f1.py
No OneTemporary
Actions

Subscribers

None

File Metadata

	import torch

	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

	def PRFromHistograms(hPos,hNeg):
	print("hPos",hPos,"hNeg",hNeg)
	positives=hPos.sum()
	negatives=hNeg.sum()
	print("positives, negatives", positives, negatives)
	truepositives =reverse(hPos.clone()).long().cumsum(dim=0)
	falsepositives=reverse(hNeg.clone()).long().cumsum(dim=0)
	predpositives=torch.add(truepositives,falsepositives)
	#protect against zero division
	predpositives[predpositives==0]=1
	precision=torch.Tensor.div_(truepositives.float(),predpositives.float())
	recall =torch.Tensor.div (truepositives.float(),positives.float())
	precision[precision<=0]=1e-12
	recall [recall <=0]=1e-12
	print("precision,recall",precision,recall)
	return precision,recall

	def PRFromOutGt(outps,targs,nbins=10000):
	hPos=torch.zeros(nbins)
	hNeg=torch.zeros(nbins)
	for o,t in zip(outps,targs):
	pos=o[lbl==1]
	neg=t[lbl==0]
	hPos+=torch.from_numpy(pos.astype(np.float32)).histc(nbins,0,1)
	hNeg+=torch.from_numpy(neg.astype(np.float32)).histc(nbins,0,1)
	precision,recall=f1.PRFromHistograms(hPos,hNeg)
	f1s=f1.F1FromPR(precision,recall)
	f=f1s.max()
	return f

	def F1FromPR(p,r):
	suminv=torch.pow(p,-1)+torch.pow(r,-1)
	f1s=torch.pow(suminv,-1).mul(2)
	print("f1s",f1s)
	return f1s