DianoiaML is an experimental Keras-like deep learning framework.
The user guide and the To-Do list can be found here.
Environment: Julia 1.6.1
Features:
- Network
- Sequential
- GAN
- ResNet
- Layer
- Flatten
- Constructive
- Dense
- Convolutional2D
- MaxPooling2D
- UpSampling2D
- Transposed Convolutional2D
- Dropout
- Residual
- Activation Function
- ReLU
- Sigmoid
- Softmax
- tanh
- Loss Function
- Quadratic Loss
- Categorical Cross Entropy Loss
- Binary Cross Entropy Loss
- Mean Squared Error
- Monitor
- Absolute
- Classification
- Optimizer
- Minibatch Gradient Descent
- Stochastic Gradient Descent
- Adam
- AdaBelief
- Genetic Algorithm
- Tools
- Model Management
- One Hot
An example of Sequential network:
using MLDatasets, DianoiaML
train_x, train_y = MNIST.traindata()
test_x, test_y = MNIST.testdata()
dict = Dict{Int64, Int64}(1=>1, 2=>2, 3=>3, 4=>4, 5=>5, 6=>6, 7=>7, 8=>8, 9=>9, 0=>10)
model = Sequential()
model.add_layer(model, Conv2D; filter=32, input_shape=(28,28,1), kernel_size=(3,3), activation_function=ReLU)
model.add_layer(model, MaxPooling2D; pool_size=(2,2))
model.add_layer(model, Conv2D; filter=64, kernel_size=(3,3), activation_function=ReLU)
model.add_layer(model, MaxPooling2D; pool_size=(2,2))
model.add_layer(model, Flatten;)
model.add_layer(model, Dense; layer_size=128, activation_function=ReLU)
model.add_layer(model, Dense; layer_size=10, activation_function=Softmax_CEL)
Adam.fit(model=model, input_data=Array{Float32}(reshape(train_x, 28,28,1,60000)), output_data=oneHot(train_y, 10, dict),
loss_function=Categorical_Cross_Entropy_Loss, monitor=Classification, epochs=10, batch=128)An example of GAN:
using MLDatasets, DianoiaML
train_x, train_y = MNIST.traindata()
test_x, test_y = MNIST.testdata()
dict = Dict{Int64, Int64}(1=>1, 2=>2, 3=>3, 4=>4, 5=>5, 6=>6, 7=>7, 8=>8, 9=>9, 0=>10)
function noise()
return reshape([rand(1.0f0:1.0f0:10.0f0), rand(Float32)], (2,1))
end
model = GAN(noise)
model.add_Glayer(model, Dense; input_shape=(2,), layer_size=16, activation_function=ReLU)
model.add_Glayer(model, Dense; layer_size=64, activation_function=ReLU)
model.add_Glayer(model, Constructive; shape=(8,8,1))
model.add_Glayer(model, UpSampling2D; size=(2,2), activation_function=None)
model.add_Glayer(model, Flatten;)
model.add_Glayer(model, Dense; layer_size=256, activation_function=ReLU)
model.add_Glayer(model, Dense; layer_size=784, activation_function=ReLU)
model.add_Glayer(model, Constructive; shape=(28,28,1))
model.add_Dlayer(model, Conv2D; filter=16, kernel_size=(3,3), activation_function=ReLU)
model.add_Dlayer(model, Conv2D; filter=32, kernel_size=(3,3), activation_function=ReLU)
model.add_Dlayer(model, MaxPooling2D; kernel_size=(2,2), activation_function=None)
model.add_Dlayer(model, Flatten;)
model.add_Dlayer(model, Dense; layer_size=128, activation_function=ReLU)
model.add_Dlayer(model, Dense; layer_size=64, activation_function=ReLU)
model.add_Dlayer(model, Dense; layer_size=2, activation_function=Sigmoid)
SGD.fit(model=model, input_data=Array{Float32}(reshape(train_x, 28,28,1,60000)), output_data=oneHot(train_y, 10, dict),
loss_function=Binary_Cross_Entropy_Loss, monitor=Classification, epochs=50, batch=128)An example of ResNet:
using MLDatasets, DianoiaML
train_x, train_y = MNIST.traindata()
test_x, test_y = MNIST.testdata()
dict = Dict{Int64, Int64}(1=>1, 2=>2, 3=>3, 4=>4, 5=>5, 6=>6, 7=>7, 8=>8, 9=>9, 0=>10)
resnet = ResNet()
resnet.add_layer(resnet, Dense; input_shape=(128,), layer_size=64, activation_function=ReLU)
resnet.add_layer(resnet, Dense; layer_size=64, activation_function=ReLU)
resnet.add_layer(resnet, Dense; layer_size=128, activation_function=ReLU)
model = Sequential()
model.add_layer(model, Conv2D; filter=32, input_shape=(28,28,1), kernel_size=(3,3), activation_function=ReLU)
model.add_layer(model, MaxPooling2D; pool_size=(2,2))
model.add_layer(model, Conv2D; filter=64, kernel_size=(3,3), activation_function=ReLU)
model.add_layer(model, MaxPooling2D; pool_size=(2,2))
model.add_layer(model, Flatten;)
model.add_layer(model, Dropout; rate=0.5)
model.add_layer(model, Dense; layer_size=128, activation_function=ReLU)
model.add_layer(model, Residual; resnet=resnet)
model.add_layer(model, Dense; layer_size=10, activation_function=Softmax_CEL)
Adam.fit(model=model, input_data=Array{Float32}(reshape(train_x, 28,28,1,60000)), output_data=oneHot(train_y, 10, dict),
loss_function=Categorical_Cross_Entropy_Loss, monitor=Classification, epochs=10, batch=128)Please feel free to leave comments, trouble-shootings or advice (which are very valuable for me).