search

shahsohil / DCC

This repository contains the source code and data for reproducing results of Deep Continuous Clustering paper
MIT License
208 stars 53 forks source link

Stopping threshold bug? #14

Closed LemonPi closed 5 years ago

LemonPi commented 5 years ago

The paper says we stop when change in assignment is below the stopping threshold, but the code implements this as:

            if change_in_assign > stopping_threshold:
                flag += 1
            if flag == 4:
                break

This is a bug right? It should be if change_in_assign < stopping_threshold: I'll fix this in my pull request if so.

LemonPi commented 5 years ago

Making the fix led to training for much more epochs with improvements on AMI and ACC. If the results in your paper was submitted with the same code with this bug in it, you might get better results with the fix!

image

shahsohil commented 5 years ago

The paper says we stop when change in assignment is below the stopping threshold, but the code implements this as:

            if change_in_assign > stopping_threshold:
                flag += 1
            if flag == 4:
                break

This is a bug right? It should be if change_in_assign < stopping_threshold: I'll fix this in my pull request if so.

No. This isn't a bug. This was put in place to let DCC run until a point it breaks down the clustering. That is why flag is updated couple of times before algorithm is halted. The original criterion is little conservative.

I believe it was added after the paper was done. Overall, we found just minor change in AMI / ACC outputs.