This project is a TensorFlow implementation of Neural Program Synthesis from Diverse Demonstration Videos, which is published in ICML 2018. We provide codes and checkpoints for our model and all baselines presented in the paper. Also, we provide scripts and codes for generating datasets as well as the datasets we used to train and test all models.
As interpreting decision making logic in demonstration videos is key to collaborating with and mimicking humans, our goal is to empower machines with this ability. To this end, we propose a neural program synthesizer that is able to explicitly synthesize underlying programs from behaviorally diverse and visually complicated demonstration videos, as illustrated in the following figure.
We introduce a summarizer module as part of our model to improve the network’s ability to integrate multiple demonstrations varying in behavior. We also employ a multi-task objective to encourage the model to learn meaningful intermediate representations for end-to-end training. Our proposed model consists three components:
- Demonstration Encoder receives a demonstration video as input and produces an embedding that captures an agent’s actions and perception.
- Summarizer Module discovers and summarizes where actions diverge between demonstrations and upon which branching conditions subsequent actions are taken.
- Program Decoder represents the summarized understanding of demonstrations as a code sequence.
The illustration of the overall architecture is as follows. For more details, please refer to the paper.
Our method is evaluated on a fully observable, third-person environment (Karel environment) and a partially observable, egocentric game (ViZDoom environment). We show that our model is able to reliably synthesize underlying programs as well as capture diverse behaviors exhibited in demonstrations.
*This code is still being developed and subject to change.
The structure of the repository:
- ./: training and evaluation scripts
- ./models: network models used for the experiments
- ./karel_env: karel environment including dsl, interpreter and dataset generator / loader
- ./vizdoom_env: vizdoom environment including dsl, interpreter and dataset generator / loader
- Python 2.7
- Tensorflow 1.3.0
- SciPy
- NumPy
- colorlog
- cv2
- h5py
- Pillow
- progressbar
- ply
- ViZDoom Deterministic
- You can find the codes for the Karel environments in this directory
- To generate a dataset for Karel environments including programs and demonstrations, use the following script.
./karel_env/generate_dataset.shDefault arguments are identical to the settings described in the paper.
- To reproduce experiments in our paper, you need to install our deterministic ViZDoom envrionment
- You can find the codes for the ViZDoom environments and detailed instructions in this directory
- To generate a dataset (vizdoom_dataset, vizdoom_dataset_ifelse) for the ViZDoom environment including programs and demonstrations, use the following script.
./vizdoom_env/generate_dataset.sh- Train the full model (with the summarizer module and the multi-task objective)
python trainer.py --model full --dataset_path /path/to/the/dataset/ --dataset_type [karel/vizdoom]
- Train the summarizer model (with the summarizer module but without multi-task objective)
python trainer.py --model summarizer --dataset_path /path/to/the/dataset/ --dataset_type [karel/vizdoom]
- Train the baseline program synthesis model (without the summarizer module and multi-task objective)
python trainer.py --model synthesis_baseline --dataset_path /path/to/the/dataset/ --dataset_type [karel/vizdoom]
- Train the baseline program induction model
python trainer.py --model induction_baseline --dataset_path /path/to/the/dataset/ --dataset_type [karel/vizdoom]
- Arguments
- --debug: set to
Trueto see debugging visualization (LSTM masks, etc.) - --prefix: a nickname for the training
- --model: specify which type of models to train/test
- --dataset_type: choose between
karelandvizdoom. You can also add your own datasets. - --dataset_path: specify the path to the dataset where you can find a HDF5 file and a .txt file
- --checkpoint: specify the path to a pre-trained checkpoint
- Logging
- --log_setp: the frequency of outputing log info ([train step 681] Loss: 0.51319 (1.896 sec/batch, 16.878 instances/sec))
- --write_summary_step: the frequency of writing TensorBoard summaries (default 100)
- --test_sample_step: the frequency of performing testing inference during training (default 100)
- Hyperparameters
- --num_k: the number of seen demonstrations (default 10)
- --batch_size: the mini-batch size (default 32)
- --learning_rate: the learning rate (default 1e-3)
- --lr_weight_decay: set to
Trueto perform expotential weight decay on the learning rate - --scheduled_sampling: set to
Trueto train models with scheduled sampling
- Architecture
- --encoder_rnn_type: the recurrent model of the demonstration encoder. Choices include RNN, GRU, and LSTM
- --num_lstm_cell_units: the size of RNN/GRU/LSTM hidden layers (default 512)
- --demo_aggregation: how to aggregate the demo features (default average pooling) for synthesis and induction baseline
- --debug: set to
- Evaluate trained models
python evaler.py --model [full/synthesis_baseline/summarizer/induction_baseline] --dataset_path /path/to/the/dataset/ --dataset_type [karel/vizdoom] [--train_dir /path/to/the/training/dir/ OR --checkpoint /path/to/the/trained/model]Because the size of the demonstrations in VizDoom is usually very large, it is difficult to use large batch size and results in very slow training. To circumvent this issue, we used two stage training; we pretrain our model with short demonstrations first and then finetune the model with the whole dataset. For the first stage of the training, we used batch size: 32, and for the second stage of the training, we used batch size: 8. Here are links to datasets that we used for the first and second stage of the training:
- First stage: vizdoom_shorter
- Second stage: vizdoom_full
To reproduce our result, you can train and evaluate models with the following command:
- First stage: training for 50000 iterations
python trainer.py --model full --dataset_path path_to_vizdoom_shorter --dataset_type vizdoom --num_k 25 --batch_size 32- Second stage: training for 50000 iterations
python trainer.py --model full --dataset_path path_to_vizdoom_full --dataset_type vizdoom --num_k 25 --batch_size 8 --checkpoint path_to_1st_step_checkpointFor evaluation, use the following command:
python evaler.py --model full --dataset_path path_to_vizdoom_full --dataset_type vizdoom --num_k 25 --checkpoint path_to_2nd_step_checkpoint| Methods | Execution | Program | Sequence |
|---|---|---|---|
| Induction baseline | 62.8% | - | - |
| Synthesis baseline | 64.1% | 42.4% | 35.7% |
| + summarizer (ours) | 68.6% | 45.3% | 38.3% |
| + multi-task loss (ours-full) | 72.1% | 48.9% | 41.0% |
- Effect of the summarizer module
To verify the effectiveness of our proposed summarizer module, we conduct experiments where models are trained on varying numbers of demonstrations (k) and compare the execution accuracy.
| Methods | k=3 | k=5 | k=10 |
|---|---|---|---|
| Synthesis baseline | 58.5% | 60.1% | 64.1% |
| + summarizer (ours) | 60.6% | 63.1% | 68.6% |
| Methods | Execution | Program | Sequence |
|---|---|---|---|
| Induction baseline | 35.1% | - | - |
| Synthesis baseline | 48.2% | 39.9% | 33.1% |
| Ours-full | 78.4% | 62.5% | 53.2% |
- If-else experiment:
To verify the importance of inferring underlying conditions, we perform evaluation only with programs containing a single if-else statement with two branching consequences. This setting is sufficiently simple to isolate other diverse factors that might affect the evaluation result.
| Methods | Execution | Program | Sequence |
|---|---|---|---|
| Induction baseline | 26.5% | - | - |
| Synthesis baseline | 59.9% | 44.4% | 36.1% |
| Ours-full | 89.4% | 69.1% | 58.8% |
- Generalization over different number of seen demonstrations
The baseline models and our model trained with 25 seen demonstration are evaluated with fewer or more seen demonstrations.
To reproduce our results, you can download our datasets and checkpoints.
While we provide the scripts and codes for generating customized datasets, we also made the datasets we used to train and test our models and baselines available.
- Karel (13GB) [link]
- For the main Karel experiment (Table 1) and the summarizer experiment (Table 2)
- ViZDoom full (505GB) [link]
- For the main ViZDoom experiment (Table 3) and the ViZDoom generalization experiment (Figure 7)
- ViZDoom shorter (248GB) [link]
- A subset of the Vizdoom full datsaet where only the shorter demos are included
- ViZDoom if-else (40GB) [link]
- For the ViZDoom if-else experiment (Table 4)
We provide checkpoints and evaluation report files of our models and baselines for all experiments.
- Karel main experiment (Table 1) [link]
- Karel summarizer experiment (Table 2) [link]
- ViZDoom main experiment (Table 3 and Figure 7) [link]
- ViZDoom if-else experiment (Table 4) [link]
- Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis in ICLR 2018
- RobustFill: Neural Program Learning under Noisy I/O in ICML 2017
- DeepCoder: Learning to Write Programs in ICLR 2017
- Neuro-Symbolic Program Synthesis in ICLR 2017
- Neural Programmer-Interpreters in ICLR 2016
- Neural Task Programming: Learning to Generalize Across Hierarchical Tasks in ICRA 2018
- Neural Program Meta-Induction in NIPS 2017
- Neural GPUs Learn Algorithms in ICLR 2016
- One-Shot Imitation Learning in NIPS 2017
- One-Shot Visual Imitation Learning via Meta-Learning in CoRL 2017
- Awesome Neural Programming
If you find this useful, please cite
@inproceedings{sun2018neural,
title = {Neural Program Synthesis from Diverse Demonstration Videos},
author = {Sun, Shao-Hua and Noh, Hyeonwoo and Somasundaram, Sriram and Lim, Joseph},
booktitle = {Proceedings of the 35th International Conference on Machine Learning},
year = {2018},
}
Shao-Hua Sun*, Hyeonwoo Noh*, Sriram Somasundaram, and Joseph J. Lim
(*Equal contribution)