autoencoder

Author: Sourland

src/Autoencoder/main.jl renamed to src/Autoencoder/Autoencoder.jl

@@ -3,6 +3,9 @@ using Images
 using Flux.Data, MLDatasets
 using Flux.Data:DataLoader
 using Noise
+
+include("train.jl")
+
 device = cpu # where will the calculations be performed?
 
 x_train, y_train = MNIST.traindata(Float32);
@@ -11,13 +14,10 @@ x_test, y_test = MNIST.testdata(Float32);
 m, n = size(x_test[:,:,1])
 
 x_train = reshape(x_train, (m*n,size(x_train)[3]))
-# add some salt pepper noise to the image
-x_train_noise = add_gauss(x_train, 0.05) 
 x_test = reshape(x_test, (m*n,size(x_test)[3]))
 
 loader = DataLoader((data = x_train, label = x_train), batchsize=512, shuffle=true)
 
-f
 
 autoencoder1 = Chain(
     #ENCODER
@@ -65,6 +65,6 @@ params1 = Flux.params(autoencoder1)
 params2 = Flux.params(autoencoder2)
 params3 = Flux.params(autoencoder3) # parameters
 
-train!(loss, params1, opt, loader, 10)
-train!(loss, params2, opt, loader, 35)
-train!(loss, params3, opt, loader, 35)
+# train!(loss,1 params1, opt, loader, 10)
+# train!(loss2, params2, opt, loader, 35)
+train!(loss3, params3, opt, loader, 35)

src/Autoencoder/CompressionAutoncoder.jl

@@ -0,0 +1,17 @@
+using Flux
+using Images
+using Flux.Data, MLDatasets
+using Flux.Data:DataLoader
+using Noise
+
+include("train.jl")
+device = cpu 
+
+x_train, y_train = CIFAR10.traindata(Float32);
+x_test, y_test = CIFAR10.testdata(Float32);
+
+
+height, width, channels, number_of_pictures = size(x_train)
+
+x_train = reshape(x_train, (m*n,size(x_train)[3]))
+x_test = reshape(x_test, (m*n,size(x_test)[3]))

src/Autoencoder/DAE.jl

@@ -0,0 +1,67 @@
+# DENOISING AUTOENCODER
+using Flux
+using Images
+using Flux.Data, MLDatasets
+using Flux.Data:DataLoader
+using Noise
+
+include("train.jl")
+device = cpu 
+
+device = cpu # where will the calculations be performed?
+
+x_train, y_train = MNIST.traindata(Float32);
+x_test, y_test = MNIST.testdata(Float32);
+
+x_train = 0.299 * x_train[:,:,1,:] + 0.587* x_train[:,:,2,:] + 0.114 * x_train[:,:,3,:]
+x_test = 0.299 * x_test[:,:,1,:] + 0.587* x_test[:,:,2,:] + 0.114 * x_test[:,:,3,:]
+
+m, n = size(x_test[:,:,1])
+
+x_train = reshape(x_train, (m*n,size(x_train)[3]))
+x_test = reshape(x_test, (m*n,size(x_test)[3]))
+
+# add some salt pepper noise to the image
+x_train_noise = add_gauss(x_train, 0.15) 
+x_test_noise = add_gauss(x_test, 0.15) 
+loader = DataLoader((data = x_train_noise, label = x_train), batchsize=512, shuffle=true)
+
+input_size = m*n
+
+DAE1 = Chain(
+    #ENCODER
+    Dense(input_size, input_size),
+    BatchNorm(input_size, relu),
+
+    Dense(input_size, input_size),
+
+    Dense(input_size, input_size),
+    BatchNorm(input_size, relu),
+
+    Dense(input_size, input_size, sigmoid)
+    
+   ) |> device
+
+
+DAE2 = Chain(
+#ENCODER
+Dense(input_size, 128),
+BatchNorm(128, relu),
+
+Dense(128, 16),
+
+Dense(16,128),
+
+# Dense(input_size, input_size),
+BatchNorm(128, relu),
+
+Dense(128, input_size, sigmoid)
+
+) |> device
+
+loss(x, y) = Flux.Losses.mse(DAE2(x), y)
+
+dae_params = Flux.params(DAE2)
+optim = ADAM(0.05)
+
+train!(loss, dae_params, optim, loader, 35)

src/Autoencoder/DenoisingAutoencoder.jl

@@ -19,9 +19,3 @@ function train!(model_loss, model_params, opt, loader, epochs = 10)
     end
     "Total train steps: $train_steps" |> println
 end
-
-function fit(model, data, optimizer, loss_func)
-Loss(x,y) = loss_func(x,y)
-model_params = Flux.params(model)
-
-end