how to converge？

[jiachen0212]

hello，I use my own data to train the network, but but it unable to converge. How can I adjust the parameters? thanks a lot！

[yangxue0827]

ANCHOR_RATIOS，BASE_ANCHOR_SIZE_LIST，LR，MINIBATCH_SIZE etc.

[jiachen0212]

ok thanks....

[DetectionIIT]

@ yangxue0827
I meet the same problem, and the data is dota. The params are as follows, Is there any question? thanks so much!!!

NET_NAME = 'resnet_v1_101'
VERSION = 'v8'
#CLASS_NUM = 1
CLASS_NUM = 16
LEVEL = ['P2', 'P3', 'P4', 'P5', 'P6']
BASE_ANCHOR_SIZE_LIST = [32, 64, 128, 256, 512]
STRIDE = [4, 8, 16, 32, 64]
ANCHOR_SCALES = [1.]
ANCHOR_RATIOS = [1 / 3., 1., 3.0]
SCALE_FACTORS = [10., 10., 5., 5., 5.]
OUTPUT_STRIDE = 16
SHORT_SIDE_LEN = 600
#DATASET_NAME = 'ship'
DATASET_NAME = 'dota'

BATCH_SIZE = 1 #add by gwf. must be 1
WEIGHT_DECAY = {'resnet_v1_50': 0.0001, 'resnet_v1_101': 0.0001}
EPSILON = 1e-5
MOMENTUM = 0.9
#MAX_ITERATION = 40000
DECAY_STEP = [100000, 200000] # 90000, 120000 #have no use
MAX_ITERATION = 300000
GPU_GROUP = "1" #add by gwf. use gpu1
LR = 0.0003