File size: 16,562 Bytes
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fac3f86
 
 
48572da
fac3f86
48572da
fac3f86
48572da
 
 
 
 
 
 
 
 
 
 
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
 
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
 
 
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
fac3f86
48572da
 
 
 
 
 
 
 
 
 
 
 
 
fac3f86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48572da
 
 
 
fac3f86
 
 
 
 
 
 
 
 
 
 
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
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
"""
Shared dataset classes and loading utilities for GAP-CLIP evaluation scripts.

Provides:
  - FashionMNISTDataset  (Fashion-MNIST grayscale images)
  - KaggleDataset        (KAGL Marqo HuggingFace dataset)
  - LocalDataset         (internal local validation dataset)
  - Matching load_* convenience functions
  - collate_fn_filter_none  (for DataLoader)
  - normalize_hierarchy_label  (text normalisation helper)
"""

from __future__ import annotations

import difflib
import hashlib
import os
import sys
from pathlib import Path
from io import BytesIO
from typing import List, Optional

import numpy as np
import pandas as pd
import torch
from PIL import Image
import requests
from torch.utils.data import Dataset
from torchvision import transforms

# Make project root importable when running evaluation scripts directly.
_PROJECT_ROOT = Path(__file__).resolve().parents[2]
if str(_PROJECT_ROOT) not in sys.path:
    sys.path.insert(0, str(_PROJECT_ROOT))

from config import (  # type: ignore
    ROOT_DIR,
    column_local_image_path,
    fashion_mnist_csv,
    local_dataset_path,
    images_dir,
)

# ---------------------------------------------------------------------------
# Fashion-MNIST helpers
# ---------------------------------------------------------------------------

def get_fashion_mnist_labels() -> dict:
    """Return the 10 Fashion-MNIST integer-to-name mapping."""
    return {
        0: "T-shirt/top",
        1: "Trouser",
        2: "Pullover",
        3: "Dress",
        4: "Coat",
        5: "Sandal",
        6: "Shirt",
        7: "Sneaker",
        8: "Bag",
        9: "Ankle boot",
    }


def create_fashion_mnist_to_hierarchy_mapping(hierarchy_classes: List[str]) -> dict:
    """Map Fashion-MNIST integer labels to nearest hierarchy class name.

    Returns dict {label_id: matched_class_name or None}.
    """
    fashion_mnist_labels = get_fashion_mnist_labels()
    hierarchy_classes_lower = [h.lower() for h in hierarchy_classes]
    mapping = {}

    for fm_label_id, fm_label in fashion_mnist_labels.items():
        fm_label_lower = fm_label.lower()
        matched_hierarchy = None

        if fm_label_lower in hierarchy_classes_lower:
            matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index(fm_label_lower)]
        elif any(h in fm_label_lower or fm_label_lower in h for h in hierarchy_classes_lower):
            for h_class in hierarchy_classes:
                if h_class.lower() in fm_label_lower or fm_label_lower in h_class.lower():
                    matched_hierarchy = h_class
                    break
        else:
            if fm_label_lower in ["t-shirt/top", "top"]:
                if "top" in hierarchy_classes_lower:
                    matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index("top")]
            elif "trouser" in fm_label_lower:
                for p in ["bottom", "pants", "trousers", "trouser", "pant"]:
                    if p in hierarchy_classes_lower:
                        matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index(p)]
                        break
            elif "pullover" in fm_label_lower:
                for p in ["sweater", "pullover"]:
                    if p in hierarchy_classes_lower:
                        matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index(p)]
                        break
            elif "dress" in fm_label_lower:
                if "dress" in hierarchy_classes_lower:
                    matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index("dress")]
            elif "coat" in fm_label_lower:
                for p in ["jacket", "outerwear", "coat"]:
                    if p in hierarchy_classes_lower:
                        matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index(p)]
                        break
            elif fm_label_lower in ["sandal", "sneaker", "ankle boot"]:
                for p in ["shoes", "shoe", "sandal", "sneaker", "boot"]:
                    if p in hierarchy_classes_lower:
                        matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index(p)]
                        break
            elif "bag" in fm_label_lower:
                if "bag" in hierarchy_classes_lower:
                    matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index("bag")]

        if matched_hierarchy is None:
            close = difflib.get_close_matches(fm_label_lower, hierarchy_classes_lower, n=1, cutoff=0.6)
            if close:
                matched_hierarchy = hierarchy_classes[hierarchy_classes_lower.index(close[0])]

        mapping[fm_label_id] = matched_hierarchy
        status = matched_hierarchy if matched_hierarchy else "NO MATCH (will be filtered out)"
        print(f"  {fm_label} ({fm_label_id}) -> {status}")

    return mapping


def convert_fashion_mnist_to_image(pixel_values) -> Image.Image:
    """Convert a flat 784-element pixel array to an RGB PIL image."""
    arr = np.array(pixel_values).reshape(28, 28).astype(np.uint8)
    arr = np.stack([arr] * 3, axis=-1)
    return Image.fromarray(arr)


class FashionMNISTDataset(Dataset):
    """PyTorch dataset wrapping Fashion-MNIST CSV rows."""

    def __init__(self, dataframe: pd.DataFrame, image_size: int = 224, label_mapping: Optional[dict] = None):
        self.dataframe = dataframe
        self.image_size = image_size
        self.labels_map = get_fashion_mnist_labels()
        self.label_mapping = label_mapping

        self.transform = transforms.Compose([
            transforms.Resize((image_size, image_size)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

    def __len__(self) -> int:
        return len(self.dataframe)

    def __getitem__(self, idx):
        row = self.dataframe.iloc[idx]
        pixel_cols = [f"pixel{i}" for i in range(1, 785)]
        image = convert_fashion_mnist_to_image(row[pixel_cols].values)
        image = self.transform(image)

        label_id = int(row["label"])
        description = self.labels_map[label_id]
        color = "unknown"
        hierarchy = (
            self.label_mapping[label_id]
            if (self.label_mapping and label_id in self.label_mapping)
            else self.labels_map[label_id]
        )
        return image, description, color, hierarchy


def load_fashion_mnist_dataset(
    max_samples: int = 10000,
    hierarchy_classes: Optional[List[str]] = None,
    csv_path: Optional[str] = None,
) -> FashionMNISTDataset:
    """Load Fashion-MNIST test CSV into a FashionMNISTDataset.

    Args:
        max_samples: Maximum number of samples to use.
        hierarchy_classes: If provided, maps Fashion-MNIST labels to these classes.
        csv_path: Path to fashion-mnist_test.csv. Defaults to config.fashion_mnist_csv.
    """
    if csv_path is None:
        csv_path = fashion_mnist_csv

    print("Loading Fashion-MNIST test dataset...")
    df = pd.read_csv(csv_path)
    print(f"Fashion-MNIST dataset loaded: {len(df)} samples")

    label_mapping = None
    if hierarchy_classes is not None:
        print("\nCreating mapping from Fashion-MNIST labels to hierarchy classes:")
        label_mapping = create_fashion_mnist_to_hierarchy_mapping(hierarchy_classes)
        valid_ids = [lid for lid, h in label_mapping.items() if h is not None]
        df = df[df["label"].isin(valid_ids)]
        print(f"\nAfter filtering to mappable labels: {len(df)} samples")

    df_sample = df.head(max_samples)
    print(f"Using {len(df_sample)} samples for evaluation")
    return FashionMNISTDataset(df_sample, label_mapping=label_mapping)


# ---------------------------------------------------------------------------
# KAGL Marqo dataset
# ---------------------------------------------------------------------------

class KaggleDataset(Dataset):
    """Dataset class for KAGL Marqo HuggingFace dataset."""

    def __init__(self, dataframe: pd.DataFrame, image_size: int = 224, include_hierarchy: bool = False):
        self.dataframe = dataframe
        self.image_size = image_size
        self.include_hierarchy = include_hierarchy

        self.transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

    def __len__(self) -> int:
        return len(self.dataframe)

    def __getitem__(self, idx):
        row = self.dataframe.iloc[idx]
        image_data = row["image_url"]

        if isinstance(image_data, dict) and "bytes" in image_data:
            image = Image.open(BytesIO(image_data["bytes"])).convert("RGB")
        elif hasattr(image_data, "convert"):
            image = image_data.convert("RGB")
        else:
            image = Image.open(BytesIO(image_data)).convert("RGB")

        image = self.transform(image)
        description = row["text"]
        color = row["color"]

        if self.include_hierarchy:
            hierarchy = row.get("hierarchy", "unknown")
            return image, description, color, hierarchy
        return image, description, color


def download_kaggle_raw_df() -> pd.DataFrame:
    """Download the raw KAGL Marqo DataFrame from HuggingFace.

    This is the expensive network operation.  Callers can cache the result
    and pass it to :func:`load_kaggle_marqo_dataset` via *raw_df* to avoid
    repeated downloads.
    """
    from datasets import load_dataset  # type: ignore

    print("Downloading KAGL Marqo dataset from HuggingFace...")
    dataset = load_dataset("Marqo/KAGL")
    df = dataset["data"].to_pandas()
    print(f"KAGL dataset downloaded: {len(df)} samples, columns: {list(df.columns)}")
    return df


def load_kaggle_marqo_dataset(
    max_samples: int = 5000,
    include_hierarchy: bool = False,
    raw_df: Optional[pd.DataFrame] = None,
) -> KaggleDataset:
    """Download and prepare the KAGL Marqo HuggingFace dataset.

    Args:
        max_samples: Maximum number of samples to return.
        include_hierarchy: If True, dataset tuples include a hierarchy element.
        raw_df: Pre-downloaded DataFrame (from :func:`download_kaggle_raw_df`).
            If *None*, the dataset is downloaded from HuggingFace.
    """
    if raw_df is not None:
        df = raw_df.copy()
        print(f"Using cached KAGL DataFrame: {len(df)} samples")
    else:
        df = download_kaggle_raw_df()

    df = df.dropna(subset=["text", "image"])

    if len(df) > max_samples:
        df = df.sample(n=max_samples, random_state=42)
        print(f"Sampled {max_samples} items")

    kaggle_df = pd.DataFrame({
        "image_url": df["image"],
        "text": df["text"],
        "color": df["baseColour"].str.lower().str.replace("grey", "gray"),
    })

    kaggle_df = kaggle_df.dropna(subset=["color"])

    print(f"Colors: {sorted(kaggle_df['color'].unique())}")

    return KaggleDataset(kaggle_df, include_hierarchy=include_hierarchy)


# ---------------------------------------------------------------------------
# Local validation dataset
# ---------------------------------------------------------------------------

class LocalDataset(Dataset):
    """Dataset class for the internal local validation dataset."""

    def __init__(self, dataframe: pd.DataFrame, image_size: int = 224, include_hierarchy: bool = False):
        self.dataframe = dataframe
        self.image_size = image_size
        self.include_hierarchy = include_hierarchy

        self.transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

    def __len__(self) -> int:
        return len(self.dataframe)

    def __getitem__(self, idx):
        row = self.dataframe.iloc[idx]
        try:
            image_path = row.get(column_local_image_path) if hasattr(row, "get") else None
            if isinstance(image_path, str) and image_path:
                if not os.path.isabs(image_path):
                    image_path = str(ROOT_DIR / image_path)
                image = Image.open(image_path).convert("RGB")
            else:
                # Fallback: download image from URL (and cache).
                image_url = row.get("image_url") if hasattr(row, "get") else None
                if isinstance(image_url, dict) and "bytes" in image_url:
                    image = Image.open(BytesIO(image_url["bytes"])).convert("RGB")
                elif isinstance(image_url, str) and image_url:
                    cache_dir = Path(images_dir)
                    cache_dir.mkdir(parents=True, exist_ok=True)
                    url_hash = hashlib.md5(image_url.encode("utf-8")).hexdigest()
                    cache_path = cache_dir / f"{url_hash}.jpg"
                    if cache_path.exists():
                        image = Image.open(cache_path).convert("RGB")
                    else:
                        resp = requests.get(image_url, timeout=10)
                        resp.raise_for_status()
                        image = Image.open(BytesIO(resp.content)).convert("RGB")
                        image.save(cache_path, "JPEG", quality=85, optimize=True)
                else:
                    raise ValueError("Missing image_path and image_url")
        except Exception as e:
            print(f"Error loading image: {e}")
            image = Image.new("RGB", (224, 224), color="gray")
        image = self.transform(image)

        description = row["text"]
        color = row["color"]

        if self.include_hierarchy:
            hierarchy = row.get("hierarchy", "unknown")
            return image, description, color, hierarchy
        return image, description, color


def load_local_validation_dataset(
    max_samples: int = 5000,
    include_hierarchy: bool = False,
    raw_df: Optional[pd.DataFrame] = None,
) -> LocalDataset:
    """Load and prepare the internal local validation dataset.

    Args:
        max_samples: Maximum number of samples to return.
        include_hierarchy: If True, dataset tuples include a hierarchy element.
        raw_df: Pre-loaded DataFrame. If *None*, the CSV is read from disk.
    """
    if raw_df is not None:
        df = raw_df.copy()
        print(f"Using cached local DataFrame: {len(df)} samples")
    else:
        print("Loading local validation dataset...")
        df = pd.read_csv(local_dataset_path)
    print(f"Dataset loaded: {len(df)} samples")

    if column_local_image_path in df.columns:
        df = df.dropna(subset=[column_local_image_path])
        print(f"After filtering NaN image paths: {len(df)} samples")
    else:
        print(f"Column '{column_local_image_path}' not found; falling back to 'image_url'.")

    if "color" in df.columns:
        print(f"After color filtering: {len(df)} samples, colors: {sorted(df['color'].unique())}")

    if len(df) > max_samples:
        df = df.sample(n=max_samples, random_state=42)
        print(f"Sampled {max_samples} items")

    print(f"Using {len(df)} samples for evaluation")
    return LocalDataset(df, include_hierarchy=include_hierarchy)


# ---------------------------------------------------------------------------
# DataLoader utilities
# ---------------------------------------------------------------------------

def collate_fn_filter_none(batch):
    """Collate function that silently drops None items from a batch."""
    original_len = len(batch)
    batch = [item for item in batch if item is not None]
    if original_len > len(batch):
        print(f"Filtered out {original_len - len(batch)} None values from batch")
    if not batch:
        print("Empty batch after filtering None values")
        return torch.tensor([]), [], []
    # Support both 3-value (image, text, color) and 4-value (image, text, color, hierarchy) tuples
    if len(batch[0]) == 4:
        images, texts, colors, hierarchies = zip(*batch)
        return torch.stack(images), list(texts), list(colors), list(hierarchies)
    images, texts, colors = zip(*batch)
    return torch.stack(images), list(texts), list(colors)


# ---------------------------------------------------------------------------
# Text normalisation helpers
# ---------------------------------------------------------------------------

def normalize_hierarchy_label(label: str) -> str:
    """Lower-case and strip a hierarchy label for consistent comparison."""
    return label.lower().strip() if label else ""