nx = npx*npx*nc # # of dimensions in X
niter = 25 # # of iter at starting learning rate
niter_decay = 0 # # of iter to linearly decay learning rate to zero
lr = 0.0002 # initial learning rate for adam
ntrain = 266251 # # of examples to train on
tr_data, te_data, tr_stream, val_stream, te_stream = faces(ntrain=ntrain)
tr_handle = tr_data.open()
vaX, = tr_data.get_data(tr_handle, slice(0, 10000))
vaX = transform(vaX)
desc = 'vc_dcgan'
model_dir = 'models/%s'%desc
samples_dir = 'samples/%s'%desc
if not os.path.exists('logs/'):
os.makedirs('logs/')
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if not os.path.exists(samples_dir):
os.makedirs(samples_dir)
l2 = 1e-5 # l2 weight decay
nvis = 196 # # of samples to visualize during training
b1 = 0.5 # momentum term of adam
nc = 3 # # of channels in image
nbatch = 128 # # of examples in batch
npx = 64 # # of pixels width/height of images
nz = 100 # # of dim for Z
ngf = 128 # # of gen filters in first conv layer
ndf = 128 # # of discrim filters in first conv layer
nx = npx * npx * nc # # of dimensions in X
niter = 25 # # of iter at starting learning rate
niter_decay = 0 # # of iter to linearly decay learning rate to zero
lr = 0.0002 # initial learning rate for adam
ntrain = 350000 # # of examples to train on
tr_data, te_data, tr_stream, val_stream, te_stream = faces(ntrain=ntrain)
tr_handle = tr_data.open()
vaX, = tr_data.get_data(tr_handle, slice(0, 10000))
vaX = transform(vaX)
desc = 'uncond_dcgan'
model_dir = 'models/%s' % desc
samples_dir = 'samples/%s' % desc
if not os.path.exists('logs/'):
os.makedirs('logs/')
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if not os.path.exists(samples_dir):
os.makedirs(samples_dir)