The notebook mnist_comparison.ipynb contains the comparison of popular deep learning libraries PyTorch and Tensorflow, along with the custom deep learning model built using only the NumPy library.
Python version - 3.9.5
Clone the repository, create a virtual environment and download the dependencies from requirements.txt file using the command
pip3 install -r requirements.txt
In this notebook, we download the MNIST dataset and prepare it so, that we have the same data fed into all the models rather than using the dataset repositories of the PyTorch or Tensorflow libraries.
We are implementing a two-layer fully connected network.
| Layer | Number of neurons | Activation Function |
|---|---|---|
| 1 | 784 | None |
| 2 | 128 | ReLu |
| 3 | 10 | Softmax |
Other parameters:
- Loss Function - Cross entropy loss
- Optimizer - Gradient descent with a learning rate of 0.01
- Number of epochs of training - 10
After training, we evaluate the model using accuracy on the testing set.
All the models were trained and tested using only the CPU.
| Model | Accuracy (%) | Training Time (min) |
|---|---|---|
| Custom model | 97.23 | 5.88 |
| PyTorch model | 97.45 | 7.41 |
| TensorFlow model | 97.56 | 19.92 |
Observing the above table, we can conclude that all three models performed almost similarly in terms of accuracy.
Well, our model is on par with the top deep learning frameworks (at least while performing all operations in CPU).