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.preproc = transform
	self.nBins = nBins
	self.saveBest = saveBest

	self.log_file = os.path.join(self.log_dir,"logF1"+self.name+".txt")
	text_file = open(self.log_file, "w")
	text_file.close()
	self.hPos = torch.zeros(self.nBins)
	self.hNeg = torch.zeros(self.nBins)
	self.bestF1 = 0


	def add(self, l, output, target, weight=None):
	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_()