QuickNet model and flip_ratio metric do not work together about zoo HOT 3 CLOSED

@AdamHillier do you think we should change zookeeper.Field to have the same semantics as @property if initialized with a lambda function or used as a decorator, so that we get a new instance on every access to prevent issues like this? I came accross a similar issue last week as well when trying to use quantizers with learnable parameters.

Another solution would be to change quantizers here to normal @property instead of having them as fields, so that they are not cached on the class instance.

@itayalfia We'll take a closer look at this next week, since this is something we've come accross before and we should fix in a proper way. For now you can use #149, which should provide an intermediate fix for the issue.

from zoo.

Thanks for the detailed issue.

My suspicion is that one quantizer is created and reused for the entire model, and flip_ratio looks at the same quantizer with inputs of different shapes and fails because of this.

I noticed that in a different case too a few days ago. I think it is because we set the input quantizers like this so they are stored on the class which will fail since the metric has internal variables:

from zoo.

@AdamHillier do you think we should change zookeeper.Field to have the same semantics as @property if initialized with a lambda function or used as a decorator, so that we get a new instance on every access to prevent issues like this? I came accross a similar issue last week as well when trying to use quantizers with learnable parameters.

I've made a Zookeeper issue for this here: larq/zookeeper#134.

from zoo.