loss cannot drop

[Bangbanggetnithecar]

I changed the backbone of centernet to mobilenet. the loss didn't drop. I trained 500 epoches on pascalVOC dataset. the hm_loss stoped at 2.6. the wh_loss stoped at 9.6 after about 350 epoches and didn't changed.
我用了您的代码运行了一下，loss始终降不到特别低，一开始我在默认设置下运行了一下，hm_loss，wh_loss分别在350个epoch时停在了2.6和9.6并且不再降了，mAP只有0.09，我尝试改了一下wh_weight 和hm_weight 以及学习率等参数，mAP也只升到了0.16。
the opt setting is listed below
==> torch version: 1.4.0 ==> cudnn version: 7605 ==> Cmd: ['main.py', 'ctdet', '--dataset', 'pascal', '--exp_id', 'voc_mbv2_0919', '--save_all', '--arch', 'mbv2_10', '--head_conv', '24', '--lr', '5e-4', '--batch_size', '128', '--gpus', '0,1,2,3,4,5,6,7'] ==> Opt: K: 100 aggr_weight: 0.0 agnostic_ex: False arch: mbv2_10 aug_ddd: 0.5 aug_rot: 0 batch_size: 128 cat_spec_wh: False center_thresh: 0.1 chunk_sizes: [16, 16, 16, 16, 16, 16, 16, 16] data_dir: /home/xiashaobang/Documents/CenterNet/src/lib/../../data dataset: pascal debug: 0 debug_dir: /home/xiashaobang/Documents/CenterNet/src/lib/../../exp/ctdet/voc_mbv2_0919/debug debugger_theme: white demo: dense_hp: False dense_wh: False dep_weight: 1 dim_weight: 1 down_ratio: 4 eval_oracle_dep: False eval_oracle_hm: False eval_oracle_hmhp: False eval_oracle_hp_offset: False eval_oracle_kps: False eval_oracle_offset: False eval_oracle_wh: False exp_dir: /home/xiashaobang/Documents/CenterNet/src/lib/../../exp/ctdet exp_id: voc_mbv2_0919 fix_res: True flip: 0.5 flip_test: False gpus: [0, 1, 2, 3, 4, 5, 6, 7] gpus_str: 0,1,2,3,4,5,6,7 head_conv: 24 heads: {'hm': 20, 'wh': 2, 'reg': 2} hide_data_time: False hm_hp: True hm_hp_weight: 1 hm_weight: 1 hp_weight: 1 input_h: 512 input_res: 512 input_w: 512 keep_res: False kitti_split: 3dop load_model: lr: 0.0005 lr_step: [100, 300] master_batch_size: 16 mean: [[[0.485 0.456 0.406]]] metric: loss mse_loss: False nms: False no_color_aug: False norm_wh: False not_cuda_benchmark: False not_hm_hp: False not_prefetch_test: False not_rand_crop: False not_reg_bbox: False not_reg_hp_offset: False not_reg_offset: False num_classes: 20 num_epochs: 500 num_iters: -1 num_stacks: 1 num_workers: 4 off_weight: 1 output_h: 128 output_res: 128 output_w: 128 pad: 31 peak_thresh: 0.2 print_iter: 0 rect_mask: False reg_bbox: True reg_hp_offset: True reg_loss: l1 reg_offset: True resume: False root_dir: /home/xiashaobang/Documents/CenterNet/src/lib/../.. rot_weight: 1 rotate: 0 save_all: True save_dir: /home/xiashaobang/Documents/CenterNet/src/lib/../../exp/ctdet/voc_mbv2_0919 scale: 0.4 scores_thresh: 0.1 seed: 317 shift: 0.1 std: [[[0.229 0.224 0.225]]] task: ctdet test: False test_scales: [1.0] trainval: False val_intervals: 50 vis_thresh: 0.3 wh_weight: 0.1
Thank you for your excellent work

[zay95]

Same issue. The loss is still high after training 320 epoches. It seems that light backbones can't fit this model net well . Is there anything advice on light net with bbox and pose detection task？Thanks!