Chinese Digit Recognition

---

Version 1.0

This is just a practice for learning pytorch. It is something I did to practice after finishing the official Pytorch tutorial. So, it's easy and very similar with the example FashionMNIST given by the official Pytorch tutorial. Here is my simple network.

class RecognitionNetwork(nn.Module):
    def __init__(self):
        super().__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(3*64 * 64, 4096),
            nn.Linear(4096, 1024),
            nn.ReLU(),
            nn.Linear(1024, 512),
            nn.ReLU(),
            nn.Linear(512, 15),
        )

    def forward(self, x):
        x = self.flatten(x)
        logits = self.linear_relu_stack(x)
        return logits

After 100 epochs we will reach about 70%

Epoch 99 
 -------------------------
loss: 0.660015 [   64/11999]
loss: 1.192717 [ 1344/11999]
loss: 0.761487 [ 2624/11999]
loss: 0.700470 [ 3904/11999]
loss: 0.984401 [ 5184/11999]
loss: 0.743671 [ 6464/11999]
loss: 0.635524 [ 7744/11999]
loss: 0.798912 [ 9024/11999]
loss: 1.092515 [10304/11999]
loss: 0.814506 [11584/11999]
Test Error: 
 Accuracy: 70.8%, Avg loss: 0.873255 

Epoch 100 
 -------------------------
loss: 0.750734 [   64/11999]
loss: 0.515754 [ 1344/11999]
loss: 0.836834 [ 2624/11999]
loss: 0.661097 [ 3904/11999]
loss: 0.719125 [ 5184/11999]
loss: 0.758513 [ 6464/11999]
loss: 0.656716 [ 7744/11999]
loss: 0.598785 [ 9024/11999]
loss: 0.849562 [10304/11999]
loss: 0.735557 [11584/11999]
Test Error: 
 Accuracy: 70.8%, Avg loss: 0.881129 

Done!

---

The Dataset

https://www.kaggle.com/datasets/gpreda/chinese-mnist

---

Run

git clone https://github.com/EMEEEEMMMM/ChineseDigitRecognition.git

Change the path in main.py and DataAdjustment.py

main.py

csv_file = (
    "/Change/it/to/your/path/ChineseDigitRecognition/data/chinese_mnist.csv"
)
project_dir = "/Change/it/to/your/path/ChineseDigitRecognition"

checkpoint = torch.load("/Change/it/to/your/path/ChineseDigitRecognition/model_checkpoint.pth")

DataAdjustment.py

project_root_dir = "/Change/it/to/your/path/ChineseDigitRecognition/"

Download the dataset from kaggle with the link put the images in data/data

Run the DataAdjustment.py

And then just run main.py

---

Grade 9 2025/06/30