Datasets:

Modalities:
Image
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
License:
Llabres commited on
Commit
8e04777
·
1 Parent(s): a30fb96

Updated Readme

Browse files
Files changed (1) hide show
  1. README.md +190 -1
README.md CHANGED
@@ -49,10 +49,199 @@ dataset = load_dataset("VLR-CVC/ComPAP", skill, split=split)
49
  ```
50
 
51
  <details>
 
52
  <summary>Map to single images</summary>
 
53
  If your model can only process single images, you can render each sample as a single image:
54
 
55
- _coming soon_
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
56
 
57
  </details>
58
 
 
49
  ```
50
 
51
  <details>
52
+
53
  <summary>Map to single images</summary>
54
+
55
  If your model can only process single images, you can render each sample as a single image:
56
 
57
+ ![Single Image Example](figures/single_image_seq_filling_example.png)
58
+
59
+
60
+ ```python
61
+ from PIL import Image, ImageDraw, ImageFont
62
+ import numpy as np
63
+ from datasets import Features, Value, Image as ImageFeature
64
+
65
+ class SingleImagePickAPanel:
66
+ def __init__(self, max_size=500, margin=10, label_space=20, font_path="Arial.ttf"):
67
+ self.max_size = max_size
68
+ self.margin = margin
69
+ self.label_space = label_space
70
+ # Add separate font sizes
71
+ self.label_font_size = 20
72
+ self.number_font_size = 24
73
+
74
+ self.font_path = font_path
75
+
76
+ def resize_image(self, img):
77
+ """Resize image keeping aspect ratio if longest edge > max_size"""
78
+ if max(img.size) > self.max_size:
79
+ ratio = self.max_size / max(img.size)
80
+ new_size = tuple(int(dim * ratio) for dim in img.size)
81
+ return img.resize(new_size, Image.Resampling.LANCZOS)
82
+ return img
83
+
84
+ def create_mask_panel(self, width, height):
85
+ """Create a question mark panel"""
86
+ mask_panel = Image.new("RGB", (width, height), (200, 200, 200))
87
+ draw = ImageDraw.Draw(mask_panel)
88
+ font_size = int(height * 0.8)
89
+ try:
90
+ font = ImageFont.truetype(self.font_path, font_size)
91
+ except:
92
+ raise ValueError("Font file not found")
93
+
94
+ text = "?"
95
+ bbox = draw.textbbox((0, 0), text, font=font)
96
+ text_x = (width - (bbox[2] - bbox[0])) // 2
97
+ text_y = (height - (bbox[3] - bbox[1])) // 2
98
+ draw.text((text_x, text_y), text, fill="black", font=font)
99
+ return mask_panel
100
+
101
+ def draw_number_on_panel(self, panel, number, font):
102
+ """Draw number on the bottom of the panel with background"""
103
+ draw = ImageDraw.Draw(panel)
104
+
105
+ # Get text size
106
+ bbox = draw.textbbox((0, 0), str(number), font=font)
107
+ text_width = bbox[2] - bbox[0]
108
+ text_height = bbox[3] - bbox[1]
109
+
110
+ # Calculate position (bottom-right corner)
111
+ padding = 2
112
+ text_x = panel.size[0] - text_width - padding
113
+ text_y = panel.size[1] - text_height - padding
114
+
115
+ # Draw semi-transparent background
116
+ bg_rect = [(text_x - padding, text_y - padding),
117
+ (text_x + text_width + padding, text_y + text_height + padding)]
118
+ draw.rectangle(bg_rect, fill=(255, 255, 255, 180))
119
+
120
+ # Draw text
121
+ draw.text((text_x, text_y), str(number), fill="black", font=font)
122
+ return panel
123
+
124
+ def map_to_single_image(self, examples):
125
+ """Process a batch of examples from a HuggingFace dataset"""
126
+ single_images = []
127
+
128
+ for i in range(len(examples['sample_id'])):
129
+ # Get context and options for current example
130
+ context = examples['context'][i] if len(examples['context'][i]) > 0 else []
131
+ options = examples['options'][i]
132
+
133
+ # Resize all images
134
+ context = [self.resize_image(img) for img in context]
135
+ options = [self.resize_image(img) for img in options]
136
+
137
+ # Calculate common panel size (use median size to avoid outliers)
138
+ all_panels = context + options
139
+ if len(all_panels) > 0:
140
+ widths = [img.size[0] for img in all_panels]
141
+ heights = [img.size[1] for img in all_panels]
142
+ panel_width = int(np.median(widths))
143
+ panel_height = int(np.median(heights))
144
+
145
+ # Resize all panels to common size
146
+ context = [img.resize((panel_width, panel_height)) for img in context]
147
+ options = [img.resize((panel_width, panel_height)) for img in options]
148
+
149
+ # Create mask panel for sequence filling tasks if needed
150
+ if 'index' in examples and len(context) > 0:
151
+ mask_idx = examples['index'][i]
152
+ mask_panel = self.create_mask_panel(panel_width, panel_height)
153
+ context.insert(mask_idx, mask_panel)
154
+
155
+ # Calculate canvas dimensions based on whether we have context
156
+ if len(context) > 0:
157
+ context_row_width = panel_width * len(context) + self.margin * (len(context) - 1)
158
+ options_row_width = panel_width * len(options) + self.margin * (len(options) - 1)
159
+ canvas_width = max(context_row_width, options_row_width)
160
+ canvas_height = (panel_height * 2 +
161
+ self.label_space * 2)
162
+ else:
163
+ # Only options row for caption_relevance
164
+ canvas_width = panel_width * len(options) + self.margin * (len(options) - 1)
165
+ canvas_height = (panel_height +
166
+ self.label_space)
167
+
168
+ # Create canvas
169
+ final_image = Image.new("RGB", (canvas_width, canvas_height), "white")
170
+ draw = ImageDraw.Draw(final_image)
171
+
172
+ try:
173
+ label_font = ImageFont.truetype(self.font_path, self.label_font_size)
174
+ number_font = ImageFont.truetype(self.font_path, self.number_font_size)
175
+ except:
176
+ raise ValueError("Font file not found")
177
+
178
+ current_y = 0
179
+
180
+ # Add context section if it exists
181
+ if len(context) > 0:
182
+ # Draw "Context" label
183
+ bbox = draw.textbbox((0, 0), "Context", font=label_font)
184
+ text_x = (canvas_width - (bbox[2] - bbox[0])) // 2
185
+ draw.text((text_x, current_y), "Context", fill="black", font=label_font)
186
+ current_y += self.label_space
187
+
188
+ # Paste context panels
189
+ x_offset = (canvas_width - (panel_width * len(context) +
190
+ self.margin * (len(context) - 1))) // 2
191
+ for panel in context:
192
+ final_image.paste(panel, (x_offset, current_y))
193
+ x_offset += panel_width + self.margin
194
+ current_y += panel_height
195
+
196
+ # Add "Options" label
197
+ bbox = draw.textbbox((0, 0), "Options", font=label_font)
198
+ text_x = (canvas_width - (bbox[2] - bbox[0])) // 2
199
+ draw.text((text_x, current_y), "Options", fill="black", font=label_font)
200
+ current_y += self.label_space
201
+
202
+ # Paste options with numbers on panels
203
+ x_offset = (canvas_width - (panel_width * len(options) +
204
+ self.margin * (len(options) - 1))) // 2
205
+ for idx, panel in enumerate(options):
206
+ # Create a copy of the panel to draw on
207
+ panel_with_number = panel.copy()
208
+ if panel_with_number.mode != 'RGBA':
209
+ panel_with_number = panel_with_number.convert('RGBA')
210
+
211
+ # Draw number on panel
212
+ panel_with_number = self.draw_number_on_panel(
213
+ panel_with_number,
214
+ idx,
215
+ number_font
216
+ )
217
+
218
+ # Paste the panel with number
219
+ final_image.paste(panel_with_number, (x_offset, current_y), panel_with_number)
220
+ x_offset += panel_width + self.margin
221
+
222
+ # Convert final_image to PIL Image format (instead of numpy array)
223
+ single_images.append(final_image)
224
+
225
+ # Prepare batch output
226
+ examples['single_image'] = single_images
227
+
228
+ return examples
229
+
230
+ from datasets import load_dataset
231
+
232
+ skill = "sequence_filling" # "sequence_filling", "char_coherence", "visual_closure", "text_closure", "caption_relevance"
233
+ split = "val" # "val", "test"
234
+ dataset = load_dataset("VLR-CVC/ComPAP", skill, split=split)
235
+
236
+ processor = SingleImagePickAPanel()
237
+ dataset = dataset.map(
238
+ processor.map_to_single_image,
239
+ batched=True,
240
+ batch_size=32,
241
+ remove_columns=['context', 'options']
242
+ )
243
+ dataset.save_to_disk(f"ComPAP_{skill}_{split}_single_images")
244
+ ```
245
 
246
  </details>
247