Upload folder using huggingface_hub

3ce19a2 verified 4 months ago

10.8 kB

# PyTorch StudioGAN: https://github.com/POSTECH-CVLab/PyTorch-StudioGAN
# The MIT License (MIT)
# See license file or visit https://github.com/POSTECH-CVLab/PyTorch-StudioGAN for details

# src/utils/ckpt.py

from os.path import join
import os
import glob

import torch
import numpy as np

import utils.log as log
try:
    import utils.misc as misc
except AttributeError:
    pass

blacklist = ["CCMGAN2048-train-2021_06_22_06_11_37"]


def make_ckpt_dir(ckpt_dir):
    if not os.path.exists(ckpt_dir):
        os.makedirs(ckpt_dir)
    return ckpt_dir


def load_ckpt(model, optimizer, ckpt_path, load_model=False, load_opt=False, load_misc=False, is_freezeD=False):
    ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage, weights_only=False)
    if load_model:
        if is_freezeD:
            mismatch_names = misc.load_parameters(src=ckpt["state_dict"],
                                                  dst=model.state_dict(),
                                                  strict=False)
            print("The following parameters/buffers do not match with the ones of the pre-trained model:", mismatch_names)
        else:
            # Inference helpers (paper viz dumpers) set STUDIOGAN_LOAD_STRICT=0 so that
            # state_dicts saved by an older code revision (e.g. extra mapping.rtm_*
            # buffers) still load. Default behaviour is unchanged.
            _strict = os.environ.get("STUDIOGAN_LOAD_STRICT", "1") != "0"
            if _strict:
                model.load_state_dict(ckpt["state_dict"], strict=True)
            else:
                missing, unexpected = model.load_state_dict(ckpt["state_dict"], strict=False)
                if missing or unexpected:
                    print("[ckpt] non-strict load: missing={n_m} unexpected={n_u}".format(
                        n_m=len(missing), n_u=len(unexpected)))

    if load_opt:
        optimizer.load_state_dict(ckpt["optimizer"])
        for state in optimizer.state.values():
            for k, v in state.items():
                if isinstance(v, torch.Tensor):
                    state[k] = v.cuda()

    if load_misc:
        seed = ckpt["seed"]
        run_name = ckpt["run_name"]
        step = ckpt["step"]
        try:
            aa_p = ckpt["aa_p"]
        except:
            aa_p = ckpt["ada_p"]
        best_step = ckpt["best_step"]
        best_fid = ckpt["best_fid"]

        try:
            epoch = ckpt["epoch"]
        except:
            epoch = 0
        try:
            topk = ckpt["topk"]
        except:
            topk = "initialize"
        try:
            best_ckpt_path = ckpt["best_fid_checkpoint_path"]
        except:
            best_ckpt_path = ckpt["best_fid_ckpt"]
        try:
            lecam_emas = ckpt["lecam_emas"]
        except:
            lecam_emas = None
        return seed, run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path, lecam_emas


def load_StudioGAN_ckpts(ckpt_dir, load_best, Gen, Dis, g_optimizer, d_optimizer, run_name, apply_g_ema, Gen_ema, ema,
                         is_train, RUN, logger, global_rank, device, cfg_file):
    when = "best" if load_best is True else "current"
    x = join(ckpt_dir, "model=G-{when}-weights-step=".format(when=when))
    y = join(ckpt_dir, "model=D-{when}-weights-step=".format(when=when))
    Gen_glob = glob.glob(glob.escape(x) + '*.pth')
    Dis_glob = glob.glob(glob.escape(y) + '*.pth')

    # Inference-only fallback for shipped checkpoint dirs that contain
    # *only* G_ema-* (e.g. StudioGAN's pretrained baseline tarballs).
    # Triggered only when both the Gen and Dis files are absent and the
    # caller is not training. Returns dummy step/epoch metadata.
    if not Gen_glob and not Dis_glob and not is_train and apply_g_ema:
        z_pat = join(ckpt_dir, "model=G_ema-{when}-weights-step=".format(when=when))
        z_glob = glob.glob(glob.escape(z_pat) + '*.pth')
        if not z_glob:
            z_glob = glob.glob(join(ckpt_dir, "model=G_ema-current-weights-step=*.pth"))
        if z_glob:
            Gen_ema_ckpt_path = sorted(z_glob)[-1]
            print("[ckpt] G/D ckpts missing -> G_ema-only inference load:", Gen_ema_ckpt_path)
            os.environ.setdefault("STUDIOGAN_LOAD_STRICT", "0")
            load_ckpt(model=Gen_ema, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
                      load_model=True, load_opt=False, load_misc=False)
            try:
                load_ckpt(model=Gen, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
                          load_model=True, load_opt=False, load_misc=False)
            except Exception as _e:
                print("[ckpt] (skipping Gen mirror; not needed for sampling):", _e)
            ema.source, ema.target = Gen, Gen_ema
            return run_name, 0, 0, "initialize", 0, 0, 0.0, "", None, logger

    Gen_ckpt_path = Gen_glob[0]
    Dis_ckpt_path = Dis_glob[0]

    prev_run_name = torch.load(Dis_ckpt_path, map_location=lambda storage, loc: storage, weights_only=False)["run_name"]
    is_freezeD = True if RUN.freezeD > -1 else False

    load_ckpt(model=Gen,
              optimizer=g_optimizer,
              ckpt_path=Gen_ckpt_path,
              load_model=True,
              load_opt=False if prev_run_name in blacklist or is_freezeD or not is_train else True,
              load_misc=False,
              is_freezeD=is_freezeD)

    seed, prev_run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path, lecam_emas =\
        load_ckpt(model=Dis,
                  optimizer=d_optimizer,
                  ckpt_path=Dis_ckpt_path,
                  load_model=True,
                  load_opt=False if prev_run_name in blacklist or is_freezeD or not is_train else True,
                  load_misc=True,
                  is_freezeD=is_freezeD)

    if apply_g_ema:
        z = join(ckpt_dir, "model=G_ema-{when}-weights-step=".format(when=when))
        Gen_ema_ckpt_path = glob.glob(glob.escape(z) + '*.pth')[0]
        load_ckpt(model=Gen_ema,
                  optimizer=None,
                  ckpt_path=Gen_ema_ckpt_path,
                  load_model=True,
                  load_opt=False,
                  load_misc=False,
                  is_freezeD=is_freezeD)

        ema.source, ema.target = Gen, Gen_ema

    if is_train and RUN.seed != seed:
        RUN.seed = seed + global_rank
        misc.fix_seed(RUN.seed)

    if device == 0:
        if not is_freezeD:
            logger = log.make_logger(RUN.save_dir, prev_run_name, None)

        logger.info("Generator checkpoint is {}".format(Gen_ckpt_path))
        if apply_g_ema:
            logger.info("EMA_Generator checkpoint is {}".format(Gen_ema_ckpt_path))
        logger.info("Discriminator checkpoint is {}".format(Dis_ckpt_path))

    if is_freezeD:
        prev_run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path =\
            run_name, 0, 0, "initialize", None, 0, None, None
    return prev_run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path, lecam_emas, logger


def load_best_model(ckpt_dir, Gen, Dis, apply_g_ema, Gen_ema, ema):
    Gen, Dis, Gen_ema = misc.peel_models(Gen, Dis, Gen_ema)
    Gen_glob = glob.glob(join(ckpt_dir, "model=G-best-weights-step*.pth"))
    Dis_glob = glob.glob(join(ckpt_dir, "model=D-best-weights-step*.pth"))
    if not Gen_glob and not Dis_glob and apply_g_ema:
        z_glob = glob.glob(join(ckpt_dir, "model=G_ema-best-weights-step*.pth"))
        if not z_glob:
            z_glob = glob.glob(join(ckpt_dir, "model=G_ema-current-weights-step*.pth"))
        if z_glob:
            Gen_ema_ckpt_path = sorted(z_glob)[-1]
            print("[ckpt] G/D best ckpts missing -> G_ema-only inference load:", Gen_ema_ckpt_path)
            os.environ.setdefault("STUDIOGAN_LOAD_STRICT", "0")
            load_ckpt(model=Gen_ema, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
                      load_model=True, load_opt=False, load_misc=False)
            try:
                load_ckpt(model=Gen, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
                          load_model=True, load_opt=False, load_misc=False)
            except Exception as _e:
                print("[ckpt] (skipping Gen mirror; not needed for sampling):", _e)
            ema.source, ema.target = Gen, Gen_ema
            try:
                step_str = os.path.basename(Gen_ema_ckpt_path).split("step=")[1].split(".")[0]
                return int(step_str)
            except Exception:
                return 0
    Gen_ckpt_path = Gen_glob[0]
    Dis_ckpt_path = Dis_glob[0]

    load_ckpt(model=Gen,
              optimizer=None,
              ckpt_path=Gen_ckpt_path,
              load_model=True,
              load_opt=False,
              load_misc=False,
              is_freezeD=False)


    _, _, _, _, _, _, best_step, _, _, _ = load_ckpt(model=Dis,
                                                  optimizer=None,
                                                  ckpt_path=Dis_ckpt_path,
                                                  load_model=True,
                                                  load_opt=False,
                                                  load_misc=True,
                                                  is_freezeD=False)

    if apply_g_ema:
        Gen_ema_ckpt_path = glob.glob(join(ckpt_dir, "model=G_ema-best-weights-step*.pth"))[0]
        load_ckpt(model=Gen_ema,
                  optimizer=None,
                  ckpt_path=Gen_ema_ckpt_path,
                  load_model=True,
                  load_opt=False,
                  load_misc=False,
                  is_freezeD=False)

        ema.source, ema.target = Gen, Gen_ema
    return best_step


def load_prev_dict(directory, file_name):
    return np.load(join(directory, file_name), allow_pickle=True).item()


def check_is_pre_trained_model(ckpt_dir, GAN_train, GAN_test):
    assert GAN_train*GAN_test == 0, "cannot conduct GAN_train and GAN_test togather."
    if GAN_train:
        mode = "fake_trained"
    else:
        mode = "real_trained"

    ckpt_list = glob.glob(join(ckpt_dir, "model=C-{mode}-best-weights.pth".format(mode=mode)))
    if len(ckpt_list) == 0:
        is_pre_train_model = False
    else:
        is_pre_train_model = True
    return is_pre_train_model, mode


def load_GAN_train_test_model(model, mode, optimizer, RUN):
    ckpt_path = join(RUN.ckpt_dir, "model=C-{mode}-best-weights.pth".format(mode=mode))
    ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage, weights_only=False)

    model.load_state_dict(ckpt["state_dict"])
    optimizer.load_state_dict(ckpt["optimizer"])
    epoch_trained = ckpt["epoch"]
    best_top1 = ckpt["best_top1"]
    best_top5 = ckpt["best_top5"]
    best_epoch = ckpt["best_epoch"]
    return epoch_trained, best_top1, best_top5, best_epoch

	# PyTorch StudioGAN: https://github.com/POSTECH-CVLab/PyTorch-StudioGAN
	# The MIT License (MIT)
	# See license file or visit https://github.com/POSTECH-CVLab/PyTorch-StudioGAN for details

	# src/utils/ckpt.py

	from os.path import join
	import os
	import glob

	import torch
	import numpy as np

	import utils.log as log
	try:
	import utils.misc as misc
	except AttributeError:
	pass

	blacklist = ["CCMGAN2048-train-2021_06_22_06_11_37"]


	def make_ckpt_dir(ckpt_dir):
	if not os.path.exists(ckpt_dir):
	os.makedirs(ckpt_dir)
	return ckpt_dir


	def load_ckpt(model, optimizer, ckpt_path, load_model=False, load_opt=False, load_misc=False, is_freezeD=False):
	ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage, weights_only=False)
	if load_model:
	if is_freezeD:
	mismatch_names = misc.load_parameters(src=ckpt["state_dict"],
	dst=model.state_dict(),
	strict=False)
	print("The following parameters/buffers do not match with the ones of the pre-trained model:", mismatch_names)
	else:
	# Inference helpers (paper viz dumpers) set STUDIOGAN_LOAD_STRICT=0 so that
	# state_dicts saved by an older code revision (e.g. extra mapping.rtm_*
	# buffers) still load. Default behaviour is unchanged.
	_strict = os.environ.get("STUDIOGAN_LOAD_STRICT", "1") != "0"
	if _strict:
	model.load_state_dict(ckpt["state_dict"], strict=True)
	else:
	missing, unexpected = model.load_state_dict(ckpt["state_dict"], strict=False)
	if missing or unexpected:
	print("[ckpt] non-strict load: missing={n_m} unexpected={n_u}".format(
	n_m=len(missing), n_u=len(unexpected)))

	if load_opt:
	optimizer.load_state_dict(ckpt["optimizer"])
	for state in optimizer.state.values():
	for k, v in state.items():
	if isinstance(v, torch.Tensor):
	state[k] = v.cuda()

	if load_misc:
	seed = ckpt["seed"]
	run_name = ckpt["run_name"]
	step = ckpt["step"]
	try:
	aa_p = ckpt["aa_p"]
	except:
	aa_p = ckpt["ada_p"]
	best_step = ckpt["best_step"]
	best_fid = ckpt["best_fid"]

	try:
	epoch = ckpt["epoch"]
	except:
	epoch = 0
	try:
	topk = ckpt["topk"]
	except:
	topk = "initialize"
	try:
	best_ckpt_path = ckpt["best_fid_checkpoint_path"]
	except:
	best_ckpt_path = ckpt["best_fid_ckpt"]
	try:
	lecam_emas = ckpt["lecam_emas"]
	except:
	lecam_emas = None
	return seed, run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path, lecam_emas


	def load_StudioGAN_ckpts(ckpt_dir, load_best, Gen, Dis, g_optimizer, d_optimizer, run_name, apply_g_ema, Gen_ema, ema,
	is_train, RUN, logger, global_rank, device, cfg_file):
	when = "best" if load_best is True else "current"
	x = join(ckpt_dir, "model=G-{when}-weights-step=".format(when=when))
	y = join(ckpt_dir, "model=D-{when}-weights-step=".format(when=when))
	Gen_glob = glob.glob(glob.escape(x) + '*.pth')
	Dis_glob = glob.glob(glob.escape(y) + '*.pth')

	# Inference-only fallback for shipped checkpoint dirs that contain
	# only G_ema-* (e.g. StudioGAN's pretrained baseline tarballs).
	# Triggered only when both the Gen and Dis files are absent and the
	# caller is not training. Returns dummy step/epoch metadata.
	if not Gen_glob and not Dis_glob and not is_train and apply_g_ema:
	z_pat = join(ckpt_dir, "model=G_ema-{when}-weights-step=".format(when=when))
	z_glob = glob.glob(glob.escape(z_pat) + '*.pth')
	if not z_glob:
	z_glob = glob.glob(join(ckpt_dir, "model=G_ema-current-weights-step=*.pth"))
	if z_glob:
	Gen_ema_ckpt_path = sorted(z_glob)[-1]
	print("[ckpt] G/D ckpts missing -> G_ema-only inference load:", Gen_ema_ckpt_path)
	os.environ.setdefault("STUDIOGAN_LOAD_STRICT", "0")
	load_ckpt(model=Gen_ema, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
	load_model=True, load_opt=False, load_misc=False)
	try:
	load_ckpt(model=Gen, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
	load_model=True, load_opt=False, load_misc=False)
	except Exception as _e:
	print("[ckpt] (skipping Gen mirror; not needed for sampling):", _e)
	ema.source, ema.target = Gen, Gen_ema
	return run_name, 0, 0, "initialize", 0, 0, 0.0, "", None, logger

	Gen_ckpt_path = Gen_glob[0]
	Dis_ckpt_path = Dis_glob[0]

	prev_run_name = torch.load(Dis_ckpt_path, map_location=lambda storage, loc: storage, weights_only=False)["run_name"]
	is_freezeD = True if RUN.freezeD > -1 else False

	load_ckpt(model=Gen,
	optimizer=g_optimizer,
	ckpt_path=Gen_ckpt_path,
	load_model=True,
	load_opt=False if prev_run_name in blacklist or is_freezeD or not is_train else True,
	load_misc=False,
	is_freezeD=is_freezeD)

	seed, prev_run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path, lecam_emas =\
	load_ckpt(model=Dis,
	optimizer=d_optimizer,
	ckpt_path=Dis_ckpt_path,
	load_model=True,
	load_opt=False if prev_run_name in blacklist or is_freezeD or not is_train else True,
	load_misc=True,
	is_freezeD=is_freezeD)

	if apply_g_ema:
	z = join(ckpt_dir, "model=G_ema-{when}-weights-step=".format(when=when))
	Gen_ema_ckpt_path = glob.glob(glob.escape(z) + '*.pth')[0]
	load_ckpt(model=Gen_ema,
	optimizer=None,
	ckpt_path=Gen_ema_ckpt_path,
	load_model=True,
	load_opt=False,
	load_misc=False,
	is_freezeD=is_freezeD)

	ema.source, ema.target = Gen, Gen_ema

	if is_train and RUN.seed != seed:
	RUN.seed = seed + global_rank
	misc.fix_seed(RUN.seed)

	if device == 0:
	if not is_freezeD:
	logger = log.make_logger(RUN.save_dir, prev_run_name, None)

	logger.info("Generator checkpoint is {}".format(Gen_ckpt_path))
	if apply_g_ema:
	logger.info("EMA_Generator checkpoint is {}".format(Gen_ema_ckpt_path))
	logger.info("Discriminator checkpoint is {}".format(Dis_ckpt_path))

	if is_freezeD:
	prev_run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path =\
	run_name, 0, 0, "initialize", None, 0, None, None
	return prev_run_name, step, epoch, topk, aa_p, best_step, best_fid, best_ckpt_path, lecam_emas, logger


	def load_best_model(ckpt_dir, Gen, Dis, apply_g_ema, Gen_ema, ema):
	Gen, Dis, Gen_ema = misc.peel_models(Gen, Dis, Gen_ema)
	Gen_glob = glob.glob(join(ckpt_dir, "model=G-best-weights-step*.pth"))
	Dis_glob = glob.glob(join(ckpt_dir, "model=D-best-weights-step*.pth"))
	if not Gen_glob and not Dis_glob and apply_g_ema:
	z_glob = glob.glob(join(ckpt_dir, "model=G_ema-best-weights-step*.pth"))
	if not z_glob:
	z_glob = glob.glob(join(ckpt_dir, "model=G_ema-current-weights-step*.pth"))
	if z_glob:
	Gen_ema_ckpt_path = sorted(z_glob)[-1]
	print("[ckpt] G/D best ckpts missing -> G_ema-only inference load:", Gen_ema_ckpt_path)
	os.environ.setdefault("STUDIOGAN_LOAD_STRICT", "0")
	load_ckpt(model=Gen_ema, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
	load_model=True, load_opt=False, load_misc=False)
	try:
	load_ckpt(model=Gen, optimizer=None, ckpt_path=Gen_ema_ckpt_path,
	load_model=True, load_opt=False, load_misc=False)
	except Exception as _e:
	print("[ckpt] (skipping Gen mirror; not needed for sampling):", _e)
	ema.source, ema.target = Gen, Gen_ema
	try:
	step_str = os.path.basename(Gen_ema_ckpt_path).split("step=")[1].split(".")[0]
	return int(step_str)
	except Exception:
	return 0
	Gen_ckpt_path = Gen_glob[0]
	Dis_ckpt_path = Dis_glob[0]

	load_ckpt(model=Gen,
	optimizer=None,
	ckpt_path=Gen_ckpt_path,
	load_model=True,
	load_opt=False,
	load_misc=False,
	is_freezeD=False)


	_, _, _, _, _, _, best_step, _, _, _ = load_ckpt(model=Dis,
	optimizer=None,
	ckpt_path=Dis_ckpt_path,
	load_model=True,
	load_opt=False,
	load_misc=True,
	is_freezeD=False)

	if apply_g_ema:
	Gen_ema_ckpt_path = glob.glob(join(ckpt_dir, "model=G_ema-best-weights-step*.pth"))[0]
	load_ckpt(model=Gen_ema,
	optimizer=None,
	ckpt_path=Gen_ema_ckpt_path,
	load_model=True,
	load_opt=False,
	load_misc=False,
	is_freezeD=False)

	ema.source, ema.target = Gen, Gen_ema
	return best_step


	def load_prev_dict(directory, file_name):
	return np.load(join(directory, file_name), allow_pickle=True).item()


	def check_is_pre_trained_model(ckpt_dir, GAN_train, GAN_test):
	assert GAN_train*GAN_test == 0, "cannot conduct GAN_train and GAN_test togather."
	if GAN_train:
	mode = "fake_trained"
	else:
	mode = "real_trained"

	ckpt_list = glob.glob(join(ckpt_dir, "model=C-{mode}-best-weights.pth".format(mode=mode)))
	if len(ckpt_list) == 0:
	is_pre_train_model = False
	else:
	is_pre_train_model = True
	return is_pre_train_model, mode


	def load_GAN_train_test_model(model, mode, optimizer, RUN):
	ckpt_path = join(RUN.ckpt_dir, "model=C-{mode}-best-weights.pth".format(mode=mode))
	ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage, weights_only=False)

	model.load_state_dict(ckpt["state_dict"])
	optimizer.load_state_dict(ckpt["optimizer"])
	epoch_trained = ckpt["epoch"]
	best_top1 = ckpt["best_top1"]
	best_top5 = ckpt["best_top5"]
	best_epoch = ckpt["best_epoch"]
	return epoch_trained, best_top1, best_top5, best_epoch