File size: 16,708 Bytes
dbcd95e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 | from __future__ import annotations
import os
import random
import re
import importlib.util
import sys
from typing import Any, Union, Mapping, List, Tuple, TYPE_CHECKING
from pathlib import Path
from typing import Callable, Optional
import importlib.resources as resources
import zipfile
# if TYPE_CHECKING:
import numpy as np
import torch
from PIL import Image
from copy import deepcopy
import tyro
import sys
def seed_everything(seed):
# accelerator.set_seed(cfg.seed, device_specific=True) TODO preferr this way!
random.seed(seed)
os.environ["PYTHONHASHSEED"] = str(seed)
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
try:
import tensorflow as tf
# unfornately, still not obtaining deterministic results for RLDS
tf.random.set_seed(seed)
except:
pass
# torch.backends.cudnn.deterministic = True
torch.use_deterministic_algorithms(True, warn_only=True) # FIXME
# torch.use_deterministic_algorithms(True) # FIXME
torch.backends.cudnn.benchmark = False
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8" # for reproducibility in CUBLAS
def flatten(dictionary, parent_key='', separator='/'):
""" https://stackoverflow.com/questions/6027558/flatten-nested-dictionaries-compressing-keys """
from collections.abc import MutableMapping
items = []
for key, value in dictionary.items():
new_key = parent_key + separator + key if parent_key else key
if isinstance(value, MutableMapping):
items.extend(flatten(value, new_key, separator=separator).items())
else:
if callable(getattr(value, 'item', None)):
value = value.item()
items.append((new_key, value))
return dict(items)
def nice(x):
if type(x) == float:
return f"{x:.1e}"
elif type(x) == list or type(x) == tuple:
return ",".join([f"{i:.1e}" for i in x])
return x
# turns long lr scheduler name short. eg: constant_with_warmup -> cww, constant-with-warmup -> cww
def shorten(x):
if "_" in x or "-" in x:
return "".join([y[0] for y in re.split(r"[-_]", x)])
return f"{x[:6]}"
def inspect(x):
return (x.max(), x.min(), x.mean(), x.std())
def pt_to_pil(x, normalized=True):
s, b = (0.5, 0.5) if normalized else (1.0, 0.0)
return Image.fromarray(
(((x.float() * s + b).clamp(0, 1))*255.0).permute(1,2,0).cpu().numpy().astype(np.uint8)
)
def rmse(l1_err_list):
squared_errors = l1_err_list ** 2
mean_squared_error = np.mean(squared_errors)
rmse_value = np.sqrt(mean_squared_error)
return rmse_value
def snake_to_pascal(snake_str: str) -> str:
parts = re.split(r"[-_]", snake_str)
return ''.join([parts[0].capitalize()] + [p.capitalize() for p in parts[1:]])
def set_global_seed(seed: int, get_worker_init_fn: bool = False) -> Optional[Callable[[int], None]]:
"""Sets seed for all randomness libraries (mostly random, numpy, torch) and produces a `worker_init_fn`"""
import numpy as np
import torch
assert np.iinfo(np.uint32).min < seed < np.iinfo(np.uint32).max, "Seed outside the np.uint32 bounds!"
# Set Seed as an Environment Variable
os.environ["EXPERIMENT_GLOBAL_SEED"] = str(seed)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
return worker_init_function if get_worker_init_fn else None
def worker_init_function(worker_id: int) -> None:
"""
Borrowed directly from PyTorch-Lightning; inspired by this issue comment in the PyTorch repo:
> Ref: https://github.com/pytorch/pytorch/issues/5059#issuecomment-817392562
Intuition: You can think of the seed sequence spawn function as a "janky" torch.Generator() or jax.PRNGKey that
you can run iterative splitting on to get new (predictable) randomness.
:param worker_id: Identifier for the given worker [0, num_workers) for the Dataloader in question.
"""
import numpy as np
import torch
# Get current `rank` (if running distributed) and `process_seed`
global_rank, process_seed = int(os.environ["LOCAL_RANK"]), torch.initial_seed()
# Back out the "base" (original) seed - the per-worker seed is set in PyTorch:
# > https://pytorch.org/docs/stable/data.html#data-loading-randomness
base_seed = process_seed - worker_id
# "Magic" code --> basically creates a seed sequence that mixes different "sources" and seeds every library...
seed_seq = np.random.SeedSequence([base_seed, worker_id, global_rank])
# Use 128 bits (4 x 32-bit words) to represent seed --> generate_state(k) produces a `k` element array!
np.random.seed(seed_seq.generate_state(4))
# Spawn distinct child sequences for PyTorch (reseed) and stdlib random
torch_seed_seq, random_seed_seq = seed_seq.spawn(2)
# Torch Manual seed takes 64 bits (so just specify a dtype of uint64
torch.manual_seed(torch_seed_seq.generate_state(1, dtype=np.uint64)[0])
# Use 128 Bits for `random`, but express as integer instead of as an array
random_seed = (random_seed_seq.generate_state(2, dtype=np.uint64).astype(list) * [1 << 64, 1]).sum()
random.seed(random_seed)
def get_asset_dir() -> Path:
return (resources.files(__package__.split('.')[0]) / ".." / ".." ).resolve() / "assets" # type: ignore
def get_cache_dir() -> Path:
if "HF_HOME" in os.environ:
return Path(os.environ["HF_HOME"])
return (resources.files(__package__.split('.')[0]) / ".." / ".." ).resolve() / ".cache" # type: ignore
def get_data_dir():
if "DATA_HOME" in os.environ:
return Path(os.environ["DATA_HOME"])
if "PSI_HOME" in os.environ:
return Path(os.environ["PSI_HOME"]) / "data"
return (resources.files(__package__.split('.')[0]) / ".." / ".." ).resolve() / ".data" # type: ignore
def get_we_dir():
# if "WE_HOME" in os.environ:
# return Path(os.environ["WE_HOME"]) / "runs"
return (resources.files(__package__.split('.')[0]) / ".." / ".." ).resolve() # type: ignore
def resolve_data_path(path: Union[str, Path], auto_download=False) -> Path:
return resolve_path(path, subdir="data", auto_download=auto_download)
def resolve_path(path: Union[str, Path], subdir="data", auto_download=False) -> Path:
if Path(path).absolute().exists():
return Path(path).absolute()
source_dir = get_asset_dir() / path
if source_dir.exists():
return source_dir
if "DATA_HOME" in os.environ and subdir == "data":
data_dir = Path(os.environ["DATA_HOME"])
filepath = data_dir / path
if filepath.exists():
return filepath
if "PSI_HOME" in os.environ:
proj_dir = Path(os.environ["PSI_HOME"])
filepath = proj_dir / subdir / path
if filepath.exists():
return filepath
if auto_download:
# auto download the file from remmote huggingface repo "USC-PSI-Lab/psi-data" and extract to the data dir
from huggingface_hub import snapshot_download
repo_id = "USC-PSI-Lab/psi-data"
filename = path if isinstance(path, str) else path.name
try:
def _parse_zip_file_from_rel_path(rel_path: str) -> str:
pattern = r"^assets/robots/([^/]+)/.+\.urdf$"
match = re.match(pattern, rel_path)
if match:
robot_name = match.group(1)
return f"assets/robots/{robot_name}.zip"
zip_file = _parse_zip_file_from_rel_path(filename)
# print(zip_file);
# auto download
snapshot_download(
repo_id="USC-PSI-Lab/psi-data",
allow_patterns=[zip_file],
local_dir=os.getcwd(),
repo_type="dataset",
# resume_download=True,
# token="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",
)
# print(data_dir);exit(0)
# zip_path = os.path.join(data_dir, zip_file)
with zipfile.ZipFile(zip_file, 'r') as zip_ref:
zip_ref.extractall(os.path.dirname(zip_file))
except Exception as e:
print(f"Failed to download {filename} from HuggingFace repo {repo_id}: {e}")
raise e
return Path(path)
def move_to_device(
batch,
device,
dtype=None
):
"""
Recursively moves a batch of data (dictionary, list, or tensor) to the specified device.
Args:
batch (dict, list, tensor): The batch of data.
device (torch.device or str): The target device ('cuda' or 'cpu').
dtype: Optional dtype to cast to. If None, preserves original dtype.
Returns:
The batch moved to the specified device.
"""
if isinstance(batch, torch.Tensor):
if dtype is not None:
return batch.to(device, dtype)
else:
return batch.to(device) # Preserve original dtype
elif isinstance(batch, dict):
return {key: move_to_device(value, device, dtype) for key, value in batch.items()}
elif isinstance(batch, list):
return [move_to_device(item, device, dtype) for item in batch]
elif isinstance(batch, tuple):
return tuple(move_to_device(item, device, dtype) for item in batch)
elif isinstance(batch, np.ndarray):
tensor = torch.from_numpy(batch)
if dtype is not None:
return tensor.to(device, dtype)
else:
return tensor.to(device) # Preserve numpy's original dtype
else:
try:
import transformers
if isinstance(batch, transformers.feature_extraction_utils.BatchFeature):
return batch.to(device)
except ImportError:
pass
return batch # Return as is if not a recognized type
def overlay(rgb, traj2d, black=0.0, eps=0.0):
"""
Overlay a 2D trajectory on an RGB image.
black: Value to consider as black in the trajectory tensor
eps: Small value to tolerate
"""
import torch
rgb = torch.clone(rgb)
lower = black + eps
rgb[traj2d > lower] = traj2d[traj2d > lower]
return rgb
def count_parameters(model, trainable: bool = False):
"""Return total and trainable parameter counts of a model."""
total_params = sum(p.numel() for p in model.parameters())
if trainable:
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
return total_params, trainable_params
else:
return total_params
# print(f"Total parameters: {total_params:,}")
# print(f"Trainable parameters: {trainable_params:,}")
def str_to_tensor(s: str):
return torch.tensor(list(s.encode('utf-8')), dtype=torch.uint8)
def tensor_to_str(t):
return bytes(t.tolist()).decode('utf-8')
def string_compatible_collate(batch):
collated = {}
keys = batch[0].keys()
for k in keys:
values = [d[k] for d in batch]
if isinstance(values[0], str):
# Convert list of strings to list of tensors, then pad
tensors = [str_to_tensor(v) for v in values]
padded = torch.nn.utils.rnn.pad_sequence(tensors, batch_first=True)
collated[f"{k}_str"] = padded
elif torch.is_tensor(values[0]):
collated[k] = torch.stack(values)
elif isinstance(values[0], np.ndarray):
collated[k] = torch.from_numpy(np.stack(values))
else:
if isinstance(values[0], dict):
collated[k] = string_compatible_collate(values)
else:
collated[k] = torch.tensor(np.array(values))
return collated
from typing import TypeVar, Dict
T = TypeVar("T")
def batch_str_to_tensor(batch: T) -> T:
import torch
if isinstance(batch, torch.Tensor):
return batch_str_to_tensor({"data": batch})["data"]
ret = {}
for k, v in batch.items():
if k.endswith("_str"):
v = [tensor_to_str(s) for s in (v)]
ret[k.replace("_str", "")] = v
return ret # type: ignore
# wrote for pytests
def extract_args_from_shell_script(script_path):
"""Extract arguments from a shell script that uses accelerate launch"""
with open(script_path, 'r') as f:
content = f.read()
# Extract from args variable if it exists
args_match = re.search(r'args="([^"]+)"', content, re.DOTALL)
if args_match:
args_content = args_match.group(1)
# Split by backslash and newline, clean up
args_list = [
arg.strip().rstrip('\\').strip()
for arg in args_content.split('\n')
if arg.strip() and not arg.strip().startswith('#')
]
# Split each argument by space to handle cases like "--train.lr_scheduler_kwargs.betas 0.9 0.999"
final_args = []
for arg in args_list:
if arg:
final_args.extend(arg.split())
return final_args
print("Could not find args variable in the script.")
return []
def parse_args_to_tyro_config(args_or_script_path: Path | str, force_rewrite_config_file=False):
assert Path(args_or_script_path).exists(), f"Path does not exist: {args_or_script_path}"
if str(args_or_script_path).endswith(".sh"):
print(f"Parsing configuration {args_or_script_path}")
argv = [sys.argv[0]] + extract_args_from_shell_script(args_or_script_path)
else:
argv = []
with open(args_or_script_path, "r") as f:
for line in f:
# Split by any whitespace (spaces, tabs, newlines)
argv.extend(line.strip().split())
assert len(argv) > 1, "Please put all arguments in a spererate variable, eg., args=..."
sys.argv = argv
try:
# By convention, the first argument after the script name is the config module name,
# eg., {trainer}_{data}_{model}_config, which corresponds to psi.config.train.pretrain_egodex_qwen3vl_config
module = importlib.import_module(f"psi.config.train.{sys.argv[1]}")
DynamicLaunchConfigClass = getattr(module, "DynamicLaunchConfig")
config = tyro.cli(DynamicLaunchConfigClass, config=(tyro.conf.ConsolidateSubcommandArgs,), args=sys.argv[2:])
except Exception as e:
print(f"Failed to import config module 'psi.config.train.{sys.argv[1]}'")
raise e
return config
def batchify(data):
data = deepcopy(data)
for k, v in data.items():
if isinstance(v, torch.Tensor):
data[k] = v[None]
elif isinstance(v, np.ndarray):
data[k] = torch.from_numpy(v)[None]
elif isinstance(v, str):
data[k] = [v]
return data
def make_image_grid(images, nrows=None, ncols=None):
# Determine grid dimensions (e.g., 2x2 for 4 images)
num_images = len(images)
if nrows is None and ncols is None:
if num_images <= 3:
nrows, ncols = 1, num_images
else:
grid_size = int(np.ceil(np.sqrt(num_images)))
nrows, ncols = grid_size, grid_size
if (nrows -1 ) * ncols >= num_images:
nrows -= 1
elif nrows is None:
nrows = int(np.ceil(num_images / (ncols or 1)))
else:
ncols = int(np.ceil(num_images / nrows))
H, W = images[0].size
grid = Image.new("RGB", size=(ncols * W, nrows * H))
for i, image in enumerate(images):
row, col = divmod(i, ncols or 1)
grid.paste(image, box = (col * W, row * H))
return grid
def pad_to_len(x, target_len, dim=1, pad_value=0.0):
"""Pads a tensor to the target length along the specified dimension.
Args:
x: np.ndarray to pad
target_len: int, target length to pad to
dim: int, dimension along which to pad
pad_value: value to use for padding
Returns:
padded: np.ndarray, padded array
mask: np.ndarray of bool, True for original data, False for padded region
"""
current_len = x.shape[dim]
if current_len >= target_len:
mask = np.ones(x.shape, dtype=bool)
return x, mask
pad_width = [(0, 0)] * x.ndim
pad_width[dim] = (0, target_len - current_len)
# np.pad pads as (before, after) for each axis
padded = np.pad(x, pad_width, mode='constant', constant_values=pad_value)
mask_shape = list(x.shape)
mask_shape[dim] = target_len
mask = np.ones(mask_shape, dtype=bool)
# Mark padded region as False (0)
idx = [slice(None)] * x.ndim
idx[dim] = slice(current_len, target_len)
mask[tuple(idx)] = False
return padded, mask |