Performance of evaluator

[tasbolat1]

Hi @arsalan-mousavian,
I have made some tests on both uploaded models (old version, i call it v1 and ACRONYM) for evaluator model. In my test, i wanted to know how good the evaluator perform for on detecting positive grasps. Using json files within splits folder, I decided evaluate train/test accuracy for positive grasps on box and cylinder categories. For each model/split I run tests three times because data loader has stochastic behavior, thus i provide standard deviation too. The results is shown below:

split	Model	box accuracy	cylinder accuracy
train	v1	0.82 (0.22)	0.82 (0.21)
test	v1	0.74 (0.24)	0.72 (0.27)
train	ACRONYM	0.66 (0.28)	0.66 (0.27)
test	ACRONYM	0.54 (0.25)	0.54 (0.25)

Generally I found these results to be poor taking into account 0.5 random choice:

1. There are substantial overfit between train and test splits
2. ACRONYM is performing worse than v1 model

Can you please confirm that this results are in line with models? May be i am doing something wrong? Thanks in advance