Given lin big brother written code that can be characterized in a single graph, not a batch, so Guaren under modified .
Here the results of comparison to see if the same. [Note that feature key frame ]
I found lin did not directly resize the input model? ? ? This model must resize the process, only a fixed bar. Let's look at the input model.
[Later discovered why not use the batch, and if the batch, then it input_shape must be fixed, the following is a detailed look at how to reshape the model, because it is the pb model may also be able to modify the specific idea is: the model the input section into a simple calculation model outside pb, pb is passed directly to the post-resize, so you can batch a] detailed in this blog post , say no more, and certainly with the results of different h5.
Another related problem can be added QQ group discussion, no micro-channel group
QQ group: 868 373 192
Voice video image depth - study group
