Missing SigmoidGrad function from tensorflow binaries

[verveguy]

I'm guessing that the relatively new SigmoidGrad function is missing from the prepared AWS builds of the tensorflow 0.9.0 packages. Current tensorflow builds are 0.11.0

My model dies with u'SigmoidGrad': KeyError

I found the instructions on creating new AWS tensorflow binary builds, and will attempt do so for the region I need (us-east-1)

However, you might want to rebuild the images soon!

[verveguy]

So I rebuilt one region (us-east-1) using the instructions from the dependencies repository and using tensorflow 0.11.0rc2.. Sadly, this seems to have blown out the size of the tensorflow package from 44.9MB to 57.4MB and this appears to now be tripping up the lambda function deploy code .. I'm getting:

Error: Request must be smaller than 69905067 bytes for the CreateFunction operation

However, I'm not sure I'm doing this right. I manually moved the new ZIP archive into the node_modules/aws-lambda-tensorflow/dependencies subdirectory. I cannot seem to figure out how this whole dep thing hangs together...

[verveguy]

OK. Figured out how the dependencies are getting downloaded. index.js has the secret. Please consider modifying this package to allow for the github repo that contains the tensorflow builds to be specified by the caller. That way, we can all build our own packages if needed and not rely on fixed URL references to contents we cannot modify.

[anandanand84]

Currently I did not work on any tensorflow project, can you give me some sample code that did not work in tensorflow 0.11.0rc2 which I can use to test once I build new image. I am planning to dynamically build new images based on version and deploy to s3 or something.

[verveguy]

Probably a good idea to move the tensorflow packages to s3. However, that's not going to fix the bigger problem: the size limit.

0.11.0rc2 is too big to be included in the lambda function payload. Unless there's a way to strip the code further, we might be barking up the wrong tree.

The other problem is that for large models, the reload time is significant.