If I understand correctly, the egraphs end with acyclic in tensat dataset are egraphs that are pruned ahead of time. However, I found there are cycles existing in some of these acyclic egraphs.
For example, in bert_acyclic.json, one can easily verify that enodes 553__1, 475__1, 7680__0 form a cycle. And there the number of cycles in some acyclic tensat egraphs is not small. According to my analysis, here is the number of cycles for each of the tensat acyclic egraphs:
bert_acyclic: 915
nasneta_acyclic: 30
nasrnn_acyclic: 10
vgg_acyclic: 0
resnet50_acyclic: 0
I am wondering if this is because my understanding is not correct or any other potential bugs?
If I understand correctly, the egraphs end with
acyclicin tensat dataset are egraphs that are pruned ahead of time. However, I found there are cycles existing in some of these acyclic egraphs.For example, in
bert_acyclic.json, one can easily verify that enodes553__1,475__1,7680__0form a cycle. And there the number of cycles in some acyclic tensat egraphs is not small. According to my analysis, here is the number of cycles for each of the tensat acyclic egraphs:I am wondering if this is because my understanding is not correct or any other potential bugs?