AI Programming with Python Project - DeepLearning Classifier

Project code for Udacity's AI Programming with Python Nanodegree program. This code is used for classifying images in the predict.py file with the model you trained with resnet, alexnet or vgg previously on the train.py file. In addition, there's a trainning, testing and plotting exercise developed into a notebook at the Image_Classifier_Proyect.ipynb.

Packages Needed

For downloading and testing out this proyect, you'll have to make sure you've installed the following python packages:

• Pytorch (at least 1.6.0)
• argparse
• json
• PIL
• matplotlib
• seaborn

Getting Started

• Open the train.py file in the terminal using the following sintax: (Just 1 line of code) python train.py --save_dir (path to the checkpoints.pth) --arch (choose the CNN model you want to use between "resnet) --learning_rate (learning rate value) --hidden_units (hidden units number) --epochs (number of epochs) --gpu (if --gpu is written, enables gpu usage)

• Open the predict.py file in the terminal using the following sintax: (Just 1 line of code) python predict.py input_image_pathfile checkpoint.pth_filepath --top_k (int of how many probability results are shown) --category_names (json filepath to convert from categories index to labels/names) --gpu (if --gpu is written, enables gpu usage)

• When openning the Image_Classifier_Proyect.ipynb file, run every cell one-by-one.

Results

• For Image_Classifier_Proyect.ipynb you should get this kind of results:

  • Training using resnet50:

    • Epochs: 5

    • LearningRate: 0.003

    • Hidden-units: 1024

    • Dropout: 0.2

    • Input units: 2048

    • Output units: 102

    • Validation accuracy result: 83.8%

    • Validation loss result: 0.574

    • Trainning loss result: 1.157

    • Testing accuracy result: 84.1%

    • Testing loss result: 0.586

  • Training using vgg16:

    • Epochs: 5

    • LearningRate: 0.003

    • Hidden-units: 1024

    • Dropout: 0.2

    • Input units: 25088

    • Output units: 102

    • Validation accuracy result: 80.6%

    • Validation loss result: 0.784

    • Trainning loss result: 1.786

    • Testing accuracy result: 74.4%

    • Testing loss result: 0.964

  • Training using alexnet:

    • Epochs: 5

    • LearningRate: 0.003

    • Hidden-units: 1024

    • Dropout: 0.2

    • Input units: 9216

    • Output units: 102

    • Validation accuracy result: 82.4%

    • Validation loss result: 0.653

    • Trainning loss result: 1.365

    • Testing accuracy result: 82.1%

    • Testing loss result: 0.703

Architectures Modifications and Criterion Explanation:

• Architecture modification to the downloaded model (resnet50, alexnet, vgg16):

• Criterion usage: For this model I selected the nn.NLLLoss() (Negative Log Likelihood Loss) criterion due to the fact that it is one of the best functions to calculate the error of a model with a lot of classes to classify. This criterion is the reason why the last layer of the network is a LogSoftmax() function.