HuggingFace commit

.gitattributes

@@ -0,0 +1 @@
+*.pth filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -4,4 +4,5 @@ train_test_images
 .ipynb_checkpoints/visualize-checkpoint.ipynb
 font_identifier_model.pth
 *.pyc
-__pycache__
+__pycache__
+google_fonts

README.md

@@ -8,5 +8,9 @@ Follow along:
 - [On Twitter](https://twitter.com/gabor/status/1722300841691103467)
 
 Generate sample images (note this will work only on Mac): [gen_sample_data.py](gen_sample_data.py)
+
 Arrange test images into test and train: [arrange_train_test_images.py](arrange_train_test_images.py)
-Train a ResNet18 on the data: [train_font_identifier.py](train_font_identifier.py)
+
+Train a ResNet18 on the data: [train_font_identifier.py](train_font_identifier.py)
+
+License: MIT

consts.py

@@ -30,3 +30,5 @@ GEN_IMAGES_DIR = './generated_images'
 TRAIN_TEST_IMAGES_DIR = './train_test_images'
 # where to grab the font files from
 FONT_FILE_DIRS = ['/System/Library/Fonts/', '/System/Library/Fonts/Supplemental/']
+# where to grab the Google Fonts, all of which are allowed
+GOOGLE_FONTS_DIR =  "./google_fonts"

gen_sample_data.py

@@ -6,66 +6,82 @@ from PIL import Image, ImageDraw, ImageFont
 import nltk
 from nltk.corpus import brown
 import random
-from consts import FONT_ALLOWLIST, IMAGES_PER_FONT, GEN_IMAGES_DIR, FONT_FILE_DIRS
+from consts import FONT_ALLOWLIST, IMAGES_PER_FONT, GEN_IMAGES_DIR, FONT_FILE_DIRS, GOOGLE_FONTS_DIR 
 
 # Download the necessary data from nltk
-nltk.download('brown')
+nltk.download('inaugural')
 
 os.makedirs(GEN_IMAGES_DIR, exist_ok=True)
 
-all_brown_words = sorted(set(brown.words(categories='news')))
-
-def wrap_text(text, line_length=10):
+def wrap_text(text, line_length=4):
     """Wraps the provided text every 'line_length' words."""
     words = text.split()
     return "\n".join([" ".join(words[i:i+line_length]) for i in range(0, len(words), line_length)])
 
-def random_prose_text(words, num_words=200):
-    """Returns a random selection of 'num_words' words from the provided list of words."""
-    random_words = " ".join(random.sample(words, num_words))
-    return wrap_text(random_words)
-
-def random_code_text(base_code, num_lines=15):
-    """Returns a random selection of 'num_lines' lines from the provided code."""
-    lines = base_code.split("\n")
-    return "\n".join(random.sample(lines, min(num_lines, len(lines))))
+def random_prose_text(line_length=4):
+    """Returns a random snippet from the Gutenberg corpus."""
+    corpus = nltk.corpus.inaugural.raw()
+    start = random.randint(0, len(corpus) - 800)
+    end = start + 800
+    return wrap_text(corpus[start:end], line_length=line_length)
 
 def main():
+    # Collect all allowed font files
+    font_files = []
+    # all of the Google fonts are allowed, no matter what
+    for font_file in os.listdir(GOOGLE_FONTS_DIR):
+        if font_file.endswith('.ttf') or font_file.endswith('.ttc'):
+            font_path = os.path.join(GOOGLE_FONTS_DIR, font_file)
+            font_name = font_file.split('.')[0]
+            font_files.append((font_path, font_name))
+
+    # for the system font directories, use the FONT_ALLOWLIST
     for font_dir in FONT_FILE_DIRS:
         for font_file in os.listdir(font_dir):
             if font_file.endswith('.ttf') or font_file.endswith('.ttc'):
                 font_path = os.path.join(font_dir, font_file)
                 font_name = font_file.split('.')[0]
-                if font_name not in FONT_ALLOWLIST:
-                    continue
-                # Output the font name so we can see the progress
-                print(font_path, font_name)
-
-                if font_file.endswith('.ttc'):
-                    # ttc fonts have multiple fonts in one file, so we need to specify which one we want
-                    font = ImageFont.truetype(font_path, random.choice(range(32, 128)), index=0)
-                else:
-                    # ttf fonts have only one font in the file
-                    font_size = random.choice(range(32, 128))  # Increased minimum font size
-                    font = ImageFont.truetype(font_path, font_size)
-
-                # Counter for the image filename
-                j = 0
-                for i in range(IMAGES_PER_FONT):  # Generate 50 images per font - reduced to 10 for now to make things faster
-                    prose_sample = random_prose_text(all_brown_words)
-
-                    for text in [prose_sample]:
-                        img = Image.new('RGB', (800, 400), color="white")  # Canvas size
-                        draw = ImageDraw.Draw(img)
-
-                        # Random offsets, but ensuring that text isn't too far off the canvas
-                        offset_x = random.randint(-20, 10)
-                        offset_y = random.randint(-20, 10)
-                        draw.text((offset_x, offset_y), text, fill="black", font=font)
-
-                        j += 1
-                        output_file = os.path.join(GEN_IMAGES_DIR, f"{font_name}_{j}.png")
-                        img.save(output_file)
+                if font_name in FONT_ALLOWLIST:
+                    font_files.append((font_path, font_name))
+
+    # Generate images for each font file
+    for font_path, font_name in font_files:
+        # Output the font name so we can see the progress
+        print(font_path, font_name)
+
+        # Random font size
+        font_size = random.choice(range(18, 72))
+
+        if font_path.endswith('.ttc'):
+            # ttc fonts have multiple fonts in one file, so we need to specify which one we want
+            font = ImageFont.truetype(font_path, font_size, index=0)
+        else:
+            # ttf fonts have only one font in the file
+            font = ImageFont.truetype(font_path, font_size)
+
+        # Counter for the image filename
+        j = 0
+        for i in range(IMAGES_PER_FONT):  # Generate 50 images per font - reduced to 10 for now to make things faster
+            # Determine the number of words that will fit on a line
+            font_avg_char_width = font.getbbox('x')[2]
+            words_per_line = int(800 / (font_avg_char_width*5))
+            prose_sample = random_prose_text(line_length=words_per_line)
+
+            for text in [prose_sample]:
+                img = Image.new('RGB', (800, 400), color="white")  # Canvas size
+                draw = ImageDraw.Draw(img)
+
+                # Random offsets, but ensuring that text isn't too far off the canvas
+                offset_x = random.randint(-20, 10)
+                offset_y = random.randint(-20, 10)
+
+                # vary the line height
+                line_height = random.uniform(0, 1.25) * font_size
+                draw.text((offset_x, offset_y), text, fill="black", font=font, spacing=line_height)
+
+                j += 1
+                output_file = os.path.join(GEN_IMAGES_DIR, f"{font_name}_{j}.png")
+                img.save(output_file)
 
 if __name__ == '__main__':
     main()