Skip to content

chore(bevfusion): update parameters for improved bevfusion-cl training #88

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 9 commits into
base: main
Choose a base branch
from
Open

chore(bevfusion): update parameters for improved bevfusion-cl training #88

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
2df7b08
updated roi params for t4dataset
SamratThapa120 Aug 26, 2025
354ff1b
remove unused param
SamratThapa120 Aug 26, 2025
dc1d30f
ci(pre-commit): autofix
pre-commit-ci[bot] Aug 26, 2025
09f35d5
added bevfusion l
SamratThapa120 Aug 26, 2025
50f35a8
update params
SamratThapa120 Sep 8, 2025
a55a907
updated docs
SamratThapa120 Sep 8, 2025
7006d5a
ci(pre-commit): autofix
pre-commit-ci[bot] Sep 8, 2025
8e74712
Merge branch 'main' into chore/bevfusion/update_rois
SamratThapa120 Sep 8, 2025
a038999
fixed lr scaler
SamratThapa120 Sep 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 8 additions & 3 deletions projects/BEVFusion/bevfusion/transforms_3d.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,12 @@ def sample_augmentation(self, results):
H, W = results["ori_shape"]
fH, fW = self.final_dim
if self.is_train:
resize = np.random.uniform(*self.resize_lim)
if isinstance(self.resize_lim, (int, float)):
aspect_ratio = min(fH / H, fW / W)
resize = np.random.uniform(aspect_ratio - self.resize_lim, aspect_ratio + self.resize_lim)
else:
resize = np.random.uniform(*self.resize_lim)

resize_dims = (int(W * resize), int(H * resize))
newW, newH = resize_dims
crop_h = int((1 - np.random.uniform(*self.bot_pct_lim)) * newH) - fH
Expand All @@ -35,7 +40,7 @@ def sample_augmentation(self, results):
flip = True
rotate = np.random.uniform(*self.rot_lim)
else:
resize = np.mean(self.resize_lim)
resize = min(fH / H, fW / W)
resize_dims = (int(W * resize), int(H * resize))
newW, newH = resize_dims
crop_h = int((1 - np.mean(self.bot_pct_lim)) * newH) - fH
Expand All @@ -52,7 +57,7 @@ def img_transform(self, img, rotation, translation, resize, resize_dims, crop, f
img = img.crop(crop)
if flip:
img = img.transpose(method=Image.FLIP_LEFT_RIGHT)
img = img.rotate(rotate)
img = img.rotate(rotate, resample=Image.BICUBIC) # Default rotation introduces artifacts.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would be nice if you have examples showing artifacts with the default rotation


# post-homography transformation
rotation *= resize
Expand Down
143 changes: 48 additions & 95 deletions ...ond_secfpn_2xb2_t4offline_no_intensity.py → ..._offline_voxel_second_secfpn_4xb8_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,9 +8,10 @@

# user setting
data_root = "data/t4dataset/"
info_directory_path = "info/user_name/"
train_gpu_size = 2
info_directory_path = "info/username/"
train_gpu_size = 4
train_batch_size = 2
test_batch_size = 2
val_interval = 5
max_epochs = 30
backend_args = None
Expand All @@ -19,43 +20,45 @@
point_cloud_range = [-122.4, -122.4, -3.0, 122.4, 122.4, 5.0]
voxel_size = [0.075, 0.075, 0.2]
grid_size = [3264, 3264, 41]

eval_class_range = {
"car": 121,
"truck": 121,
"bus": 121,
"bicycle": 121,
"pedestrian": 121,
"car": 120,
"truck": 120,
"bus": 120,
"bicycle": 120,
"pedestrian": 120,
}

# model parameter
input_modality = dict(use_lidar=True, use_camera=True)
point_load_dim = 5 # x, y, z, intensity, ring_id
point_use_dim = 5
point_intensity_dim = 3
sweeps_num = 1
max_num_points = 10
max_voxels = [120000, 160000]
num_proposals = 500
image_size = [256, 704]
lidar_sweep_dims = [0, 1, 2, 4]
num_workers = 1
sweeps_num = 1
image_size = [576, 864] # height, width
num_workers = 32
lidar_sweep_dims = [0, 1, 2, 4] # x, y, z, time_lag
lidar_feature_dims = 4

model = dict(
type="BEVFusion",
data_preprocessor=dict(
type="Det3DDataPreprocessor",
pad_size_divisor=32,
voxelize_cfg=dict(
max_num_points=max_num_points,
point_cloud_range=point_cloud_range,
voxel_size=voxel_size,
point_cloud_range=point_cloud_range,
max_voxels=max_voxels,
deterministic=True,
voxelize_reduce=True,
),
type="Det3DDataPreprocessor",
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
bgr_to_rgb=False,
),
pts_voxel_encoder=dict(type="HardSimpleVFE", num_features=4),
pts_middle_encoder=dict(in_channels=4, sparse_shape=grid_size),
pts_middle_encoder=dict(sparse_shape=grid_size, in_channels=lidar_feature_dims),
img_backbone=dict(
type="mmdet.SwinTransformer",
embed_dims=96,
Expand Down Expand Up @@ -92,16 +95,11 @@
in_channels=256,
out_channels=80,
image_size=image_size,
feature_size=[32, 88],
# xbound=[-54.0, 54.0, 0.3],
# ybound=[-54.0, 54.0, 0.3],
# xbound=[-122.4, 122.4, 0.68],
# ybound=[-122.4, 122.4, 0.68],
feature_size=[72, 108],
xbound=[-122.4, 122.4, 0.3],
ybound=[-122.4, 122.4, 0.3],
zbound=[-10.0, 10.0, 20.0],
# dbound=[1.0, 60.0, 0.5],
dbound=[1.0, 166.2, 1.4],
dbound=[1.0, 134, 1.4],
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any reason we change it to 134, I am thinking we should make the depth and bin size even smaller, and make sure it's evenly divided by the bin size?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bin size: 1.4 could be a little big too large

downsample=2,
),
fusion_layer=dict(type="ConvFuser", in_channels=[80, 256], out_channels=256),
Expand All @@ -115,6 +113,7 @@
code_weights=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.2, 0.2],
),
test_cfg=dict(
dataset="t4datasets",
grid_size=grid_size,
voxel_size=voxel_size[0:2],
pc_range=point_cloud_range[0:2],
Expand All @@ -124,49 +123,13 @@
voxel_size=voxel_size[0:2],
),
),
# Lidar pipeline
pts_voxel_encoder=dict(num_features=lidar_feature_dims),
)

# TODO: support object sample
# db_sampler = dict(
# data_root=data_root,
# info_path=data_root +'nuscenes_dbinfos_train.pkl',
# rate=1.0,
# prepare=dict(
# filter_by_difficulty=[-1],
# filter_by_min_points=dict(
# car=5,
# truck=5,
# bus=5,
# trailer=5,
# construction_vehicle=5,
# traffic_cone=5,
# barrier=5,
# motorcycle=5,
# bicycle=5,
# pedestrian=5)),
# classes=class_names,
# sample_groups=dict(
# car=2,
# truck=3,
# construction_vehicle=7,
# bus=4,
# trailer=6,
# barrier=2,
# motorcycle=6,
# bicycle=6,
# pedestrian=2,
# traffic_cone=2),
# points_loader=dict(
# type='LoadPointsFromFile',
# coord_type='LIDAR',
# load_dim=5,
# use_dim=[0, 1, 2, 3, 4],
# backend_args=backend_args))

train_pipeline = [
dict(
type="BEVLoadMultiViewImageFromFiles",
data_root=data_root,
to_float32=True,
color_type="color",
backend_args=backend_args,
Expand All @@ -175,39 +138,31 @@
type="LoadPointsFromFile",
coord_type="LIDAR",
load_dim=point_load_dim,
use_dim=point_use_dim,
use_dim=point_load_dim,
backend_args=backend_args,
),
# TODO: add feature
# dict(
# type="IntensityNorm",
# alpha=10.0,
# intensity_dim=point_intensity_dim,
# div_factor=255.0,
# ),
dict(
type="LoadPointsFromMultiSweeps",
sweeps_num=sweeps_num,
load_dim=5,
load_dim=point_load_dim,
use_dim=lidar_sweep_dims,
pad_empty_sweeps=True,
remove_close=True,
backend_args=backend_args,
test_mode=False,
),
dict(type="LoadAnnotations3D", with_bbox_3d=True, with_label_3d=True, with_attr_label=False),
# TODO: support object sample
# dict(type='ObjectSample', db_sampler=db_sampler),
dict(
type="ImageAug3D",
final_dim=image_size,
resize_lim=[0.38, 0.55],
resize_lim=0.02,
bot_pct_lim=[0.0, 0.0],
rot_lim=[-5.4, 5.4],
rand_flip=True,
is_train=True,
),
dict(
type="GlobalRotScaleTrans",
type="BEVFusionGlobalRotScaleTrans",
rot_range=[-1.571, 1.571],
scale_ratio_range=[0.8, 1.2],
translation_std=[1.0, 1.0, 0.2],
Expand Down Expand Up @@ -259,23 +214,21 @@
test_pipeline = [
dict(
type="BEVLoadMultiViewImageFromFiles",
data_root=data_root,
to_float32=True,
color_type="color",
backend_args=backend_args,
test_mode=True,
),
dict(
type="LoadPointsFromFile",
coord_type="LIDAR",
load_dim=5,
use_dim=5,
load_dim=point_load_dim,
use_dim=point_load_dim,
backend_args=backend_args,
),
dict(
type="LoadPointsFromMultiSweeps",
sweeps_num=sweeps_num,
load_dim=5,
load_dim=point_load_dim,
use_dim=lidar_sweep_dims,
pad_empty_sweeps=True,
remove_close=True,
Expand All @@ -285,7 +238,7 @@
dict(
type="ImageAug3D",
final_dim=image_size,
resize_lim=[0.48, 0.48],
resize_lim=0.02,
bot_pct_lim=[0.0, 0.0],
rot_lim=[0.0, 0.0],
rand_flip=False,
Expand Down Expand Up @@ -313,27 +266,31 @@
),
]

filter_cfg = dict(filter_frames_with_missing_image=True)

train_dataloader = dict(
batch_size=train_batch_size,
num_workers=num_workers,
persistent_workers=True,
sampler=dict(type="DefaultSampler", shuffle=True),
dataset=dict(
type=_base_.dataset_type,
pipeline=train_pipeline,
modality=input_modality,
backend_args=backend_args,
data_root=data_root,
ann_file=info_directory_path + _base_.info_train_file_name,
pipeline=train_pipeline,
metainfo=_base_.metainfo,
class_names=_base_.class_names,
modality=input_modality,
data_prefix=_base_.data_prefix,
test_mode=False,
data_prefix=_base_.data_prefix,
box_type_3d="LiDAR",
backend_args=backend_args,
filter_cfg=filter_cfg,
),
)

val_dataloader = dict(
batch_size=2,
batch_size=test_batch_size,
num_workers=num_workers,
persistent_workers=True,
sampler=dict(type="DefaultSampler", shuffle=False),
Expand All @@ -351,8 +308,9 @@
backend_args=backend_args,
),
)

test_dataloader = dict(
batch_size=2,
batch_size=test_batch_size,
num_workers=num_workers,
persistent_workers=True,
sampler=dict(type="DefaultSampler", shuffle=False),
Expand Down Expand Up @@ -392,6 +350,7 @@
name_mapping=_base_.name_mapping,
eval_class_range=eval_class_range,
filter_attributes=_base_.filter_attributes,
save_csv=True,
)

# learning rate
Expand Down Expand Up @@ -456,14 +415,8 @@
clip_grad=dict(max_norm=35, norm_type=2),
)

# Default setting for scaling LR automatically
# - `enable` means enable scaling LR automatically
# or not by default.
# - `base_batch_size` = (8 GPUs) x (4 samples per GPU).
# auto_scale_lr = dict(enable=False, base_batch_size=32)
auto_scale_lr = dict(enable=False, base_batch_size=train_gpu_size * train_batch_size)
auto_scale_lr = dict(enable=True, base_batch_size=4)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Keep the comment, and any reason we set it to True? Does it show any significant improvement/stability for training?


# Only set if the number of train_gpu_size more than 1
if train_gpu_size > 1:
sync_bn = "torch"

randomness = dict(seed=0, diff_rank_seed=False, deterministic=True)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any reason we delete it? I believe we need to keep it for reproducibility

Loading