This project is my course project for "An introduction to Artificial Intelligence"

The code relies on the deep learning framework of Pytorch

• Anaconda
• numpy
• torch
• torchvision
• tqdm
• matplotlib
• Pillow
• bidict
• torchcam

Please download the dataset from here and unzip it. You are recommand to rename the dataset directory into marine-creature-dataset.

You can download the pretrained model from here.

You are recommanded to put the pretrained model under the ./pretrained_models/ directory.

Run the following command to train on ConvNet:

python train.py --epochs 100 --image_dir /path/to/your/dataset

Or you can train on ResNet101 or DenseNet121 with the folowing command:

python train.py --model resnet101 --epochs 20 --image_dir /path/to/your/dataset

python train.py --model densenet --epochs 20 --image_dir /path/to/your/dataset

During training, you can supervise the process with tensorboard. The best model on validation set will be recorded under the ./logs directory.

Please refer to the options.py for detailed parameters of training.

Run the following command to inference on trained model:

python inference.py --pretrained_model /path/to/your/model --image_dir /path/to/your/dataset

For example:

python inference.py --pretrained_model ./pretrained_models/densenet.pth --image_dir ./marine-creature-dataset

You can specifify network layer for creating the heatmap by adding the --layer options like:

python inference.py --pretrained_model ./pretrained_models/densenet.pth --image_dir ./marine-creature-dataset --layer net.features

After inference, you can see the classfication precision and recall value on terminal. The visualization output will be recorded under the ./results directory.

Also, please refer to the options.py for detailed parameters of inferencing.