File size: 12,566 Bytes
8919f98
 
 
 
 
 
 
 
 
 
 
 
5174029
8919f98
 
 
 
 
 
 
 
 
 
5174029
8919f98
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5174029
8919f98
 
5174029
 
 
 
8919f98
5174029
 
 
 
 
8919f98
 
 
 
 
 
 
5174029
8919f98
 
5174029
8919f98
 
5174029
 
8919f98
 
 
 
 
 
 
 
5174029
 
 
8919f98
 
 
5174029
 
 
 
 
8919f98
5174029
8919f98
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6fa476d
8919f98
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5174029
8919f98
 
 
 
 
5174029
8919f98
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5174029
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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
"""
ChromaForge β€” Grayscale Image Colorization API
FastAPI backend with async inference, proper error handling, and HF Spaces compatibility.
"""

from __future__ import annotations

import io
import logging
import os
import time
import uuid
import warnings
from contextlib import asynccontextmanager
from pathlib import Path
from typing import Optional

import numpy as np
import torch
import torch.nn as nn
from fastapi import FastAPI, File, HTTPException, UploadFile
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import Response, JSONResponse
from PIL import Image, ImageOps
from skimage.color import lab2rgb, rgb2lab
from skimage.transform import resize

# ── Logging ─────────────────────────────────────────────────────────────────
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s | %(levelname)s | %(message)s",
)
logger = logging.getLogger(__name__)

# ── Constants ────────────────────────────────────────────────────────────────
MODEL_PATH = Path(os.getenv("MODEL_PATH", "weights/generator.pth"))
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
IMG_SIZE = 256
MAX_UPLOAD_BYTES = 10 * 1024 * 1024  # 10 MB
ALLOWED_CONTENT_TYPES = {"image/jpeg", "image/png", "image/webp", "image/bmp"}

# ── Generator Architecture (U-Net with skip connections) ─────────────────────

class ConvBlock(nn.Module):
    """Encoder block: Conv β†’ BatchNorm β†’ LeakyReLU"""
    def __init__(self, in_ch: int, out_ch: int, use_bn: bool = True):
        super().__init__()
        layers: list[nn.Module] = [
            nn.Conv2d(in_ch, out_ch, 4, stride=2, padding=1, bias=not use_bn)
        ]
        if use_bn:
            layers.append(nn.BatchNorm2d(out_ch))
        layers.append(nn.LeakyReLU(0.2, inplace=True))
        self.block = nn.Sequential(*layers)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.block(x)


class DeconvBlock(nn.Module):
    """Decoder block: ConvTranspose β†’ BatchNorm β†’ ReLU (+ optional Dropout)"""
    def __init__(self, in_ch: int, out_ch: int, dropout: bool = False):
        super().__init__()
        layers: list[nn.Module] = [
            nn.ConvTranspose2d(in_ch, out_ch, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True),
        ]
        if dropout:
            layers.append(nn.Dropout(0.5))
        self.block = nn.Sequential(*layers)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.block(x)


class UNetGenerator(nn.Module):
    """
    U-Net Generator for image colorization.
    Input:  (B, 1, 256, 256)  β€” L channel, normalized to [-1, 1]
    Output: (B, 2, 256, 256)  β€” AB channels, normalized to [-1, 1]
    
    Architecture follows Isola et al. (2017) "Image-to-Image Translation
    with Conditional Adversarial Networks" (pix2pix).
    """

    def __init__(self):
        super().__init__()
        # ── Encoder ─────────────────────────────────────────────────────────
        self.enc1 = nn.Sequential(
            nn.Conv2d(1, 64, 4, stride=2, padding=1),
            nn.LeakyReLU(0.2, inplace=True),
        )                                        # 256 β†’ 128
        self.enc2 = ConvBlock(64,  128)          # 128 β†’ 64
        self.enc3 = ConvBlock(128, 256)          # 64  β†’ 32
        self.enc4 = ConvBlock(256, 512)          # 32  β†’ 16
        self.enc5 = ConvBlock(512, 512)          # 16  β†’ 8
        self.enc6 = ConvBlock(512, 512)          # 8   β†’ 4
        self.enc7 = ConvBlock(512, 512)          # 4   β†’ 2
        self.enc8 = ConvBlock(512, 512, use_bn=False)  # 2 β†’ 1 (bottleneck)

        # ── Decoder (with skip connections) ─────────────────────────────────
        self.dec1 = DeconvBlock(512,  512, dropout=True)   # 1  β†’ 2
        self.dec2 = DeconvBlock(1024, 512, dropout=True)   # 2  β†’ 4
        self.dec3 = DeconvBlock(1024, 512, dropout=True)   # 4  β†’ 8
        self.dec4 = DeconvBlock(1024, 512)                 # 8  β†’ 16
        self.dec5 = DeconvBlock(1024, 256)                 # 16 β†’ 32
        self.dec6 = DeconvBlock(512,  128)                 # 32 β†’ 64
        self.dec7 = DeconvBlock(256,  64)                  # 64 β†’ 128
        self.dec8 = nn.Sequential(
            nn.ConvTranspose2d(128, 2, 4, stride=2, padding=1),
            nn.Tanh(),
        )                                                  # 128 β†’ 256

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # Encode
        e1 = self.enc1(x)
        e2 = self.enc2(e1)
        e3 = self.enc3(e2)
        e4 = self.enc4(e3)
        e5 = self.enc5(e4)
        e6 = self.enc6(e5)
        e7 = self.enc7(e6)
        e8 = self.enc8(e7)
        # Decode with skip connections (U-Net concat)
        d1 = self.dec1(e8)
        d2 = self.dec2(torch.cat([d1, e7], dim=1))
        d3 = self.dec3(torch.cat([d2, e6], dim=1))
        d4 = self.dec4(torch.cat([d3, e5], dim=1))
        d5 = self.dec5(torch.cat([d4, e4], dim=1))
        d6 = self.dec6(torch.cat([d5, e3], dim=1))
        d7 = self.dec7(torch.cat([d6, e2], dim=1))
        return self.dec8(torch.cat([d7, e1], dim=1))


# ── Image Processing ─────────────────────────────────────────────────────────

def preprocess(image_bytes: bytes) -> tuple[torch.Tensor, np.ndarray, tuple[int, int]]:
    """
    Convert raw image bytes β†’ model-ready tensor.
    Applies EXIF orientation correction (phone photos store rotation as
    metadata rather than rotating pixels, so this must run before any
    other processing or output will appear rotated).
    Returns (tensor, original_L_channel, original_size) for reconstruction.
    """
    img = Image.open(io.BytesIO(image_bytes))
    img = ImageOps.exif_transpose(img)  # apply stored orientation, strip EXIF
    img = img.convert("RGB")
    original_size = img.size  # (W, H), used to restore output to input resolution

    img_np = np.array(img, dtype=np.float32) / 255.0
    img_resized = resize(img_np, (IMG_SIZE, IMG_SIZE), anti_aliasing=True)
    img_lab = rgb2lab(img_resized).astype(np.float32)

    L = img_lab[:, :, 0]                        # [0, 100]
    L_norm = (L / 50.0) - 1.0                   # β†’ [-1, 1]
    tensor = torch.from_numpy(L_norm).unsqueeze(0).unsqueeze(0)  # (1,1,H,W)
    return tensor, L, original_size


def postprocess(pred_ab: torch.Tensor, L: np.ndarray, original_size: tuple[int, int]) -> bytes:
    """
    Merge predicted AB channels with original L β†’ RGB image bytes (JPEG).
    Upscales back to the input resolution so output quality is not
    capped at the model's fixed 256x256 working resolution.
    """
    ab = pred_ab.squeeze(0).permute(1, 2, 0).cpu().numpy()  # (H,W,2)
    ab = ab * 128.0                              # [-1,1] β†’ [-128, 128]

    lab = np.zeros((IMG_SIZE, IMG_SIZE, 3), dtype=np.float32)
    lab[:, :, 0] = L
    lab[:, :, 1:] = ab

    with warnings.catch_warnings():
        warnings.simplefilter("ignore")  # LAB→RGB clipping warnings are expected, not errors
        rgb = lab2rgb(lab)               # [0, 1]
    rgb_uint8 = (rgb * 255).clip(0, 255).astype(np.uint8)

    out = Image.fromarray(rgb_uint8)
    # Restore original aspect ratio / resolution rather than returning a
    # fixed 256x256 image, which previously made output look low quality
    # regardless of the input's actual size.
    out = out.resize(original_size, Image.LANCZOS)

    buf = io.BytesIO()
    out.save(buf, format="JPEG", quality=95)
    return buf.getvalue()


# ── App Lifecycle ─────────────────────────────────────────────────────────────

generator: Optional[UNetGenerator] = None


@asynccontextmanager
async def lifespan(app: FastAPI):
    """Load model weights on startup; clean up on shutdown."""
    global generator
    logger.info(f"Loading generator from {MODEL_PATH} on {DEVICE}")
    generator = UNetGenerator().to(DEVICE)

    if MODEL_PATH.exists():
        state = torch.load(MODEL_PATH, map_location=DEVICE, weights_only=False)
        generator.load_state_dict(state)
        logger.info("Weights loaded successfully.")
    else:
        logger.warning(
            "No weights file found β€” running with random weights. "
            "Place a trained generator.pth in weights/ to enable real colorization."
        )

    generator.eval()
    yield
    logger.info("Shutting down ChromaForge API.")


# ── FastAPI App ──────────────────────────────────────────────────────────────

app = FastAPI(
    title="ChromaForge API",
    description="Automatic grayscale image colorization using a conditional GAN.",
    version="2.0.0",
    lifespan=lifespan,
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=os.getenv("ALLOWED_ORIGINS", "*").split(","),
    allow_methods=["GET", "POST"],
    allow_headers=["*"],
)


@app.get("/", tags=["Health"])
async def root():
    return {
        "service": "ChromaForge",
        "version": "2.0.0",
        "device": str(DEVICE),
        "model_loaded": generator is not None and MODEL_PATH.exists(),
    }


@app.get("/health", tags=["Health"])
async def health():
    return {"status": "ok", "timestamp": time.time()}


@app.post("/colorize", tags=["Inference"])
async def colorize(file: UploadFile = File(...)):
    """
    Colorize a grayscale image.

    - **file**: JPEG, PNG, WebP, or BMP image (max 10 MB)
    - Returns: colorized JPEG image as binary response
    """
    # ── Validate ─────────────────────────────────────────────────────────────
    if file.content_type not in ALLOWED_CONTENT_TYPES:
        raise HTTPException(
            status_code=415,
            detail=f"Unsupported file type '{file.content_type}'. "
                   f"Accepted: {', '.join(ALLOWED_CONTENT_TYPES)}",
        )

    raw = await file.read()
    if len(raw) > MAX_UPLOAD_BYTES:
        raise HTTPException(
            status_code=413,
            detail=f"File too large ({len(raw) / 1e6:.1f} MB). Max: 10 MB.",
        )

    if generator is None:
        raise HTTPException(status_code=503, detail="Model not initialized.")

    # ── Inference ─────────────────────────────────────────────────────────────
    request_id = str(uuid.uuid4())[:8]
    logger.info(f"[{request_id}] Colorizing image ({len(raw) / 1024:.1f} KB)")
    t0 = time.perf_counter()

    try:
        tensor, L, original_size = preprocess(raw)
        tensor = tensor.to(DEVICE)

        with torch.inference_mode():
            pred_ab = generator(tensor)

        result_bytes = postprocess(pred_ab, L, original_size)

    except Exception as exc:
        logger.exception(f"[{request_id}] Inference failed: {exc}")
        raise HTTPException(status_code=500, detail=f"Inference error: {str(exc)}")

    elapsed = time.perf_counter() - t0
    logger.info(f"[{request_id}] Done in {elapsed:.3f}s")

    return Response(
        content=result_bytes,
        media_type="image/jpeg",
        headers={
            "X-Request-Id": request_id,
            "X-Inference-Time": f"{elapsed:.3f}",
        },
    )


@app.post("/colorize/base64", tags=["Inference"])
async def colorize_base64(file: UploadFile = File(...)):
    """Same as /colorize but returns JSON with base64-encoded image."""
    import base64
    response = await colorize(file)
    encoded = base64.b64encode(response.body).decode()
    return JSONResponse({
        "image": f"data:image/jpeg;base64,{encoded}",
        "inference_time": response.headers.get("X-Inference-Time"),
    })