Source code for modulus.datapipes.gnn.ahmed_body_dataset
# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
# SPDX-FileCopyrightText: All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import concurrent.futures as cf
import os
import re
from dataclasses import dataclass
from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
import numpy as np
import torch
import yaml
from torch import Tensor
from modulus.datapipes.datapipe import Datapipe
from modulus.datapipes.meta import DatapipeMetaData
from .utils import load_json, read_vtp_file, save_json
try:
import dgl
from dgl.data import DGLDataset
except ImportError:
raise ImportError(
"Ahmed Body Dataset requires the DGL library. Install the "
+ "desired CUDA version at: \n https://www.dgl.ai/pages/start.html"
)
try:
import pyvista as pv
import vtk
except ImportError:
raise ImportError(
"Ahmed Body Dataset requires the vtk and pyvista libraries. Install with "
+ "pip install vtk pyvista"
)
[docs]@dataclass
class FileInfo:
"""VTP file info storage."""
velocity: float
reynolds_number: float
length: float
width: float
height: float
ground_clearance: float
slant_angle: float
fillet_radius: float
[docs]@dataclass
class MetaData(DatapipeMetaData):
name: str = "AhmedBody"
# Optimization
auto_device: bool = True
cuda_graphs: bool = False
# Parallel
ddp_sharding: bool = True
[docs]class AhmedBodyDataset(DGLDataset, Datapipe):
"""
In-memory Ahmed body Dataset
Parameters
----------
data_dir: str
The directory where the data is stored.
split: str, optional
The dataset split. Can be 'train', 'validation', or 'test', by default 'train'.
num_samples: int, optional
The number of samples to use, by default 10.
invar_keys: Iterable[str], optional
The input node features to consider. Default includes 'pos', 'velocity', 'reynolds_number', 'length', 'width', 'height', 'ground_clearance', 'slant_angle', and 'fillet_radius'.
outvar_keys: Iterable[str], optional
The output features to consider. Default includes 'p' and 'wallShearStress'.
normalize_keys Iterable[str], optional
The features to normalize. Default includes 'p', 'wallShearStress', 'velocity', 'length', 'width', 'height', 'ground_clearance', 'slant_angle', and 'fillet_radius'.
normalization_bound: Tuple[float, float], optional
The lower and upper bounds for normalization. Default is (-1, 1).
force_reload: bool, optional
If True, forces a reload of the data, by default False.
name: str, optional
The name of the dataset, by default 'dataset'.
verbose: bool, optional
If True, enables verbose mode, by default False.
compute_drag: bool, optional
If True, also returns the coefficient and mesh area and normals that are required for computing the drag coefficient.
num_workers: int, optional
Number of dataset pre-loading workers. If None, will be chosen automatically.
"""
def __init__(
self,
data_dir: str,
split: str = "train",
num_samples: int = 10,
invar_keys: Iterable[str] = (
"pos",
"velocity",
"reynolds_number",
"length",
"width",
"height",
"ground_clearance",
"slant_angle",
"fillet_radius",
),
outvar_keys: Iterable[str] = ("p", "wallShearStress"),
normalize_keys: Iterable[str] = (
"p",
"wallShearStress",
"velocity",
"reynolds_number",
"length",
"width",
"height",
"ground_clearance",
"slant_angle",
"fillet_radius",
),
normalization_bound: Tuple[float, float] = (-1.0, 1.0),
force_reload: bool = False,
name: str = "dataset",
verbose: bool = False,
compute_drag: bool = False,
num_workers: Optional[int] = None,
):
DGLDataset.__init__(
self,
name=name,
force_reload=force_reload,
verbose=verbose,
)
Datapipe.__init__(
self,
meta=MetaData(),
)
self.split = split
self.num_samples = num_samples
self.data_dir = os.path.join(data_dir, self.split)
self.info_dir = os.path.join(data_dir, self.split + "_info")
self.input_keys = list(invar_keys)
self.output_keys = list(outvar_keys)
self.normalize_keys = list(normalize_keys)
self.normalization_bound = normalization_bound
self.compute_drag = compute_drag
# get the list of all files in the data_dir
all_entries = os.listdir(self.data_dir)
all_info = os.listdir(self.info_dir)
data_list = [
os.path.join(self.data_dir, entry)
for entry in all_entries
if os.path.isfile(os.path.join(self.data_dir, entry))
]
info_list = [
os.path.join(self.info_dir, entry)
for entry in all_info
if os.path.isfile(os.path.join(self.info_dir, entry))
]
numbers = []
for directory in data_list:
match = re.search(r"\d+", directory)
if match:
numbers.append(int(match.group()))
numbers_info = []
for directory in info_list:
match = re.search(r"\d+", directory)
if match:
numbers_info.append(int(match.group()))
numbers = [int(n) for n in numbers]
numbers_info = [int(n) for n in numbers_info]
# sort the data_list and info_list according to the numbers
args = np.argsort(numbers)
data_list = [data_list[index] for index in args]
numbers = [numbers[index] for index in args]
args = np.argsort(numbers_info)
info_list = [info_list[index] for index in args]
numbers_info = [numbers_info[index] for index in args]
if sorted(numbers) != sorted(numbers_info):
raise AssertionError
self.numbers = numbers
# create the graphs and add the node and features
self.length = min(len(data_list), self.num_samples)
if self.num_samples > self.length:
raise ValueError(
f"Number of available {self.split} dataset entries "
f"({self.length}) is less than the number of samples "
f"({self.num_samples})"
)
self.graphs = [None] * self.length
if self.compute_drag:
self.normals = [None] * self.length
self.areas = [None] * self.length
self.coeff = [None] * self.length
# create graphs from VTP files using multiprocessing.
if num_workers is None or num_workers <= 0:
def get_num_workers():
# Make sure we don't oversubscribe CPUs on a node.
# TODO(akamenev): this should be in DistributedManager.
local_node_size = max(
int(os.environ.get("OMPI_COMM_WORLD_LOCAL_SIZE", 1)), 1
)
num_workers = len(os.sched_getaffinity(0)) // local_node_size
return max(num_workers - 1, 1)
num_workers = get_num_workers()
with cf.ProcessPoolExecutor(
max_workers=num_workers,
mp_context=torch.multiprocessing.get_context("spawn"),
) as executor:
for (i, graph, coeff, normal, area) in executor.map(
self.create_graph,
range(self.length),
data_list[: self.length],
info_list[: self.length],
chunksize=max(1, self.length // num_workers),
):
self.graphs[i] = graph
if self.compute_drag:
self.coeff[i] = coeff
self.normals[i] = normal
self.areas[i] = area
# add the edge features
self.graphs = self.add_edge_features()
# normalize the node and edge features
if self.split == "train":
self.node_stats = self._get_node_stats(keys=self.normalize_keys)
self.edge_stats = self._get_edge_stats()
else:
if not os.path.exists("node_stats.json"):
raise FileNotFoundError(
"node_stats.json not found! Node stats must be computed on the training set."
)
if not os.path.exists("edge_stats.json"):
raise FileNotFoundError(
"edge_stats.json not found! Edge stats must be computed on the training set."
)
self.node_stats = load_json("node_stats.json")
self.edge_stats = load_json("edge_stats.json")
self.graphs = self.normalize_node()
self.graphs = self.normalize_edge()
[docs] def create_graph(self, index: int, file_path: str, info_path: str) -> None:
"""Creates a graph from VTP file.
This method is used in parallel loading of graphs.
Returns
-------
Tuple that contains graph index, graph, and optionally coeff, normal and area values.
"""
polydata = read_vtp_file(file_path)
graph = self._create_dgl_graph(polydata, self.output_keys, dtype=torch.int32)
info = self._read_info_file(info_path)
for v in vars(info):
if v not in self.input_keys:
continue
graph.ndata[v] = getattr(info, v) * torch.ones_like(
graph.ndata["pos"][:, [0]]
)
coeff = None
normal = None
area = None
if "normals" in self.input_keys or self.compute_drag:
mesh = pv.read(file_path)
mesh.compute_normals(cell_normals=True, point_normals=False, inplace=True)
if "normals" in self.input_keys:
graph.ndata["normals"] = torch.from_numpy(
mesh.cell_data_to_point_data()["Normals"]
)
if self.compute_drag:
mesh = mesh.compute_cell_sizes()
mesh = mesh.cell_data_to_point_data()
frontal_area = info.width * info.height / 2 * (10 ** (-6))
coeff = 2.0 / ((info.velocity**2) * frontal_area)
normal = torch.from_numpy(mesh["Normals"])
area = torch.from_numpy(mesh["Area"])
return index, graph, coeff, normal, areadef __getitem__(self, idx):
graph = self.graphs[idx]
if self.compute_drag:
sid = self.numbers[idx]
return graph, sid, self.normals[idx], self.areas[idx], self.coeff[idx]
return graph
def __len__(self):
return self.length
[docs] def add_edge_features(self) -> List[dgl.DGLGraph]:
"""
Add relative displacement and displacement norm as edge features for each graph
in the list of graphs. The calculations are done using the 'pos' attribute in the
node data of each graph. The resulting edge features are stored in the 'x' attribute
in the edge data of each graph.
This method will modify the list of graphs in-place.
Returns
-------
List[dgl.DGLGraph]
The list of graphs with updated edge features.
"""
if not hasattr(self, "graphs") or not self.graphs:
raise ValueError("The list 'graphs' is empty.")
for graph in self.graphs:
pos = graph.ndata.get("pos")
if pos is None:
raise ValueError(
"'pos' does not exist in the node data of one or more graphs."
)
row, col = graph.edges()
row = row.long()
col = col.long()
disp = pos[row] - pos[col]
disp_norm = torch.linalg.norm(disp, dim=-1, keepdim=True)
graph.edata["x"] = torch.cat((disp, disp_norm), dim=-1)
return self.graphs
[docs] def normalize_node(self) -> List[dgl.DGLGraph]:
"""
Normalize node data in each graph in the list of graphs.
Returns
-------
List[dgl.DGLGraph]
The list of graphs with normalized and concatenated node data.
"""
if not hasattr(self, "graphs") or not self.graphs:
raise ValueError("The list 'graphs' is empty.")
if not hasattr(self, "node_stats") or not isinstance(self.node_stats, dict):
raise ValueError(
"The 'node_stats' attribute does not exist or is not a dictionary."
)
invar_keys = set(
[
key.replace("_mean", "").replace("_std", "")
for key in self.node_stats.keys()
]
)
for i in range(len(self.graphs)):
for key in invar_keys:
self.graphs[i].ndata[key] = (
self.graphs[i].ndata[key] - self.node_stats[key + "_mean"]
) / self.node_stats[key + "_std"]
self.graphs[i].ndata["x"] = torch.cat(
[self.graphs[i].ndata[key] for key in self.input_keys], dim=-1
)
self.graphs[i].ndata["y"] = torch.cat(
[self.graphs[i].ndata[key] for key in self.output_keys], dim=-1
)
return self.graphs
[docs] def normalize_edge(self) -> List[dgl.DGLGraph]:
"""
Normalize edge data 'x' in each graph in the list of graphs.
Returns
-------
List[dgl.DGLGraph]
The list of graphs with normalized edge data 'x'.
"""
if not hasattr(self, "graphs") or not self.graphs:
raise ValueError("The list 'graphs' is empty.")
if not hasattr(self, "edge_stats") or not isinstance(self.edge_stats, dict):
raise ValueError(
"The 'edge_stats' attribute does not exist or is not a dictionary."
)
for i in range(len(self.graphs)):
self.graphs[i].edata["x"] = (
self.graphs[i].edata["x"] - self.edge_stats["edge_mean"]
) / self.edge_stats["edge_std"]
return self.graphs
[docs] def denormalize(self, pred, gt, device) -> Tuple[Tensor, Tensor]:
"""
Denormalize the graph node data.
Parameters:
-----------
pred: Tensor
Normalized prediction
gt: Tensor
Normalized ground truth
device: Any
The device
Returns:
--------
Tuple(Tensor, Tensor)
Denormalized prediction and ground truth
"""
stats = self.node_stats
stats = {key: val.to(device) for key, val in stats.items()}
p_pred = pred[:, [0]]
s_pred = pred[:, 1:]
p_gt = gt[:, [0]]
s_gt = gt[:, 1:]
p_pred = p_pred * stats["p_std"] + stats["p_mean"]
s_pred = s_pred * stats["wallShearStress_std"] + stats["wallShearStress_mean"]
p_gt = p_gt * stats["p_std"] + stats["p_mean"]
s_gt = s_gt * stats["wallShearStress_std"] + stats["wallShearStress_mean"]
pred = torch.cat((p_pred, s_pred), dim=-1)
gt = torch.cat((p_gt, s_gt), dim=-1)
return pred, gtdef _get_edge_stats(self) -> Dict[str, Any]:
"""
Computes the mean and standard deviation of each edge attribute 'x' in the
graphs, and saves to a JSON file.
Returns
-------
dict
A dictionary with keys 'edge_mean' and 'edge_std' and the corresponding values being
1-D tensors containing the mean or standard deviation value for each dimension of the edge attribute 'x'.
"""
if not self.graphs:
raise ValueError("The list 'graphs' is empty.")
stats = {
"edge_mean": 0,
"edge_meansqr": 0,
}
for i in range(self.length):
stats["edge_mean"] += (
torch.mean(self.graphs[i].edata["x"], dim=0) / self.length
)
stats["edge_meansqr"] += (
torch.mean(torch.square(self.graphs[i].edata["x"]), dim=0) / self.length
)
stats["edge_std"] = torch.sqrt(
stats["edge_meansqr"] - torch.square(stats["edge_mean"])
)
stats.pop("edge_meansqr")
# save to file
save_json(stats, "edge_stats.json")
return stats
def _get_node_stats(self, keys: List[str]) -> Dict[str, Any]:
"""
Computes the mean and standard deviation values of each node attribute
for the list of keys in the graphs, and saves to a JSON file.
Parameters
----------
keys : list of str
List of keys for the node attributes.
Returns
-------
dict
A dictionary with each key being a string of format '[key]_mean' or '[key]_std'
and each value being a 1-D tensor containing the mean or standard deviation for each
dimension of the node attribute.
"""
if not self.graphs:
raise ValueError("The list 'graphs' is empty.")
stats = {}
for key in keys:
stats[key + "_mean"] = 0
stats[key + "_meansqr"] = 0
for i in range(self.length):
for key in keys:
stats[key + "_mean"] += (
torch.mean(self.graphs[i].ndata[key], dim=0) / self.length
)
stats[key + "_meansqr"] += (
torch.mean(torch.square(self.graphs[i].ndata[key]), dim=0)
/ self.length
)
for key in keys:
stats[key + "_std"] = torch.sqrt(
stats[key + "_meansqr"] - torch.square(stats[key + "_mean"])
)
stats.pop(key + "_meansqr")
# save to file
save_json(stats, "node_stats.json")
return stats
@staticmethod
def _read_info_file(file_path: str) -> FileInfo:
"""
Parse the values of specific parameters from a given text file.
Parameters
----------
file_path : str
Path to the text file.
Returns
-------
FileInfo
A FileInfo object.
"""
with open(file_path, mode="rt", encoding="utf-8") as file:
info = yaml.safe_load(file)
return FileInfo(
info["Velocity"],
info["Re (based on length)"],
info["Length"],
info["Width"],
info["Height"],
info["GroundClearance"],
info["SlantAngle"],
info["FilletRadius"],
)
@staticmethod
def _create_dgl_graph(
polydata: Any,
outvar_keys: List[str],
to_bidirected: bool = True,
add_self_loop: bool = False,
dtype: Union[torch.dtype, str] = torch.int32,
) -> dgl.DGLGraph:
"""
Create a DGL graph from vtkPolyData.
Parameters
----------
polydata : vtkPolyData
vtkPolyData from which the DGL graph is created.
outvar_keys : list of str
List of keys for the node attributes to be extracted from the vtkPolyData.
to_bidirected : bool, optional
Whether to make the graph bidirected. Default is True.
add_self_loop : bool, optional
Whether to add self-loops in the graph. Default is False.
dtype : torch.dtype or str, optional
Data type for the graph. Default is torch.int32.
Returns
-------
dgl.DGLGraph
The DGL graph created from the vtkPolyData.
"""
# Extract point data and connectivity information from the vtkPolyData
points = polydata.GetPoints()
if points is None:
raise ValueError("Failed to get points from the polydata.")
vertices = np.array(
[points.GetPoint(i) for i in range(points.GetNumberOfPoints())]
)
polys = polydata.GetPolys()
if polys is None:
raise ValueError("Failed to get polygons from the polydata.")
polys.InitTraversal()
edge_list = []
for i in range(polys.GetNumberOfCells()):
id_list = vtk.vtkIdList()
polys.GetNextCell(id_list)
for j in range(id_list.GetNumberOfIds() - 1):
edge_list.append( # noqa: PERF401
(id_list.GetId(j), id_list.GetId(j + 1))
)
# Create DGL graph using the connectivity information
graph = dgl.graph(edge_list, idtype=dtype)
if to_bidirected:
graph = dgl.to_bidirected(graph)
if add_self_loop:
graph = dgl.add_self_loop(graph)
# Assign node features using the vertex data
graph.ndata["pos"] = torch.tensor(vertices, dtype=torch.float32)
# Extract node attributes from the vtkPolyData
point_data = polydata.GetPointData()
if point_data is None:
raise ValueError("Failed to get point data from the polydata.")
for i in range(point_data.GetNumberOfArrays()):
array = point_data.GetArray(i)
array_name = array.GetName()
if array_name in outvar_keys:
array_data = np.zeros(
(points.GetNumberOfPoints(), array.GetNumberOfComponents())
)
for j in range(points.GetNumberOfPoints()):
array.GetTuple(j, array_data[j])
# Assign node attributes to the DGL graph
graph.ndata[array_name] = torch.tensor(array_data, dtype=torch.float32)
return graph