Made solely by Aevin Jais
This project utilises the Keras sequential model to create a custom deep convoluted neural network in order to classify images into multiple classes. In this specific case, there are three classes: Cats, Dogs and Fish.
I utilised an initial model composed of 3 convolutional layers in order to create the model. Upon training for 10 epochs, the val_accuracy converged to 73%. This was a solid start.
From here I increased the complexity of the model and added more 2D convolutional layers. Initially, I trained for 10 epochs but noticed the val_accuracy was dropping around 8, indicating overfitting was occurring. So I reran on 8 epochs and got the following:
X-Axis: epochs | Y-Axis: accuracy (max is 1)
This change resulted in an improved accuracy of 76%.
After, this I altered the learning rate via my optimizer (Adam) from the default of 0.001 to 0.0001. Here are the results:
X-Axis: epochs | Y-Axis: accuracy (max is 1)
This change resulted in an improved accuracy of 81%.
However, I believe there to be some degree of overfitting being introduced even in this model given that the gradient of val_accuracy plateaus around epoch 6. Nevertheless, the val_accuracy of this model was 81% and my highest-achieving model.
main.py offers the ability to:
- Create the model (given that you have the necessary libraries installed - and preferably have your GPU enabled for TensorFlow) (using the model template in the code of course)
- Build the model
- Train the model
- Evaluate the model
- Test on data that is unseen to the model
- Load an existing model (avoids the need to create a model)
Python 3.9.4 was used to run this code. I suggest you use Python 3.9. as well**
There is a requirements.txt file in the root directory. Use this along with pip in order to install the necessary libraries:
pip install -r requirements.txt
NOTE: requirements.txt will not enable GPU usage. You have to do that yourself. Read below:
This project was made on my laptop. I have a mobile RTX 3060, therefore I was able to train my models in minutes as opposed to hours. If you do not configure GPU usage or have a GPU usage the code will still work but just slowly.
To enable GPU usage, follow this guide on installing WSL on Windows machines: https://www.tensorflow.org/install/pip#step-by-step_instructions
Once installing WSL, you will need to run the following lines:
conda activate tf
export XLA_FLAGS=--xla_gpu_cuda_data_dir=/usr/lib/cuda
To be implemented
Language:
Python
Libraries:
- TensorFlow
- matplotlib
- cv2
- numpy
If you're open to contributing to this project please contact me via email: ajjaevinjais@gmail.com.
None
Email: ajjaevinjais@gmail.com IG: aevin.j
@NicholasRenotte @KGPTalkie @SimplilearnOfficial