nesl / FlexLoc

FlexLoc

Create Environment

1. Clone the repository to any directory
2. Create a conda environment with the following command conda create --env flexloc python=3.10
3. In the CondConvResNet directory, there is a requirements.txt with the necessary packages

cd CondConvResNet
pip3 install -r requirements.txt

4. Separately install pytorch

conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.7 -c pytorch -c nvidia

Run Small Scale Test

We include 30 files from our test set to run a small scale test verifying that all the libraries are correctly installed.

In either conditional convolution (CondConvResNet) or conditional layer normalization (Final_CLN_Resnet) folders, run

cd CondConvResNet OR cd Final_CLN_Resnet
python3 batch_test.py --folder 1 --checkpoint best_val.pt

to run a small scale test utilizing our provided checkpoints.

In the logs folder, under the folder 1, it generates two .txt files. predictions.txt contains the predicted coordinates vs. the ground truth coordinates, while test_loss.txt contains the evaluation metrics. We utilize the Average Distance metric for our evaluations.

Run Large Scale Test

1. Download data from this Google Drive Link.
2. After unzipping the data, there will be 5 different test folders (test29, test70, test73, test74, test77) each containing 500 cached pickle files representing ~30 seconds of data
3. Place these folders into the top level respository directory
4. Rename a given viewpoint's folder name to test. For example, if we want to evaluate the model on viewpoint 29, rename test29 to test.
5. Navigate to the appropriate CLN or CondConv directory, and run python3 batch_test.py --folder 1 --checkpoint best_val.pt. The results will be under test_loss.txt in logs/1
6. Revert back to the original name of the test data, e.g., rename test back to test29.
7. You can test on other viewpoints by repeating steps 4-6. Note that successive calls to the batch.test.py script will append to the test_loss.txt file instead of overwriting previous data.

Train the Model

1. Refer to the GTDM work to download the entire dataset and run necessary preprocessing
2. Once we have the files in hdf5 form separated into appropriate train, test, and val directories, we must cache these into .pickle files for training. Provide the root directory to these hdf5 files by modifying the base_root variable in line 2 of the configs/data_configs.py, and adjusting the subsequent data_root variables. Adjust the cache_dir as you see fit to decide where these files will be cached
3. Running python3 train.py 100 to begin training with a seed of 100. This will check the specified cache directory for train, test, and val folders containing the pickle files. If they are not found, it will search the specified data_root directories for appropriate hdf5 files to begin the process of caching to .pickle. Training will commence after the one-time caching process.
4. The training logs and checkpoints will be saved in logs under a folder with the current timestamp