Handwritten Digit Classifier
A PyTorch image classification model that recognizes handwritten digits (0–9), built on a pretrained ResNet-18 backbone (ImageNet weights) fine-tuned on a combined MNIST + EMNIST dataset with aggressive data augmentation. Achieves 99.46% accuracy on the combined test set.
Model Details
Property Value Architecture ResNet-18 (pretrained on ImageNet) Framework PyTorch Task Image Classification (10 classes, digits 0–9) Input Size 32 × 32 (grayscale, converted to 3-channel) Output Softmax probabilities over digits 0–9 Test Accuracy 99.46% Training Device CUDA (GPU) Epochs 7 Batch Size 256 Optimizer Adam (differential learning rates) Loss Function CrossEntropyLoss LR Scheduler StepLR (step=2, gamma=0.5)
Architecture
The model uses a ResNet-18 backbone pretrained on ImageNet, with the default classification head replaced by a custom fully-connected head:
ResNet-18 Backbone (pretrained on ImageNet1K)
↓
Linear(512 → 128)
↓
ReLU()
↓
Dropout(0.3)
↓
Linear(128 → 10)
↓
Softmax (at inference)Differential learning rates were used to preserve pretrained features while allowing the new head to learn faster:
Pretrained backbone layers: lr = 0.0001
New classification head (last 4 param groups): lr = 0.001
The dropout layer (p=0.3) reduces overfitting given the simplicity of digit images relative to the model's capacity.
Dataset
The model was trained on a combined MNIST + EMNIST (digits split) dataset for greater diversity and robustness.
MNIST
Property Value Classes 10 (digits 0–9) Training set 60,000 grayscale images Test set 10,000 grayscale images Image size 28 × 28 pixels Source yann.lecun.com/exdb/mnist
EMNIST (digits split)
Property Value Classes 10 (digits 0–9) Training set 240,000 grayscale images Test set 40,000 grayscale images Image size 28 × 28 pixels Source NIST Special Database 19
Combined total: 300,000 training images and 50,000 test images.
Training
The model was trained for 7 epochs on CUDA with a StepLR scheduler (halving LR every 2 epochs). Loss decreased consistently across all epochs.
Epoch Loss 1 0.1732 2 0.0635 3 0.0446 4 0.0409 5 0.0340 6 0.0307 7 0.0279
Final Test Accuracy: 99.46%
Data Augmentation
Aggressive augmentation was applied during training to improve generalization to real-world handwriting styles:
Augmentation Parameters Random Rotation ±15° Random Affine (translate) ±15% horizontal and vertical Random Affine (shear) 10° Random Perspective distortion scale 0.3, p=0.3 Color Jitter brightness ±0.3, contrast ±0.3 Normalization mean (0.5, 0.5, 0.5), std (0.5, 0.5, 0.5)
No augmentation was applied to the test set (only resize + normalize).
Preprocessing
At inference, input images go through the following pipeline:
Convert to grayscale
Invert colors (white background → black background to match MNIST format)Resize to 32 × 32Convert to 3-channel (grayscale replicated across RGB channels for ResNet compatibility)
Normalize with mean (0.5, 0.5, 0.5) and std (0.5, 0.5, 0.5)
Usage
1 import torch
2 import torch . nn as nn
3 from torchvision import transforms , models
4 from huggingface_hub import hf_hub_download
5 from PIL import Image
6 import numpy as np
7
8 # Load model
9 model = models . resnet18 ( weights = None )
10 model . fc = nn . Sequential (
11 nn . Linear ( 512 , 128 ) , nn . ReLU ( ) , nn . Dropout ( 0.3 ) , nn . Linear ( 128 , 10 )
12 )
13
14 weights_path = hf_hub_download (
15 repo_id = "AdityaManojShinde/handwritten_digit_classifier" ,
16 filename = "mnist_model.pth"
17 )
18 model . load_state_dict ( torch . load ( weights_path , map_location = "cpu" ) )
19 model . eval ( )
20
21 # Preprocessing pipeline
22 transform = transforms . Compose ( [
23 transforms . Grayscale ( num_output_channels = 3 ) ,
24 transforms . Resize ( ( 32 , 32 ) ) ,
25 transforms . ToTensor ( ) ,
26 transforms . Normalize ( ( 0.5 , 0.5 , 0.5 ) , ( 0.5 , 0.5 , 0.5 ) ) ,
27 ] )
28
29 # Inference
30 image = Image . open ( "your_digit.png" ) . convert ( "L" )
31 img_array = 255 - np . array ( image ) # invert: white bg → black bg
32 image = Image . fromarray ( img_array )
33 img_tensor = transform ( image ) . unsqueeze ( 0 )
34
35 with torch . no_grad ( ) :
36 output = model ( img_tensor )
37 probs = torch . nn . functional . softmax ( output , dim = 1 ) [ 0 ]
38 predicted = probs . argmax ( ) . item ( )
39
40 print ( f"Predicted digit: { predicted } ( { probs [ predicted ] * 100 : .1f } % confidence)" )
Limitations
Works best with centered, clearly written single digits on a plain background.
Not suitable for multi-digit recognition or digit detection in natural scenes.
May struggle with highly stylized or non-standard digit handwriting not represented in MNIST/EMNIST.
License
This model is released under the
MIT License .
