Kaixin WANG

Batch size affects the behaviour of Masked Average Pooling

7

In [onemodel_sg-one.py](https://github.com/xiaomengyc/SG-One/blob/master/oneshot/onemodel_sg-one.py#L37) line 37, the behaviour of `torch.sum(pos_mask)` will be different if the batch size is not 1. It is better to write it batchsize-agnostic.

Bias in EvalCollector

4

`EvalCollector` (e.g. `GpuEvalCollector`) seems to bias towards short trajectories, leading to inaccurate evaluation. The problem occurs when `eval_n_envs > 1`. In [GpuEvalCollector.collect_evaluation()](https://github.com/astooke/rlpyt/blob/master/rlpyt/samplers/parallel/gpu/collectors.py#L125), the list of `TrajInfoCls` is created as a...

train_locally.py error

6

When running `train_locally.py`, the script print "{'state': 'ERROR', 'score': {}, 'instances': [], 'reason': 'You started more instances (2) then allowed limit (1).'}". Is this an intended behavior?

parser inside train.py takes no effect

4

it seems that changes to the parser object in `train.py` have no effect. The script always use parser in the `utility/parser.py` ( called by `python3 utility/parser.py || true` in the...

When will you release the competition baseline?

2

There is only a README in the directory.

Hi @bmazoure , thank you for this great repo! Just curious about how JAX compares to TensorFlow 1.x (the official OpenAI's train-procgen repo) and PyTorch, in terms of training speed...