-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconfig.py
235 lines (201 loc) · 13.5 KB
/
config.py
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
#!/usr/bin/env python3
# -*- coding = utf-8 -*-
# @Time = 2021/12/24 下午12:56
# @Author = chenxb
# @FileName = config.py
# @Software = PyCharm
# @mail ：[email protected]
from pathlib import Path
import yaml
from easydict import EasyDict
cfg = EasyDict()
# model
cfg.model = EasyDict()
cfg.model.voxelnet = EasyDict()
cfg.model.voxelnet.voxel_generator = EasyDict()
cfg.model.voxelnet.voxel_generator.point_cloud_range = [0, -40, -3, 70.4, 40, 1]
cfg.model.voxelnet.voxel_generator.voxel_size = [0.2, 0.2, 0.4]
cfg.model.voxelnet.voxel_generator.max_number_of_points_per_voxel = 35
cfg.model.voxelnet.num_class = 1
cfg.model.voxelnet.lidar_input = False # just a flag can ignore it
cfg.model.voxelnet.voxel_feature_extractor = EasyDict()
cfg.model.voxelnet.voxel_feature_extractor.module_class_name = "VoxelFeatureExtractor"
cfg.model.voxelnet.voxel_feature_extractor.num_filters = [32, 128]
cfg.model.voxelnet.voxel_feature_extractor.with_distance = False
cfg.model.voxelnet.middle_feature_extractor = EasyDict()
cfg.model.voxelnet.middle_feature_extractor.module_class_name = "MiddleExtractor"
cfg.model.voxelnet.middle_feature_extractor.num_filters_down1 = [64]
cfg.model.voxelnet.middle_feature_extractor.num_filters_down2 = [64, 64]
cfg.model.voxelnet.rpn = EasyDict()
cfg.model.voxelnet.rpn.module_class_name = "RPN"
cfg.model.voxelnet.rpn.layer_nums = [3, 5, 5]
cfg.model.voxelnet.rpn.layer_strides = [2, 2, 2]
cfg.model.voxelnet.rpn.num_filters = [128, 128, 256]
cfg.model.voxelnet.rpn.upsample_strides = [1, 2, 4]
cfg.model.voxelnet.rpn.num_upsample_filters = [256, 256, 256]
cfg.model.voxelnet.rpn.use_groupnorm = False
cfg.model.voxelnet.rpn.num_groups = 32
cfg.model.voxelnet.loss = EasyDict()
cfg.model.voxelnet.loss.classification_weight = 1.0
cfg.model.voxelnet.loss.localization_weight = 2.0 # 2.0 get better performance
cfg.model.voxelnet.loss.classification_loss = EasyDict()
cfg.model.voxelnet.loss.classification_loss.classification_loss_type = "weighted_sigmoid" # "weighted_sigmoid" | "weighted_sigmoid_focal"
cfg.model.voxelnet.loss.classification_loss.weighted_sigmoid_focal = EasyDict()
cfg.model.voxelnet.loss.classification_loss.weighted_sigmoid_focal.alpha = 0.25
cfg.model.voxelnet.loss.classification_loss.weighted_sigmoid_focal.gamma = 2.0
cfg.model.voxelnet.loss.classification_loss.weighted_sigmoid_focal.anchorwise_output = True
cfg.model.voxelnet.loss.localization_loss = EasyDict()
cfg.model.voxelnet.loss.localization_loss.localization_loss_type = "weighted_smooth_l1"
cfg.model.voxelnet.loss.localization_loss.weighted_smooth_l1 = EasyDict()
cfg.model.voxelnet.loss.localization_loss.weighted_smooth_l1.sigma = 3.0
cfg.model.voxelnet.loss.localization_loss.weighted_smooth_l1.code_weight = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
# Outputs
cfg.model.voxelnet.use_sigmoid_score = True
cfg.model.voxelnet.encode_background_as_zeros = True
cfg.model.voxelnet.encode_rad_error_by_sin = True
cfg.model.voxelnet.use_direction_classifier = False # this can help for orientation benchmark.
cfg.model.voxelnet.direction_loss_weight = 0.2 # enough.
cfg.model.voxelnet.use_aux_classifier = False
# Loss
cfg.model.voxelnet.pos_class_weight = 1.5
cfg.model.voxelnet.neg_class_weight = 1.0
cfg.model.voxelnet.loss_norm_type = 1 #1 = NormByNumPositives 0 = LossNormType.NormByNumExamples, 2 = LossNormType.NormByNumPosNeg,
# Postprocess
cfg.model.voxelnet.post_center_limit_range = [0, -40, -3, 70.4, 40, 1]
cfg.model.voxelnet.use_rotate_nms = True
cfg.model.voxelnet.use_multi_class_nms = False
cfg.model.voxelnet.nms_pre_max_size = 1000
cfg.model.voxelnet.nms_post_max_size = 100
cfg.model.voxelnet.nms_score_threshold = 0.3
cfg.model.voxelnet.nms_iou_threshold = 0.01
cfg.model.voxelnet.use_bev = False
cfg.model.voxelnet.num_point_features = 4
cfg.model.voxelnet.without_reflectivity = False
cfg.model.voxelnet.box_coder = EasyDict()
cfg.model.voxelnet.box_coder.box_coder_type = "ground_box3d_coder"
cfg.model.voxelnet.box_coder.ground_box3d_coder = EasyDict()
cfg.model.voxelnet.box_coder.ground_box3d_coder.linear_dim = False
cfg.model.voxelnet.box_coder.ground_box3d_coder.encode_angle_vector = False
cfg.model.voxelnet.target_assigner = EasyDict()
cfg.model.voxelnet.target_assigner.anchor_generators = EasyDict()
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_type = "anchor_generator_range"
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range = EasyDict()
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range.sizes = [1.6, 3.9, 1.56] # wlh
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range.anchor_ranges = [0, -40, -1, 70.4, 40, -1] # carefully set z center -1
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range.rotations = [0, 1.57] # 0-pi/2
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range.matched_threshold = 0.6
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range.unmatched_threshold = 0.45
cfg.model.voxelnet.target_assigner.anchor_generators.anchor_generator_range.class_name = "Car"
cfg.model.voxelnet.target_assigner.sample_positive_fraction = -1
cfg.model.voxelnet.target_assigner.sample_size = 512
cfg.model.voxelnet.target_assigner.region_similarity_calculator = EasyDict()
cfg.model.voxelnet.target_assigner.region_similarity_calculator.region_similarity_type = "nearest_iou_similarity"
cfg.model.voxelnet.target_assigner.region_similarity_calculator.nearest_iou_similarity = EasyDict()
# train dataset
cfg.train_input_reader = EasyDict()
cfg.train_input_reader.record_file_path = "/root/paddlejob/workspace/train_data/datasets/kitti/kitti_train.log"
cfg.train_input_reader.class_names = ["Car"]
cfg.train_input_reader.max_num_epochs = 160
cfg.train_input_reader.batch_size = 2 # use 14.2 GB GPU memory when batch_size=3
cfg.train_input_reader.prefetch_size = 25
cfg.train_input_reader.max_number_of_voxels = 20000 # 6500 to support batchsize=2
cfg.train_input_reader.shuffle_points = True
cfg.train_input_reader.num_workers = 2
cfg.train_input_reader.groundtruth_localization_noise_std = [1.0, 1.0, 1.0] #1.0, 1.0, 1.0 in paper #1.0 1.0 0.5 in second
cfg.train_input_reader.groundtruth_rotation_uniform_noise = [-0.3141592654, 0.3141592654] # -pi/10-pi/10 in paper #-0.78539816, 0.78539816 in second
cfg.train_input_reader.global_rotation_uniform_noise = [-0.78539816, 0.78539816] # -pi/4-pi/4
cfg.train_input_reader.global_scaling_uniform_noise = [0.95, 1.05]
cfg.train_input_reader.global_random_rotation_range_per_object = [0, 0] # pi/4 ~ 3pi/4
cfg.train_input_reader.anchor_area_threshold = 1
cfg.train_input_reader.remove_points_after_sample = False
cfg.train_input_reader.groundtruth_points_drop_percentage = 0.0
cfg.train_input_reader.groundtruth_drop_max_keep_points = 15
cfg.train_input_reader.use_group_id = False
#the "group_id" is used for data augmentation of objects like tractor/trailer pair. we can use group_id to sample/rotate them together. This feature is deprecated, may be removed in future. just ignore it.
cfg.train_input_reader.database_sampler = EasyDict()
cfg.train_input_reader.database_sampler.database_info_path = "/root/paddlejob/workspace/train_data/datasets/kitti/kitti_dbinfos_train.pkl"
cfg.train_input_reader.database_sampler.sample_groups = EasyDict()
cfg.train_input_reader.database_sampler.sample_groups.name_to_max_num = EasyDict()
cfg.train_input_reader.database_sampler.sample_groups.name_to_max_num.key = "Car"
cfg.train_input_reader.database_sampler.sample_groups.name_to_max_num.value = 15
cfg.train_input_reader.database_sampler.database_prep_steps = EasyDict()
cfg.train_input_reader.database_sampler.database_prep_steps.database_preprocessing_step_type = ["filter_by_min_num_points","filter_by_difficulty"] # add prep step by list
cfg.train_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points = EasyDict()
cfg.train_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points.min_num_point_pairs = EasyDict()
cfg.train_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points.min_num_point_pairs.key = "Car"
cfg.train_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points.min_num_point_pairs.value = 5
cfg.train_input_reader.database_sampler.database_prep_steps.filter_by_difficulty = EasyDict()
cfg.train_input_reader.database_sampler.database_prep_steps.filter_by_difficulty.removed_difficulties = [-1]
cfg.train_input_reader.database_sampler.global_random_rotation_range_per_object = [0,0]
cfg.train_input_reader.database_sampler.rate = 1.0
cfg.train_input_reader.remove_unknown_examples = False
cfg.train_input_reader.remove_environment = False
cfg.train_input_reader.kitti_info_path = "/root/paddlejob/workspace/train_data/datasets/kitti/kitti_infos_train.pkl"
cfg.train_input_reader.kitti_root_path = "/root/paddlejob/workspace/train_data/datasets/kitti"
# eval dataset
cfg.eval_input_reader = EasyDict()
cfg.eval_input_reader.record_file_path = "/root/paddlejob/workspace/train_data/datasets/kitti/kitti_val.log"
cfg.eval_input_reader.class_names = ["Car"]
cfg.eval_input_reader.batch_size = 2
cfg.eval_input_reader.max_num_epochs = 160
cfg.eval_input_reader.prefetch_size = 25
cfg.eval_input_reader.max_number_of_voxels = 20000
cfg.eval_input_reader.shuffle_points = False
cfg.eval_input_reader.num_workers = 2
cfg.eval_input_reader.anchor_area_threshold = 1
cfg.eval_input_reader.remove_unknown_examples = False
cfg.eval_input_reader.remove_environment = False
cfg.eval_input_reader.kitti_info_path = "/root/paddlejob/workspace/train_data/datasets/kitti/kitti_infos_val.pkl"
cfg.eval_input_reader.kitti_root_path = "/root/paddlejob/workspace/train_data/datasets/kitti"
cfg.eval_input_reader.database_sampler = EasyDict()
cfg.eval_input_reader.database_sampler.database_info_path = "/root/paddlejob/workspace/train_data/datasets/kitti/kitti_dbinfos_train.pkl"
cfg.eval_input_reader.database_sampler.sample_groups = EasyDict()
cfg.eval_input_reader.database_sampler.sample_groups.name_to_max_num = EasyDict()
cfg.eval_input_reader.database_sampler.sample_groups.name_to_max_num.key = "Car"
cfg.eval_input_reader.database_sampler.sample_groups.name_to_max_num.value = 15
cfg.eval_input_reader.database_sampler.database_prep_steps = EasyDict()
cfg.eval_input_reader.database_sampler.database_prep_steps.database_preprocessing_step_type = ["filter_by_min_num_points","filter_by_difficulty"] # add prep step by list
cfg.eval_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points = EasyDict()
cfg.eval_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points.min_num_point_pairs = EasyDict()
cfg.eval_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points.min_num_point_pairs.key = "Car"
cfg.eval_input_reader.database_sampler.database_prep_steps.filter_by_min_num_points.min_num_point_pairs.value = 5
cfg.eval_input_reader.database_sampler.database_prep_steps.filter_by_difficulty = EasyDict()
cfg.eval_input_reader.database_sampler.global_random_rotation_range_per_object = [0,0]
cfg.eval_input_reader.database_sampler.database_prep_steps.filter_by_difficulty.removed_difficulties = [-1]
cfg.eval_input_reader.database_sampler.rate = 1.0
cfg.eval_input_reader.groundtruth_localization_noise_std = [1.0, 1.0, 1.0]
cfg.eval_input_reader.groundtruth_rotation_uniform_noise = [-0.3141592654, 0.3141592654]
cfg.eval_input_reader.global_rotation_uniform_noise = [-0.78539816, 0.78539816]
cfg.eval_input_reader.global_scaling_uniform_noise = [0.95, 1.05]
cfg.eval_input_reader.global_random_rotation_range_per_object = [0, 0] # pi/4 ~ 3pi/4
cfg.eval_input_reader.groundtruth_points_drop_percentage = 0.0
cfg.eval_input_reader.groundtruth_drop_max_keep_points = 15
cfg.eval_input_reader.use_group_id = False
cfg.eval_input_reader.remove_points_after_sample = False
# train config
cfg.train_config = EasyDict()
cfg.train_config.optimizer = EasyDict()
cfg.train_config.optimizer.optimizer_type = "momentum_optimizer" # SGD
cfg.train_config.optimizer.use_moving_average = False
cfg.train_config.optimizer.momentum_optimizer = EasyDict()
cfg.train_config.optimizer.momentum_optimizer.weight_decay = 0.0001
cfg.train_config.optimizer.momentum_optimizer.momentum_optimizer_value = 0.9
cfg.train_config.optimizer.momentum_optimizer.learning_rate = EasyDict()
cfg.train_config.optimizer.momentum_optimizer.learning_rate.learning_rate_type = "polynomial_decay_learning_rate"
cfg.train_config.optimizer.momentum_optimizer.learning_rate.polynomial_decay_learning_rate = EasyDict()
cfg.train_config.optimizer.momentum_optimizer.learning_rate.polynomial_decay_learning_rate.initial_learning_rate = 0.0015 # 0.01(bs=16) | 0.002(bs=3)
#0.01*3/16 ~=0.002 #0.0002 in second
cfg.train_config.optimizer.momentum_optimizer.learning_rate.polynomial_decay_learning_rate.decay_steps = 296960 # (bs=16) | 198080(bs=3)
cfg.train_config.optimizer.momentum_optimizer.learning_rate.polynomial_decay_learning_rate.decay_factor = 0.9
cfg.train_config.optimizer.momentum_optimizer.learning_rate.polynomial_decay_learning_rate.end_lr = 0
cfg.train_config.inter_op_parallelism_threads = 4
cfg.train_config.intra_op_parallelism_threads = 4
cfg.train_config.steps = 296960 # 37120: 232 (bs=16) * 160 | 198080: 1238(bs=3) *160 | 296960: 1856(bs=2)*160
cfg.train_config.steps_per_eval = 9280 # 232 *5 | 1238 * 5 |1856 *5
# steps = 296960 # 1857 * 160
# steps_per_eval = 9280 # 1856 * 5
cfg.train_config.save_checkpoints_secs = 3600 # one hour
cfg.train_config.save_summary_steps = 10
cfg.train_config.enable_mixed_precision = False # dont support now
cfg.train_config.loss_scale_factor = 512.0 #dont support now
cfg.train_config.clear_metrics_every_epoch = False