File size: 10,023 Bytes
f52a8f9
 
 
9639dd1
f52a8f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fb10678
f52a8f9
 
 
 
fb10678
f52a8f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
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
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
import os
import gc
import time
import gradio as gr
import torch
from PIL import Image

# -----------------------
# Device + CPU perf knobs
# -----------------------
device = "cuda" if torch.cuda.is_available() else "cpu"

# Threads (tune for HF CPU Space)
os.environ.setdefault("OMP_NUM_THREADS", "4")
os.environ.setdefault("MKL_NUM_THREADS", "4")
torch.set_num_threads(int(os.environ["OMP_NUM_THREADS"]))
torch.set_num_interop_threads(max(1, int(int(os.environ["OMP_NUM_THREADS"]) // 2)))

INFER = torch.inference_mode if hasattr(torch, "inference_mode") else torch.no_grad

# -----------------------
# Stable Diffusion 1.5 (img2img) for style transfer
# -----------------------
from diffusers import StableDiffusionImg2ImgPipeline, EulerAncestralDiscreteScheduler

def load_sd15_pipe():
    pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
        "runwayml/stable-diffusion-v1-5",
        safety_checker=None,
        requires_safety_checker=False,
    )
    pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
    pipe = pipe.to(device)
    pipe.enable_attention_slicing()
    pipe.enable_vae_tiling()
    pipe.enable_vae_slicing()
    if device == "cuda":
        pipe.unet.to(memory_format=torch.channels_last)
    return pipe

_sd_pipe = None

def sd_style_transfer(input_image, prompt, strength=0.55, guidance=5.5, steps=18, width=512, height=512, seed=0):
    global _sd_pipe
    if input_image is None:
        raise gr.Error("Please upload an input image.")
    if not prompt or not prompt.strip():
        raise gr.Error("Please provide a style prompt.")

    if _sd_pipe is None:
        t0 = time.time()
        _sd_pipe = load_sd15_pipe()
        print(f"[SD] Pipeline loaded in {time.time()-t0:.2f}s on {device}.", flush=True)

    generator = torch.Generator(device=device) if device == "cuda" else torch.Generator()
    if isinstance(seed, (int, float)) and int(seed) > 0:
        generator = generator.manual_seed(int(seed))

    img = input_image.convert("RGB").resize((int(width), int(height)), Image.LANCZOS)

    with INFER():
        out = _sd_pipe(
            prompt=str(prompt),
            image=img,
            strength=float(strength),
            guidance_scale=float(guidance),
            num_inference_steps=int(steps),
            generator=generator,
        ).images[0]

    if device == "cuda":
        torch.cuda.empty_cache()
    gc.collect()
    return out

# -----------------------
# Grammar correction models
# T5-small (prithivida), T5-base (vennify), GECToR (optional), Llama-3.1-8B-GEC (GGUF)
# -----------------------
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

T5_SMALL = "prithivida/grammar_error_correcter_v1"      # T5-small
T5_BASE  = "vennify/t5-base-grammar-correction"         # T5-base

_t5_tok = {}
_t5_mdl = {}

def load_t5(model_name: str):
    if model_name not in _t5_mdl:
        tok = AutoTokenizer.from_pretrained(model_name)
        mdl = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
        _t5_tok[model_name] = tok
        _t5_mdl[model_name] = mdl
    return _t5_tok[model_name], _t5_mdl[model_name]

def t5_correct(text: str, model_name: str, max_new_tokens=128):
    tok, mdl = load_t5(model_name)
    prefix = "gec: " if "prithivida" in model_name else "grammar: "
    inputs = tok(prefix + text, return_tensors="pt").to(device)
    with INFER():
        out = mdl.generate(**inputs, max_length=max_new_tokens)
    return tok.decode(out[0], skip_special_tokens=True)

# ---- Optional: GECToR (lazy load) ----
_gector_predictor = None
_gector_error = None
_gector_tried = False

def try_load_gector():
    global _gector_predictor, _gector_error, _gector_tried
    if _gector_tried:
        return _gector_predictor, _gector_error
    _gector_tried = True
    try:
        from gector.gec_model import GECModel  # requires allennlp + pretrained artifacts
        model_paths = os.environ.get("GEC_MODEL_PATHS", "").strip()
        vocab_path = os.environ.get("GEC_VOCAB_PATH", "").strip()
        if not model_paths or not vocab_path:
            raise RuntimeError(
                "GECToR selected but model artifacts are not configured. "
                "Set GEC_MODEL_PATHS (space-separated .th files) and GEC_VOCAB_PATH (vocab dir)."
            )
        taggers = model_paths.split()
        _gector_predictor = GECModel(
            model_paths=taggers,
            vocab_path=vocab_path,
            device=("cuda" if device == "cuda" else "cpu"),
            min_error_probability=0.0,
            confidence=0.0,
            iterations=2,
            special_tokens_fix=1,
        )
    except Exception as e:
        _gector_error = str(e)
        _gector_predictor = None
    return _gector_predictor, _gector_error

def gector_correct(text: str):
    predictor, err = try_load_gector()
    if err or predictor is None:
        return f"[GECToR not active] {err or 'Unknown error.'}\n" \
               f"Enable by setting GEC_MODEL_PATHS and GEC_VOCAB_PATH to pretrained files."
    tokens = text.strip().split()
    corrected = predictor.handle_batch([tokens])[0]
    return " ".join(corrected)

# ---- Llama-3.1-8B GEC (GGUF via llama-cpp-python) ----
_llama_model = None
_llama_err = None
_llama_tried = False

# Choose a sensible quant filename; adjust if you upload a different one to your Space.
LLAMA_REPO = "mradermacher/Llama-3.1-8B-Instruct-Grammatical-Error-Correction-2-GGUF"
LLAMA_FILE = os.environ.get("LLAMA_GGUF_FILE", "llama-3.1-8b-instruct-gec.Q4_K_S.gguf")

def try_load_llama():
    global _llama_model, _llama_err, _llama_tried
    if _llama_tried:
        return _llama_model, _llama_err
    _llama_tried = True
    try:
        from llama_cpp import Llama
        # Load directly from Hub (no need to manually download)
        _llama_model = Llama.from_pretrained(
            repo_id=LLAMA_REPO,
            filename=LLAMA_FILE,
            n_ctx=2048,
            n_threads=int(os.environ.get("OMP_NUM_THREADS", "4")),
            n_batch=128,
            verbose=False
        )
    except Exception as e:
        _llama_model = None
        _llama_err = str(e)
    return _llama_model, _llama_err

def llama_gec_correct(text: str, max_new_tokens=256):
    mdl, err = try_load_llama()
    if err or mdl is None:
        return f"[Llama GGUF not active] {err or 'Unknown error.'}\n" \
               f"Check model availability or set LLAMA_GGUF_FILE to a valid filename."
    prompt = (
        "You are a precise grammatical error corrector. "
        "Return only the corrected text without explanations.\n\n"
        f"Input: {text}\n"
        "Corrected:"
    )
    out = mdl(prompt, max_tokens=max_new_tokens, stop=["\n\n", "\nCorrected:"])
    return out["choices"][0]["text"].strip()

# -----------------------
# Router
# -----------------------
MODEL_OPTIONS = [
    "T5-small (prithivida)",
    "T5-base (vennify)",
    "GECToR (tagging)",
    "Llama-3.1-8B-GEC (GGUF)"
]

def correct_text_router(text: str, model_choice: str, max_new_tokens=128):
    text = (text or "").strip()
    if not text:
        raise gr.Error("Please enter text to correct.")
    if model_choice == "T5-small (prithivida)":
        return t5_correct(text, T5_SMALL, max_new_tokens=max_new_tokens)
    if model_choice == "T5-base (vennify)":
        return t5_correct(text, T5_BASE, max_new_tokens=max_new_tokens)
    if model_choice == "GECToR (tagging)":
        return gector_correct(text)
    if model_choice == "Llama-3.1-8B-GEC (GGUF)":
        return llama_gec_correct(text, max_new_tokens=max_new_tokens)
    return "Unknown model selection."

# -----------------------
# UI
# -----------------------
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        f"# 🎨 Style transfer (SD 1.5 img2img) + ✍️ English correction\n"
        f"- Device detected: **{device.upper()}**\n"
        f"- Models: T5-small, T5-base, GECToR, Llama-3.1-8B-GEC (GGUF)\n"
    )

    with gr.Tab("Image style transfer"):
        with gr.Row():
            img_in = gr.Image(label="Input image", type="pil")
            img_out = gr.Image(label="Styled output")
        prompt = gr.Textbox(label="Style prompt", placeholder="e.g., watercolor wash, halftone dots, 1960s comic shading")
        with gr.Row():
            strength = gr.Slider(0.1, 0.95, value=0.55, step=0.05, label="Style strength")
            guidance = gr.Slider(1.0, 12.0, value=5.5, step=0.5, label="Guidance")
            steps = gr.Slider(5, 40, value=18, step=1, label="Steps")
        with gr.Row():
            width = gr.Slider(256, 768, value=512, step=64, label="Width")
            height = gr.Slider(256, 768, value=512, step=64, label="Height")
            seed = gr.Number(value=0, precision=0, label="Seed (0 = random)")
        run_btn = gr.Button("Transfer style", variant="primary")
        run_btn.click(
            fn=sd_style_transfer,
            inputs=[img_in, prompt, strength, guidance, steps, width, height, seed],
            outputs=[img_out]
        )

    with gr.Tab("English grammar correction"):
        model_choice = gr.Dropdown(MODEL_OPTIONS, value="T5-small (prithivida)", label="Model")
        txt_in = gr.Textbox(lines=6, label="Input text")
        max_new = gr.Slider(32, 512, value=128, step=16, label="Max tokens (generation models)")
        txt_out = gr.Textbox(lines=6, label="Corrected text")
        corr_btn = gr.Button("Correct", variant="primary")
        corr_btn.click(
            fn=correct_text_router,
            inputs=[txt_in, model_choice, max_new],
            outputs=[txt_out]
        )

    gr.Markdown(
        "Tips:\n"
        "- On CPU: steps 12–20, guidance 4–7, 512×512 for SD speed.\n"
        "- T5-small = fastest, T5-base = more accurate.\n"
        "- GECToR needs AllenNLP and pretrained tagger files (set GEC_MODEL_PATHS & GEC_VOCAB_PATH).\n"
        "- Llama GGUF loads from Hub (Q4_K_S by default). Adjust LLAMA_GGUF_FILE if needed."
    )

if __name__ == "__main__":
    demo.launch()