loggerF1.py
No OneTemporary
Actions

Subscribers

None

File Metadata

	import os
	import torch
	import numpy as np
	from .f1 import PRFromHistograms, F1FromPR

	def reverse(t):
	idx = [i for i in range(t.size(0)-1, -1, -1)]
	idx = torch.LongTensor(idx)
	it = t.index_select(0, idx)
	return it

	class LoggerF1:

	def __init__(self,logdir,fname,transform,nBins=10000,saveBest=False):
	self.log_dir=logdir
	self.name=fname
	self.log_file=os.path.join(self.log_dir,"logF1"+self.name+".txt")
	text_file = open(self.log_file, "w")
	text_file.close()
	self.preproc=transform
	self.nBins=nBins
	self.hPos=torch.zeros(self.nBins)
	self.hNeg=torch.zeros(self.nBins)
	self.bestF1=0
	self.saveBest=saveBest

	def add(self,l,output,target):
	o,t=self.preproc(output.cpu().data,target.cpu().data)
	pos=o[t==1]
	neg=o[t==0]
	self.hPos+=pos.histc(self.nBins,0,1)
	self.hNeg+=neg.histc(self.nBins,0,1)

	def logEpoch(self,net):
	precision,recall=PRFromHistograms(self.hPos,self.hNeg)
	f1s=F1FromPR(precision,recall)
	f=f1s.max()
	text_file=open(self.log_file, "a")
	text_file.write('{}\n'.format(f))
	text_file.close()
	if self.saveBest and f > self.bestF1:
	self.bestF1=f
	torch.save({'state_dict': net.state_dict()},
	os.path.join(self.log_dir,
	'net_'+self.name+'_bestF1.pth'))


	self.hPos.zero_()
	self.hNeg.zero_()