RuntimeError due to input size mismatch when running evaluation of goat_skill_chain

[fylk317]

Input Size Mismatch in goat_skill_chain Evaluation

Problem Description

When running the evaluation of goat_skill_chain, two main issues are encountered:

1. The checkpoint file ckpt.121.pth is missing.
2. After commenting out the missing checkpoint, the following error occurs when loading cache:

Cache dir: data/goat-assets/goal_cache/iin/val_seen_embeddings/wcojb4TFT35_embedding.pkl
  0%|                                                                                                                                                                                               | 0/360 [00:00<?, ?it/s]

RuntimeError: input.size(-1) must be equal to input_size. Expected 1376, got 1632

The error might stem from a mismatch in RNN input configurations:

Configuration 1:

Observation space info:
  compass: Box(-3.1415927, 3.1415927, (1,), float32)
  current_subtask: Box(0, 4, (1,), int32)
  gps: Box(-3.4028235e+38, 3.4028235e+38, (2,), float32)
  language_goal: Box(-inf, inf, (768,), float32)
  rgb: Box(0, 255, (224, 224, 3), uint8)

Near objectgoal policy: objectgoal
Near CLIP obj/img policy: clip_objectgoal -  clip_imagegoal
In CLIP policy: language_goal - True
Language embedding: 768, Add Language linear: False

RNN input size info: 
  prev_action: 32
  language_goal: 768
  gps_embedding: 32
  compass_embedding: 32
  visual_feats: 512
  --------------------------
  Total RNN input size: 1376

Configuration 2:

Observation space info:
  cache_instance_imagegoal: Box(-inf, inf, (1024,), float32)
  compass: Box(-3.1415927, 3.1415927, (1,), float32)
  current_subtask: Box(0, 4, (1,), int32)
  gps: Box(-3.4028235e+38, 3.4028235e+38, (2,), float32)
  rgb: Box(0, 255, (224, 224, 3), uint8)

Near objectgoal policy: objectgoal
Near CLIP obj/img policy: clip_objectgoal -  clip_imagegoal
In CLIP policy: language_goal - False
InstanceImage embedding: 1024, Add Instance linear: False

RNN input size info: 
  prev_action: 32
  instance_goal: 1024
  gps_embedding: 32
  compass_embedding: 32
  visual_feats: 512
  --------------------------
  Total RNN input size: 1632

Request for Assistance

I would greatly appreciate any guidance on how to resolve this input size mismatch and ensure compatibility between the two configurations. Thank you in advance for your help!

[Ram81]

@fylk317 can you share the command you are using to run this eval? If you are using this command with paths to correct embedding files it should be working

[fylk317]

Hi @Ram81,

Here’s the command I’m using for eval:

DATA_PATH="data/datasets/goat_bench/hm3d/v1/"
tensorboard_dir="data/goat-assets/tensorboard/_logs/"
eval_ckpt_path_dir="data/goat-assets/checkpoints/sense_act_nn_skill_chain/"
split="val_seen"

python -um goat_bench.run \
  --run-type eval \
  --exp-config config/experiments/ver_goat_skill_chain.yaml \
  habitat_baselines.num_environments=1 \
  habitat_baselines.trainer_name="goat_ppo" \
  habitat_baselines.rl.policy.name=GoatHighLevelPolicy \
  habitat_baselines.tensorboard_dir=$tensorboard_dir \
  habitat_baselines.eval_ckpt_path_dir=$eval_ckpt_path_dir \
  habitat_baselines.checkpoint_folder=$eval_ckpt_path_dir \
  habitat.dataset.data_path="${DATA_PATH}/${split}/${split}.json.gz" \
  +habitat/task/lab_sensors@habitat.task.lab_sensors.clip_objectgoal_sensor=clip_objectgoal_sensor \
  +habitat/task/lab_sensors@habitat.task.lab_sensors.language_goal_sensor=language_goal_sensor \
  +habitat/task/lab_sensors@habitat.task.lab_sensors.cache_instance_imagegoal_sensor=cache_instance_imagegoal_sensor \
  ~habitat.task.lab_sensors.goat_goal_sensor \
  habitat.task.lab_sensors.cache_instance_imagegoal_sensor.cache=data/goat-assets/goal_cache/iin/${split}_embeddings/ \
  habitat.task.lab_sensors.language_goal_sensor.cache=data/goat-assets/goal_cache/language_nav/${split}_instruction_clip_embeddings.pkl \
  habitat_baselines.load_resume_state_config=False \
  habitat_baselines.eval.use_ckpt_config=False \
  habitat_baselines.eval.split=$split \
  habitat_baselines.eval.should_load_ckpt=False \
  habitat_baselines.should_load_agent_state=False

I’m using the instruction_clip_embeddings.pkl instead of bert_embedding.pkl since the bert-embedding file is missing in goat-assets. Could this difference be causing the error? If so, could you let me know where the bert_embedding.pkl file is located?

Thanks!