make_random_init_ckpt.py

#!/usr/bin/env python3
"""Save an UNTRAINED stage-2 checkpoint, as the iter-0 control for eval_vsl.py.

Without this baseline, a flat generation-metric curve is ambiguous: it could mean
"training converged in the first few thousand iters" or "the transformer barely
contributes and the number reflects the pose prior". Scoring a randomly
initialized GPT through the *trained* VQ-VAE decoder separates the two.

Reuses the trained checkpoint's args so the architecture and eval path are
identical -- only the transformer weights are random.
"""
import argparse

import torch

import models.t2m_trans as trans
from models.text_encoder_vi import ViTextEncoder


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument('--like', required=True, help='trained stage-2 ckpt to copy args from')
    ap.add_argument('--out', required=True)
    ap.add_argument('--seed', type=int, default=1234)
    args = ap.parse_args()

    ref = torch.load(args.like, map_location='cpu')
    gargs = argparse.Namespace(**ref['args'])
    torch.manual_seed(args.seed)

    # text encoder is only needed for its hidden size (it is frozen either way)
    dim = ViTextEncoder(gargs.text_model, device='cpu').dim
    net = trans.Text2Motion_Transformer(
        num_vq=gargs.nb_code, embed_dim=gargs.embed_dim_gpt, clip_dim=dim,
        block_size=gargs.max_tokens + 1, num_layers=gargs.num_layers,
        n_head=gargs.n_head_gpt, drop_out_rate=gargs.drop_out_rate, fc_rate=gargs.ff_rate)

    torch.save({'trans': net.state_dict(), 'args': ref['args'], 'iter': 0,
                'val_loss': float('nan'), 'val_acc': float('nan'),
                'note': 'UNTRAINED random init - iter-0 control'}, args.out)
    print(f'wrote untrained control checkpoint -> {args.out}')


if __name__ == '__main__':
    main()