Code for Multimodal Neural SLAM for Interactive Instruction Following

Code for Multimodal Neural SLAM for Interactive Instruction Following

Code structure

The code is adapted from E.T. and most training as well as data processing files are in currently in the ET/notebooks folder and the et_train folder.

Dependency

Inherited from the E.T. repo, the package is depending on:

• numpy
• pandas
• opencv-python
• tqdm
• vocab
• revtok
• numpy
• Pillow
• sacred
• etaprogress
• scikit-video
• lmdb
• gtimer
• filelock
• networkx
• termcolor
• torch==1.7.1
• torchvision==0.8.2
• tensorboardX==1.8
• ai2thor==2.1.0
• E.T. (https://github.com/alexpashevich/E.T.)

MaskRCNN Fine-tuning

To fine-tune the MaskRCNN module used in solving the Alfred challenge, we provide the code adapted from the official PyTorch tutorial.

Setup

We assume the environment and the code structure as in the E.T. model is set up, with this repo served as an extension. Although the fine-tuning code should be a standalone unit.

Training Data Geneation

Given a traj_data.json file (e.g., the 45K one used in E.T. joint-training here), run python -m alfred.gen.render_trajs as in E.T. to render the training inputs (raw images) and the ground truth labels (instance segmentation masks) for all the frames recorded in the traj_data.json files. Make sure the flag for generating instance level segmentation masks is set to True.

Pre-processing Instance Segmentation Masks

The rendered instance segmentation masks need to be preprocessed so that the data format is aligned with the one used in the official PyTorch tutorial. In specific, each generated mask is of a different RGB color per instance, which is mapped to the unique instance index in the frame as well as a label index for its semantic class. The mapping is constructed by looking up the traj['scene']['color_to_object_type'] in each of the json dictionaries. The code also supports the functionality to only collect training data from certain subgoal data (such as for PickupObject in Alfred). Notice that there are some bugs in the rendering process of the masks which creates some artifacts (small regions in the ground truth labels that correspond to no actual objects). This can be fixed by only selecting instance masks that are larger than certain area (e.g., > 10 as in alfred/data/maskrcnn.py).

Training

Run python -m alfred.maskrcnn.train which first loads the pre-trained model provided by E.T. and then fine-tunes it on the pre-processed data mentioned above.

Evaluation

We follow the MSCOCO evaluation protocal which is widely used for object detection and instance segmentation, which output average precision and recall at multiple scales. The evaluation function call evaluate(model, data_loader_test, device=device) in alfred/maskrcnn/train.py serves as an example.