Add flop counter hook for rnn, gru and lstm #38
Conversation
|
Hi!. Thanks for that contribution. RNNs support is really missing in this project. |
|
If anything, just let me know. Glad to help. If there is something wrong with the RNN flop computation, please do not hesitate to tell me, I will be glad to know it, since I am using it for my own purposes. |
…eal flops by input or hidden size
sovrasov
left a comment
sovrasov
left a comment
There was a problem hiding this comment.
Now the hook seems to be OK. Let's merge it and then I'm going to update the readme and resolve the shape inconsistency problem between RNNs and CNNs (LNH vs NCHW): currently without using
the input_constructor argument the first dimension of the input is forced to 1.
|
I have taken into account that RNNs were built as NLH, contrary to pytorch standard scheme; I mean, change it if you would like to. It is stated under the header of the function. If you want me to do something, do not doubt to ask :) |
|
It seems to me this code is N-L transition agnostic, so here's nothing to change. Now I have to fix some common things related to the whole project and I need some time to figure it out. |
2 participants
Hello,
First of all: very nice repo and application. I am user.
I am using now RNNs, so I thought that I could expand the functionality for those. I have added a hook and a function for computing the flops. I think the params are automatically added by your already made computations and functionalities.
I have checked its functionality, but I would be more than glad if you would take a look at it.
Thank you very much!