def run(**args):
args = EasyDict(args)
train = EasyDict(run_func_name = "training.training_loop.training_loop") # training loop options
sched = EasyDict() # TrainingSchedule options
vis = EasyDict() # visualize.eval() options
grid = EasyDict(size = "1080p", layout = "random") # setup_snapshot_img_grid() options
sc = dnnlib.SubmitConfig() # dnnlib.submit_run() options
# Environment configuration
tf_config = {
"rnd.np_random_seed": 1000,
"allow_soft_placement": True,
"gpu_options.per_process_gpu_memory_fraction": 1.0
}
if args.gpus != "":
num_gpus = len(args.gpus.split(","))
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
assert num_gpus in [1, 2, 4, 8]
sc.num_gpus = num_gpus
# Networks configuration
cG = set_net("G", reg_interval = 4)
cD = set_net("D", reg_interval = 16)
# Dataset configuration
ratios = {
"clevr": 0.75,
"lsun-bedrooms": 0.72,
"cityscapes": 0.5,
"ffhq": 1.0
}
args.ratio = ratios.get(args.dataset, args.ratio)
dataset_args = EasyDict(tfrecord_dir = args.dataset, max_imgs = args.max_images, ratio = args.ratio,
num_threads = args.num_threads)
for arg in ["data_dir", "mirror_augment", "total_kimg"]:
cset(train, arg, args[arg])
# Training and Optimizations configuration
for arg in ["eval", "train", "recompile", "last_snapshots"]:
cset(train, arg, args[arg])
# Round to the closest multiply of minibatch size for validity
args.batch_size -= args.batch_size % args.minibatch_size
args.minibatch_std_size -= args.minibatch_std_size % args.minibatch_size
args.latent_size -= args.latent_size % args.component_num
if args.latent_size == 0:
print(bcolored("Error: latent-size is too small. Must best a multiply of component-num.", "red"))
exit()
sched_args = {
"G_lrate": "g_lr",
"D_lrate": "d_lr",
"minibatch_size": "batch_size",
"minibatch_gpu": "minibatch_size"
}
for arg, cmd_arg in sched_args.items():
cset(sched, arg, args[cmd_arg])
cset(train, "clip", args.clip)
# Logging and metrics configuration
metrics = [metric_defaults[x] for x in args.metrics]
cset(cG.args, "truncation_psi", args.truncation_psi)
for arg in ["summarize", "keep_samples"]:
cset(train, arg, args[arg])
# Visualization
args.imgs = args.images
args.ltnts = args.latents
vis_types ["imgs", "ltnts", "maps", "layer_maps", "interpolations", "noise_var", "style_mix"]:
# Set of all the set visualization types option
vis.vis_types = {arg for arg in vis_types if args[arg]}
vis_args = {
"grid": "vis_grid" ,
"num": "vis_num" ,
"rich_num": "vis_rich_num",
"section_size": "vis_section_size",
"intrp_density": "intrpolation_density",
"intrp_per_component": "intrpolation_per_component",
"alpha": "blending_alpha"
}
for arg, cmd_arg in vis_args.items():
cset(vis, arg, args[cmd_arg])
# Networks architecture
cset(cG.args, "architecture", args.g_arch)
cset(cD.args, "architecture", args.d_arch)
cset([cG.args, cD.args], "resnet_mlp", args.resnet_mlp)
cset(cG.args, "tanh", args.tanh)
# Latent sizes
if args.component_num > 1
if not (args.attention or args.merge):
print(bcolored("Error: component-num > 1 but the model is not using components.", "red"))
print(bcolored("Either add --attention for GANsformer or --merge for k-GAN).", "red"))
exit()
args.latent_size = int(args.latent_size / args.component_num)
cD.args.latent_size = cG.args.latent_size = cG.args.dlatent_size = args.latent_size
cset([cG.args, cD.args, train, vis], "component_num", args.component_num)
# Mapping network
for arg in ["layersnum", "lrmul", "dim", "shared"]:
cset(cG.args, arg, args["mapping_{}".formt(arg)])
# StyleGAN settings
for arg in ["style", "latent_stem", "fused_modconv", "local_noise"]:
cset(cG.args, arg, args[arg])
cD.args.mbstd_group_size = args.minibatch_std_size
# GANsformer
cset([cG.args, train], "attention", args.transformer)
cset(cD.args, "attention", args.d_transformer)
cset([cG.args, cD.args], "num_heads", args.num_heads)
args.norm = args.normalize
for arg in ["norm", "integration", "ltnt_gate", "img_gate", "kmeans",
"kmeans_iters", "asgn_direct", "mapping_ltnt2ltnt"]:
cset(cG.args, arg, args[arg])
for arg in ["attention_inputs", "use_pos"]:
cset([cG.args, cD.args], arg, args[arg])
# Positional encoding
for arg in ["dim", "init", "directions_num"]:
field = "pos_{}".format(arg)
cset([cG.args, cD.args], field, args[field])
# k-GAN
for arg in ["layer", "type", "channelwise"]:
field = "merge_{}".format(arg)
cset(cG.args, field, args[field])
cset([cG.args, train], "merge", args.merge)
# Attention
for arg in ["start_res", "end_res", "ltnt2ltnt", "img2img", "local_attention"]:
cset(cG.args, arg, args["g_{}".format(arg)])
cset(cD.args, arg, args["d_{}".format(arg)])
cset(cG.args, "img2ltnt", args.g_img2ltnt)
cset(cD.args, "ltnt2img", args.d_ltnt2img)
# Mixing and dropout
for arg in ["style_mixing", "component_mixing", "component_dropout", "attention_dropout"]:
cset(cG.args, arg, args[arg])
# Loss and regularization
gloss_args = {
"loss_type": "g_loss",
"reg_weight": "g_reg_weight"
"pathreg": "pathreg",
}
dloss_args = {
"loss_type": "d_loss",
"reg_type": "d_reg",
"gamma": "gamma"
}
for arg, cmd_arg in gloss_args.items():
cset(cG.loss_args, arg, args[cmd_arg])
for arg, cmd_arg in dloss_args.items():
cset(cD.loss_args, arg, args[cmd_arg])
##### Experiments management:
# Whenever we start a new experiment we store its result in a directory named 'args.expname:000'.
# When we rerun a training or evaluation command it restores the model from that directory by default.
# If we wish to restart the model training, we can set --restart and then we will store data in a new
# directory: 'args.expname:001' after the first restart, then 'args.expname:002' after the second, etc.
# Find the latest directory that matches the experiment
exp_dir = sorted(glob.glob("{}/{}:*".format(args.result_dir, args.expname)))[-1]
run_id = int(exp_dir.split(":")[-1])
# If restart, then work over a new directory
if args.restart:
run_id += 1
run_name = "{}:{0:03d}".format(args.expname, run_id)
train.printname = "{} ".format(misc.bold(args.expname))
snapshot, kimg, resume = None, 0, False
pkls = sorted(glob.glob("{}/{}/network*.pkl".format(args.result_dir, run_name)))
# Load a particular snapshot is specified
if args.pretrained_pkl:
# Soft links support
snapshot = glob.glob(args.pretrained_pkl)[0]
if os.path.islink(snapshot):
snapshot = os.readlink(snapshot)
# Extract training step from the snapshot if specified
try:
kimg = int(snapshot.split("-")[-1].split(".")[0])
except:
pass
# Find latest snapshot in the directory
elif len(pkls) > 0:
snapshot = pkls[-1]
kimg = int(snapshot.split("-")[-1].split(".")[0])
resume = True
if snapshot:
print(misc.bcolored("Resuming {}, kimg {}".format(snapshot, kimg), "white"))
train.resume_pkl = snapshot
train.resume_kimg = kimg
else:
print("Start model training from scratch.", "white")
# Run environment configuration
sc.run_dir_root = args.result_dir
sc.run_desc = args.expname
sc.run_id = run_id
sc.run_name = run_name
sc.submit_target = dnnlib.SubmitTarget.LOCAL
sc.local.do_not_copy_source_files = True
kwargs = EasyDict(train)
kwargs.update(cG = cG, cD = cD)
kwargs.update(dataset_args = dataset_args, vis_args = vis, sched_args = sched, grid_args = grid, metric_arg_list = metrics, tf_config = tf_config)
kwargs.submit_config = copy.deepcopy(sc)
kwargs.resume = resume
# If reload new options from the command line, no need to load the original configuration file
kwargs.load_config = not args.reload
dnnlib.submit_run(**kwargs)
def run(opt):
"""Sets-up all of the parameters necessary to start a ProgressiveGAN training job."""
desc = build_job_name(
opt) # Description string included in result subdir name.
train = EasyDict(run_func_name='training.training_loop.training_loop'
) # Options for training loop.
G = EasyDict(func_name='training.networks_progan.G_paper'
) # Options for generator network.
D = EasyDict(func_name='training.networks_progan.D_paper'
) # 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.
G_loss = EasyDict(
func_name='training.loss.G_wgan') # Options for generator loss.
D_loss = EasyDict(
func_name='training.loss.D_wgan_gp') # Options for discriminator loss.
sched = EasyDict() # Options for TrainingSchedule.
grid = EasyDict(
size='1080p',
layout='random') # Options for setup_snapshot_image_grid().
submit_config = dnnlib.SubmitConfig() # Options for dnnlib.submit_run().
tf_config = {
'rnd.np_random_seed': opt.seed
} # Options for tflib.init_tf().
metrics = [] # Metrics to run during training.
if 'FID' in opt.metrics:
metrics.append(metric_base.fid50k)
if 'PPL' in opt.metrics:
metrics.append(metric_base.ppl_zend_v2)
train.network_metric_ticks = opt.compute_metrics_ticks
train.interp_snapshot_ticks = opt.compute_interp_ticks
find_dataset(opt.dataset)
# Optionally resume from checkpoint:
if opt.resume_exp is not None:
results_dir = os.path.join(os.getcwd(), config.result_dir)
_resume_pkl = find_pkl(results_dir, opt.resume_exp,
opt.resume_snapshot)
train.resume_run_id = opt.resume_exp
train.resume_snapshot = _resume_pkl
train.resume_kimg = int(_resume_pkl.split('.pkl')[0][-6:])
if f'hessian_penalty_{opt.dataset}' not in _resume_pkl and opt.hp_lambda > 0:
print(
'When fine-tuning a job that was originally trained without the Hessian Penalty, '
'hp_start_kimg is relative to the kimg of the checkpoint being resumed from. '
'Hessian Penalty will be phased-in starting at absolute '
f'kimg={opt.hp_start_kimg + train.resume_kimg}.')
opt.hp_start_kimg += train.resume_kimg
# Set up dataset hyper-parameters:
dataset = EasyDict(tfrecord_dir=os.path.join(os.getcwd(), config.data_dir,
opt.dataset),
resolution=opt.resolution)
train.mirror_augment = False
# Set up network hyper-parameters:
G.latent_size = opt.nz
D.infogan_nz = opt.infogan_nz
G.infogan_lambda = opt.infogan_lambda
D.infogan_lambda = opt.infogan_lambda
# When computing the multi-layer Hessian Penalty, we retrieve intermediate activations by accessing the
# corresponding tensor's name. Below are the names of various activations in G that we can retrieve:
activation_type = 'norm'
progan_generator_layer_index_to_name = {
1: f'4x4/Dense/Post_{activation_type}',
2: f'4x4/Conv/Post_{activation_type}',
3: f'8x8/Conv0_up/Post_{activation_type}',
4: f'8x8/Conv1/Post_{activation_type}',
5: f'16x16/Conv0_up/Post_{activation_type}',
6: f'16x16/Conv1/Post_{activation_type}',
7: f'32x32/Conv0_up/Post_{activation_type}',
8: f'32x32/Conv1/Post_{activation_type}',
9: f'64x64/Conv0_up/Post_{activation_type}',
10: f'64x64/Conv1/Post_{activation_type}',
11: f'128x128/Conv0_up/Post_{activation_type}',
12: f'128x128/Conv1/Post_{activation_type}',
13: 'images_out' # final full-resolution RGB activation
}
# Convert from layer indices to layer names (which we'll need to compute the Hessian Penalty):
layers_to_reg = [
progan_generator_layer_index_to_name[layer_ix]
for layer_ix in sorted(opt.layers_to_reg)
]
# Store the Hessian Penalty parameters in their own dictionary:
HP = EasyDict(hp_lambda=opt.hp_lambda,
epsilon=opt.epsilon,
num_rademacher_samples=opt.num_rademacher_samples,
layers_to_reg=layers_to_reg,
warmup_nimg=opt.warmup_kimg * 1000,
hp_start_nimg=opt.hp_start_kimg * 1000)
# How long to train for (as measured by thousands of real images processed, not gradient steps):
train.total_kimg = opt.total_kimg
# We ran the original experiments using 4 GPUs per job. If using a different number,
# we try to scale batch sizes up or down accordingly in the for-loop below. Note that
# using other batch sizes is somewhat untested, though!
submit_config.num_gpus = opt.num_gpus
sched.minibatch_base = 32
sched.minibatch_dict = {
4: 2048,
8: 1024,
16: 512,
32: 256,
64: 128,
128: 96,
256: 32,
512: 16
}
for res, batch_size in sched.minibatch_dict.items():
sched.minibatch_dict[res] = int(batch_size * opt.num_gpus / 4)
# Set-up WandB if optionally using it instead of TensorBoard:
if opt.dashboard_api == 'wandb':
init_wandb(opt=opt,
name=desc,
group=opt.dataset,
entity=opt.wandb_entity)
# Start the training job:
kwargs = EasyDict(train)
kwargs.update(HP_args=HP,
G_args=G,
D_args=D,
G_opt_args=G_opt,
D_opt_args=D_opt,
G_loss_args=G_loss,
D_loss_args=D_loss)
kwargs.update(dataset_args=dataset,
sched_args=sched,
grid_args=grid,
metric_arg_list=metrics,
tf_config=tf_config)
kwargs.submit_config = copy.deepcopy(submit_config)
kwargs.submit_config.run_dir_root = dnnlib.submission.submit.get_template_from_path(
config.result_dir)
kwargs.submit_config.run_dir_ignore += config.run_dir_ignore
kwargs.submit_config.run_desc = desc
dnnlib.submit_run(**kwargs)
def run(dataset, data_dir, result_dir, config_id, num_gpus, total_kimg, gamma,
mirror_augment, metrics):
train = EasyDict(run_func_name='training.training_loop.training_loop'
) # Options for training loop.
G = EasyDict(
func_name='training.invGAN.G_main') # Options for generator network.
D = EasyDict(func_name='training.invGAN.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.
G_loss = EasyDict(func_name='training.loss.G_logistic_ns_pathreg_inv'
) # Options for generator loss.
D_loss = EasyDict(func_name='training.loss.D_logistic_r1'
) # Options for discriminator loss.
sched = EasyDict() # Options for TrainingSchedule.
grid = EasyDict(
size='8k', 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().
train.data_dir = data_dir
train.total_kimg = total_kimg
train.mirror_augment = mirror_augment
train.image_snapshot_ticks = train.network_snapshot_ticks = 10
sched.G_lrate_base = sched.D_lrate_base = 0.002
sched.minibatch_size_base = 32
sched.minibatch_gpu_base = 4
D_loss.gamma = 10
G.synthesis_func = 'G_quotient'
# G.latents_size = 4096 * 3
G.dlatent_size = 4096 * 3
G.latent_size = 512
G.mapping_fmaps = 512
G.fmap_final = 3
metrics = [metric_defaults[x] for x in metrics]
desc = 'InvGan'
desc += '-' + dataset
dataset_args = EasyDict(tfrecord_dir=dataset)
assert num_gpus in [1, 2, 4, 8]
sc.num_gpus = num_gpus
desc += '-%dgpu' % num_gpus
assert config_id in _valid_configs
desc += '-' + config_id
# Configs A-E: Shrink networks to match original StyleGAN.
if config_id != 'config-f':
G.fmap_base = D.fmap_base = 8 << 10
# Config E: Set gamma to 100 and override G & D architecture.
if config_id.startswith('config-e'):
D_loss.gamma = 100
# Configs A-D: Enable progressive growing and switch to networks that support it.
if config_id in ['config-a', 'config-b', 'config-c', 'config-d']:
sched.lod_initial_resolution = 8
sched.G_lrate_base = sched.D_lrate_base = 0.001
sched.G_lrate_dict = sched.D_lrate_dict = {
128: 0.0015,
256: 0.002,
512: 0.003,
1024: 0.003
}
sched.minibatch_size_base = 32 # (default)
sched.minibatch_size_dict = {8: 256, 16: 128, 32: 64, 64: 32}
sched.minibatch_gpu_base = 4 # (default)
sched.minibatch_gpu_dict = {8: 32, 16: 16, 32: 8, 64: 4}
G.synthesis_func = 'G_synthesis_stylegan_revised'
D.func_name = 'training.networks_stylegan2.D_stylegan'
# Configs A-C: Disable path length regularization.
if config_id in ['config-a', 'config-b', 'config-c']:
G_loss = EasyDict(func_name='training.loss.G_logistic_ns')
# Configs A-B: Disable lazy regularization.
if config_id in ['config-a', 'config-b']:
train.lazy_regularization = False
# Config A: Switch to original StyleGAN networks.
if config_id == 'config-a':
pass
# G = EasyDict(func_name='training.networks_stylegan.G_style')
# D = EasyDict(func_name='training.networks_stylegan.D_basic')
if gamma is not None:
D_loss.gamma = gamma
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,
G_loss_args=G_loss,
D_loss_args=D_loss)
kwargs.update(dataset_args=dataset_args,
sched_args=sched,
grid_args=grid,
metric_arg_list=metrics,
tf_config=tf_config)
kwargs.submit_config = copy.deepcopy(sc)
kwargs.submit_config.run_dir_root = result_dir
kwargs.submit_config.run_desc = desc
dnnlib.submit_run(**kwargs)
def run(**args):
args = EasyDict(args)
train = EasyDict(run_func_name="training.training_loop.training_loop"
) # training loop options
sched = EasyDict() # TrainingSchedule options
vis = EasyDict() # visualize.eval() options
grid = EasyDict(size="1080p",
layout="random") # setup_snapshot_img_grid() options
sc = dnnlib.SubmitConfig() # dnnlib.submit_run() options
# If the flag is specified without arguments (--arg), set to True
for arg in [
"summarize", "keep_samples", "style", "fused_modconv",
"local_noise"
]:
if args[arg] is None:
args[arg] = True
if not args.train and not args.eval:
misc.log(
"Warning: Neither --train nor --eval are provided. Therefore, we only print network shapes",
"red")
if args.gansformer_default:
task = args.dataset
pretrained = "gdrive:{}-snapshot.pkl".format(task)
if pretrained not in pretrained_networks.gdrive_urls:
pretrained = None
nset(args, "recompile", pretrained is not None)
nset(args, "pretrained_pkl", pretrained)
nset(args, "mirror_augment", task in ["cityscapes", "ffhq"])
nset(args, "transformer", True)
nset(args, "components_num", {"clevr": 8}.get(task, 16))
nset(args, "latent_size", {"clevr": 128}.get(task, 512))
nset(args, "normalize", "layer")
nset(args, "integration", "mul")
nset(args, "kmeans", True)
nset(args, "use_pos", True)
nset(args, "mapping_ltnt2ltnt", task != "clevr")
nset(args, "style", task != "clevr")
nset(args, "g_arch", "resnet")
nset(args, "mapping_resnet", True)
gammas = {"ffhq": 10, "cities": 20, "clevr": 40, "bedrooms": 100}
nset(args, "gamma", gammas.get(task, 10))
if args.baseline == "GAN":
nset(args, "style", False)
nset(args, "latent_stem", True)
if args.baseline == "SAGAN":
nset(args, "style", False)
nset(args, "latent_stem", True)
nset(args, "g_img2img", 5)
if args.baseline == "kGAN":
nset(args, "kgan", True)
nset(args, "merge_layer", 5)
nset(args, "merge_type", "softmax")
nset(args, "components_num", 8)
# Environment configuration
tf_config = {
"rnd.np_random_seed": 1000,
"allow_soft_placement": True,
"gpu_options.per_process_gpu_memory_fraction": 1.0
}
if args.gpus != "":
num_gpus = len(args.gpus.split(","))
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
assert num_gpus in [1, 2, 4, 8]
sc.num_gpus = num_gpus
# Networks configuration
cG = set_net("G", reg_interval=4)
cD = set_net("D", reg_interval=16)
# Dataset configuration
# For bedrooms, we choose the most common ratio in the
# dataset and crop the other images into that ratio.
ratios = {
"clevr": 0.75,
"bedrooms": 188 / 256,
"cityscapes": 0.5,
"ffhq": 1.0
}
args.ratio = ratios.get(args.dataset, args.ratio)
dataset_args = EasyDict(tfrecord_dir=args.dataset,
max_imgs=args.train_images_num,
num_threads=args.num_threads)
for arg in ["data_dir", "mirror_augment", "total_kimg", "ratio"]:
cset(train, arg, args[arg])
# Training and Optimizations configuration
for arg in ["eval", "train", "recompile", "last_snapshots"]:
cset(train, arg, args[arg])
# Round to the closest multiply of minibatch size for validity
args.batch_size -= args.batch_size % args.minibatch_size
args.minibatch_std_size -= args.minibatch_std_size % args.minibatch_size
args.latent_size -= args.latent_size % args.components_num
if args.latent_size == 0:
misc.error(
"--latent-size is too small. Must best a multiply of components-num"
)
sched_args = {
"G_lrate": "g_lr",
"D_lrate": "d_lr",
"minibatch_size": "batch_size",
"minibatch_gpu": "minibatch_size"
}
for arg, cmd_arg in sched_args.items():
cset(sched, arg, args[cmd_arg])
cset(train, "clip", args.clip)
# Logging and metrics configuration
metrics = [metric_defaults[x] for x in args.metrics]
cset(cG.args, "truncation_psi", args.truncation_psi)
for arg in ["keep_samples", "num_heads"]:
cset(vis, arg, args[arg])
for arg in ["summarize", "eval_images_num"]:
cset(train, arg, args[arg])
# Visualization
args.vis_imgs = args.vis_images
args.vis_ltnts = args.vis_latents
vis_types = [
"imgs", "ltnts", "maps", "layer_maps", "interpolations", "noise_var",
"style_mix"
]
# Set of all the set visualization types option
vis.vis_types = {arg for arg in vis_types if args["vis_{}".format(arg)]}
vis_args = {
"attention": "transformer",
"grid": "vis_grid",
"num": "vis_num",
"rich_num": "vis_rich_num",
"section_size": "vis_section_size",
"intrp_density": "interpolation_density",
# "intrp_per_component": "interpolation_per_component",
"alpha": "blending_alpha"
}
for arg, cmd_arg in vis_args.items():
cset(vis, arg, args[cmd_arg])
# Networks architecture
cset(cG.args, "architecture", args.g_arch)
cset(cD.args, "architecture", args.d_arch)
cset(cG.args, "tanh", args.tanh)
# Latent sizes
if args.components_num > 1:
if not (args.transformer or args.kgan):
misc.error(
"--components-num > 1 but the model is not using components. "
+
"Either add --transformer for GANsformer or --kgan for k-GAN.")
args.latent_size = int(args.latent_size / args.components_num)
cD.args.latent_size = cG.args.latent_size = cG.args.dlatent_size = args.latent_size
cset([cG.args, cD.args, vis], "components_num", args.components_num)
# Mapping network
for arg in ["layersnum", "lrmul", "dim", "resnet", "shared_dim"]:
field = "mapping_{}".format(arg)
cset(cG.args, field, args[field])
# StyleGAN settings
for arg in ["style", "latent_stem", "fused_modconv", "local_noise"]:
cset(cG.args, arg, args[arg])
cD.args.mbstd_group_size = args.minibatch_std_size
# GANsformer
cset(cG.args, "transformer", args.transformer)
cset(cD.args, "transformer", args.d_transformer)
args.norm = args.normalize
for arg in [
"norm", "integration", "ltnt_gate", "img_gate", "iterative",
"kmeans", "kmeans_iters", "mapping_ltnt2ltnt"
]:
cset(cG.args, arg, args[arg])
for arg in ["use_pos", "num_heads"]:
cset([cG.args, cD.args], arg, args[arg])
# Positional encoding
for arg in ["dim", "init", "directions_num"]:
field = "pos_{}".format(arg)
cset([cG.args, cD.args], field, args[field])
# k-GAN
for arg in ["layer", "type", "same"]:
field = "merge_{}".format(arg)
cset(cG.args, field, args[field])
cset([cG.args, train], "merge", args.kgan)
if args.kgan and args.transformer:
misc.error(
"Either have --transformer for GANsformer or --kgan for k-GAN, not both"
)
# Attention
for arg in ["start_res", "end_res", "ltnt2ltnt",
"img2img"]: # , "local_attention"
cset(cG.args, arg, args["g_{}".format(arg)])
cset(cD.args, arg, args["d_{}".format(arg)])
cset(cG.args, "img2ltnt", args.g_img2ltnt)
# cset(cD.args, "ltnt2img", args.d_ltnt2img)
# Mixing and dropout
for arg in [
"style_mixing", "component_mixing", "component_dropout",
"attention_dropout"
]:
cset(cG.args, arg, args[arg])
# Loss and regularization
gloss_args = {
"loss_type": "g_loss",
"reg_weight": "g_reg_weight",
# "pathreg": "pathreg",
}
dloss_args = {"loss_type": "d_loss", "reg_type": "d_reg", "gamma": "gamma"}
for arg, cmd_arg in gloss_args.items():
cset(cG.loss_args, arg, args[cmd_arg])
for arg, cmd_arg in dloss_args.items():
cset(cD.loss_args, arg, args[cmd_arg])
##### Experiments management:
# Whenever we start a new experiment we store its result in a directory named 'args.expname:000'.
# When we rerun a training or evaluation command it restores the model from that directory by default.
# If we wish to restart the model training, we can set --restart and then we will store data in a new
# directory: 'args.expname:001' after the first restart, then 'args.expname:002' after the second, etc.
# Find the latest directory that matches the experiment
exp_dir = sorted(glob.glob("{}/{}-*".format(args.result_dir,
args.expname)))
run_id = 0
if len(exp_dir) > 0:
run_id = int(exp_dir[-1].split("-")[-1])
# If restart, then work over a new directory
if args.restart:
run_id += 1
run_name = "{}-{:03d}".format(args.expname, run_id)
train.printname = "{} ".format(misc.bold(args.expname))
snapshot, kimg, resume = None, 0, False
pkls = sorted(
glob.glob("{}/{}/network*.pkl".format(args.result_dir, run_name)))
# Load a particular snapshot is specified
if args.pretrained_pkl is not None and args.pretrained_pkl != "None":
# Soft links support
if args.pretrained_pkl.startswith("gdrive"):
if args.pretrained_pkl not in pretrained_networks.gdrive_urls:
misc.error(
"--pretrained_pkl {} not available in the catalog (see pretrained_networks.py)"
)
snapshot = args.pretrained_pkl
else:
snapshot = glob.glob(args.pretrained_pkl)[0]
if os.path.islink(snapshot):
snapshot = os.readlink(snapshot)
# Extract training step from the snapshot if specified
try:
kimg = int(snapshot.split("-")[-1].split(".")[0])
except:
pass
# Find latest snapshot in the directory
elif len(pkls) > 0:
snapshot = pkls[-1]
kimg = int(snapshot.split("-")[-1].split(".")[0])
resume = True
if snapshot:
misc.log(
"Resuming {}, from {}, kimg {}".format(run_name, snapshot, kimg),
"white")
train.resume_pkl = snapshot
train.resume_kimg = kimg
else:
misc.log("Start model training from scratch", "white")
# Run environment configuration
sc.run_dir_root = args.result_dir
sc.run_desc = args.expname
sc.run_id = run_id
sc.run_name = run_name
sc.submit_target = dnnlib.SubmitTarget.LOCAL
sc.local.do_not_copy_source_files = True
kwargs = EasyDict(train)
kwargs.update(cG=cG, cD=cD)
kwargs.update(dataset_args=dataset_args,
vis_args=vis,
sched_args=sched,
grid_args=grid,
metric_arg_list=metrics,
tf_config=tf_config)
kwargs.submit_config = copy.deepcopy(sc)
kwargs.resume = resume
kwargs.load_config = args.reload
dnnlib.submit_run(**kwargs)
def setup_config(run_dir, **args):
args = EasyDict(args) # command-line options
train = EasyDict(run_dir=run_dir) # training loop options
vis = EasyDict(run_dir=run_dir) # visualization loop options
if args.reload:
config_fn = os.path.join(run_dir, "training_options.json")
if os.path.exists(config_fn):
# Load config form the experiment existing file (and so ignore command-line arguments)
with open(config_fn, "rt") as f:
config = json.load(f)
return config
misc.log(
f"Warning: --reload is set for a new experiment {args.expname}," +
f" but configuration file to reload from {config_fn} doesn't exist.",
"red")
# GANformer and baselines default settings
# ----------------------------------------------------------------------------
if args.ganformer_default:
task = args.dataset
nset(args, "mirror_augment", task in ["cityscapes", "ffhq"])
nset(args, "transformer", True)
nset(args, "components_num", {"clevr": 8}.get(task, 16))
nset(args, "latent_size", {"clevr": 128}.get(task, 512))
nset(args, "normalize", "layer")
nset(args, "integration", "mul")
nset(args, "kmeans", True)
nset(args, "use_pos", True)
nset(args, "mapping_ltnt2ltnt", task != "clevr")
nset(args, "style", task != "clevr")
nset(args, "g_arch", "resnet")
nset(args, "mapping_resnet", True)
gammas = {"ffhq": 10, "cityscapes": 20, "clevr": 40, "bedrooms": 100}
nset(args, "gamma", gammas.get(task, 10))
if args.baseline == "GAN":
nset(args, "style", False)
nset(args, "latent_stem", True)
## k-GAN and SAGAN are not currently supported in the pytorch version.
## See the TF version for implementation of these baselines!
# if args.baseline == "SAGAN":
# nset(args, "style", False)
# nset(args, "latent_stem", True)
# nset(args, "g_img2img", 5)
# if args.baseline == "kGAN":
# nset(args, "kgan", True)
# nset(args, "merge_layer", 5)
# nset(args, "merge_type", "softmax")
# nset(args, "components_num", 8)
# General setup
# ----------------------------------------------------------------------------
# If the flag is specified without arguments (--arg), set to True
for arg in [
"cuda_bench", "allow_tf32", "keep_samples", "style", "local_noise"
]:
if args[arg] is None:
args[arg] = True
if not any([args.train, args.eval, args.vis]):
misc.log(
"Warning: None of --train, --eval or --vis are provided. Therefore, we only print network shapes",
"red")
for arg in ["train", "eval", "vis", "last_snapshots"]:
cset(train, arg, args[arg])
if args.gpus != "":
num_gpus = len(args.gpus.split(","))
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
if not (num_gpus >= 1 and num_gpus & (num_gpus - 1) == 0):
misc.error("Number of GPUs must be a power of two")
args.num_gpus = num_gpus
# CUDA settings
for arg in ["batch_size", "batch_gpu", "allow_tf32"]:
cset(train, arg, args[arg])
cset(train, "cudnn_benchmark", args.cuda_bench)
# Data setup
# ----------------------------------------------------------------------------
# For bedrooms, we choose the most common ratio in the
# dataset and crop the other images into that ratio.
ratios = {
"clevr": 0.75,
"bedrooms": 188 / 256,
"cityscapes": 0.5,
"ffhq": 1.0
}
args.ratio = args.ratio or ratios.get(args.dataset, 1.0)
args.crop_ratio = 0.5 if args.resolution > 256 and args.ratio < 0.5 else None
args.printname = args.expname
for arg in ["total_kimg", "printname"]:
cset(train, arg, args[arg])
dataset_args = EasyDict(class_name="training.dataset.ImageFolderDataset",
path=f"{args.data_dir}/{args.dataset}",
max_items=args.train_images_num,
resolution=args.resolution,
ratio=args.ratio,
mirror_augment=args.mirror_augment)
dataset_args.loader_args = EasyDict(num_workers=args.num_threads,
pin_memory=True,
prefetch_factor=2)
# Optimization setup
# ----------------------------------------------------------------------------
cG = set_net("Generator", ["mapping", "synthesis"], args.g_lr, 4)
cD = set_net("Discriminator", ["mapping", "block", "epilogue"], args.d_lr,
16)
cset([cG, cD], "crop_ratio", args.crop_ratio)
mbstd = min(
args.batch_gpu, 4
) # other hyperparams behave more predictably if mbstd group size remains fixed
cset(cD.epilogue_kwargs, "mbstd_group_size", mbstd)
# Automatic tuning
if args.autotune:
batch_size = max(
min(args.num_gpus * min(4096 // args.resolution, 32), 64),
args.num_gpus) # keep gpu memory consumption at bay
batch_gpu = args.batch_size // args.num_gpus
nset(args, "batch_size", batch_size)
nset(args, "batch_gpu", batch_gpu)
fmap_decay = 1 if args.resolution >= 512 else 0.5 # other hyperparams behave more predictably if mbstd group size remains fixed
lr = 0.002 if args.resolution >= 1024 else 0.0025
gamma = 0.0002 * (args.resolution**
2) / args.batch_size # heuristic formula
cset([cG.synthesis_kwargs, cD], "dim_base", int(fmap_decay * 32768))
nset(args, "g_lr", lr)
cset(cG.opt_args, "lr", args.g_lr)
nset(args, "d_lr", lr)
cset(cD.opt_args, "lr", args.d_lr)
nset(args, "gamma", gamma)
train.ema_rampup = 0.05
train.ema_kimg = batch_size * 10 / 32
if args.batch_size % (args.batch_gpu * args.num_gpus) != 0:
misc.error(
"--batch-size should be divided by --batch-gpu * 'num_gpus'")
# Loss and regularization settings
loss_args = EasyDict(class_name="training.loss.StyleGAN2Loss",
g_loss=args.g_loss,
d_loss=args.d_loss,
r1_gamma=args.gamma,
pl_weight=args.pl_weight)
# if args.fp16:
# cset([cG.synthesis_kwargs, cD], "num_fp16_layers", 4) # enable mixed-precision training
# cset([cG.synthesis_kwargs, cD], "conv_clamp", 256) # clamp activations to avoid float16 overflow
# cset([cG.synthesis_kwargs, cD.block_args], "fp16_channels_last", args.nhwc)
# Evaluation and visualization
# ----------------------------------------------------------------------------
from metrics import metric_main
for metric in args.metrics:
if not metric_main.is_valid_metric(metric):
misc.error(
f"Unknown metric: {metric}. The valid metrics are: {metric_main.list_valid_metrics()}"
)
for arg in ["num_gpus", "metrics", "eval_images_num", "truncation_psi"]:
cset(train, arg, args[arg])
for arg in ["keep_samples", "num_heads"]:
cset(vis, arg, args[arg])
args.vis_imgs = args.vis_images
args.vis_ltnts = args.vis_latents
vis_types = [
"imgs", "ltnts", "maps", "layer_maps", "interpolations", "noise_var",
"style_mix"
]
# Set of all the set visualization types option
vis.vis_types = list({arg for arg in vis_types if args[f"vis_{arg}"]})
vis_args = {
"attention": "transformer",
"grid": "vis_grid",
"num": "vis_num",
"rich_num": "vis_rich_num",
"section_size": "vis_section_size",
"intrp_density": "interpolation_density",
# "intrp_per_component": "interpolation_per_component",
"alpha": "blending_alpha"
}
for arg, cmd_arg in vis_args.items():
cset(vis, arg, args[cmd_arg])
# Networks setup
# ----------------------------------------------------------------------------
# Networks architecture
cset(cG.synthesis_kwargs, "architecture", args.g_arch)
cset(cD, "architecture", args.d_arch)
# Latent sizes
if args.components_num > 0:
if not args.transformer: # or args.kgan):
misc.error(
"--components-num > 0 but the model is not using components. "
+
"Add --transformer for GANformer (which uses latent components)."
)
if args.latent_size % args.components_num != 0:
misc.error(
f"--latent-size ({args.latent_size}) should be divisible by --components-num (k={k})"
)
args.latent_size = int(args.latent_size / args.components_num)
cG.z_dim = cG.w_dim = args.latent_size
cset([cG, vis], "k", args.components_num +
1) # We add a component to modulate features globally
# Mapping network
args.mapping_layer_dim = args.mapping_dim
for arg in ["num_layers", "layer_dim", "resnet", "shared", "ltnt2ltnt"]:
field = f"mapping_{arg}"
cset(cG.mapping_kwargs, arg, args[field])
# StyleGAN settings
for arg in ["style", "latent_stem", "local_noise"]:
cset(cG.synthesis_kwargs, arg, args[arg])
# GANformer
cset([cG.synthesis_kwargs, cG.mapping_kwargs], "transformer",
args.transformer)
# Attention related settings
for arg in ["use_pos", "num_heads", "ltnt_gate", "attention_dropout"]:
cset([cG.mapping_kwargs, cG.synthesis_kwargs], arg, args[arg])
# Attention types and layers
for arg in ["start_res", "end_res"
]: # , "local_attention" , "ltnt2ltnt", "img2img", "img2ltnt"
cset(cG.synthesis_kwargs, arg, args[f"g_{arg}"])
# Mixing and dropout
for arg in ["style_mixing", "component_mixing"]:
cset(loss_args, arg, args[arg])
cset(cG, "component_dropout", args["component_dropout"])
# Extra transformer options
args.norm = args.normalize
for arg in [
"norm", "integration", "img_gate", "iterative", "kmeans",
"kmeans_iters"
]:
cset(cG.synthesis_kwargs, arg, args[arg])
# Positional encoding
# args.pos_dim = args.pos_dim or args.latent_size
for arg in ["dim", "type", "init", "directions_num"]:
field = f"pos_{arg}"
cset(cG.synthesis_kwargs, field, args[field])
# k-GAN
# for arg in ["layer", "type", "same"]:
# field = "merge_{}".format(arg)
# cset(cG.args, field, args[field])
# cset(cG.synthesis_kwargs, "merge", args.kgan)
# if args.kgan and args.transformer:
# misc.error("Either have --transformer for GANformer or --kgan for k-GAN, not both")
config = EasyDict(train)
config.update(cG=cG,
cD=cD,
loss_args=loss_args,
dataset_args=dataset_args,
vis_args=vis)
# Save config file
with open(os.path.join(run_dir, "training_options.json"), "wt") as f:
json.dump(config, f, indent=2)
return config
