mallport / tdt4265-project

SSD in PyTorch with ResNet-34 backbone

Hyperparameters

All the parameters we used can be found in train_waymo.yaml and train_tdt4265.yaml

How to reproduce our results

Set up the Python/Anaconda environment as done in assignment 4, and install the requirements:

pip install -r requirements.txt

Run setup_waymo.py and update_tdt4265_dataset.py.

Train on the Waymo dataset:

python train.py configs/train_waymo.yaml

Transfer the model:

python transfer_learn.py configs/train_waymo.yaml

Train on the TDT4265 dataset:

python train.py configs/train_tdt4265.yaml

Finally run the evaluation script.

Model architecture

For the backbone we use a ResNet-34 (pretrained on ImageNet), with some modifications, for the first three output feature banks. For the last three feature banks we used additional layers of ResNet BasicBlocks, each of size 2 and with the first block downsampling.

Layer	Output
Conv2d(3, 64) kernel 7x7, stride 2x2, padding 3x3
BatchNorm2d(64)
ReLU
ResNet-34 layer 1-2 (extra downsampling BasicBlock at the end)	Feature bank 1 (128 x 38 x 38)
ResNet-34 layer 3	Feature bank 2 (256 x 19 x 19)
ResNet-34 layer 4	Feature bank 3 (512 x 10 x 10)
Additional layer of two BasicBlocks (first downsampling)	Feature bank 4 (256 x 5 x 5)
Additional layer of two BasicBlocks (first downsampling)	Feature bank 5 (256 x 3 x 3)
Additional layer of two BasicBlocks (first downsampling)
Conv2d(128, 128) kernel 3x3, stride 2x2, padding 1x1
BatchNorm2d(128)
ReLU	Feature bank 6 (128 x 1 x 1)