File size: 2,548 Bytes
c6f3061 | 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 | import argparse
import json
import torch
from .data import make_corpus
from .engine import train, evaluate_checkpoint, load_checkpoint
from .model import generate
def main():
parser = argparse.ArgumentParser(description="Train and explain a small diffusion text model")
sub = parser.add_subparsers(dest="command", required=True)
corpus = sub.add_parser("prepare", help="Generate the original controlled English grammar corpus")
corpus.add_argument("--output", default="data/synthetic.txt")
training = sub.add_parser("train")
training.add_argument("--data", default="data/synthetic.txt")
training.add_argument("--output", default="runs/demo")
for name, default in [("steps", 600), ("batch-size", 32), ("eval-every", 100), ("seed", 42), ("width", 64), ("heads", 4), ("layers", 2), ("length", 64), ("threads", 4)]:
training.add_argument("--" + name, type=int, default=default)
training.add_argument("--lr", type=float, default=0.001)
training.add_argument("--device", choices=["cpu", "cuda", "auto"], default="cpu")
evaluation = sub.add_parser("evaluate")
evaluation.add_argument("--data", default="data/synthetic.txt")
sampling = sub.add_parser("sample")
sampling.add_argument("--template", default="The cat ~~~~~ in the garden today.")
sampling.add_argument("--steps", type=int, default=12)
sampling.add_argument("--temperature", type=float, default=0.0)
sampling.add_argument("--seed", type=int, default=42)
serving = sub.add_parser("serve")
serving.add_argument("--port", type=int, default=7860)
for cmd in (evaluation, sampling, serving):
cmd.add_argument("--checkpoint", default="runs/demo/best.pt")
cmd.add_argument("--device", choices=["cpu", "cuda"], default="cpu")
cmd.add_argument("--threads", type=int, default=4)
args = parser.parse_args()
if args.command == "prepare":
print(f"Wrote {make_corpus(args.output)} unique sentences to {args.output}")
elif args.command == "train":
train(args)
elif args.command == "evaluate":
evaluate_checkpoint(args)
else:
torch.set_num_threads(args.threads)
model, checkpoint = load_checkpoint(args.checkpoint, args.device)
if args.command == "sample":
print(json.dumps(generate(model, args.template, args.steps, args.temperature, args.seed), indent=2, ensure_ascii=False))
else:
from .demo import serve
serve(model, checkpoint, args.port)
if __name__ == "__main__":
main()
|