Fine
A keras-like neural network framework built purely using Python and Numpy that's just that, fine.
Table of Contents
1- How to use
2- Demo
3- Technical Specifications
How to use
git clone git@github.com:haidousm/fine.git
cd fine
python3 -m pip install -r requirements.txt
Demo
MNIST Demo Link
Demo was built using javascript for the frontend, and a flask server to serve predictions from the model.
Demo model creation & training:
from datasets import load_mnist
from models import Sequential
from layers import Conv2D
from layers import MaxPool2D
from layers import Flatten
from layers import Dense
from activations import ReLU
from activations import Softmax
from loss import CategoricalCrossEntropy
from models.model_utils import Categorical
from optimizers import Adam
X_train, y_train, X_test, y_test = load_mnist()
model = Sequential(
layers=[
Conv2D(16, (1, 3, 3)),
ReLU(),
Conv2D(16, (16, 3, 3)),
ReLU(),
MaxPool2D((2, 2)),
Conv2D(32, (16, 3, 3)),
ReLU(),
Conv2D(32, (32, 3, 3)),
ReLU(),
MaxPool2D((2, 2)),
Flatten(),
Dense(1568, 64),
ReLU(),
Dense(64, 64),
ReLU(),
Dense(64, 10),
Softmax()
],
loss=CategoricalCrossEntropy(),
optimizer=Adam(decay=1e-3),
accuracy=Categorical()
)
model.train(X_train, y_train, epochs=5, batch_size=120, print_every=100)
model.evaluate(X_test, y_test, batch_size=120)
Technical Specifications
Layers
- Dense Layer
- Dropout Layer
- Flatten Layer
- 2D Convolutional Layer
- Max Pool Layer
Activation Functions
- Rectified Linear (ReLU)
- Sigmoid
- Softmax
- Linear
Loss Functions
- Categorical Cross Entropy
- Binary Cross Entropy
- Mean Squared Error
Optimizers
- Stochastic Gradient Descent (SGD) with rate decay and momentum
- Adaptive Moment Estimation (ADAM)
