Skip to content

Latest commit

 

History

History
87 lines (76 loc) · 4.34 KB

README.md

File metadata and controls

87 lines (76 loc) · 4.34 KB
Raw

$Font^2$

This repo contains the code to generate and download the $Font^2$ dataset presented in the paper Evaluating Synthetic Pre-Training for Handwriting Processing Tasks

If you find it useful, please cite it as:

@article{pippi2023evaluating,
  title={Evaluating Synthetic Pre-Training for Handwriting Processing Tasks},
  author={Pippi, Vittorio and Cascianelli, Silvia and Baraldi, Lorenzo and Cucchiara, Rita},
  journal={Pattern Recognition Letters},
  year={2023},
  publisher={Elsevier}
}

To obtain the dataset, please run:

dataset = Font2('path/to/dataset', store_on_disk=True, auto_download=True)
loader = DataLoader(dataset, batch_size=32, num_workers=1, collate_fn=dataset.collate_fn)

for i, (imgs, widths, font_ids, words) in enumerate(loader):
    print(imgs.shape, widths.shape, font_ids.shape, len(words))

This way, the loader will automatically download (auto_download=True) and save (store_on_disk=True) the data inside the blocks folder, at the same level as the fonts.json and words.json files.

Please note that you can also download the dataset (already shuffled and with the same augmentations as presented in the paper) manually from the releases, where can also find the checkpoints of a ResNet-18 and a VGG-19 trained on it.

Load checkpoints

You can find the checkpoints of the models trained on $Font^2$ in the releases

ResNet-18

import torch
from torchvision import models

model = models.resnet18(num_classes=10400)
checkpoint = torch.hub.load_state_dict_from_url('https://github.com/aimagelab/font_square/releases/download/ResNet-18/RN18_class_10400.pth')
model.load_state_dict(checkpoint)

VGG-16

import torch
from torchvision import models

model = models.vgg16(num_classes=10400)
checkpoint = torch.hub.load_state_dict_from_url('https://github.com/aimagelab/font_square/releases/download/VGG-16/VGG16_class_10400.pth')
model.load_state_dict(checkpoint)

Inception-v3

import torch
from torchvision import models

model = models.inception_v3(num_classes=10400, init_weights=False, aux_logits=False)
checkpoint = torch.hub.load_state_dict_from_url('https://github.com/aimagelab/font_square/releases/download/Inception-v3/IV3_class_10400.pth')
model.load_state_dict(checkpoint, strict=False)

WARNING: Since the input size of the Inception-v3 is $299 \times 299$ we implemented the following transform to adapt the $Font^2$ samples to the network input:

class ToInceptionV3Input:
    def __init__(self, size=299):
        self.size = size

    def __call__(self, x):
        h_rep = math.ceil(self.size / x.shape[1])
        w_rep = math.ceil(self.size / x.shape[2])
        return x.repeat(1, h_rep, w_rep)[:, :self.size, :self.size]

The ToInceptionV3Input transform repeats the input image vertically and horizontally to cover the whole input image and trims the excess to fit the desired shape.

Results

In the following table there are the Top-1, Top-5, and Top-10 accuracy scores on the Font2 Test set.

dataset = Font2('font_square', store_on_disk=True, auto_download=True, nameset='test')
loader = DataLoader(dataset, batch_size=32, num_workers=1, collate_fn=dataset.collate_fn)
Classes Top-1 Top-5 Top-10
ResNet-18 10400 79.81 94.55 96.99
VGG-16 10400 81.72 96.24 98.12
Inception-v3 10400 82.44 95.75 97.71

Train with a subset of the dataset

If you want to train on a subset of the dataset, you can define fonts_ids and words_ids as follows:

dataset = Font2('path/to/dataset', store_on_disk=True, auto_download=True, fonts_ids=[0, 1, 2], words_ids=[0, 1, 2])
loader = DataLoader(dataset, batch_size=32, num_workers=1, collate_fn=dataset.collate_fn)

WARNING: Since each block is evenly distributed among the fonts and the words, training on a subset of the dataset reduces the number of available samples per block. (e.g., if you train on 1000 fonts and 1000 words, it means that you are using $(1000 \times 1000) / (10400 * 10400) \approx 0.009245$ of the dataset. Therefore, each block contains roughly $0.009245 \times 200000 \approx 1,849.11$ available samples).