Fix random (#36)

* pass random seed to get_data

* fix

* add setuptools

* Update train.py

* pin setuptools

* unmute Wandb

* unpin gym

* fix formatting

* change n_proc to n_proc_per_worker

---------

Co-authored-by: Ben Eisner <ben.a.eisner@gmail.com>

.github/workflows/build-site.yaml

@@ -22,6 +22,9 @@ jobs:
  with:
  python-version: '3.8'
  ##############################################
+ - name: Update pip and setuptools.
+ run: pip install --upgrade pip setuptools==57.1.0 wheel
+
  - name: Install pyg special-like.
  run: pip install pyg-lib torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.13.0+cpu.html
 

.github/workflows/run-tests.yaml

@@ -32,6 +32,8 @@ jobs:
  with:
  python-version: '3.8'
  ##############################################
+ - name: Update pip and setuptools.
+ run: pip install --upgrade pip setuptools==57.1.0 wheel
 
  - name: Install pyg special-like.
  run: pip install pyg-lib torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.13.0+cpu.html

flowbot3d/datasets/flow_dataset_pyg.py

@@ -23,6 +23,7 @@ def __init__(
  randomize_joints: bool = True,
  randomize_camera: bool = True,
  n_points: Optional[int] = 1200,
+ seed: int = 42,
  ) -> None:
  super().__init__()
 
@@ -33,21 +34,22 @@ def __init__(
  randomize_camera,
  n_points,
  )
+ self.seed = seed
 
  def len(self) -> int:
  return len(self.dataset)
 
  def get(self, index) -> tgd.Data:
- return self.get_data(self.dataset._dataset._ids[index], seed=None)
+ return self.get_data(self.dataset._dataset._ids[index])
 
  @staticmethod
  def get_processed_dir(randomize_joints, randomize_camera):
  joint_chunk = "rj" if randomize_joints else "sj"
  camera_chunk = "rc" if randomize_camera else "sc"
  return f"processed_{joint_chunk}_{camera_chunk}"
 
- def get_data(self, obj_id: str, seed=None) -> Flowbot3DTGData:
- data_dict = self.dataset.get_data(obj_id, seed)
+ def get_data(self, obj_id: str) -> Flowbot3DTGData:
+ data_dict = self.dataset.get_data(obj_id, self.seed)
 
  data = tgd.Data(
  id=data_dict["id"],

flowbot3d/evaluate_flow_models.py

@@ -40,7 +40,7 @@ def make_eval_dataset(
  randomize_camera,
  ),
  n_repeat=1,
- n_proc=n_proc,
+ n_proc_per_worker=n_proc,
  seed=42,
  )
 

flowbot3d/train.py

@@ -42,6 +42,7 @@ def create_flowbot_datasets(
  dataset: str,
  n_proc=-1,
  randomize_camera: bool = True,
+ seed=42,
 ) -> Tuple[tgd.Dataset, tgd.Dataset, tgd.Dataset]:
  if dataset == "umpnet":
  train_dset = CachedByKeyDataset(
@@ -50,6 +51,7 @@ def create_flowbot_datasets(
  root=root / "raw",
  split="umpnet-train-train",
  randomize_camera=randomize_camera,
+ seed=seed,
  ),
  data_keys=rpd.UMPNET_TRAIN_TRAIN_OBJ_IDS,
  root=root,
@@ -58,7 +60,8 @@ def create_flowbot_datasets(
  randomize_camera,
  ),
  n_repeat=100,
- n_proc=n_proc,
+ n_workers=32,
+ n_proc_per_worker=n_proc,
  )
 
  test_dset = CachedByKeyDataset(
@@ -67,6 +70,7 @@ def create_flowbot_datasets(
  root=root / "raw",
  split="umpnet-train-test",
  randomize_camera=randomize_camera,
+ seed=seed,
  ),
  data_keys=rpd.UMPNET_TRAIN_TEST_OBJ_IDS,
  root=root,
@@ -75,7 +79,8 @@ def create_flowbot_datasets(
  randomize_camera,
  ),
  n_repeat=1,
- n_proc=n_proc,
+ n_workers=32,
+ n_proc_per_worker=n_proc,
  )
 
  unseen_dset = CachedByKeyDataset(
@@ -84,6 +89,7 @@ def create_flowbot_datasets(
  root=root / "raw",
  split="umpnet-test",
  randomize_camera=randomize_camera,
+ seed=seed,
  ),
  data_keys=rpd.UMPNET_TEST_OBJ_IDS,
  root=root,
@@ -92,7 +98,8 @@ def create_flowbot_datasets(
  randomize_camera,
  ),
  n_repeat=1,
- n_proc=n_proc,
+ n_workers=32,
+ n_proc_per_worker=n_proc,
  )
  elif dataset == "single":
  dset = CachedByKeyDataset(
@@ -101,6 +108,7 @@ def create_flowbot_datasets(
  root=root / "raw",
  split=["7179"],
  randomize_camera=randomize_camera,
+ seed=seed,
  ),
  data_keys=["7179"],
  root=root,
@@ -109,7 +117,8 @@ def create_flowbot_datasets(
  randomize_camera,
  ),
  n_repeat=1,
- n_proc=n_proc,
+ n_workers=32,
+ n_proc_per_worker=n_proc,
  )
  train_dset = dset
  test_dset = dset
@@ -257,6 +266,7 @@ def train(
  dataset,
  n_proc,
  randomize_camera=randomize_camera,
+ seed=seed,
  )
 
  train_loader = tgl.DataLoader(train_dset, batch_size, shuffle=True, num_workers=0)
@@ -276,6 +286,7 @@ def train(
  project = "flowbot3d"
  logger = plog.WandbLogger(
  project=project,
+ entity="r-pad",
  config={
  "dataset": dataset,
  "model": model_type,
@@ -306,7 +317,8 @@ def train(
 
  # Create the trainer, which we'll train on only 1 gpu.
  trainer = pl.Trainer(
- gpus=1,
+ accelerator="gpu",
+ devices=[0],
  logger=logger,
  callbacks=cbs,
  log_every_n_steps=log_every_n_steps,

pyproject.toml

@@ -12,7 +12,8 @@ authors = [
 ]
 dependencies = [
  "gif",
- "gym <= 0.20.0",
+ # "gym <= 0.20.0", # This specific version is probably desired, but the CI no longer seems to build :( will fix later.
+ "gym",
  "h5py",
  "numpy <= 1.23.1",
  "open3d",