DeepLearning_NanoDegree

Project overview for the deep learning nanodegree from Udacity. This page is just a brief overview over some selected results. For further details please feel free to have a look into the project folders.

---

Fully connected networks: Bike sharing predictor

Project to implement a fully conncted net without external frameworks like Pytorch or Tensorflow. This includes both forward and backwards pass of the network.

Architecture visualized:

With this architecture the following results were archived on the bike sharing dataset:

---

Convolutional neural networks

Dog breed classifier: Custom network architecture

The project task was to design a custom CNN for dog breed classification and achieve a accuracy greater than 10%. My approach features 4 convolutional layers with 3x3 kernels and max-pooling with stride 2 after each convolution. The final classifier vector contains 133 values for the 133 dog breed classes. The final accuracy was 12%.

Net(
  (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
  (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (conv4): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (pools2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  (linear): Linear(in_features=12544, out_features=133, bias=True)
  (dropout): Dropout(p=0.3)
)

Architecture visualized:

Dog breed classifier: Transfer learning

The second task was to create the same dog breed classifier via transfer learning. As reference model I choose ResNet-50 Added a linear layer with 133 outputs as final classifier. Without further data augmentation this resulted in a accuracy of 67% on the test set.

Some classification results of the final app:

Style transfer with VGG

During the chapter I implemented a CNN which uses style transfer to blend 2 different images into one. The network learns to only extract the style from on of the pictures and to keep to content in the other one. Here is the result of a photo I took myself back in 2017 mixed with the style of the artist Hundertwasser:

---

Reccurent neural network: TV script generation

Implement a recurrent neural network with Long short-term memory to create new tv scripts. Architechture of my implementation looks like this:

RNN(
  (embedding): Embedding(21384, 150)
  (lstm): LSTM(150, 250, num_layers=2, batch_first=True, dropout=0.5)
  (dropout): Dropout(p=0.3)
  (fc): Linear(in_features=250, out_features=21384, bias=True)
)

Some newly invented tv scripts by the RNN can be found in the corresponding sub-folder.

---

Generative adverserial network (GAN): Face generation

Project to implement a GAN with generator and discriminator for face generation based on the CelebA dataset. Training process after 35 epochs:

Here are some faces the network generated:

---

AWS Sagemaker deployment: IMDB sentiment analysis

Project to deploy a sentiment analysis model by using AWS Sagemaker and a REST API. The result was a deployed ML endpoint integrated into a intuitive website by using Lambda functions and the AWS API gateway.

---