Update instructions for supercloud cpu-only iGibson usage! (#76)

* Update instructions for supercloud cpu-only iGibson usage!

* fix tests by porting fix from predicators

* remove CNNRegressor

* fix

* test collecting aluminum cans data collection speed

* revert commenting out stuff in the yaml

behavior.md

@@ -49,7 +49,10 @@ python predicators/main.py --env behavior --approach oracle --option_model_name
 ## Installing on MIT Supercloud
 First, follow steps in our [Supercloud guide](supercloud.md) to get an account and setup this repository on Supercloud.
 
-Next, simply follow the steps linked in the [above section](#installation) (though ignore the first step; supercloud already has the iGibson pre-reqs installed)! Note that for various driver-related reasons, this code only works on GPU-machines with supercloud (so always remember to request a GPU when submitting jobs involving this codebase).
+Next, simply follow the steps linked in the [above section](#installation) (though ignore the first step; supercloud already has the iGibson pre-reqs installed)! Importantly, we have a separate branch of iGibson that *completely disables* rendering so that (1) the simulator runs on CPU-only nodes on SuperCloud, and (2) planning is extremely fast. To use this branch (recommended), `cd` to the iGibson repo and run:
+`git checkout no-render`
+
+Note that if you want to use the master branch of iGibson on SuperCloud, you will need to exclusively request GPU nodes. 
 
 To test installation, do:
 ```
@@ -86,7 +89,7 @@ python predicators/main.py --env behavior --approach oracle --option_model_name
 * Set `--behavior_scene_name` to the name of the house setting (e.g. `Pomaria_1_int`) you want to try running the particular task in. Note that not all tasks are available in all houses (e.g. `re-shelving_library_books` might only be available with `Pomaria_1_int`).
 * If you'd like to see a visual of the agent planning in iGibson, set the command line argument `--behavior_mode simple`. If you want to run in headless mode without any visuals, leave the default (i.e `--behavior_mode headless`).
 * Be sure to set `--plan_only_eval True`: this is necessary to account for the fact that the iGibson simulator is non-deterministic when saving and loading states (which is currently an unresolved bug).
-* Example command: `python predicators/main.py --env behavior --approach oracle --option_model_name oracle_behavior --num_train_tasks 0 --num_test_tasks 1 --behavior_scene_name Pomaria_2_int --behavior_task_list "[opening_packages]" --seed 1000 --offline_data_planning_timeout 500.0 --timeout 500.0 --behavior_option_model_eval True --plan_only_eval True`.
+* Example command: `python predicators/main.py --env behavior --approach oracle --option_model_name oracle_behavior --num_train_tasks 0 --num_test_tasks 1 --behavior_train_scene_name Pomaria_2_int --behavior_test_scene_name Pomaria_2_int --behavior_task_list "[opening_packages]" --seed 1000 --offline_data_planning_timeout 500.0 --timeout 500.0 --behavior_option_model_eval True --plan_only_eval True`.
 
 ## Troubleshooting
 ### Visualizing what the robot is doing

scripts/configs/behavior_20_evaluation.yaml

@@ -157,4 +157,4 @@ FLAGS:  # general flags
   num_test_tasks: 5
 START_SEED: 0
 NUM_SEEDS: 1
-USE_GPU: True
+USE_GPU: False

scripts/configs/behavior_all_tasks_oracle.yaml

@@ -717,4 +717,4 @@ FLAGS:  # general flags
   num_test_tasks: 5
 START_SEED: 0
 NUM_SEEDS: 1
-USE_GPU: True
+USE_GPU: False

scripts/configs/behavior_example.yaml

@@ -30,4 +30,4 @@ FLAGS:  # general flags
   num_test_tasks: 10
 START_SEED: 456
 NUM_SEEDS: 3
-USE_GPU: True
+USE_GPU: False

scripts/configs/behavior_pick_place_data_collection.yaml

@@ -49,5 +49,5 @@ FLAGS:  # general flags
   num_train_tasks: 10
   num_test_tasks: 0
 START_SEED: 456
-NUM_SEEDS: 10
-USE_GPU: True
+NUM_SEEDS: 1
+USE_GPU: False

scripts/configs/behavior_pick_place_evaluation.yaml

@@ -87,4 +87,4 @@ FLAGS:  # general flags
   num_test_tasks: 0
 START_SEED: 463
 NUM_SEEDS: 2
-USE_GPU: True
+USE_GPU: False

scripts/configs/behavior_pick_place_learning.yaml

@@ -84,4 +84,4 @@ FLAGS:  # general flags
   num_test_tasks: 10
 START_SEED: 456
 NUM_SEEDS: 3
-USE_GPU: True
+USE_GPU: False

tests/test_ml_models.py

@@ -1,6 +1,8 @@
 """Tests for models."""
 
+import logging
 import time
+from unittest.mock import patch
 
 import numpy as np
 import pytest
@@ -172,17 +174,17 @@ def test_mlp_classifier():
     assert prediction
     assert model.predict_proba(np.ones(input_size)) > 0.5
     # Test for early stopping
-    start_time = time.perf_counter()
     model = MLPBinaryClassifier(seed=123,
                                 balance_data=True,
                                 max_train_iters=100000,
                                 learning_rate=1e-2,
-                                n_iter_no_change=1,
+                                n_iter_no_change=-1,
                                 hid_sizes=[32, 32],
                                 n_reinitialize_tries=1,
                                 weight_init="default")
-    model.fit(X, y)
-    assert time.perf_counter() - start_time < 3, "Didn't early stop"
+    with patch.object(logging, "info", return_value=None) as mock_logging_info:
+        model.fit(X, y)
+    assert mock_logging_info.call_count < 5
     # Test with no positive examples.
     num_class_samples = 1000
     X = np.concatenate([