#desc += '-cond'; dataset.max_label_size = 'full' # conditioned on full label
#desc += '-cond1'; dataset.max_label_size = 1 # conditioned on first component of the label
#desc += '-g4k'; grid.size = '4k'
#desc += '-grpc'; grid.layout = 'row_per_class'
# Config presets (choose one).
desc += '-preset-v1-1gpu'
submit_config.num_gpus = 1
D.mbstd_group_size = 16
sched.minibatch_base = 16
sched.minibatch_dict = {
256: 14,
512: 6,
1024: 3
}
sched.lod_training_kimg = 800
sched.lod_transition_kimg = 800
train.total_kimg = 19000
#desc += '-preset-v2-1gpu'; submit_config.num_gpus = 1; sched.minibatch_base = 4; sched.minibatch_dict = {4: 128, 8: 128, 16: 128, 32: 64, 64: 32, 128: 16, 256: 8, 512: 4}; sched.G_lrate_dict = {1024: 0.0015}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
#desc += '-preset-v2-2gpus'; submit_config.num_gpus = 2; sched.minibatch_base = 8; sched.minibatch_dict = {4: 256, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16, 256: 8}; sched.G_lrate_dict = {512: 0.0015, 1024: 0.002}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
#desc += '-preset-v2-4gpus'; submit_config.num_gpus = 4; sched.minibatch_base = 16; sched.minibatch_dict = {4: 512, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16}; sched.G_lrate_dict = {256: 0.0015, 512: 0.002, 1024: 0.003}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
#desc += '-preset-v2-8gpus'; submit_config.num_gpus = 8; sched.minibatch_base = 32; sched.minibatch_dict = {4: 512, 8: 256, 16: 128, 32: 64, 64: 32}; sched.G_lrate_dict = {128: 0.0015, 256: 0.002, 512: 0.003, 1024: 0.003}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
# Numerical precision (choose one).
desc += '-fp32'
sched.max_minibatch_per_gpu = {
256: 16,
512: 8,
1024: 4
}
#desc += '-fp16'; G.dtype = 'float16'; D.dtype = 'float16'; G.pixelnorm_epsilon=1e-4; G_opt.use_loss_scaling = True; D_opt.use_loss_scaling = True; sched.max_minibatch_per_gpu = {512: 16, 1024: 8}
64: 32,
128: 16,
256: 8,
512: 4
}
#desc += '-2gpu'; submit_config.num_gpus = 2; sched.minibatch_base = 8; sched.minibatch_dict = {4: 256, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16, 256: 8}
#desc += '-4gpu'; submit_config.num_gpus = 4; sched.minibatch_base = 16; sched.minibatch_dict = {4: 512, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16}
#desc += '-8gpu'; submit_config.num_gpus = 8; sched.minibatch_base = 32; sched.minibatch_dict = {4: 512, 8: 256, 16: 128, 32: 64, 64: 32}
# Default options.
train_kimg = 120 # Should be enough in most cases.
train.total_kimg = train_kimg
train.resume_run_id = "models/2019-02-26-stylegan-faces-network-02048-016041.pkl"
train.network_snapshot_ticks = 1 # Save more frequently.
sched.lod_initial_resolution = 512 # Train from the highest resolution.
sched.lod_training_kimg = train_kimg
sched.lod_transition_kimg = train_kimg
sched.G_lrate_dict = {128: 0.0015, 256: 0.002, 512: 0.003, 1024: 0.003}
sched.D_lrate_dict = EasyDict(sched.G_lrate_dict)
sched.tick_kimg_dict = {
4: 160,
8: 140,
16: 120,
32: 100,
64: 80,
128: 60,
256: 40,
#512:30,
512: 2, # 2k img / tick
1024: 20
}
def run(dataset, train_dir, config, d_aug, diffaug_policy, cond, ops, jpg_data, mirror, mirror_v, \
lod_step_kimg, batch_size, resume, resume_kimg, finetune, num_gpus, ema_kimg, gamma, freezeD):
# training functions
if d_aug: # https://github.com/mit-han-lab/data-efficient-gans
train = EasyDict(
run_func_name='training.training_loop_diffaug.training_loop'
) # Options for training loop (Diff Augment method)
loss_args = EasyDict(
func_name='training.loss_diffaug.ns_DiffAugment_r1',
policy=diffaug_policy) # Options for loss (Diff Augment method)
else: # original nvidia
train = EasyDict(run_func_name='training.training_loop.training_loop'
) # Options for training loop (original from NVidia)
G_loss = EasyDict(func_name='training.loss.G_logistic_ns_pathreg'
) # Options for generator loss.
D_loss = EasyDict(func_name='training.loss.D_logistic_r1'
) # Options for discriminator loss.
# network functions
G = EasyDict(func_name='training.networks_stylegan2.G_main'
) # Options for generator network.
D = EasyDict(func_name='training.networks_stylegan2.D_stylegan2'
) # Options for discriminator network.
G_opt = EasyDict(beta1=0.0, beta2=0.99,
epsilon=1e-8) # Options for generator optimizer.
D_opt = EasyDict(beta1=0.0, beta2=0.99,
epsilon=1e-8) # Options for discriminator optimizer.
sched = EasyDict() # Options for TrainingSchedule.
grid = EasyDict(
size='1080p',
layout='random') # Options for setup_snapshot_image_grid().
sc = dnnlib.SubmitConfig() # Options for dnnlib.submit_run().
tf_config = {'rnd.np_random_seed': 1000} # Options for tflib.init_tf().
G.impl = D.impl = ops
# dataset (tfrecords) - get or create
tfr_files = file_list(os.path.dirname(dataset), 'tfr')
tfr_files = [
f for f in tfr_files if basename(f).split('-')[0] == basename(dataset)
]
if len(tfr_files) == 0 or os.stat(tfr_files[0]).st_size == 0:
tfr_file, total_samples = create_from_images(dataset, jpg=jpg_data)
else:
tfr_file = tfr_files[0]
dataset_args = EasyDict(tfrecord=tfr_file, jpg_data=jpg_data)
# resolutions
with tf.Graph().as_default(), tflib.create_session().as_default(): # pylint: disable=not-context-manager
dataset_obj = _dataset.load_dataset(
**dataset_args) # loading the data to see what comes out
resolution = dataset_obj.resolution
init_res = dataset_obj.init_res
res_log2 = dataset_obj.res_log2
dataset_obj.close()
dataset_obj = None
if list(init_res) == [4, 4]:
desc = '%s-%d' % (basename(dataset), resolution)
else:
print(' custom init resolution', init_res)
desc = basename(tfr_file)
G.init_res = D.init_res = list(init_res)
train.savenames = [desc.replace(basename(dataset), 'snapshot'), desc]
desc += '-%s' % config
# training schedule
sched.lod_training_kimg = lod_step_kimg
sched.lod_transition_kimg = lod_step_kimg
sched.tick_kimg_base = 2 # if finetune else 0.2
train.total_kimg = lod_step_kimg * res_log2 * 3 # 1.5 * ProGAN
train.image_snapshot_ticks = 1
train.network_snapshot_ticks = 5
train.mirror_augment = mirror
train.mirror_augment_v = mirror_v
# learning rate
if config == 'e':
if finetune: # uptrain 1024
sched.G_lrate_base = 0.001
sched.lrate_step = 150 # period for stepping to next lrate, in kimg
else: # train 1024
sched.G_lrate_base = 0.001
sched.G_lrate_dict = {0: 0.001, 1: 0.0007, 2: 0.0005, 3: 0.0003}
sched.lrate_step = 1500 # period for stepping to next lrate, in kimg
if config == 'f':
sched.G_lrate_base = 0.001 # 0.0003 for few-shot datasets
sched.D_lrate_base = sched.G_lrate_base # *2 - not used anyway
sched.minibatch_gpu_base = batch_size
sched.minibatch_size_base = num_gpus * sched.minibatch_gpu_base
sc.num_gpus = num_gpus
if config == 'e':
G.fmap_base = D.fmap_base = 8 << 10
if d_aug: loss_args.gamma = 100 if gamma is None else gamma
else: D_loss.gamma = 100 if gamma is None else gamma
elif config == 'f':
G.fmap_base = D.fmap_base = 16 << 10
else:
print(' Only configs E and F are implemented')
exit()
if cond:
desc += '-cond'
dataset_args.max_label_size = 'full' # conditioned on full label
if freezeD:
D.freezeD = True
train.resume_with_new_nets = True
if d_aug:
desc += '-daug'
sc.submit_target = dnnlib.SubmitTarget.LOCAL
sc.local.do_not_copy_source_files = True
kwargs = EasyDict(train)
kwargs.update(G_args=G, D_args=D, G_opt_args=G_opt, D_opt_args=D_opt)
kwargs.update(dataset_args=dataset_args,
sched_args=sched,
grid_args=grid,
tf_config=tf_config)
kwargs.update(resume_pkl=resume,
resume_kimg=resume_kimg,
resume_with_new_nets=True)
if ema_kimg is not None:
kwargs.update(G_ema_kimg=ema_kimg)
if d_aug:
kwargs.update(loss_args=loss_args)
else:
kwargs.update(G_loss_args=G_loss, D_loss_args=D_loss)
kwargs.submit_config = copy.deepcopy(sc)
kwargs.submit_config.run_dir_root = train_dir
kwargs.submit_config.run_desc = desc
dnnlib.submit_run(**kwargs)
def run(dataset, train_dir, config, d_aug, diffaug_policy, cond, ops, jpg_data, mirror, mirror_v, \
lod_step_kimg, batch_size, resume, resume_kimg, finetune, num_gpus, ema_kimg, gamma, freezeD):
# dataset (tfrecords) - preprocess or get
tfr_files = file_list(os.path.dirname(dataset), 'tfr')
tfr_files = [f for f in tfr_files if basename(dataset) in f]
if len(tfr_files) == 0:
tfr_file, total_samples = create_from_images(dataset, jpg=jpg_data)
else:
tfr_file = tfr_files[0]
dataset_args = EasyDict(tfrecord=tfr_file, jpg_data=jpg_data)
desc = basename(tfr_file).split('-')[0]
# training functions
if d_aug: # https://github.com/mit-han-lab/data-efficient-gans
train = EasyDict(
run_func_name='training.training_loop_diffaug.training_loop'
) # Options for training loop (Diff Augment method)
loss_args = EasyDict(
func_name='training.loss_diffaug.ns_DiffAugment_r1',
policy=diffaug_policy) # Options for loss (Diff Augment method)
else: # original nvidia
train = EasyDict(run_func_name='training.training_loop.training_loop'
) # Options for training loop (original from NVidia)
G_loss = EasyDict(func_name='training.loss.G_logistic_ns_pathreg'
) # Options for generator loss.
D_loss = EasyDict(func_name='training.loss.D_logistic_r1'
) # Options for discriminator loss.
# network functions
G = EasyDict(func_name='training.networks_stylegan2.G_main'
) # Options for generator network.
D = EasyDict(func_name='training.networks_stylegan2.D_stylegan2'
) # Options for discriminator network.
G_opt = EasyDict(beta1=0.0, beta2=0.99,
epsilon=1e-8) # Options for generator optimizer.
D_opt = EasyDict(beta1=0.0, beta2=0.99,
epsilon=1e-8) # Options for discriminator optimizer.
sched = EasyDict() # Options for TrainingSchedule.
grid = EasyDict(
size='1080p',
layout='random') # Options for setup_snapshot_image_grid().
sc = dnnlib.SubmitConfig() # Options for dnnlib.submit_run().
tf_config = {'rnd.np_random_seed': 1000} # Options for tflib.init_tf().
G.impl = D.impl = ops
# resolutions
data_res = basename(tfr_file).split('-')[-1].split(
'x') # get resolution from dataset filename
data_res = list(reversed([int(x)
for x in data_res])) # convert to int list
init_res, resolution, res_log2 = calc_init_res(data_res)
if init_res != [4, 4]:
print(' custom init resolution', init_res)
G.init_res = D.init_res = list(init_res)
train.setname = desc + config
desc = '%s-%d-%s' % (desc, resolution, config)
# training schedule
sched.lod_training_kimg = lod_step_kimg
sched.lod_transition_kimg = lod_step_kimg
train.total_kimg = lod_step_kimg * res_log2 * 2 # a la ProGAN
if finetune:
train.total_kimg = 15000 # should start from ~10k kimg
train.image_snapshot_ticks = 1
train.network_snapshot_ticks = 5
train.mirror_augment = mirror
train.mirror_augment_v = mirror_v
# learning rate
if config == 'e':
if finetune: # uptrain 1024
sched.G_lrate_base = 0.001
else: # train 1024
sched.G_lrate_base = 0.001
sched.G_lrate_dict = {0: 0.001, 1: 0.0007, 2: 0.0005, 3: 0.0003}
sched.lrate_step = 1500 # period for stepping to next lrate, in kimg
if config == 'f':
# sched.G_lrate_base = 0.0003
sched.G_lrate_base = 0.001
sched.D_lrate_base = sched.G_lrate_base # *2 - not used anyway
sched.minibatch_gpu_base = batch_size
sched.minibatch_size_base = num_gpus * sched.minibatch_gpu_base
sc.num_gpus = num_gpus
if config == 'e':
G.fmap_base = D.fmap_base = 8 << 10
if d_aug: loss_args.gamma = 100 if gamma is None else gamma
else: D_loss.gamma = 100 if gamma is None else gamma
elif config == 'f':
G.fmap_base = D.fmap_base = 16 << 10
else:
print(' Only configs E and F are implemented')
exit()
if cond:
desc += '-cond'
dataset_args.max_label_size = 'full' # conditioned on full label
if freezeD:
D.freezeD = True
train.resume_with_new_nets = True
if d_aug:
desc += '-daug'
sc.submit_target = dnnlib.SubmitTarget.LOCAL
sc.local.do_not_copy_source_files = True
kwargs = EasyDict(train)
kwargs.update(G_args=G, D_args=D, G_opt_args=G_opt, D_opt_args=D_opt)
kwargs.update(dataset_args=dataset_args,
sched_args=sched,
grid_args=grid,
tf_config=tf_config)
kwargs.update(resume_pkl=resume,
resume_kimg=resume_kimg,
resume_with_new_nets=True)
if ema_kimg is not None:
kwargs.update(G_ema_kimg=ema_kimg)
if d_aug:
kwargs.update(loss_args=loss_args)
else:
kwargs.update(G_loss_args=G_loss, D_loss_args=D_loss)
kwargs.submit_config = copy.deepcopy(sc)
kwargs.submit_config.run_dir_root = train_dir
kwargs.submit_config.run_desc = desc
dnnlib.submit_run(**kwargs)
#desc += '-lsun-restaurant'; dataset = EasyDict(tfrecord_dir='lsun-restaurant-100k'); train.mirror_augment = True
#desc += '-lsun-sheep'; dataset = EasyDict(tfrecord_dir='lsun-sheep-100k'); train.mirror_augment = True
#desc += '-lsun-sofa'; dataset = EasyDict(tfrecord_dir='lsun-sofa-100k'); train.mirror_augment = True
#desc += '-lsun-tower'; dataset = EasyDict(tfrecord_dir='lsun-tower-100k'); train.mirror_augment = True
#desc += '-lsun-train'; dataset = EasyDict(tfrecord_dir='lsun-train-100k'); train.mirror_augment = True
#desc += '-lsun-tvmonitor'; dataset = EasyDict(tfrecord_dir='lsun-tvmonitor-100k'); train.mirror_augment = True
desc += '-test'; dataset = EasyDict(tfrecord_dir='test-2k', resolution=256); train.mirror_augment = True
# Conditioning & snapshot options.
#desc += '-cond'; dataset.max_label_size = 'full' # conditioned on full label
#desc += '-cond1'; dataset.max_label_size = 1 # conditioned on first component of the label
#desc += '-g4k'; grid.size = '4k'
#desc += '-grpc'; grid.layout = 'row_per_class'
# Config presets (choose one).
desc += '-preset-v1-1gpu'; submit_config.num_gpus = 1; D.mbstd_group_size = 16; sched.minibatch_base = 16; sched.minibatch_dict = {256: 14, 512: 6, 1024: 3}; sched.lod_training_kimg = 800; sched.lod_transition_kimg = 800; train.total_kimg = 19000
#desc += '-preset-v2-1gpu'; submit_config.num_gpus = 1; sched.minibatch_base = 4; sched.minibatch_dict = {4: 128, 8: 128, 16: 128, 32: 64, 64: 32, 128: 16, 256: 8, 512: 4}; sched.G_lrate_dict = {1024: 0.0015}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
#desc += '-preset-v2-2gpus'; submit_config.num_gpus = 2; sched.minibatch_base = 8; sched.minibatch_dict = {4: 256, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16, 256: 8}; sched.G_lrate_dict = {512: 0.0015, 1024: 0.002}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
#desc += '-preset-v2-4gpus'; submit_config.num_gpus = 4; sched.minibatch_base = 16; sched.minibatch_dict = {4: 512, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16}; sched.G_lrate_dict = {256: 0.0015, 512: 0.002, 1024: 0.003}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
#desc += '-preset-v2-8gpus'; submit_config.num_gpus = 8; sched.minibatch_base = 32; sched.minibatch_dict = {4: 512, 8: 256, 16: 128, 32: 64, 64: 32}; sched.G_lrate_dict = {128: 0.0015, 256: 0.002, 512: 0.003, 1024: 0.003}; sched.D_lrate_dict = EasyDict(sched.G_lrate_dict); train.total_kimg = 12000
# Numerical precision (choose one).
desc += '-fp32'; sched.max_minibatch_per_gpu = {256: 16, 512: 8, 1024: 4}
#desc += '-fp16'; G.dtype = 'float16'; D.dtype = 'float16'; G.pixelnorm_epsilon=1e-4; G_opt.use_loss_scaling = True; D_opt.use_loss_scaling = True; sched.max_minibatch_per_gpu = {512: 16, 1024: 8}
# Disable individual features.
#desc += '-nogrowing'; sched.lod_initial_resolution = 1024; sched.lod_training_kimg = 0; sched.lod_transition_kimg = 0; train.total_kimg = 10000
#desc += '-nopixelnorm'; G.use_pixelnorm = False
#desc += '-nowscale'; G.use_wscale = False; D.use_wscale = False
#desc += '-noleakyrelu'; G.use_leakyrelu = False
#desc += '-nosmoothing'; train.G_smoothing_kimg = 0.0