Skip to content

Latest commit

 

History

History
79 lines (72 loc) · 6.88 KB

HELP.md

File metadata and controls

79 lines (72 loc) · 6.88 KB
Raw

Parameter Configuration

This document contains detailed configuration for parameters using in training and testing.

  • 🔥 Update: [Adding Function] Parameter --norm_fun specify different normalizing function and --percentile specify percentile value for Pas or IVR.
  • 🔥 Update: [Repairing Function] Parameter --debug can be used to visualize the localization result.

🌳 Train

  • --snapshot_dir: prefixes: ../snapshots/ilsvrc/ for ILSVRC and ../snapshots/ for CUB, and the following format is recommended: Base_Mode_#ExpNumber, such as ../snapshots/ilsvrc/vgg16_sos+sa_v3_#1.
  • --log_dir: prefixes: ../log/ilsvrc/ for ILSVRC and ../log/ for CUB, and the following dir name is the same as snapshot_dir, such as ../log/ilsvrc/vgg16_sos+sa_v3_#1.
  • --lr: learning rate setting for backbone and classification-head, default 0.001 (cls_head will be ×10 when training, i.e. equals to 0.01).
  • --warmup: using warmup strategy or not, 'True' or 'False'. (See the end for more details)
  • --weight_decay: SGD optimizer used, default is 0.0005.
  • --load_finetune: 'True' or 'False' for loading fine-tuned classification model(includes cls_head) on ILSVRC. (To solve the problem of square activation)
  • --pretrained_model: fine-tuned model name, default is 'ilsvrc_epoch_20.pth.tar'. Here provide the fine-tuned model of VGG16.
  • --pretrained_model_dir: fine_tuned model path, default is '../snapshots/ilsvrc/vgg16_spa_#1'
  • --resume: resume training from checkpoint or not, 'False' or 'True', default is 'False'.
  • --restore_from: default is 'none', or need to specify the path of checkpoint.
  • --ram_th_bg : when using SCGv2 in SPA on ILSVRC, 0.1 is better, but 0.4 is better for SCGv1 on ILSVRC.
  • --ram_bg_fg_gap: when using SCGv2 in SPA on ILSVRC, 0.2 is better, but 0.1 is better for SCGv1 on ILSVRC.
  • --scg_fosc_th: when using SCGv2 in SPA on ILSVRC, 0.2 is better, but 0.1 is default setting in paper for SCGv1 on ILSVRC.
  • --scg_sosc_th: when using SCGv2 in SPA on ILSVRC, 1 is better, but 0.5 is default setting in paper for SCGv1 on ILSVRC.
  • --sos_lr: learning rate of sos_head, defalut is 5e-05.
  • --sos_gt_seg: using segmentation or not when generate pasedo ground-truth for predicting sos_map. 'True' or 'False', default is 'True'.
  • --sos_seg_method: when setting sos_gt_seg=True , sos_gt_seg specify the segmetation method, 'TC' and 'BC' is supported for double-thresholding and single-thresholding, respectively. Default is 'TC'.
  • --sos_loss_method: loss function for SOS branch, BCE loss function and MSE loss function are supported, default is 'BCE'.
  • --sos_fg_th: foreground segmentation threshold, default is 0.2 on CUB, it is not clear the value on ILSVRC.
  • --sos_bg_th: background segmentation threshold, default is 0.1 on CUB, it is not clear the value on ILSVRC. (Note: when sos_seg_method is set to 'BC', it is invalid)
  • --sos_start: SOS branch start training epoch, default is 0 on CUB and ILSVRC.
  • --sa_lr: learning rate of self-attention module, defalut is 0.005 on CUB, 0.001 on ILSVRC.
  • --sa_use_edge: using edge-encoding or not, 'True' or 'False', default is 'True'.
  • --sa_edge_weight: weight of edge-encoding when add with attention weight. Default is 1.
  • --sa_edge_stage: specifing the features from certain stages to calculate edge-encoding. Default is '4,5'.
  • --sa_start: epoch to add SA module for training, default is 20 on CUB, 3 on ILSVRC.
  • --sa_head: the number of multi-heads of SA module, default is 8.
  • --sa_neu_num: channel number of SA module. Default is 512 in VGG16.
  • --norm_fun: default is norm_min_max, imp for different normalize function, includes:
 'norm_max(_batch)' / 'norm_min_max(_batch)'(default) / 'norm_pas(_batch)' / 'norm_ivr(_batch)'
  • --watch_cam: save CAM, sos_map and gt_sos_map during training. The specific visualized image is set in exper/train_sst.py, and the default setting is 'True'.
  • --mode: current code support modes include:
'spa' / 'sos' / 'spa+sa' / 'sos+sa_v1' / 'sos+sa_v2' / 'sos+sa_v3' (default)

📚 Explanation of Warmup in the code

If you use warmup strategy, you need also to set it on lines 56~69 of ./exper/train_sst.py before each training. Among them, wp_period and wp_node respectively specify the period and training node of each warm-up phase, and wp_ps specifies the parameters that need to be executed in each warm-up phase.

For example, when wp_period = [2,2,2], wp_node = [0,3,6], wp_ps = [['sos_weight','sos_bias'], ['sa_weight','sa_bias'], op_params_list], that means the learning rate warmup will be performed on the two parameters'sos_weight' and 'sos_bias' in the 0th to 2nd epoch, and the learning rate warmup will be performed on the two parameters of 'sa_weight' and'sa_bias' in the 3rd to 5th epoch. op_params_list represents all parameters in the model, the learning rate will be warmup in the 6th to 8th epoch in this example. In addition, when warmup is performed on a certain set of parameters (such as'sos_weight' and'sos_bias'), the learning rate of all other parameters will be set to 0 to limit the updating of other modules in training .

All parameter field 'op_params_list' of the model sos+sa include:

'sos_weight' / 'sos_bias' / 'sa_weight' / 'sa_bias' / 'other_weight' / 'other_bias' / 'cls_weight' / 'cls_bias'

where 'other' refers to the parameters of the backbone (such as vgg16) ,'cls' refers to the parameters of the cls_head, 'sos' refers to the parameters of sos_head, and 'sa' refers to the parameters of the self-attention module.

🌳 Test (quick version by default here)

  • --debug_dir: a suffix is usually added to the test directory: _#t1 to distinguish different test results of the same model, such as ../debug/ilsvrc/vgg16_sos+sa_v3_wp_#3_t1 , ../debug/ilsvrc/vgg16_sos+sa_v3_wp_#3_t2 , etc.
  • --scg_version: specify the version of SCG algorithm, 'v2' is the default setting.
  • --sos_seg_method: need to align with the settings during training, default is 'TC'.
  • --sos_loss_method: need to align with the settings during training, default is 'BCE'.
  • --sa_use_edge: using edge-encoding or not, default is 'True'.
  • --sa_edge_weight: need to align with the settings during training, default is 1.
  • --sa_edge_stage: need to align with the settings during training, default is '4,5'.
  • --sa_head: need to align with the settings during training, default is 8.
  • --sa_neu_num: need to align with the settings during training, default is 512 for VGG16.
  • --mode: need to align with the settings during training, default is sos+sa_v3.
  • --debug: it is supported for quick version script now! (Repaired Date:2021.10.13, 16:00)