File size: 2,574 Bytes
af8c481 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 | """Parameter budget of a FACT configuration (meta device, no weights allocated).
Usage::
python scripts/param_count.py --config configs/fact_base_text.json
python scripts/param_count.py --config-json '{"dim": 768, "num_layers": 8}'
The T-Rex Track-Force baselines (Wan2.2-TI2V-5B DiT + LoRA r=4 + force stage)
are reported as fixed reference numbers; their model code is not part of this
repository.
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import torch
from torch import nn
from fact import FactConfig, FactPolicy
# Deployed / trainable parameter counts of the baselines (same counting rule,
# frozen umT5-XXL, CLIP ViT-H and the VAE excluded), in millions.
BASELINE_PARAMS_M = {
"trex_track_force (A, track+force)": {"deployed_total": 5686.4, "trainable": 120.2},
"trex notrack (B)": {"deployed_total": 5686.4, "trainable": 109.0},
}
def count(module: nn.Module, trainable_only: bool = False) -> int:
return sum(p.numel() for p in module.parameters() if p.requires_grad or not trainable_only)
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument("--config", default=None, help="FactConfig JSON file (configs/*.json)")
parser.add_argument("--config-json", default=None, help="inline JSON overrides applied after --config")
args = parser.parse_args()
cfg = FactConfig()
overrides: dict = {}
if args.config:
overrides.update(json.loads(Path(args.config).read_text()))
if args.config_json:
overrides.update(json.loads(args.config_json))
for key, value in overrides.items():
if not hasattr(cfg, key):
raise KeyError(f"unknown FactConfig field: {key}")
setattr(cfg, key, value)
with torch.device("meta"):
fact = FactPolicy(cfg)
report = {
"fact": {
"coarse_total": count(fact.coarse),
"force_stage": 0 if fact.force is None else count(fact.force),
"deployed_total": count(fact),
"trainable": count(fact),
}
}
print("fact")
for key, value in report["fact"].items():
print(f" {key:20s} {value / 1e6:10.2f} M")
for name, rec in BASELINE_PARAMS_M.items():
print(name)
for key, value in rec.items():
print(f" {key:20s} {value:10.2f} M (reference)")
print(json.dumps({**report, "baselines_M": BASELINE_PARAMS_M, "config": overrides}, indent=1))
if __name__ == "__main__":
main()
|