This repositor contains some projects that i've made on my studies about neural network
On this one i got to make a image classifier, i used the CIFAR10 dataset for this. For validation, i pick up the holdout. 80% train, 10% test and 10% validation. For reduce the overfitting i choose the data augmentation technique, so i flip, rotate and normalized the picture during the training.
| Metrics | Results |
|---|---|
| Epochs | 20 |
| Loss error | 0.9148 |
| Val Loss: | 0.8675 |
| Acuracy | 70% |
| F1-score avg | 70% |
On LSTM i desing a price estimator for bitcoin price. I used the bitcoin historic price change, but i only took data from 2023 onwards. Why? Because my experiment was taking more than 6 hours to complete. And yes using the VGA on google colab. I normalizes the atributes using the z-score, because is (as far as I know) the best way to normalize without experiencing significant fluctuations in the data's means and standard deviations, and without the data being influenced by extreme variations.
| Metrics | Results |
|---|---|
| Lookback | 72 |
| Batch size | 2048 |
| Batches | 12 |
| Hidden size | 128 |
| Numbers of layers | 3 |
| Output Size | 1 |
| Epochs | 25 |
| Loss error | 0.001212 |
| Val Loss | 0.001100 |
| MAE | 0.76% |
On VAEs i created an anomaly detector. So the neural network learn how to tell the difference, between an image she recognizes and one she doesn't. For this one i got the fashionmnist dataset. Again i used holdout for separate the dataset 80/10/10. The model compression is from 748 to 32 dimensions.
| Metrics | Results |
|---|---|
| Epochs | 40 |
| Loss error | 272.1 |
| Val Loss | 272.5 |
| Threshold MSE | 0.0405 |
| MCC (mathews correlation coeficient) Max | 0.7631 |
| F1-score avg | 88% |
These are the "simplest" ones. I built two neural network implementations using PyTorch to understand how the network changes and learns by adjusting its parameters such as layers, neurons, and epochs and by employing techniques to prevent overfitting, like dropout. These two implementations were designed for linear regression and a sine wave.