How to use AR?

[GuoPingPan]

If I want to use Anticipation Reward(AR), should I set reward_type = map_accuracy ?
But I found it wasn't set in
https://github.com/facebookresearch/OccupancyAnticipation/blob/aea6a2c0d9/configs/exploration/gibson_train_w_ar.yaml

[srama2512]

@GuoPingPan - It is set in the baseline config, not the task config.

OccupancyAnticipation/configs/model_configs/occant_rgbd/ppo_exploration.yaml

Lines 24 to 25 in aea6a2c

	# Uncomment this for anticipation reward
	# reward_type: "map_accuracy"

[GuoPingPan]

Thanks you so much! But I got another question which have the multi-gpu trainning failed. Have you ever meet the error:

Process ForkServerProcess-74:
Traceback (most recent call last):
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/multiprocessing/process.py", line 315, in _bootstrap
    self.run()
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/supervised/map_update.py", line 325, in map_update_worker
    losses = map_update_fn(ps_args)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/supervised/map_update.py", line 219, in map_update_fn
    mapper_outputs = mapper(mapper_inputs, method_name="predict_deltas")
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py", line 168, in forward
    outputs = self.parallel_apply(replicas, inputs, kwargs)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py", line 178, in parallel_apply
    return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/parallel/parallel_apply.py", line 86, in parallel_apply
    output.reraise()
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/_utils.py", line 457, in reraise
    raise exception
RuntimeError: Caught RuntimeError in replica 0 on device 1.
Original Traceback (most recent call last):
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/parallel/parallel_apply.py", line 61, in _worker
    output = module(*input, **kwargs)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/rl/policy.py", line 633, in forward
    outputs = self.predict_deltas(*args, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/rl/policy.py", line 316, in predict_deltas
    pu_outputs = self.projection_unit(pu_inputs)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/rl/policy_utils.py", line 49, in forward
    x_full = self.main(x)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/models/occant.py", line 453, in forward
    return self.main(x)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/models/occant.py", line 51, in forward
    gp_outputs = self._do_gp_anticipation(x)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/models/occant.py", line 278, in _do_gp_anticipation
    x_enc = self.gp_depth_proj_encoder(
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/models/unet.py", line 107, in forward
    x1 = self.inc(x)  # (bs, nsf, ..., ...)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/models/unet.py", line 41, in forward
    x = self.conv(x)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/workspace/OccAnt/occant_baselines/models/unet.py", line 31, in forward
    x = self.conv(x)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/container.py", line 141, in forward
    input = module(input)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 447, in forward
    return self._conv_forward(input, self.weight, self.bias)
  File "/home/yzc1/miniconda3/envs/occ/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 443, in _conv_forward
    return F.conv2d(input, weight, bias, self.stride,
RuntimeError: cuDNN error: CUDNN_STATUS_INTERNAL_ERROR

And I only can use one gpu and ensure that they are the same between mapper_copy and MapLargeRolloutStorageMP, as bellow

if mapper_cfg.use_data_parallel and len(mapper_cfg.gpu_ids) > 0:
    self.mapper_copy.to(self.mapper.config.gpu_ids[0]) # cuda:1
    self.mapper_copy = nn.DataParallel(
        self.mapper_copy,
        device_ids=self.mapper.config.gpu_ids, # device = 1, 2, 3, 4, 5, 6, 7
        output_device=self.mapper.config.gpu_ids[0],
        # device_ids=[self.mapper.config.gpu_ids[0]]
    )

...

if ans_cfg.MAPPER.use_data_parallel and len(ans_cfg.MAPPER.gpu_ids) > 0:
    mapper_device = ans_cfg.MAPPER.gpu_ids[0]
    # mapper_device = torch.device("cuda:1")
mapper_rollouts = MapLargeRolloutStorageMP(
    ans_cfg.MAPPER.replay_size,
    mapper_observation_space,
    mapper_device,
    mapper_manager,
)

But I found that the MapLargeRolloutStorageMP, which will easily cause OOM Error.
The most importance thing is how to use multi-gpu or it will be very slow for training

[srama2512]

@GuoPingPan - the mapper training was intended to work only on a single GPU. The other GPUs are used primarily for data collection via habitat-sim/lab. So, GPU-0 uses most memory for mapper training and the other GPUs use most memory for habitat simulator instances. If you plan to use multi-GPU training for the mapper, you may have to appropriately modify how resources are allocated across GPUs.