ShapeAdaptor issue about shape-adaptor HOT 4 CLOSED

Hi,

The code is correct.

The ShapeAdaptor.current_dim_true is defined in each network forward function. For example in VGG network, it's in line 174 - 176. This current_dim_true indicates the true output dimension dynamically updated for each iteration without the memory penalty, which is required to compute the corresponding penalty term to make it grow no more than the required limit_dim.

As such, this current_dim_true will only be available if we have at least one network pass using the forward function, i.e. running predict = model(data) at least once.

If you make sure you did not modify any part of the code, just try python model_training.py --gpu ID to test. It should give no error.

Hope it helps.

Sk.

from shape-adaptor.

At the very first iteration, which current_limit_dim has not yet defined, python would directly give True if one of the conditions in or is true.

For example, if you try True or a even if a is not defined, it would output a True.

from shape-adaptor.

ok, Maybe I add some bugs while modified the code to distributed training

from shape-adaptor.

Ok. Make sure the model has at least one pass on the training data.. Should be an easy fix.

from shape-adaptor.