[BUG] Code for real-world cvrplib instance inference.

[Elon-Lau]

Hello, I can't solve the instances in cvrplib with the trained model.

Traceback (most recent call last): File "/NCO/RL4CO/examples/test_cvrplib_demo.py", line 91, in out = policy(td.clone(), decode_type='greedy', return_actions=True) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl return self._call_impl(*args, *kwargs) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl return forward_call(args, **kwargs) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/rl4co/models/common/constructive/base.py", line 237, in forward td = env.step(td)["next"] File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/rl4co/envs/common/base.py", line 129, in step td = self._step(td) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/rl4co/envs/routing/cvrp/env.py", line 89, in _step visited = td["visited"].scatter(-1, current_node, 1) RuntimeError: Index tensor must have the same number of dimensions as self tensor

[fedebotu]

Have you checked out this notebook?

https://github.com/ai4co/rl4co/blob/main/examples/datasets/2-test-on-cvrplib.ipynb

There was some small problem with squeezing dimensions for evaluating with batch size of 1, so we made some monkey patching (that should be solved in 0.5.1)

Also, here is an example of a recent project based on RL4CO in which we evaluate CVRPLIB (with some extra utilities):

https://github.com/ai4co/routefinder/blob/main/examples/2.eval-cvrplib.ipynb

[Elon-Lau]

Follow the https://github.com/ai4co/rl4co/blob/main/examples/datasets/2-test-on-cvrplib.ipynb, i set batch_size=2. However, I still have this problem. If the current version can be used for CVRPlib instance inference？

[Elon-Lau]

I use the same code in https://github.com/ai4co/rl4co/blob/main/examples/2d-meta_train.py for training and the same code in https://github.com/ai4co/rl4co/blob/main/examples/datasets/2-test-on-cvrplib.ipynb for cvrplib instance inference. But the following bug still occur.

Traceback (most recent call last): File "/NCO/RL4CO/examples/test_cvrplib_demo.py", line 91, in out = policy(td.clone(), decode_type='greedy', return_actions=True) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl return self._call_impl(*args, *kwargs) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl return forward_call(args, **kwargs) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/rl4co/models/common/constructive/base.py", line 237, in forward td = env.step(td)["next"] File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/rl4co/envs/common/base.py", line 129, in step td = self._step(td) File "/miniconda3/envs/rl4co/lib/python3.9/site-packages/rl4co/envs/routing/cvrp/env.py", line 89, in _step visited = td["visited"].scatter(-1, current_node, 1) RuntimeError: Index tensor must have the same number of dimensions as self tensor

[fedebotu]

Thanks for reporting the bug - you may check the most recent improvements! Please let us know if now it works well on your end ~

[Elon-Lau]

Thanks for the update, it looks like the patch will fix the current bug.