Shape2Force / experiments /configs.py
kaveh's picture
S2F v1.2.0
d2a9154
Raw History Blame Contribute Delete
4.1 kB
"""
Training protocol shared by the baseline and ablation runs, and the variants.
phase 1: weighted MSE + soft Dice (lambda_rec 100), lr 1e-4
phase 2: RME (lambda_rec 10) + traction-stress consistency (lambda_force 100), lr 1e-5,
starting from the best phase-1 generator
"""
import copy
PROTOCOL = dict(
train_manifest="alltune_train",
val_manifest="alltune_val",
phases=[
dict(name="phase1", epochs=100, rec_loss="wmd", lambda_rec=100.0, lambda_force=0.0, g_lr=1e-4, d_lr=1e-4),
dict(name="phase2", epochs=40, rec_loss="rme", lambda_rec=10.0, lambda_force="from_variant", g_lr=1e-5, d_lr=1e-5),
],
batch_size=2,
img_size=1024,
augment=True,
beta1=0.5, beta2=0.999,
warmup_frac=0.1,
lambda_gan=1.0,
lambda_force=100.0,
arch="s2f", in_channels=3, use_multi_scale_input=True, bridge="cbam", attention_gates=True,
substrate_mode="legacy",
input_mode="bf",
seed=0,
)
VARIANTS = {
# reference
"s2f_full": dict(),
# baselines
"unet": dict(arch="unet", lambda_gan=0.0),
"unet_gan": dict(arch="unet"),
"mask_only": dict(input_mode="mask"),
# loss ablations
"s2f_no_gan": dict(lambda_gan=0.0),
"s2f_no_force": dict(lambda_force=0.0, share_phase1_with="s2f_full"),
"s2f_gan_0.1": dict(lambda_gan=0.1),
"s2f_gan_5": dict(lambda_gan=5.0),
# input ablations
"s2f_no_substrate": dict(in_channels=1, substrate_mode="none"),
"s2f_nominal_substrate": dict(substrate_mode="nominal"),
# architecture ablations
"s2f_no_multiscale": dict(use_multi_scale_input=False),
"s2f_bridge_no_cbam": dict(bridge="cnn"),
"s2f_bridge_plain": dict(bridge="plain"),
"s2f_no_attention_gates": dict(attention_gates=False),
# bright-field only
"unet_no_substrate": dict(arch="unet", lambda_gan=0.0, in_channels=1, substrate_mode="none"),
"unet_gan_no_substrate": dict(arch="unet", in_channels=1, substrate_mode="none"),
"s2f_no_substrate_no_gan": dict(in_channels=1, substrate_mode="none", lambda_gan=0.0),
"s2f_no_substrate_no_multiscale": dict(in_channels=1, substrate_mode="none", use_multi_scale_input=False),
"s2f_no_substrate_bridge_plain": dict(in_channels=1, substrate_mode="none", bridge="plain"),
"s2f_no_substrate_no_attention_gates": dict(in_channels=1, substrate_mode="none", attention_gates=False),
# substrate channels added in phase 2 (phase 1 shared with s2f_no_substrate)
"s2f_staged": dict(share_phase1_with="s2f_no_substrate", expand_from_image_only=True),
"s2f_staged_no_force": dict(share_phase1_with="s2f_no_substrate", expand_from_image_only=True,
lambda_force=0.0),
}
PRIORITY = ["s2f_full", "unet", "unet_gan", "s2f_no_gan", "s2f_no_substrate", "s2f_no_force",
"s2f_no_attention_gates", "s2f_no_multiscale", "s2f_bridge_plain",
"s2f_bridge_no_cbam", "s2f_nominal_substrate", "s2f_gan_0.1", "s2f_gan_5"]
def get_config(variant, epochs=None, g_lr=None, d_lr=None, **overrides):
"""epochs / g_lr / d_lr override every phase (smoke tests); other overrides are top-level keys."""
if variant not in VARIANTS:
raise KeyError(f"unknown variant '{variant}'. Options: {', '.join(VARIANTS)}")
cfg = copy.deepcopy(PROTOCOL)
cfg.update(VARIANTS[variant])
cfg.update({k: v for k, v in overrides.items() if v is not None})
cfg["variant"] = variant
if cfg["substrate_mode"] == "none":
cfg["in_channels"] = 1
for ph in cfg["phases"]:
if ph["lambda_force"] == "from_variant":
ph["lambda_force"] = cfg["lambda_force"]
for k, v in (("epochs", epochs), ("g_lr", g_lr), ("d_lr", d_lr)):
if v is not None:
ph[k] = v
cfg["epochs"] = sum(ph["epochs"] for ph in cfg["phases"]) # total, used to detect finished runs
return cfg