| """Per-head CE (text + 12 codebooks) of a checkpoint on a fixed set of single-speaker batches. |
| usage: per_head_ce.py <model_path> [n_batches]""" |
| import json, random, sys |
|
|
| sys.path.insert(0, "/data/mukil/datasets/fused/tts/moss_tts_local_v1.5/finetuning") |
| import numpy as np |
| import torch |
| import torch.nn.functional as F |
| from transformers import AutoModel |
|
|
| import sft |
| from dataset import SingleSpeakerContextDataset, load_prompt_index |
|
|
| model_path = sys.argv[1] |
| n_batches = int(sys.argv[2]) if len(sys.argv) > 2 else 16 |
| random.seed(0); torch.manual_seed(0) |
| processor = sft.load_processor_with_codec_kwargs( |
| "OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5", codec_path="OpenMOSS-Team/MOSS-Audio-Tokenizer-v2", |
| codec_weight_dtype="fp32", codec_compute_dtype="bf16", codec_attention_implementation="sdpa", |
| load_audio_tokenizer=False) |
| path = "/data/mukil/datasets/fused/jsonl/train-00120-of-00121.jsonl" |
| records = [dict(json.loads(l), _base_dir="/data/mukil/datasets/fused/jsonl") for l in open(path)] |
| records = random.Random(0).sample(records, n_batches * 4) |
| |
| ds = SingleSpeakerContextDataset(records, processor, load_prompt_index("/data/mukil/datasets/fused/prompt_index")) |
| random.seed(1) |
| batches = [ds.collate_fn([ds[i] for i in range(b * 4, b * 4 + 4)]) for b in range(n_batches)] |
|
|
| model = AutoModel.from_pretrained(model_path, trust_remote_code=True, dtype=torch.bfloat16, |
| attn_implementation="sdpa").cuda().eval() |
| losses = [] |
| orig_ce = F.cross_entropy |
| def recording_ce(*a, **k): |
| out = orig_ce(*a, **k); losses.append(float(out)); return out |
| F.cross_entropy = recording_ce |
| rows = [] |
| with torch.no_grad(): |
| for batch in batches: |
| losses.clear() |
| total = sft.compute_supervised_loss( |
| model, input_ids=batch["input_ids"].cuda(), attention_mask=batch["attention_mask"].cuda(), |
| labels=batch["labels"].cuda(), channelwise_loss_weight=[1.0] + [32 / 12] * 12) |
| rows.append([float(total)] + losses) |
| r = np.array(rows).mean(0) |
| print(f"{model_path}\n weighted_total={r[0]:.4f} text/eos={r[1]:.4f}") |
| print(" codebooks: " + " ".join(f"vq{i}={v:.3f}" for i, v in enumerate(r[2:]))) |
| print(f" uniform over 1024 codes = {np.log(1024):.3f}") |
|
|
