How to use it just for testing one normal image?

[oneTaken]

Don't know hot to use it.

[zhiqwang]

Hi oneTaken, you can refer test script for one image below.

import torch
import torchvision.transforms as transforms
from models.crnn import init_network
from datasets.datahelpers import default_loader
from utils.converter import LabelConverter

if __name__ == '__main__':

    img_name = './data/dev/00129_19515496090345.jpg'
    device = torch.device("cpu")
    model_path = './data/model/densenet_cifar.pth'

    alphabet = "0123456789"
    model_params = {}
    model_params['architecture'] = "densenet_cifar"
    model_params['num_classes'] = len(alphabet) + 1
    model_params['mean'] = [0.396, 0.576, 0.562]
    model_params['std'] = [0.154, 0.128, 0.130]
    model = init_network(model_params)
    model = model.to(device)

    # load checkpoint
    checkpoint = torch.load(model_path)
    model.load_state_dict(checkpoint['state_dict'])

    converter = LabelConverter(alphabet)

    transform = transforms.Compose([
        transforms.Resize((32, 200)),
        transforms.ToTensor(),
        transforms.Normalize(mean=model.meta['mean'], std=model.meta['std']),
    ])
    img = default_loader(img_name)
    img = transform(img)
    img = img.unsqueeze(0)

    log_probs = model(img)
    preds = converter.best_path_decode(log_probs)

    print(preds)

[oneTaken]

Thanks a lot.

And I do not find the model path in the repo. Can you provide your pre-trained model to let me dive into your code?

Thanks very much.

[zhiqwang]

Hi oneTaken, Thanks for your attention. My previous model was trained on a private datasets. Now I am training a public datasets with Chinese character, and I will push the model in GitHub when I am done.

[oneTaken]

So sad to hear this.

[zhiqwang]

This work will be completed about in 1 day.

[zhiqwang]

Hi @oneTaken, I have updated the readme, load the pre-trained model, and make some difference in the training strategy, you can refer to the last commit.