| """Extract SGPT-pooled final pre-logit representations + mean sentence NLL for one model. |
| hidden_states[-1] from a HF causal LM is the final pre-logit hidden state (post final |
| layer-norm for every architecture used here), which is the layer the paper reports: |
| per-layer trajectories are not comparable when depths differ (16/22/24/28/32). |
| """ |
| import argparse, json, os, sys |
| import numpy as np, torch |
| torch.set_num_threads(8) |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
| def sgpt_pool(h, mask): |
| """SGPT position-weighted mean pooling: weight token i by its 1-based position.""" |
| w = torch.arange(1, h.shape[1] + 1, device=h.device, dtype=h.dtype)[None, :] * mask |
| return (h * w[..., None]).sum(1) / w.sum(1, keepdim=True).clamp(min=1e-6) |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--repo"); ap.add_argument("--revision", default=None) |
| ap.add_argument("--sents"); ap.add_argument("--out") |
| ap.add_argument("--bs", type=int, default=16); ap.add_argument("--maxlen", type=int, default=256) |
| a = ap.parse_args() |
| if os.path.exists(a.out): |
| print("SKIP", a.out); return |
| sents = json.load(open(a.sents)) |
| tok = AutoTokenizer.from_pretrained(a.repo, revision=a.revision) |
| if tok.pad_token is None: tok.pad_token = tok.eos_token or tok.unk_token |
| model = AutoModelForCausalLM.from_pretrained( |
| a.repo, revision=a.revision, torch_dtype=torch.float16, low_cpu_mem_usage=True).cuda().eval() |
| reps, nlls, ntok = [], [], [] |
| with torch.no_grad(): |
| for i in range(0, len(sents), a.bs): |
| b = tok(sents[i:i + a.bs], return_tensors="pt", padding=True, |
| truncation=True, max_length=a.maxlen) |
| ids = b["input_ids"].cuda(); m = b["attention_mask"].cuda().to(torch.float16) |
| o = model(input_ids=ids, attention_mask=b["attention_mask"].cuda(), |
| output_hidden_states=True) |
| reps.append(sgpt_pool(o.hidden_states[-1].float(), m).cpu().numpy()) |
| lg = o.logits[:, :-1].float().log_softmax(-1) |
| tgt = ids[:, 1:] |
| lp = lg.gather(-1, tgt[..., None])[..., 0] |
| mm = m[:, 1:] |
| nlls.append((-(lp * mm).sum(1)).cpu().numpy()); ntok.append(mm.sum(1).cpu().numpy()) |
| reps = np.concatenate(reps); nll = np.concatenate(nlls); nt = np.concatenate(ntok) |
| os.makedirs(os.path.dirname(a.out), exist_ok=True) |
| np.savez(a.out, reps=reps.astype(np.float32), nll_sum=nll, ntok=nt) |
| print("DONE", a.out, reps.shape, "tok_nll", float(nll.sum() / nt.sum()), |
| "sent_nll", float(nll.mean()), flush=True) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
