DeepStorm: Deep Learning Framework

Summary:

Deep Learning Framework from scratch, with an API of a combination of pytorch and keras APIs, only uses numpy for tensor operations.

Pip install:

pip install DeepStorm

Layers & DL classes in framework:

• Conv2d
• MaxPool2d
• BatchNorm2d
• Flatten
• Dropout
• Linear
• ReLU
• Sigmoid
• Softmax
• SgdWithMomentum
• Adam
• CrossEntropyLoss
• Xavier
• He

Model building:

layers = [
    Conv2d(in_channels=1, out_channels=32,
           kernel_size=3, stride=1, padding='same'),
    BatchNorm2d(32),
    Dropout(probability=0.3),
    ReLU(),

    Conv2d(in_channels=32, out_channels=64,
           kernel_size=3, stride=1, padding='same'),
    BatchNorm2d(64),
    ReLU(),
    MaxPool2d(kernel_size=2, stride=2),

    Conv2d(in_channels=64, out_channels=64,
           kernel_size=3, stride=1, padding='same'),
    BatchNorm2d(64),
    ReLU(),
    MaxPool2d(kernel_size=2, stride=2),

    Flatten(),

    Linear(in_features=64*7*7, out_features=128),
    ReLU(),
    Linear(128, 64),
    ReLU(),
    Linear(64, 10),
    SoftMax(),
]

model = Model(layers)

Or

model = Model()

model.append_layer(Conv2d(in_channels=1, out_channels=32,
                          kernel_size=3, stride=1, padding='same'))
model.append_layer(BatchNorm2d(32))
model.append_layer(ReLU())
model.append_layer(Conv2d(in_channels=32, out_channels=64,
                          kernel_size=3, stride=1, padding='same'))
model.append_layer(BatchNorm2d(64))
model.append_layer(ReLU())
model.append_layer(MaxPool2d(kernel_size=2, stride=2))

model.append_layer(Conv2d(in_channels=64, out_channels=64,
                          kernel_size=3, stride=1, padding='same'))
model.append_layer(BatchNorm2d(64))
model.append_layer(ReLU())
model.append_layer(MaxPool2d(kernel_size=2, stride=2))
model.append_layer(Flatten())
model.append_layer(Linear(in_features=64*7*7, out_features=128))
model.append_layer(ReLU())
model.append_layer(Linear(in_features=128, out_features=64))
model.append_layer(ReLU())
model.append_layer(Linear(in_features=64, out_features=10))
model.append_layer(SoftMax())

Model compile:

batch_size = 16
model.compile(optimizer=Adam(learning_rate=5e-3, mu=0.98, rho=0.999), loss=CrossEntropyLoss(),
              batch_size=batch_size, metrics=['accuracy'])

Model training:

epochs = 25
history = model.fit(x_train=train_images, y_train=train_labels, x_val=val_images, y_val=val_labels, epochs=epochs)

Model performance:

plt.plot(history['accuracy'])
plt.plot(history['val_accuracy'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'val'], loc='upper left')
plt.show()