Gabor Cselle
commited on
Commit
•
99f802a
1
Parent(s):
95ccd40
Train a Font Identifier using ResNet18
Browse files- README.md +3 -2
- arrange_train_test_images.py +2 -2
- gen_sample_data.py +3 -1
- requirements.txt +5 -0
- train_font_identifier.py +122 -0
README.md
CHANGED
@@ -7,5 +7,6 @@ Follow along:
|
|
7 |
- [On Threads.net](https://www.threads.net/@gaborcselle/post/CzZJpJCpxTz)
|
8 |
- [On Twitter](https://twitter.com/gabor/status/1722300841691103467)
|
9 |
|
10 |
-
Generate sample images (note this will work only on Mac): [gen_sample_data.py]
|
11 |
-
Arrange test images into test and train: [arrange_train_test_images.py]
|
|
|
|
7 |
- [On Threads.net](https://www.threads.net/@gaborcselle/post/CzZJpJCpxTz)
|
8 |
- [On Twitter](https://twitter.com/gabor/status/1722300841691103467)
|
9 |
|
10 |
+
Generate sample images (note this will work only on Mac): [gen_sample_data.py](gen_sample_data.py)
|
11 |
+
Arrange test images into test and train: [arrange_train_test_images.py](arrange_train_test_images.py)
|
12 |
+
Train a ResNet18 on the data: [train_font_identifier.py](train_font_identifier.py)
|
arrange_train_test_images.py
CHANGED
@@ -29,10 +29,10 @@ for font in fonts:
|
|
29 |
train_files = font_files[:int(0.8 * len(font_files))]
|
30 |
test_files = font_files[int(0.8 * len(font_files)):]
|
31 |
|
32 |
-
#
|
33 |
for train_file in train_files:
|
34 |
shutil.move(os.path.join(source_dir, train_file), font_train_dir)
|
35 |
|
36 |
-
#
|
37 |
for test_file in test_files:
|
38 |
shutil.move(os.path.join(source_dir, test_file), font_test_dir)
|
|
|
29 |
train_files = font_files[:int(0.8 * len(font_files))]
|
30 |
test_files = font_files[int(0.8 * len(font_files)):]
|
31 |
|
32 |
+
# Move training files
|
33 |
for train_file in train_files:
|
34 |
shutil.move(os.path.join(source_dir, train_file), font_train_dir)
|
35 |
|
36 |
+
# Move test files
|
37 |
for test_file in test_files:
|
38 |
shutil.move(os.path.join(source_dir, test_file), font_test_dir)
|
gen_sample_data.py
CHANGED
@@ -7,6 +7,8 @@ import nltk
|
|
7 |
from nltk.corpus import brown
|
8 |
import random
|
9 |
|
|
|
|
|
10 |
# Download the necessary data from nltk
|
11 |
nltk.download('brown')
|
12 |
|
@@ -55,7 +57,7 @@ for font_dir in font_dirs:
|
|
55 |
|
56 |
# Counter for the image filename
|
57 |
j = 0
|
58 |
-
for i in range(
|
59 |
prose_sample = random_prose_text(all_brown_words)
|
60 |
|
61 |
for text in [prose_sample]:
|
|
|
7 |
from nltk.corpus import brown
|
8 |
import random
|
9 |
|
10 |
+
IMAGES_PER_FONT = 50
|
11 |
+
|
12 |
# Download the necessary data from nltk
|
13 |
nltk.download('brown')
|
14 |
|
|
|
57 |
|
58 |
# Counter for the image filename
|
59 |
j = 0
|
60 |
+
for i in range(IMAGES_PER_FONT): # Generate 50 images per font - reduced to 10 for now to make things faster
|
61 |
prose_sample = random_prose_text(all_brown_words)
|
62 |
|
63 |
for text in [prose_sample]:
|
requirements.txt
CHANGED
@@ -1 +1,6 @@
|
|
|
|
1 |
Pillow==9.5.0
|
|
|
|
|
|
|
|
|
|
1 |
+
nltk==3.8.1
|
2 |
Pillow==9.5.0
|
3 |
+
torch==2.0.0
|
4 |
+
torchaudio==2.0.1
|
5 |
+
torchvision==0.15.1
|
6 |
+
tqdm==4.65.0
|
train_font_identifier.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import os
|
3 |
+
import time
|
4 |
+
import torch
|
5 |
+
import torch.optim as optim
|
6 |
+
from torch.optim import lr_scheduler
|
7 |
+
from torchvision import datasets, models, transforms
|
8 |
+
from tqdm import tqdm
|
9 |
+
|
10 |
+
# Directory with organized font images
|
11 |
+
data_dir = './train_test_images'
|
12 |
+
|
13 |
+
# Define transformations for the image data
|
14 |
+
data_transforms = {
|
15 |
+
'train': transforms.Compose([
|
16 |
+
transforms.Resize((224, 224)), # Resize to the input size expected by the model
|
17 |
+
transforms.ToTensor(),
|
18 |
+
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) # ImageNet standards
|
19 |
+
]),
|
20 |
+
'test': transforms.Compose([
|
21 |
+
transforms.Resize((224, 224)), # Resize to the input size expected by the model
|
22 |
+
transforms.ToTensor(),
|
23 |
+
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
24 |
+
]),
|
25 |
+
}
|
26 |
+
|
27 |
+
|
28 |
+
# Create datasets
|
29 |
+
image_datasets = {
|
30 |
+
x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x])
|
31 |
+
for x in ['train', 'test']
|
32 |
+
}
|
33 |
+
|
34 |
+
# Create dataloaders
|
35 |
+
dataloaders = {
|
36 |
+
'train': torch.utils.data.DataLoader(image_datasets['train'], batch_size=4),
|
37 |
+
'test': torch.utils.data.DataLoader(image_datasets['test'], batch_size=4)
|
38 |
+
}
|
39 |
+
|
40 |
+
# Define the model
|
41 |
+
model = models.resnet18(weights=models.ResNet18_Weights.DEFAULT)
|
42 |
+
|
43 |
+
# Define the loss function
|
44 |
+
criterion = torch.nn.CrossEntropyLoss()
|
45 |
+
|
46 |
+
# Optimizer (you can replace 'model.parameters()' with specific parameters to optimize if needed)
|
47 |
+
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
|
48 |
+
|
49 |
+
# Decay LR by a factor of 0.1 every 7 epochs
|
50 |
+
exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)
|
51 |
+
|
52 |
+
# Number of epochs to train for
|
53 |
+
num_epochs = 25
|
54 |
+
|
55 |
+
def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
|
56 |
+
since = time.time()
|
57 |
+
|
58 |
+
best_model_wts = copy.deepcopy(model.state_dict())
|
59 |
+
best_acc = 0.0
|
60 |
+
|
61 |
+
for epoch in range(num_epochs):
|
62 |
+
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
|
63 |
+
print('-' * 10)
|
64 |
+
|
65 |
+
# Each epoch has a training and validation phase
|
66 |
+
for phase in ['train', 'test']:
|
67 |
+
if phase == 'train':
|
68 |
+
model.train() # Set model to training mode
|
69 |
+
else:
|
70 |
+
model.eval() # Set model to evaluate mode
|
71 |
+
|
72 |
+
running_loss = 0.0
|
73 |
+
running_corrects = 0
|
74 |
+
|
75 |
+
# Iterate over data.
|
76 |
+
# Here we wrap the dataloader with tqdm for a progress bar
|
77 |
+
for inputs, labels in tqdm(dataloaders[phase], desc=f"Epoch {epoch} - {phase}"):
|
78 |
+
# Zero the parameter gradients
|
79 |
+
optimizer.zero_grad()
|
80 |
+
|
81 |
+
# Forward
|
82 |
+
# Track history if only in train
|
83 |
+
with torch.set_grad_enabled(phase == 'train'):
|
84 |
+
outputs = model(inputs)
|
85 |
+
_, preds = torch.max(outputs, 1)
|
86 |
+
loss = criterion(outputs, labels)
|
87 |
+
|
88 |
+
# Backward + optimize only if in training phase
|
89 |
+
if phase == 'train':
|
90 |
+
loss.backward()
|
91 |
+
optimizer.step()
|
92 |
+
|
93 |
+
# Statistics
|
94 |
+
running_loss += loss.item() * inputs.size(0)
|
95 |
+
running_corrects += torch.sum(preds == labels.data)
|
96 |
+
if phase == 'train':
|
97 |
+
scheduler.step()
|
98 |
+
|
99 |
+
epoch_loss = running_loss / len(image_datasets[phase])
|
100 |
+
epoch_acc = running_corrects.double() / len(image_datasets[phase])
|
101 |
+
|
102 |
+
print('{} Loss: {:.4f} Acc: {:.4f}'.format(
|
103 |
+
phase, epoch_loss, epoch_acc))
|
104 |
+
|
105 |
+
# Deep copy the model
|
106 |
+
if phase == 'test' and epoch_acc > best_acc:
|
107 |
+
best_acc = epoch_acc
|
108 |
+
best_model_wts = copy.deepcopy(model.state_dict())
|
109 |
+
|
110 |
+
print()
|
111 |
+
|
112 |
+
time_elapsed = time.time() - since
|
113 |
+
print('Training complete in {:.0f}m {:.0f}s'.format(
|
114 |
+
time_elapsed // 60, time_elapsed % 60))
|
115 |
+
print('Best test Acc: {:4f}'.format(best_acc))
|
116 |
+
|
117 |
+
# Load best model weights
|
118 |
+
model.load_state_dict(best_model_wts)
|
119 |
+
return model
|
120 |
+
|
121 |
+
# Train the model
|
122 |
+
model = train_model(model, criterion, optimizer, exp_lr_scheduler, num_epochs=num_epochs)
|