code for Selected Topics in Visual Recognition using Deep Learning Homework 2

• Ubuntu 16.04 LTS
• NVIDIA 1080ti

1. Clone this repo
2. Install the required packages:

apt-get install tk-dev python-tk

3. Install the python packages:

pip install pandas
pip install pycocotools
pip install opencv-python
pip install requests

4. Train model by running Train.py:
5. Output predicted result by running Test.py and upload to Google drive submission

The network can be trained using the train.py script. We use HW2 as our dataset.

python train.py --dataset HW2 --HW2_path ../train --depth 50

In Train.py, we preprocess the data and train our model.

A pre-trained model is available at:

• https://drive.google.com/open?id=1yLmjq3JtXi841yXWBxst0coAgR26MNBS The state dict model can be loaded using:

retinanet = model.resnet50(num_classes=dataset_train.num_classes(),)
retinanet.load_state_dict(torch.load(PATH_TO_WEIGHTS))

Then we do image preprocessing in the following order.

• Normalize
• Augmentation
• Resize And remember to resize the annotation when you resize the corresponding image.
  After preprocessing, we use dataloader with batch_size=10 to feed training data into our model.

python train.py --dataset HW2 --HW2_path ../train --depth 50

We use RetinaNet as backbone network, and modify the class numbers to 10.
10 is the total class numbers of our training_data (0~9).

We use focalLoss as loss function and Adam optimizer, and save the model every single epochs in total 100 epochs.
Set learning rate = 1e-4.
The model parameters would be saved in folder saved_models_3/.

python test.py

For testing data, we first use simliar dataloader to load testing images.
Then we import the saved model.

retinanet = torch.load('saved_models_3/HW2_retinanet_xx.pt')
retinanet = retinanet.cuda()
retinanet.eval()

For each testing image, we normalize and resize it.
In the end, write image id and predicted result to a .josn file.
Upload google drive submission.
Best accuracy : mAP : 0.46387.