This is the official repository for the TMLR paper Self-Improvement for Neural Combinatorial Optimization: Sample Without Replacement, but Improvement, using PyTorch. Parallelization is done with ray.io.
To provide a low barrier to get started, we assembled a reproduction script that loads the model checkpoints for TSP, CVRP, and JSSP and reproduces the greedy results reported in the paper. Run with
$ python reproduce.py
The reproduction script is set up to perform evaluation with four parallel workers that do model inference on the CPU. To align this with your hardware, change the variables CUDA_VISIBLE_DEVICES and devices_for_eval_workers and beams_with_batch_sizes at the beginning of the script. For example, to perform evaluation with 16 workers on two GPUs and a batch size of 64 for greedy evaluation, we set:
CUDA_VISIBLE_DEVICES = "0,1" # Must be set for ray, as it can have difficulties detecting multiple GPUs
devices_for_eval_workers = ["cuda:0"] * 8 + ["cuda:1"] * 8 # i-th entry is the device that the i-th worker should use for evaluation
beams_with_batch_sizes = {
1: 64 # Beam width 1 (= greedy) with a batch size of 64 per worker.
}
When using our method, we would advise to start with the sampling method "wor" (see section Gumbeldore config below). Here, solutions to a problem instance are sampled in multiple rounds without replacement, but without the advantage-based trie update derived in the paper. In the paper we show that this method alone gives very good results. So you can see how far you can get with this method alone, and then switch to "gumbeldore" and start tuning the step size that controls the advantage-based update. So, using sampling without replacement as a baseline is always a good idea.
The directory core contains the, well, core of our method. In the file core/train.py is the method main_train_cycle, which is used both for supervised and Gumbeldore training. It sets up the whole data loading/checkpoint saving/iterate over epochs cycle, and the files jssp_main.py, tsp_main.py, cvrp_main.py all pass their problem-specific methods to main_train_cycle.
When training with Gumbeldore, core.GumbeldoreDataset is the general class which generates random instances, and distributes them to multiple parallel workers. The workers sample with Gumbeldore and maintain the trie via core.IncrementalSBS. In particular, the files core/incremental_sbs.py and core/stochastic_beam_search.py contain everything regarding the sampling mechanism.
Once sampled, trajectories are saved as a pseudo-expert solution to disk in a pickle-file. All dataset loading and preparation is performed in <tsp,cvrp,jssp>/dataset.py. For information about the structure of the pickle files, see the README under ./data.
For each problem class tsp, cvrp and jssp, we specify all options in config.py. The different options should be well-commented, let us know if there are any problems to get something running!
While the implementation performance can certainly be improved, we want to put focus on the following options which are crucial to get a good performance from your hardware:
CUDA_VISIBLE_DEVICES: ray.io sometimes needs help to recognize multiple GPUs. This variable will be set as the env var of the same name in each worker.num_dataloader_workers: Number of parallel dataloader workers to prepare batches.training_device: Device on which network training is performed.devices_for_eval_workers: List of devices on which evaluation is performed. Here, the i-th worker uses the i-th device.
In each config object, there's the dictionary gumbeldore_config that specifies how Gumbeldore sampling is performed. As there are a lot of options to finetune the sampling, we go through the options here in detail:
active_search[str]: Optional string, defaults toNone. If given, the instances for which solutions are sampled are not created randomly, but loaded from the given path. Can be used for a kind of active search, where the model is finetuned to the given instances at hand. This had no relevance in the paper.use_best_model_for_generation[bool]: If this is True - which is the way the paper uses it - the current best greedy model is used for trajectory generation. If False, simply the current model in each epoch is used.append_if_not_new_best[bool]: If True (as in paper), we don't empty the dataset after an epoch if the model has not improved. For debugging purposes on new problem classes, it can be helpful to turn this off until a good balance of number of instances/solution quality/training time was found, so that the model does not overfit to the generated data.devices_for_workers[List[int]]: List of devices where the i-th data generation worker uses the i-th device. Length of the list specifies number of workers then.num_instances_to_generate[int]: Number of instances to generate in each epoch. For JSSP, see the corresponding config, as it's structured a bit differently.destination_path[str]: Path to save the generated pickle with the instances and best found solution.batch_size_per_worker[int]: Batch size for each worker on a GPU. For JSSP, as above, see comments in config file.batch_size_per_cpu_worker[int]: Batch size for each worker on a CPU (usually smaller than for GPU, e.g., 1).search_type[str]: Sampling and policy update type. Must be "wor" (sample without replacement using round-wise SBS), "gumbeldore" (our method), or "theory_gumbeldore" (locally estimated advantage with theoretical policy improvement).beam_width[int]: Beam width k for Stochastic Beam Search in each round.num_rounds[int]: Number of sampling rounds.pin_workers_to_core[bool]: Defaults is False. If True, workers are pinned to single CPU threads, which can help with many workers on the CPU to prevent them from jamming each other with their numpy and pytorch operations.advantage_constant[float]: Policy update step size 'sigma' in the paper. For policy_improvement, we update the logits by adding the normalized advantage timesadvantage_constant. This should be tuned manually specifically to the problem at hand.min_max_normalize_advantage[bool]: If True, we perform a min max normalization on the advantages so the highest/lowest advantage is 1/-1. Default is False.expected_value_use_simple_mean[bool]: If True, we do not compute the expected value using Gumbels and importance sampling, but simply calculate the arithmetic mean of the leaf outcomes. Defaults to False.use_pure_outcomes[bool]: If True, we do not subtract the expected value from the actual outcomes, but simply return the outcome uncentered. This can be desirable in, e.g., a board game situation, where we want to value wins/losses regardless of the expected value of the trajectories. In particular, we only use it for Gomoku.normalize_advantage_by_visit_count[bool]: If True, we divide the accumulated advantage of a node by its visit count. This shifts the search method further towards a Gumbel AlphaZero-type update. Can be beneficial in conjunction withuse_pure_outcomesset to True. Can be used for Gomoku. For the paper, we use it nowhere.perform_first_round_deterministic[bool]: If True, the first Gumbeldore round is a simple deterministic beam search round, to get the exploitation of the policy "out of the system". The probability mass of the beam search trajectories is subtracted from the tree, but no advantage is propagated. In the paper's Appendix, we mention that a combination of Gumbeldore and deterministic beam search can be a powerful inference technique. This can be done with this option for example.min_nucleus_top_p[float]: p_min in the paper. If smaller than 1, couples SBS with Top-p (nucleus) sampling, where we increase p linearly to 1 starting frommin_nucleus_top_pover the rounds.
Trained model checkpoints mentioned in the paper can be found in the directory ./model_checkpoints.
Due to size constraints, only validation and test data can be found under ./data. For supervised training for CVRP and TSP, find the datasets with one million training instances here.
See README in directory ./data.
UPDATE: This behavior should be fixed now, i.e., errors in workers are shown while the progress bar still works. Please open an issue if something doesn't work as you think it should.
When generating data in an epoch with core.GumbeldoreDataset, we use tqdm to track the progress with a visual progress bar while collecting the data from parallel workers. However, as ray.get is called on the workers after all data has been collected, the program will stall without throwing an error if an error occurs in one of the workers. If you think you are experiencing that, place a ray.get after the future tasks have been created and before tqdm in core/gumbeldore_dataset.py as follows:
# Kick off workers
future_tasks = [
async_sbs_worker.remote(
self.config, job_pool, network_weights, device,
batch_size_gpu if device != "cpu" else batch_size_cpu,
self.beam_leaves_to_result_fn,
self.get_network_fn, self.trajectory_cls,
cpu_cores[i], memory_aggressive
)
for i, device in enumerate(self.devices_for_workers)
]
+ ray.get(future_tasks) # PLACE THIS LINE HERE
with tqdm(total=len(problem_instances)) as progress_bar:
while None in results:
This will turn off live updates in the progress bar, but errors are thrown correctly.
Thanks to the following repositories: