grandlantern 0.0.7

This is the small library for deep learning.

This is Grand Lantern neural network library рџЏ®

The main purpose of this library is to understand - what is happening under the hood of popular deep learning frameworks. Also, it can be used to build your own ideas - no special knowledge for thst you need!

This library is written only using numpy (maybe cupy in future), you can see every calculation in neural network!

How to install в›ЏпёЏ

Put just this command to terminal

pip install grand_lantern

Get started рџљЂ

1. Import library

import grand_lantern as gl

2. Define Data Iterator

from grand_lantern.dataiterators import DatasetIterator

batch_size = 100
my_dataset_iterator = DatasetIterator(dataset = gl.TableDataset(), batch_size = batch_size)

3. Choose optimizer

from grand_lantern.optimizers import SGD

my_optimizer = SGD(learning_rate = 0.01)

4. Build model

from grand_lantern import model
from grand_lantern.metrics import CrossEntropy, Accuracy

NN = model(n_epochs=100,
           dataset_iterator = my_dataset_iterator,
           loss_function = CrossEntropy(),
           metric_function = Accuracy(),
           optimizer = my_optimizer)

5. Add layers

from grand_lantern.layers import LinearLayer
from grand_lantern.layers.Activation import Sigmoid, ReLU, SoftMax

NN.add_layer(LinearLayer(n_neurons=100, activation=Sigmoid(), biased=True))
NN.add_layer(LinearLayer(n_neurons=50, activation=ReLU(), biased=True))
NN.add_layer(LinearLayer(n_neurons=10, activation=SoftMax(), biased=True))

6. Train model

NN.fit(X_train, y_train.reshape(-1, 1), X_test, y_test.reshape(-1, 1))

7. Use model

y_pred = NN.predict(X_test)

What is supported now вњ…

Layers

• Linear Layer
• Image Convolutional Layer (Conv2DLayer)
• Batch Normalization Layer (BatchNormLayer)
• RNN Layer (RNNLayer)

Optimizers

• SGD
• NAG
• Adagrad
• Adam

What will be done рџ“ќ

New Layers

• LSTM, GRU layers
• Transformers

GPU Accelaration using cupy

Some more optimizers

Regularization

• L1, L2 regularization
• Dropout
• Pooling

Preprocessing modules