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
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=(3, 2),
layout="random") # setup_snapshot_img_grid() options
sc = dnnlib.SubmitConfig() # dnnlib.submit_run() options
# GANformer and baselines default settings
# ----------------------------------------------------------------------------
if args.ganformer_default:
task = args.dataset
nset(args, "recompile", args.pretrained_pkl is not None)
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)
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 [
"summarize", "keep_samples", "style", "fused_modconv",
"local_noise"
]:
if args[arg] is None:
args[arg] = True
# 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
# Data setup
# ----------------------------------------------------------------------------
# Dataset configuration
if not os.path.exists(f"{args.data_dir}/{args.dataset}"):
misc.error(
f"The dataset {args.data_dir}/{args.dataset} directory does not exist"
)
# 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
dataset_args = EasyDict(tfrecord_dir=args.dataset,
max_imgs=args.train_images_num,
num_threads=args.num_threads,
resolution=args.resolution,
crop_ratio=args.crop_ratio,
ratio=args.ratio)
for arg in ["data_dir", "mirror_augment", "total_kimg"]:
cset(train, arg, args[arg])
# Optimization setup
# ----------------------------------------------------------------------------
# Networks configuration
cG = set_net("G", reg_interval=4)
cD = set_net("D", reg_interval=16)
cset([cG, cD], "crop_ratio", args.crop_ratio)
# Training and Optimizations configuration
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", "recompile", "last_snapshots"]:
cset(train, arg, args[arg])
if args.batch_size % (args.batch_gpu * num_gpus) != 0:
misc.error(
"--batch-size should be divided by --batch-gpu * 'num_gpus'")
if args.latent_size % args.components_num != 0:
misc.error(
f"--latent-size ({args.latent_size}) should be divisible by --components-num (k={k})"
)
sched_args = {
"G_lrate": "g_lr",
"D_lrate": "d_lr",
"batch_size": "batch_size",
"batch_gpu": "batch_gpu"
}
for arg, cmd_arg in sched_args.items():
cset(sched, arg, args[cmd_arg])
cset(train, "clip", args.clip)
# Evaluation and visualization
# ----------------------------------------------------------------------------
# Logging and metrics configuration
for metric in args.metrics:
if metric not in metric_defaults:
misc.error(f"Unknown metric: {metric}")
metrics = [metric_defaults[x] for x in args.metrics]
for arg in ["summarize", "eval_images_num"]:
cset(train, arg, args[arg])
cset(cG.args, "truncation_psi", args.truncation_psi)
for arg in ["keep_samples", "num_heads"]:
cset(vis, 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[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.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 GANformer or --kgan for k-GAN.")
args.latent_size = int(args.latent_size / args.components_num)
cD.args.a_dim = cG.args.z_dim = cG.args.w_dim = 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 = f"mapping_{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 = min(args.batch_size, 4)
# GANformer
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", "type", "init", "directions_num"]:
field = f"pos_{arg}"
cset([cG.args, cD.args], field, args[field])
# k-GAN
for arg in ["layer", "type", "same"]:
field = f"merge_{arg}"
cset(cG.args, field, args[field])
cset(cG.args, "merge", args.kgan)
if args.kgan and args.transformer:
misc.error(
"Either have --transformer for GANformer or --kgan for k-GAN, not both"
)
# Attention
for arg in ["start_res", "end_res", "ltnt2ltnt",
"img2img"]: # , "local_attention"
cset(cG.args, arg, args[f"g_{arg}"])
cset(cD.args, arg, args[f"d_{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])
# Setup and launching
# ----------------------------------------------------------------------------
##### 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(f"{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 = f"{args.expname}-{run_id:03d}"
train.printname = f"{misc.bold(args.expname)} "
snapshot, kimg, resume = None, 0, False
pkls = sorted(glob.glob(f"{args.result_dir}/{run_name}/network*.pkl"))
# 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(f"Resuming {run_name}, from {snapshot}, kimg {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)
