search

adatao / tensorspark

TensorFlow on Spark
297 stars 101 forks source link

Csmith #1

Closed illuzen closed 8 years ago

phvu commented 8 years ago

Good work @snakecharmer1024. To check if the parameter server is really working, I would recommend to try with a feedforward network of 1024-1024-1024. With this architecture on the whole MNIST dataset, you should get ~5% error rate on the test set after a few epochs. If it doesn't, then there is something wrong with the parameter server.

Using a straightforward feedforward architecture would eliminate any potential problem with CNN in tensorflow that you might be not aware of.

phvu commented 8 years ago

@ushnish okay so that's not on spark yet. You might need to implement the spark part of the thing. @snakecharmer1024 Since you both are on the same branch, please merge this and fix the issues in separated branches.