BowieHsu
commented
1 year ago @yeyan00 您好，您还有更多的log和超参能共享一下吗，比较重要的参数是生成伪标签的conf和nms阈值
Open yeyan00 opened 1 year ago
BowieHsu
commented
1 year ago @yeyan00 您好，您还有更多的log和超参能共享一下吗，比较重要的参数是生成伪标签的conf和nms阈值
yeyan00
commented
1 year ago @yeyan00 您好，您还有更多的log和超参能共享一下吗，比较重要的参数是生成伪标签的conf和nms阈值
非常感谢您的回复！
config和log如下： 训练参数
# EfficientTeacher by Alibaba Cloud
project: 'yolov5_ssod_pw'
adam: False
epochs: 100
weights: 'exp11/weights/last.pt'
prune_finetune: False
linear_lr: True
find_unused_parameters: True
hyp:
lr0: 0.01
hsv_h: 0.015
hsv_s: 0.7
hsv_v: 0.4
lrf: 1.0
scale: 0.9
burn_epochs: 200
no_aug_epochs: 0
mosaic: 1.0
# mixup: 0.1
warmup_epochs: 3
Model:
depth_multiple: 1.00 # model depth multiple
width_multiple: 1.00 # layer channel multiple
Backbone:
name: 'YoloV5'
activation: 'SiLU'
Neck:
name: 'YoloV5'
in_channels: [256, 512, 1024]
out_channels: [256, 512, 1024]
activation: 'SiLU'
Head:
name: 'YoloV5'
activation: 'SiLU'
anchors: [[10,13, 16,30, 33,23],[30,61, 62,45, 59,119],[116,90, 156,198, 373,326]] # P5/32]
Loss:
type: 'ComputeLoss'
cls: 0.3
obj: 0.7
anchor_t: 4.0
Dataset:
data_name: 'coco'
train: datasets/test_target/train.txt
val: datasets/test_target/val.txt
test: datasets/test_target/val.txt
target: datasets/target.txt
nc: 7 # number of classes
np: 0 #number of keypoints
names: [ '1', '2', '3', '4', '5', '6', '7', ]
img_size: 1024
batch_size: 8 #12
SSOD:
train_domain: True
nms_conf_thres: 0.1
nms_iou_thres: 0.65
teacher_loss_weight: 1.0
cls_loss_weight: 0.3
box_loss_weight: 0.05
obj_loss_weight: 0.7
loss_type: 'ComputeStudentMatchLoss'
ignore_thres_low: 0.1
ignore_thres_high: 0.6
uncertain_aug: True
use_ota: False
multi_label: False
ignore_obj: False
pseudo_label_with_obj: True
pseudo_label_with_bbox: True
pseudo_label_with_cls: False
with_da_loss: False
da_loss_weights: 0.01
epoch_adaptor: True
resample_high_percent: 0.25
resample_low_percent: 0.99
ema_rate: 0.999
cosine_ema: True
imitate_teacher: False
# dynamic_thres: True
ssod_hyp:
with_gt: False
mosaic: 1.0
cutout: 0.5
autoaugment: 0.5
scale: 0.8
degrees: 0.0
shear: 0.0log:
epoch, train/box_loss, train/obj_loss, train/cls_loss, metrics/precision, metrics/recall, metrics/mAP_0.5,metrics/mAP_0.5:0.95, val/box_loss, val/obj_loss, val/cls_loss, x/lr0, x/lr1, x/lr2
200, 0.058321, 0.0075394, 0.010123, 0.77295, 0.57089, 0.65443, 0.45817, 0, 0, 0, 0.01, 0.01, 0.01
201, 0.052145, 0.0069274, 0.007829, 0.73603, 0.57944, 0.64247, 0.45159, 0, 0, 0, 0.01, 0.01, 0.01
202, 0.048028, 0.0064725, 0.0061186, 0.72618, 0.54796, 0.61716, 0.43428, 0, 0, 0, 0.01, 0.01, 0.01
203, 0.043944, 0.00602, 0.0047832, 0.72813, 0.52685, 0.59519, 0.41158, 0, 0, 0, 0.01, 0.01, 0.01
204, 0.042173, 0.0057919, 0.0043807, 0.75121, 0.50088, 0.5751, 0.39154, 0, 0, 0, 0.01, 0.01, 0.01
205, 0.041191, 0.0056821, 0.004147, 0.72235, 0.50613, 0.56577, 0.38433, 0, 0, 0, 0.01, 0.01, 0.01
206, 0.040881, 0.005655, 0.0040669, 0.72657, 0.5016, 0.56242, 0.37758, 0, 0, 0, 0.01, 0.01, 0.01
207, 0.039792, 0.0054416, 0.0036513, 0.72811, 0.50103, 0.56049, 0.37467, 0, 0, 0, 0.01, 0.01, 0.01
208, 0.040218, 0.0055046, 0.0038934, 0.68189, 0.49388, 0.55154, 0.37384, 0, 0, 0, 0.01, 0.01, 0.01
209, 0.040424, 0.0055795, 0.0039208, 0.6868, 0.49531, 0.55343, 0.37651, 0, 0, 0, 0.01, 0.01, 0.01
210, 0.040183, 0.0055007, 0.0038468, 0.67679, 0.50268, 0.55736, 0.37618, 0, 0, 0, 0.01, 0.01, 0.01
211, 0.039495, 0.005447, 0.0036891, 0.67071, 0.50672, 0.55611, 0.37446, 0, 0, 0, 0.01, 0.01, 0.01
212, 0.040163, 0.0054785, 0.0038868, 0.69708, 0.49645, 0.55594, 0.37297, 0, 0, 0, 0.01, 0.01, 0.01
213, 0.040442, 0.0055599, 0.0040067, 0.71779, 0.49104, 0.55733, 0.37369, 0, 0, 0, 0.01, 0.01, 0.01
214, 0.040356, 0.0055748, 0.0040357, 0.70968, 0.49747, 0.55785, 0.3758, 0, 0, 0, 0.01, 0.01, 0.01
215, 0.040113, 0.0055368, 0.0038601, 0.73359, 0.48886, 0.55911, 0.3779, 0, 0, 0, 0.01, 0.01, 0.01
216, 0.040519, 0.0056348, 0.0040296, 0.73027, 0.49079, 0.55974, 0.37739, 0, 0, 0, 0.01, 0.01, 0.01
217, 0.04211, 0.0058181, 0.0046134, 0.72598, 0.4941, 0.56208, 0.37913, 0, 0, 0, 0.01, 0.01, 0.01
218, 0.040778, 0.0056085, 0.0040758, 0.78336, 0.49196, 0.57009, 0.3828, 0, 0, 0, 0.01, 0.01, 0.01
219, 0.040072, 0.0055178, 0.0038117, 0.77906, 0.49113, 0.57438, 0.38569, 0, 0, 0, 0.01, 0.01, 0.01
220, 0.039904, 0.0054627, 0.0037906, 0.77396, 0.4903, 0.57428, 0.38615, 0, 0, 0, 0.01, 0.01, 0.01
221, 0.039558, 0.0054479, 0.0037251, 0.77501, 0.5084, 0.58562, 0.39149, 0, 0, 0, 0.01, 0.01, 0.01
222, 0.039841, 0.0054796, 0.0038705, 0.78104, 0.50777, 0.58867, 0.39407, 0, 0, 0, 0.01, 0.01, 0.01
223, 0.040341, 0.0055374, 0.0040509, 0.68206, 0.55447, 0.5923, 0.39558, 0, 0, 0, 0.01, 0.01, 0.01
224, 0.040295, 0.0056236, 0.0040332, 0.68204, 0.55051, 0.59192, 0.3941, 0, 0, 0, 0.01, 0.01, 0.01
225, 0.040649, 0.0056158, 0.0040314, 0.69214, 0.5489, 0.59185, 0.39529, 0, 0, 0, 0.01, 0.01, 0.01
226, 0.042779, 0.0059315, 0.0047266, 0.69578, 0.55081, 0.59398, 0.39638, 0, 0, 0, 0.01, 0.01, 0.01
227, 0.041858, 0.0057899, 0.0044368, 0.68412, 0.56147, 0.59354, 0.39755, 0, 0, 0, 0.01, 0.01, 0.01
228, 0.040571, 0.0055551, 0.0041417, 0.6965, 0.55951, 0.59272, 0.39671, 0, 0, 0, 0.01, 0.01, 0.01
229, 0.039785, 0.0055068, 0.0038227, 0.76667, 0.52193, 0.59231, 0.3948, 0, 0, 0, 0.01, 0.01, 0.01
230, 0.03998, 0.0055113, 0.0039404, 0.76032, 0.52615, 0.59029, 0.39385, 0, 0, 0, 0.01, 0.01, 0.01
231, 0.04042, 0.0055823, 0.0039816, 0.73844, 0.53282, 0.59177, 0.39466, 0, 0, 0, 0.01, 0.01, 0.01
232, 0.040161, 0.0055292, 0.0039988, 0.68458, 0.55917, 0.59314, 0.39495, 0, 0, 0, 0.01, 0.01, 0.01
233, 0.041172, 0.0057276, 0.0043228, 0.77819, 0.51305, 0.5941, 0.39629, 0, 0, 0, 0.01, 0.01, 0.01
234, 0.039622, 0.0054349, 0.0037551, 0.76874, 0.51686, 0.59404, 0.39708, 0, 0, 0, 0.01, 0.01, 0.01
235, 0.039399, 0.0053831, 0.0036508, 0.7258, 0.52973, 0.5928, 0.39991, 0, 0, 0, 0.01, 0.01, 0.01
236, 0.039229, 0.0053484, 0.0035704, 0.74353, 0.5188, 0.59367,
数据集是自有数据集，图像大小1024*1024，训练集6K，验证集500,无标签数据40K 修改训练参数 image size:1024 batch size:12(supervised)，8(ssod) burn_epochs:200
先监督训练200个epoch,p/r value:0.76813/0.58473
然后用ssod训练，dataset_ssod无标签，走的是fake_iamge_label；
1.请问ssod训练过程还要注意哪些参数吗？