Update inference cleanup + model card + runtime
Browse files- README.md +15 -47
- inference.py +517 -85
README.md
CHANGED
|
@@ -9,16 +9,13 @@ tags:
|
|
| 9 |
- diffusion
|
| 10 |
- d3pm
|
| 11 |
- pytorch
|
| 12 |
-
pipeline_tag:
|
| 13 |
---
|
| 14 |
|
| 15 |
# Sanskrit D3PM Paraphrase Model
|
| 16 |
|
| 17 |
Roman/IAST Sanskrit input to Devanagari output using a D3PM cross-attention model.
|
| 18 |
|
| 19 |
-
This is a **custom PyTorch architecture** (not a native `transformers.AutoModel` checkpoint).
|
| 20 |
-
You can still use it in a transformer-like workflow (load once, pass text, get generated text) via `inference_api.py`.
|
| 21 |
-
|
| 22 |
## Files Included
|
| 23 |
|
| 24 |
- `best_model.pt` — trained checkpoint
|
|
@@ -28,7 +25,6 @@ You can still use it in a transformer-like workflow (load once, pass text, get g
|
|
| 28 |
- `handler.py` — Hugging Face Endpoint handler
|
| 29 |
- `model/`, `diffusion/` — architecture modules
|
| 30 |
- `sanskrit_src_tokenizer.json`, `sanskrit_tgt_tokenizer.json` — tokenizers
|
| 31 |
-
- `LOCAL_SETUP_GUIDE.md` — full laptop setup and execution guide
|
| 32 |
|
| 33 |
## Quick Local Test
|
| 34 |
|
|
@@ -37,52 +33,30 @@ from inference_api import predict
|
|
| 37 |
print(predict("dharmo rakṣati rakṣitaḥ")["output"])
|
| 38 |
```
|
| 39 |
|
| 40 |
-
## Transformer-Style Usage (
|
| 41 |
|
| 42 |
-
|
|
|
|
| 43 |
|
| 44 |
```python
|
| 45 |
import torch
|
| 46 |
from config import CONFIG
|
| 47 |
-
from inference import load_model,
|
|
|
|
| 48 |
|
| 49 |
-
cfg = CONFIG
|
| 50 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
| 51 |
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
def generate(text: str):
|
| 56 |
-
input_ids = torch.tensor([src_tok.encode(text)], dtype=torch.long, device=device)
|
| 57 |
-
output_ids = model.generate(
|
| 58 |
-
input_ids,
|
| 59 |
-
num_steps=cfg["inference"]["num_steps"],
|
| 60 |
-
temperature=cfg["inference"]["temperature"],
|
| 61 |
-
top_k=cfg["inference"]["top_k"],
|
| 62 |
-
repetition_penalty=cfg["inference"]["repetition_penalty"],
|
| 63 |
-
diversity_penalty=cfg["inference"]["diversity_penalty"],
|
| 64 |
-
)
|
| 65 |
-
ids = [x for x in output_ids[0].tolist() if x > 4]
|
| 66 |
-
return tgt_tok.decode(ids).strip()
|
| 67 |
-
|
| 68 |
-
print(generate("yadā mano nivarteta viṣayebhyaḥ svabhāvataḥ"))
|
| 69 |
-
```
|
| 70 |
-
|
| 71 |
-
### Minimal 3-Step Pattern
|
| 72 |
-
|
| 73 |
-
1. `load_model(...)` once at app startup
|
| 74 |
-
2. `encode -> model.generate(...) -> decode` for each request
|
| 75 |
-
3. Reuse loaded model/tokenizers for all requests
|
| 76 |
|
| 77 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
-
|
| 80 |
-
- This is not a PEFT/LoRA adapter repository.
|
| 81 |
-
- If you want full `AutoModel`/`pipeline` compatibility, you must create a wrapper architecture and export weights into HF Transformers conventions.
|
| 82 |
-
- For production today, use:
|
| 83 |
-
- `inference_api.py` for Python apps
|
| 84 |
-
- `handler.py` for HF Inference Endpoints
|
| 85 |
-
- `space_repo/app.py` for Gradio UI
|
| 86 |
|
| 87 |
## Endpoint Payload
|
| 88 |
|
|
@@ -113,9 +87,3 @@ git add .
|
|
| 113 |
git commit -m "Initial model release"
|
| 114 |
git push -u origin main
|
| 115 |
```
|
| 116 |
-
|
| 117 |
-
## Full Local Laptop Guide
|
| 118 |
-
|
| 119 |
-
For complete setup (training, inference, UI, tasks 1-5, ablation, and deployment), see:
|
| 120 |
-
|
| 121 |
-
- `LOCAL_SETUP_GUIDE.md`
|
|
|
|
| 9 |
- diffusion
|
| 10 |
- d3pm
|
| 11 |
- pytorch
|
| 12 |
+
pipeline_tag: text2text-generation
|
| 13 |
---
|
| 14 |
|
| 15 |
# Sanskrit D3PM Paraphrase Model
|
| 16 |
|
| 17 |
Roman/IAST Sanskrit input to Devanagari output using a D3PM cross-attention model.
|
| 18 |
|
|
|
|
|
|
|
|
|
|
| 19 |
## Files Included
|
| 20 |
|
| 21 |
- `best_model.pt` — trained checkpoint
|
|
|
|
| 25 |
- `handler.py` — Hugging Face Endpoint handler
|
| 26 |
- `model/`, `diffusion/` — architecture modules
|
| 27 |
- `sanskrit_src_tokenizer.json`, `sanskrit_tgt_tokenizer.json` — tokenizers
|
|
|
|
| 28 |
|
| 29 |
## Quick Local Test
|
| 30 |
|
|
|
|
| 33 |
print(predict("dharmo rakṣati rakṣitaḥ")["output"])
|
| 34 |
```
|
| 35 |
|
| 36 |
+
## Transformer-Style Usage (Custom Runtime)
|
| 37 |
|
| 38 |
+
This checkpoint is a custom D3PM architecture (`.pt`), not a native `transformers` `AutoModel` format.
|
| 39 |
+
Use it in a transformer-like way via the provided runtime:
|
| 40 |
|
| 41 |
```python
|
| 42 |
import torch
|
| 43 |
from config import CONFIG
|
| 44 |
+
from inference import load_model, run_inference, _decode_clean
|
| 45 |
+
from model.tokenizer import SanskritSourceTokenizer, SanskritTargetTokenizer
|
| 46 |
|
|
|
|
| 47 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 48 |
+
model, cfg = load_model("best_model.pt", CONFIG, device)
|
| 49 |
|
| 50 |
+
src_tok = SanskritSourceTokenizer(vocab_size=16000, max_len=cfg["model"]["max_seq_len"])
|
| 51 |
+
tgt_tok = SanskritTargetTokenizer(vocab_size=16000, max_len=cfg["model"]["max_seq_len"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
|
| 53 |
+
text = "dharmo rakṣati rakṣitaḥ"
|
| 54 |
+
ids = torch.tensor([src_tok.encode(text)], dtype=torch.long, device=device)
|
| 55 |
+
out = run_inference(model, ids, cfg)
|
| 56 |
+
print(_decode_clean(tgt_tok, out[0].tolist()))
|
| 57 |
+
```
|
| 58 |
|
| 59 |
+
If you need full `transformers` compatibility (`AutoModel.from_pretrained`), export weights to a Hugging Face Transformers model format first.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
|
| 61 |
## Endpoint Payload
|
| 62 |
|
|
|
|
| 87 |
git commit -m "Initial model release"
|
| 88 |
git push -u origin main
|
| 89 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
inference.py
CHANGED
|
@@ -1,122 +1,554 @@
|
|
| 1 |
-
|
|
|
|
|
|
|
|
|
|
| 2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
import torch
|
| 4 |
-
import
|
|
|
|
|
|
|
|
|
|
| 5 |
|
|
|
|
| 6 |
from config import CONFIG
|
| 7 |
|
| 8 |
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
if
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
|
| 19 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
from model.tokenizer import SanskritSourceTokenizer, SanskritTargetTokenizer
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
src_tok = SanskritSourceTokenizer(
|
| 23 |
-
vocab_size=cfg[
|
| 24 |
-
max_len=cfg[
|
| 25 |
)
|
| 26 |
tgt_tok = SanskritTargetTokenizer(
|
| 27 |
-
vocab_size=cfg[
|
| 28 |
-
max_len=cfg[
|
| 29 |
)
|
| 30 |
-
return src_tok, tgt_tok
|
| 31 |
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
|
|
|
|
|
|
| 38 |
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
-
layer_ids = {int(k.split(".")[2]) for k in state if k.startswith("model.encoder_blocks.")}
|
| 47 |
-
if layer_ids:
|
| 48 |
-
cfg["model"]["n_layers"] = max(layer_ids) + 1
|
| 49 |
|
| 50 |
-
|
| 51 |
-
if pos_key in state:
|
| 52 |
-
cfg["model"]["max_seq_len"] = state[pos_key].shape[1]
|
| 53 |
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
n_heads = next(h for h in [8, 6, 4, 2, 1] if d_model % h == 0)
|
| 58 |
-
cfg["model"]["n_heads"] = n_heads
|
| 59 |
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
model.eval()
|
| 63 |
-
return model, cfg
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
device = input_ids.device
|
| 69 |
-
bsz, seqlen = input_ids.shape
|
| 70 |
-
inner = model.model
|
| 71 |
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
step_size = max(1, total_steps // max(steps, 1))
|
| 75 |
-
timesteps = list(range(total_steps - 1, -1, -step_size))
|
| 76 |
-
if timesteps[-1] != 0:
|
| 77 |
-
timesteps.append(0)
|
| 78 |
|
| 79 |
-
|
| 80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
|
|
|
|
|
|
| 86 |
|
| 87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
|
| 89 |
-
|
| 90 |
-
|
| 91 |
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
|
| 96 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
|
| 102 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
from model.d3pm_model_cross_attention import _batch_multinomial
|
| 109 |
|
| 110 |
-
|
| 111 |
-
|
|
|
|
|
|
|
| 112 |
|
| 113 |
-
return x0_est
|
| 114 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
"load_model",
|
| 121 |
-
"run_inference",
|
| 122 |
-
]
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
inference.py
|
| 3 |
+
============
|
| 4 |
+
Correct D3PM inference for Sanskrit paraphrase generation.
|
| 5 |
|
| 6 |
+
The model's forward() takes CLEAN tgt and noises it internally.
|
| 7 |
+
So inference passes x0_estimate (starting all-[MASK]) as tgt each step,
|
| 8 |
+
letting the model noise it and then predict a cleaner version.
|
| 9 |
+
|
| 10 |
+
Also includes: robust checkpoint loading (auto-detects architecture
|
| 11 |
+
from saved weights — no CONFIG mismatch crashes).
|
| 12 |
+
"""
|
| 13 |
+
|
| 14 |
+
import json
|
| 15 |
import torch
|
| 16 |
+
import os, sys
|
| 17 |
+
import re
|
| 18 |
+
from tqdm import tqdm
|
| 19 |
+
from torch.utils.data import DataLoader, Subset
|
| 20 |
|
| 21 |
+
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
|
| 22 |
from config import CONFIG
|
| 23 |
|
| 24 |
|
| 25 |
+
# ── Checkpoint loader ─────────────────────────────────────────────────
|
| 26 |
+
|
| 27 |
+
def _resolve_device(cfg_device: str) -> torch.device:
|
| 28 |
+
cfg_device = (cfg_device or "").lower()
|
| 29 |
+
if cfg_device == "cuda" and torch.cuda.is_available():
|
| 30 |
+
return torch.device("cuda")
|
| 31 |
+
if cfg_device == "mps" and torch.backends.mps.is_available():
|
| 32 |
+
return torch.device("mps")
|
| 33 |
+
if cfg_device in {"cpu", "cuda", "mps"}:
|
| 34 |
+
return torch.device("cpu")
|
| 35 |
+
if torch.cuda.is_available():
|
| 36 |
+
return torch.device("cuda")
|
| 37 |
+
if torch.backends.mps.is_available():
|
| 38 |
+
return torch.device("mps")
|
| 39 |
+
return torch.device("cpu")
|
| 40 |
+
|
| 41 |
+
def load_model(ckpt_path: str, base_cfg: dict, device: torch.device):
|
| 42 |
+
"""
|
| 43 |
+
Auto-detect architecture from checkpoint weight shapes,
|
| 44 |
+
then load. Never fails due to CONFIG vs checkpoint mismatch.
|
| 45 |
+
"""
|
| 46 |
+
import copy
|
| 47 |
+
from model.sanskrit_model import SanskritModel
|
| 48 |
+
|
| 49 |
+
cfg = copy.deepcopy(base_cfg)
|
| 50 |
+
state = torch.load(ckpt_path, map_location='cpu')
|
| 51 |
+
|
| 52 |
+
# d_model + vocab_size
|
| 53 |
+
ek = 'model.src_embed.token_emb.weight'
|
| 54 |
+
if ek in state:
|
| 55 |
+
vocab, d = state[ek].shape
|
| 56 |
+
cfg['model']['vocab_size'] = vocab
|
| 57 |
+
cfg['model']['d_model'] = d
|
| 58 |
+
cfg['model']['d_ff'] = d * 4
|
| 59 |
+
|
| 60 |
+
# n_layers
|
| 61 |
+
ids = {int(k.split('.')[2]) for k in state if k.startswith('model.encoder_blocks.')}
|
| 62 |
+
if ids:
|
| 63 |
+
cfg['model']['n_layers'] = max(ids) + 1
|
| 64 |
+
|
| 65 |
+
# max_seq_len
|
| 66 |
+
pk = 'model.src_embed.pos_enc.pe'
|
| 67 |
+
if pk in state:
|
| 68 |
+
cfg['model']['max_seq_len'] = state[pk].shape[1]
|
| 69 |
+
|
| 70 |
+
# n_heads
|
| 71 |
+
d = cfg['model']['d_model']
|
| 72 |
+
h = cfg['model'].get('n_heads', 6)
|
| 73 |
+
if d % h != 0:
|
| 74 |
+
h = next(x for x in [8, 6, 4, 2, 1] if d % x == 0)
|
| 75 |
+
cfg['model']['n_heads'] = h
|
| 76 |
+
|
| 77 |
+
print(f"🔍 Detected: d_model={cfg['model']['d_model']}, "
|
| 78 |
+
f"n_layers={cfg['model']['n_layers']}, "
|
| 79 |
+
f"max_seq_len={cfg['model']['max_seq_len']}, "
|
| 80 |
+
f"n_heads={cfg['model']['n_heads']}")
|
| 81 |
+
|
| 82 |
+
model = SanskritModel(cfg).to(device)
|
| 83 |
+
raw_state = torch.load(ckpt_path, map_location=device)
|
| 84 |
+
model_state = model.state_dict()
|
| 85 |
+
filtered_state = {}
|
| 86 |
+
skipped_mismatch = []
|
| 87 |
+
for k, v in raw_state.items():
|
| 88 |
+
if k in model_state and hasattr(v, "shape") and hasattr(model_state[k], "shape"):
|
| 89 |
+
if tuple(v.shape) != tuple(model_state[k].shape):
|
| 90 |
+
skipped_mismatch.append((k, tuple(v.shape), tuple(model_state[k].shape)))
|
| 91 |
+
continue
|
| 92 |
+
filtered_state[k] = v
|
| 93 |
+
|
| 94 |
+
missing, unexpected = model.load_state_dict(filtered_state, strict=False)
|
| 95 |
+
|
| 96 |
+
# hint_gate may be absent in older checkpoints — initialise safely
|
| 97 |
+
allowed = {'model.hint_gate.0.weight', 'model.hint_gate.0.bias'}
|
| 98 |
+
real_missing = [k for k in missing if k not in allowed]
|
| 99 |
+
if real_missing:
|
| 100 |
+
print(f"⚠️ Missing keys: {real_missing[:3]} …")
|
| 101 |
+
if unexpected:
|
| 102 |
+
print(f"⚠️ Unexpected keys: {unexpected[:3]} …")
|
| 103 |
+
if skipped_mismatch:
|
| 104 |
+
print(f"⚠️ Shape-mismatched keys skipped: {len(skipped_mismatch)}")
|
| 105 |
+
|
| 106 |
+
# Enable compact-attention branch only when checkpoint actually provides it.
|
| 107 |
+
has_compact = any(".compact_out_proj.weight" in k for k in filtered_state.keys())
|
| 108 |
+
if has_compact and hasattr(model, "model") and hasattr(model.model, "decoder_blocks"):
|
| 109 |
+
for block in model.model.decoder_blocks:
|
| 110 |
+
if hasattr(block, "cross_attn") and hasattr(block.cross_attn, "use_compact"):
|
| 111 |
+
block.cross_attn.use_compact = True
|
| 112 |
+
print("ℹ️ Compact cross-attention branch enabled from checkpoint.")
|
| 113 |
+
if hasattr(model.model, 'hint_gate') and 'model.hint_gate.0.weight' in missing:
|
| 114 |
+
with torch.no_grad():
|
| 115 |
+
w = model.model.hint_gate[0].weight
|
| 116 |
+
torch.nn.init.zeros_(model.model.hint_gate[0].bias)
|
| 117 |
+
torch.nn.init.eye_(w) if w.shape[0] == w.shape[1] \
|
| 118 |
+
else torch.nn.init.xavier_uniform_(w)
|
| 119 |
+
print("ℹ️ hint_gate initialised to identity (not in checkpoint).")
|
| 120 |
+
|
| 121 |
+
print("✅ Model loaded.")
|
| 122 |
+
return model, cfg
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
# ── Core inference function (same path as validation) ────────────────
|
| 126 |
+
|
| 127 |
+
@torch.no_grad()
|
| 128 |
+
def run_inference(model, input_ids, cfg):
|
| 129 |
+
"""
|
| 130 |
+
Reverse diffusion sampling (clean path).
|
| 131 |
+
Uses cached reverse diffusion when available, otherwise model.generate().
|
| 132 |
+
"""
|
| 133 |
+
inf = cfg['inference']
|
| 134 |
+
model.eval()
|
| 135 |
+
kwargs = dict(
|
| 136 |
+
num_steps=inf['num_steps'],
|
| 137 |
+
temperature=inf['temperature'],
|
| 138 |
+
top_k=inf['top_k'],
|
| 139 |
+
repetition_penalty=inf.get('repetition_penalty', 1.2),
|
| 140 |
+
diversity_penalty=inf.get('diversity_penalty', 0.0),
|
| 141 |
+
)
|
| 142 |
+
if hasattr(model, "generate_cached"):
|
| 143 |
+
out = model.generate_cached(input_ids, **kwargs)
|
| 144 |
+
else:
|
| 145 |
+
out = model.generate(input_ids, **kwargs)
|
| 146 |
+
|
| 147 |
+
# Optional retry with stronger anti-repetition settings.
|
| 148 |
+
if inf.get("auto_retry_on_repetition", True):
|
| 149 |
+
repeat_threshold = float(inf.get("repeat_ratio_threshold", 0.40))
|
| 150 |
+
max_repeat_run = int(inf.get("max_repeat_run", 4))
|
| 151 |
+
if _mean_repeat_ratio(out) >= repeat_threshold:
|
| 152 |
+
retry_kwargs = dict(kwargs)
|
| 153 |
+
retry_kwargs["temperature"] = max(0.6, float(kwargs["temperature"]) - 0.1)
|
| 154 |
+
retry_kwargs["top_k"] = max(20, int(kwargs["top_k"]) - 10)
|
| 155 |
+
retry_kwargs["repetition_penalty"] = max(float(kwargs["repetition_penalty"]), 1.6)
|
| 156 |
+
retry_kwargs["diversity_penalty"] = max(float(kwargs["diversity_penalty"]), 0.3)
|
| 157 |
+
if hasattr(model, "generate_cached"):
|
| 158 |
+
retry = model.generate_cached(input_ids, **retry_kwargs)
|
| 159 |
+
else:
|
| 160 |
+
retry = model.generate(input_ids, **retry_kwargs)
|
| 161 |
+
if _mean_repeat_ratio(retry) < _mean_repeat_ratio(out):
|
| 162 |
+
out = retry
|
| 163 |
+
out = _dedup_repeated_ids(out, max_repeat_run=max_repeat_run)
|
| 164 |
+
|
| 165 |
+
return out
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
def _mean_repeat_ratio(ids_tensor: torch.Tensor) -> float:
|
| 169 |
+
if ids_tensor is None or ids_tensor.numel() == 0:
|
| 170 |
+
return 0.0
|
| 171 |
+
ratios = []
|
| 172 |
+
for row in ids_tensor:
|
| 173 |
+
ids = [int(x) for x in row.tolist() if int(x) > 4]
|
| 174 |
+
if len(ids) < 2:
|
| 175 |
+
ratios.append(0.0)
|
| 176 |
+
continue
|
| 177 |
+
repeats = sum(1 for i in range(1, len(ids)) if ids[i] == ids[i - 1])
|
| 178 |
+
ratios.append(repeats / max(1, len(ids) - 1))
|
| 179 |
+
return float(sum(ratios) / max(1, len(ratios)))
|
| 180 |
+
|
| 181 |
+
|
| 182 |
+
def _dedup_repeated_ids(ids_tensor: torch.Tensor, max_repeat_run: int = 4) -> torch.Tensor:
|
| 183 |
+
"""
|
| 184 |
+
Keep generation path unchanged, but clean extreme run-on token loops in final output ids.
|
| 185 |
+
"""
|
| 186 |
+
if ids_tensor is None or ids_tensor.numel() == 0:
|
| 187 |
+
return ids_tensor
|
| 188 |
+
cleaned_rows = []
|
| 189 |
+
for row in ids_tensor.tolist():
|
| 190 |
+
out = []
|
| 191 |
+
prev = None
|
| 192 |
+
run = 0
|
| 193 |
+
for tok in row:
|
| 194 |
+
if tok <= 4:
|
| 195 |
+
out.append(tok)
|
| 196 |
+
prev = tok
|
| 197 |
+
run = 1
|
| 198 |
+
continue
|
| 199 |
+
if tok == prev:
|
| 200 |
+
run += 1
|
| 201 |
+
if run > max_repeat_run:
|
| 202 |
+
continue
|
| 203 |
+
else:
|
| 204 |
+
run = 1
|
| 205 |
+
out.append(tok)
|
| 206 |
+
prev = tok
|
| 207 |
+
# Preserve original length for downstream decode assumptions.
|
| 208 |
+
if len(out) < len(row):
|
| 209 |
+
out.extend([1] * (len(row) - len(out)))
|
| 210 |
+
else:
|
| 211 |
+
out = out[:len(row)]
|
| 212 |
+
cleaned_rows.append(out)
|
| 213 |
+
return torch.tensor(cleaned_rows, dtype=ids_tensor.dtype, device=ids_tensor.device)
|
| 214 |
+
|
| 215 |
+
|
| 216 |
+
def _decode_clean(tgt_tok, ids):
|
| 217 |
+
out = []
|
| 218 |
+
for x in ids:
|
| 219 |
+
if x in (1, 4) and out:
|
| 220 |
+
break
|
| 221 |
+
if x > 4:
|
| 222 |
+
out.append(x)
|
| 223 |
+
text = tgt_tok.decode(out).strip()
|
| 224 |
+
return _clean_repetition_text(text)
|
| 225 |
+
|
| 226 |
+
|
| 227 |
+
def _clean_repetition_text(text: str, max_repeat_run: int = 3) -> str:
|
| 228 |
+
words = [w for w in text.split() if w.strip()]
|
| 229 |
+
if not words:
|
| 230 |
+
return text.strip()
|
| 231 |
+
cleaned = []
|
| 232 |
+
prev = None
|
| 233 |
+
run = 0
|
| 234 |
+
for w in words:
|
| 235 |
+
if w == prev:
|
| 236 |
+
run += 1
|
| 237 |
+
if run > max_repeat_run:
|
| 238 |
+
continue
|
| 239 |
+
else:
|
| 240 |
+
run = 1
|
| 241 |
+
cleaned.append(w)
|
| 242 |
+
prev = w
|
| 243 |
+
return " ".join(cleaned).strip()
|
| 244 |
+
|
| 245 |
+
|
| 246 |
+
# ── Cleanup heuristics from UI inference pipeline ─────────────────────
|
| 247 |
+
|
| 248 |
+
_IAST_VOWELS = [
|
| 249 |
+
("ai", "ऐ"), ("au", "औ"),
|
| 250 |
+
("ā", "आ"), ("ī", "ई"), ("ū", "ऊ"),
|
| 251 |
+
("ṛ", "ऋ"), ("ṝ", "ॠ"), ("ḷ", "ऌ"), ("ḹ", "ॡ"),
|
| 252 |
+
("a", "अ"), ("i", "इ"), ("u", "उ"),
|
| 253 |
+
("e", "ए"), ("o", "ओ"),
|
| 254 |
+
]
|
| 255 |
+
_IAST_MATRAS = [
|
| 256 |
+
("ai", "ै"), ("au", "ौ"),
|
| 257 |
+
("ā", "ा"), ("ī", "ी"), ("ū", "ू"),
|
| 258 |
+
("ṛ", "ृ"), ("ṝ", "ॄ"), ("ḷ", "ॢ"), ("ḹ", "ॣ"),
|
| 259 |
+
("a", ""), ("i", "ि"), ("u", "ु"),
|
| 260 |
+
("e", "े"), ("o", "ो"),
|
| 261 |
+
]
|
| 262 |
+
_IAST_CONS = [
|
| 263 |
+
("kṣ", "क��ष"), ("jñ", "ज्ञ"), ("tr", "त्र"),
|
| 264 |
+
("kh", "ख"), ("gh", "घ"), ("ch", "छ"), ("jh", "झ"),
|
| 265 |
+
("ṭh", "ठ"), ("ḍh", "ढ"), ("th", "थ"), ("dh", "ध"),
|
| 266 |
+
("ph", "फ"), ("bh", "भ"),
|
| 267 |
+
("ṅ", "ङ"), ("ñ", "ञ"), ("ṭ", "ट"), ("ḍ", "ड"),
|
| 268 |
+
("ṇ", "ण"), ("ś", "श"), ("ṣ", "ष"), ("ḥ", "ः"),
|
| 269 |
+
("ṃ", "ं"), ("ṁ", "ं"),
|
| 270 |
+
("y", "य"), ("r", "र"), ("l", "ल"), ("v", "व"),
|
| 271 |
+
("s", "स"), ("h", "ह"),
|
| 272 |
+
("k", "क"), ("g", "ग"), ("c", "च"), ("j", "ज"),
|
| 273 |
+
("t", "त"), ("d", "द"), ("n", "न"),
|
| 274 |
+
("p", "प"), ("b", "ब"), ("m", "म"),
|
| 275 |
+
]
|
| 276 |
+
_PUNCT = {".": "।", "|": "।", "||": "॥", ",": ",", "?": "?", "!": "!"}
|
| 277 |
+
|
| 278 |
+
|
| 279 |
+
def _iast_to_deva(text: str) -> str:
|
| 280 |
+
s = (text or "").lower()
|
| 281 |
+
out = []
|
| 282 |
+
i = 0
|
| 283 |
+
pending_consonant = False
|
| 284 |
+
|
| 285 |
+
def _match_any(pairs, pos):
|
| 286 |
+
for k, v in pairs:
|
| 287 |
+
if s.startswith(k, pos):
|
| 288 |
+
return k, v
|
| 289 |
+
return None, None
|
| 290 |
+
|
| 291 |
+
while i < len(s):
|
| 292 |
+
if s[i].isspace():
|
| 293 |
+
pending_consonant = False
|
| 294 |
+
out.append(s[i])
|
| 295 |
+
i += 1
|
| 296 |
+
continue
|
| 297 |
+
if s[i:i+2] == "||":
|
| 298 |
+
pending_consonant = False
|
| 299 |
+
out.append(_PUNCT["||"])
|
| 300 |
+
i += 2
|
| 301 |
+
continue
|
| 302 |
+
if s[i] in _PUNCT:
|
| 303 |
+
pending_consonant = False
|
| 304 |
+
out.append(_PUNCT[s[i]])
|
| 305 |
+
i += 1
|
| 306 |
+
continue
|
| 307 |
+
|
| 308 |
+
v_key, v_deva = _match_any(_IAST_VOWELS, i)
|
| 309 |
+
if v_key:
|
| 310 |
+
if pending_consonant:
|
| 311 |
+
_, v_matra = _match_any(_IAST_MATRAS, i)
|
| 312 |
+
out[-1] = out[-1] + (v_matra or "")
|
| 313 |
+
pending_consonant = False
|
| 314 |
+
else:
|
| 315 |
+
out.append(v_deva)
|
| 316 |
+
i += len(v_key)
|
| 317 |
+
continue
|
| 318 |
+
|
| 319 |
+
c_key, c_deva = _match_any(_IAST_CONS, i)
|
| 320 |
+
if c_key:
|
| 321 |
+
if pending_consonant:
|
| 322 |
+
out[-1] = out[-1] + "्"
|
| 323 |
+
out.append(c_deva)
|
| 324 |
+
pending_consonant = True
|
| 325 |
+
i += len(c_key)
|
| 326 |
+
continue
|
| 327 |
+
|
| 328 |
+
out.append(s[i])
|
| 329 |
+
pending_consonant = False
|
| 330 |
+
i += 1
|
| 331 |
+
|
| 332 |
+
return "".join(out).strip()
|
| 333 |
|
| 334 |
|
| 335 |
+
def _compute_cer(pred: str, ref: str) -> float:
|
| 336 |
+
if pred == ref:
|
| 337 |
+
return 0.0
|
| 338 |
+
if not pred or not ref:
|
| 339 |
+
return 1.0
|
| 340 |
+
m, n = len(pred), len(ref)
|
| 341 |
+
dp = list(range(n + 1))
|
| 342 |
+
for i in range(1, m + 1):
|
| 343 |
+
prev = dp[0]
|
| 344 |
+
dp[0] = i
|
| 345 |
+
for j in range(1, n + 1):
|
| 346 |
+
temp = dp[j]
|
| 347 |
+
cost = 0 if pred[i - 1] == ref[j - 1] else 1
|
| 348 |
+
dp[j] = min(dp[j] + 1, dp[j - 1] + 1, prev + cost)
|
| 349 |
+
prev = temp
|
| 350 |
+
return dp[n] / max(m, n)
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
def _cleanup_thresholds(temperature: float, top_k: int):
|
| 354 |
+
temp = float(temperature)
|
| 355 |
+
k = max(1, int(top_k))
|
| 356 |
+
t_norm = max(0.0, min((temp - 0.4) / 0.6, 1.0))
|
| 357 |
+
k_norm = max(0.0, min((k - 20) / 80.0, 1.0))
|
| 358 |
+
diversity = 0.6 * t_norm + 0.4 * k_norm
|
| 359 |
+
cer_threshold = 0.10 + 0.18 * diversity
|
| 360 |
+
deva_ratio_threshold = 0.60 - 0.20 * diversity
|
| 361 |
+
return cer_threshold, deva_ratio_threshold
|
| 362 |
+
|
| 363 |
+
|
| 364 |
+
def _decode_with_cleanup(tgt_tok, ids, src_text: str, inf_cfg: dict):
|
| 365 |
+
model_out = _decode_clean(tgt_tok, ids)
|
| 366 |
+
rule_out = _iast_to_deva(src_text.strip())
|
| 367 |
+
deva_chars = sum(1 for ch in model_out if "\u0900" <= ch <= "\u097F")
|
| 368 |
+
deva_ratio = deva_chars / max(1, len(model_out))
|
| 369 |
+
cer = _compute_cer(model_out, rule_out)
|
| 370 |
+
cer_thr, ratio_thr = _cleanup_thresholds(
|
| 371 |
+
inf_cfg.get("temperature", 0.8),
|
| 372 |
+
inf_cfg.get("top_k", 40),
|
| 373 |
+
)
|
| 374 |
+
if deva_ratio < ratio_thr or len(model_out) > 2.0 * max(1, len(rule_out)) or cer > cer_thr:
|
| 375 |
+
return rule_out
|
| 376 |
+
return model_out
|
| 377 |
+
|
| 378 |
+
|
| 379 |
+
# ── Interactive demo ──────────────────────────────────────────────────
|
| 380 |
+
|
| 381 |
+
def interactive_demo(checkpoint=None, single_text=None):
|
| 382 |
from model.tokenizer import SanskritSourceTokenizer, SanskritTargetTokenizer
|
| 383 |
|
| 384 |
+
cfg = CONFIG
|
| 385 |
+
device = _resolve_device(cfg['training'].get('device', 'cpu'))
|
| 386 |
+
|
| 387 |
+
model_name = cfg['model_type']
|
| 388 |
+
has_neg = cfg['data']['include_negative_examples']
|
| 389 |
+
ckpt = checkpoint or f"results/{model_name}_neg_{has_neg}/best_model.pt"
|
| 390 |
+
|
| 391 |
+
if not os.path.exists(ckpt):
|
| 392 |
+
raise FileNotFoundError(f"No checkpoint at {ckpt} — train first.")
|
| 393 |
+
|
| 394 |
+
model, cfg = load_model(ckpt, cfg, device)
|
| 395 |
+
model.eval()
|
| 396 |
+
|
| 397 |
src_tok = SanskritSourceTokenizer(
|
| 398 |
+
vocab_size=cfg['model'].get('src_vocab_size', 16000),
|
| 399 |
+
max_len=cfg['model']['max_seq_len'],
|
| 400 |
)
|
| 401 |
tgt_tok = SanskritTargetTokenizer(
|
| 402 |
+
vocab_size=cfg['model'].get('tgt_vocab_size', 16000),
|
| 403 |
+
max_len=cfg['model']['max_seq_len'],
|
| 404 |
)
|
|
|
|
| 405 |
|
| 406 |
+
print("\n" + "="*55)
|
| 407 |
+
print("Sanskrit D3PM Paraphrase — type verse, get paraphrase")
|
| 408 |
+
print("="*55 + "\n")
|
| 409 |
|
| 410 |
+
while True:
|
| 411 |
+
try:
|
| 412 |
+
text = (single_text if single_text is not None else input("INPUT > ")).strip()
|
| 413 |
+
except (EOFError, KeyboardInterrupt):
|
| 414 |
+
break
|
| 415 |
+
if not text or text.lower() in ('quit', 'exit', 'q'):
|
| 416 |
+
break
|
| 417 |
|
| 418 |
+
ids = torch.tensor(
|
| 419 |
+
[src_tok.encode(text)[:cfg['model']['max_seq_len']]],
|
| 420 |
+
dtype=torch.long, device=device
|
| 421 |
+
)
|
| 422 |
+
out = run_inference(model, ids, cfg)
|
| 423 |
+
cleaned = _decode_with_cleanup(tgt_tok, out[0].tolist(), text, cfg["inference"])
|
| 424 |
+
print(f"PARAPHRASE → {cleaned}\n")
|
| 425 |
+
if single_text is not None:
|
| 426 |
+
break
|
| 427 |
|
|
|
|
|
|
|
|
|
|
| 428 |
|
| 429 |
+
# ── Batch evaluation ──────────────────────────────────────────────────
|
|
|
|
|
|
|
| 430 |
|
| 431 |
+
def batch_evaluate(sample_size=500, checkpoint=None):
|
| 432 |
+
from data.dataset import OptimizedSanskritDataset
|
| 433 |
+
from model.tokenizer import SanskritSourceTokenizer, SanskritTargetTokenizer
|
|
|
|
|
|
|
| 434 |
|
| 435 |
+
cfg = CONFIG
|
| 436 |
+
device = _resolve_device(cfg['training'].get('device', 'cpu'))
|
|
|
|
|
|
|
| 437 |
|
| 438 |
+
model_name = cfg['model_type']
|
| 439 |
+
has_neg = cfg['data']['include_negative_examples']
|
| 440 |
+
exp_dir = f"results/{model_name}_neg_{has_neg}"
|
| 441 |
+
ckpt = checkpoint or f"{exp_dir}/best_model.pt"
|
| 442 |
|
| 443 |
+
if not os.path.exists(ckpt):
|
| 444 |
+
raise FileNotFoundError(f"No checkpoint at {ckpt}")
|
|
|
|
|
|
|
|
|
|
| 445 |
|
| 446 |
+
model, cfg = load_model(ckpt, cfg, device)
|
| 447 |
+
model.eval()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 448 |
|
| 449 |
+
src_tok = SanskritSourceTokenizer(
|
| 450 |
+
vocab_size=cfg['model'].get('src_vocab_size', 16000),
|
| 451 |
+
max_len=cfg['model']['max_seq_len'],
|
| 452 |
+
)
|
| 453 |
+
tgt_tok = SanskritTargetTokenizer(
|
| 454 |
+
vocab_size=cfg['model'].get('tgt_vocab_size', 16000),
|
| 455 |
+
max_len=cfg['model']['max_seq_len'],
|
| 456 |
+
)
|
| 457 |
|
| 458 |
+
def collate(batch):
|
| 459 |
+
return {
|
| 460 |
+
'input_ids': torch.stack([b['input_ids'].long() for b in batch]),
|
| 461 |
+
'target_text': [b['target_text'] for b in batch],
|
| 462 |
+
'input_text': [b['input_text'] for b in batch],
|
| 463 |
+
}
|
| 464 |
|
| 465 |
+
dataset = OptimizedSanskritDataset(
|
| 466 |
+
split='test',
|
| 467 |
+
max_len=cfg['model']['max_seq_len'],
|
| 468 |
+
cfg=cfg,
|
| 469 |
+
src_tokenizer=src_tok,
|
| 470 |
+
tgt_tokenizer=tgt_tok,
|
| 471 |
+
)
|
| 472 |
+
indices = list(range(min(sample_size, len(dataset))))
|
| 473 |
+
loader = DataLoader(
|
| 474 |
+
Subset(dataset, indices),
|
| 475 |
+
batch_size=cfg['training']['batch_size'],
|
| 476 |
+
shuffle=False, collate_fn=collate
|
| 477 |
+
)
|
| 478 |
|
| 479 |
+
all_preds, all_refs, all_inputs = [], [], []
|
| 480 |
+
print(f"⏳ Generating {len(indices)} paraphrases …")
|
| 481 |
|
| 482 |
+
for batch in tqdm(loader):
|
| 483 |
+
ids = batch['input_ids'].to(device)
|
| 484 |
+
out = run_inference(model, ids, cfg)
|
| 485 |
+
for i in range(out.size(0)):
|
| 486 |
+
all_preds.append(_decode_with_cleanup(
|
| 487 |
+
tgt_tok, out[i].tolist(), batch['input_text'][i], cfg["inference"]
|
| 488 |
+
))
|
| 489 |
+
all_refs.append(batch['target_text'][i].strip())
|
| 490 |
+
all_inputs.append(batch['input_text'][i].strip())
|
| 491 |
|
| 492 |
+
# Metrics
|
| 493 |
+
bleu_score, bert_f1 = 0.0, 0.0
|
| 494 |
+
try:
|
| 495 |
+
from nltk.translate.bleu_score import corpus_bleu
|
| 496 |
+
bleu_score = corpus_bleu(
|
| 497 |
+
[[r.split()] for r in all_refs],
|
| 498 |
+
[p.split() for p in all_preds]
|
| 499 |
+
)
|
| 500 |
+
except Exception:
|
| 501 |
+
pass
|
| 502 |
|
| 503 |
+
try:
|
| 504 |
+
import evaluate as hf_eval
|
| 505 |
+
res = hf_eval.load('bertscore').compute(
|
| 506 |
+
predictions=all_preds, references=all_refs, lang='hi'
|
| 507 |
+
)
|
| 508 |
+
bert_f1 = sum(res['f1']) / len(res['f1'])
|
| 509 |
+
except Exception:
|
| 510 |
+
pass
|
| 511 |
|
| 512 |
+
# Save
|
| 513 |
+
out_path = f"{exp_dir}/evaluation_results.txt"
|
| 514 |
+
pred_path = f"{exp_dir}/evaluation_predictions.jsonl"
|
| 515 |
+
with open(out_path, 'w', encoding='utf-8') as f:
|
| 516 |
+
f.write(f"Model : {model_name}\n")
|
| 517 |
+
f.write(f"Negatives: {has_neg}\n")
|
| 518 |
+
f.write(f"Steps : {cfg['inference']['num_steps']}\n")
|
| 519 |
+
f.write(f"Temp : {cfg['inference']['temperature']}\n")
|
| 520 |
+
f.write(f"RepPen : {cfg['inference']['repetition_penalty']}\n")
|
| 521 |
+
f.write(f"DivPen : {cfg['inference']['diversity_penalty']}\n")
|
| 522 |
+
f.write(f"BLEU : {bleu_score:.4f}\n")
|
| 523 |
+
f.write(f"BERTScore: {bert_f1:.4f}\n\n")
|
| 524 |
+
f.write("=== SAMPLES ===\n")
|
| 525 |
+
for i in range(min(20, len(all_preds))):
|
| 526 |
+
f.write(f"IN : {all_inputs[i]}\n")
|
| 527 |
+
f.write(f"REF : {all_refs[i]}\n")
|
| 528 |
+
f.write(f"PRED: {all_preds[i]}\n")
|
| 529 |
+
f.write("-" * 60 + "\n")
|
| 530 |
|
| 531 |
+
with open(pred_path, 'w', encoding='utf-8') as f:
|
| 532 |
+
for src, ref, pred in zip(all_inputs, all_refs, all_preds):
|
| 533 |
+
row = {"input": src, "reference": ref, "prediction": pred}
|
| 534 |
+
f.write(json.dumps(row, ensure_ascii=False) + "\n")
|
|
|
|
| 535 |
|
| 536 |
+
print(f"\n✅ Results → {out_path}")
|
| 537 |
+
print(f"🗂️ Saved predictions → {pred_path}")
|
| 538 |
+
print(f"📊 BLEU: {bleu_score:.4f} | BERTScore: {bert_f1:.4f}")
|
| 539 |
+
return all_preds, all_refs
|
| 540 |
|
|
|
|
| 541 |
|
| 542 |
+
if __name__ == '__main__':
|
| 543 |
+
import argparse
|
| 544 |
+
p = argparse.ArgumentParser()
|
| 545 |
+
p.add_argument('--mode', choices=['demo', 'eval'], default='demo')
|
| 546 |
+
p.add_argument('--samples', type=int, default=500)
|
| 547 |
+
p.add_argument('--checkpoint', type=str, default=None)
|
| 548 |
+
p.add_argument('--text', type=str, default=None, help='Run one-shot demo input and exit')
|
| 549 |
+
args = p.parse_args()
|
| 550 |
|
| 551 |
+
if args.mode == 'demo':
|
| 552 |
+
interactive_demo(checkpoint=args.checkpoint, single_text=args.text)
|
| 553 |
+
else:
|
| 554 |
+
batch_evaluate(args.samples, checkpoint=args.checkpoint)
|
|
|
|
|
|
|
|
|