Hi, thanks for your wonderful work! Where can I find the code about Adaptive Filtering

<a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/us

Some questions about your code about dsl HOT 1 CLOSED

chenbinghui1 commented on July 28, 2024

Some questions about your code

from dsl.

Comments (1)

chenbinghui1 commented on July 28, 2024

@qdd1234

The implementation of AF is composed of several parts:
(1) ignore-label assignment:

DSL/mmdet/datasets/semicoco.py

Line 232 in 72338c3

if 'scores' in data.keys() and self.thres is not None:

Because the DSL-style data annotation file contains bboxes, tags, scores information, when we load the annos we can assign each instance into FG, Ignorable Region or BG according to the scores and thresholds. Scores and thresholds \tao_{2} will be updated by mmdet/runner/hooks/unlabel_pred_hook.py.

(2) loss computation:

DSL/mmdet/models/dense_heads/fcos_head.py

Line 209 in 72338c3

ig_labels = None

Ignored region will be weighted by 0 when computing losses.

(3) Adaptive \tao_{2} update:

DSL/mmdet/runner/hooks/unlabel_pred_hook.py

Line 295 in 72338c3

    
           def adathres(rank, flag, filename, id2cat, cat2id, input_list, input_path, settings):

After each epoch, we will update \tao_{2} for each class, and the thresholds for each classes will be saved in json file "adathres.json" which is defined in config_file(

DSL/configs/fcos_semi/RLA_r50_caffe_mslonger_tricks_0.Xdata_unlabel_dynamic_lw_nofuse_iterlabel_si-soft_singlestage.py

Line 141 in 72338c3

thres="adathres.json",

)

(4) Instance score update:

DSL/mmdet/runner/hooks/unlabel_pred_hook.py

Lines 455 to 469 in 72338c3

    
           def after_train_iter(self, runner): 
        
               if self.interval_mode == "iteration" and runner.iter+1>=(self.start_point*runner.iter_tol_epoch)+1: 
        
                   if not self.every_n_iters_with_startpoint(runner, self.interval, self.start_point): 
        
                       return 
        
                   if self.iter_fuse_flag == True: 
        
                       self.after_train_iter_func(runner) 
        
                   else: 
        
                       # the first fuse will be the same as epoch manner 
        
                       self.after_train_epoch_iter(runner) 
        
                       self.iter_fuse_flag=True 
        
                       next(runner.ITER) 
        
                       # handling the preload mechanism in torch.dataloader 
        
                       # runner.ITER is an independent iterator for iter-style pseudo-label update 
        
                       for i in range(int(self.preload_num)): 
        
                           next(runner.ITER)

Here, we will update the annotation file of each image in iter-manner(or epoch-manner, you can try by yourself, we also provide other kinds of update, e.g. fuse with pseudo-labels in history like in ISMT. )

from dsl.

Recommend Projects

Some questions about your code about dsl HOT 1 CLOSED

Comments (1)

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	def after_train_iter(self, runner):
	if self.interval_mode == "iteration" and runner.iter+1>=(self.start_point*runner.iter_tol_epoch)+1:
	if not self.every_n_iters_with_startpoint(runner, self.interval, self.start_point):
	return
	if self.iter_fuse_flag == True:
	self.after_train_iter_func(runner)
	else:
	# the first fuse will be the same as epoch manner
	self.after_train_epoch_iter(runner)
	self.iter_fuse_flag=True
	next(runner.ITER)
	# handling the preload mechanism in torch.dataloader
	# runner.ITER is an independent iterator for iter-style pseudo-label update
	for i in range(int(self.preload_num)):
	next(runner.ITER)