is it a dog???

For this project, I trained a Deep Neural Network for item/animal detection which is in the make-model.ipynb file.

The aim for the model is to determine the probability that there is a dog label: 0 means a dog is not there, 1 means there is certainly a dog in there.

What I have done:

1. Exploratory Data Analysis. I have showed the following:

   • The shape of the datasets
   • Rendered 5 images from each label to show that I know how to assemble an image from the dataset.
   • Chosen the dog label for probabability that the item/animal is there.

2. Data Processing. I have done the following:

   • Obtained training and testing dataset.
   • Ensured that there is no data imbalance between the images. In other words, there are 10000 data entries for training and 2000 for testing.

3. Data Modelling.

   • Created my Model.
     • Added hidden layers with their node counts and activation types.
     • Chosen a optimizer, loss and the metric for the output.
   • Fit my training and validation (testing in this case) dataset to the model.

4. Written a report about my model. This is in the report.ipynb file. The report includes the following:

   • Introduction.
     • Which label I picked for item/animal detection.
     • Why I picked the label.
   • Modelling Process.
     • What the shape of my model is.
     • The reason why I pick the selected loss and optimizer.
   • Model Performance.
     • Shown the results of my training through the use of graphs. The graphs can be obtained from tensorboard. These include:
       • Accuracy
       • Loss
       • Validation Accuracy
       • Validation Loss
     • Any comments about the results.
   • Conclusion.
     • Summary of the report without adding too much information.

5. Applied hyperparameter tuning to my model so I could find out the best possible model compilation parameters for accuracy.

6. Made a python application/function that can evaluate my model when given any PNG and/or JPEG image. This application is able to resize images to 32x32 before having the model evaluate said image.

Required Software:

• Python 3.9+
• Text Editor for Python (Visual Studio Code with Python extension installed is recommended)

Recommended:

• Anaconda
• Virtualenv

Setting Up:

Download the CIFAR-10 dataset from here and place in your project directory.

Install the libraries listed in the requirements.txt via the command:

pip install -r requirements.txt

Recommended to use a virtual environment either from either virtualenv or anaconda. This setup helps isolate project dependencies from user dependencies as to avoid version conflicts.

Tensorboard:

Run

tensorboard --logdir output/logs/

to run tensorboard. The main points of interests are accuracy and loss graphs. The tensorboard application can then be opened via the site http://localhost:6006/.