BRNN examles fails with Keras 1.0.1

Question

BRNN examles fails with Keras 1.0.1

sun9700 opened this issue 9 years ago · 16 comments

I tired seya with the lasted keras(1.0.1),
unfortunately， Seya fails with the following outputs:

/usr/bin/python2.7 /home/sun/seya/examples/imdb_brnn.py
Using TensorFlow backend.
Loading data...
20000 train sequences
5000 test sequences
Pad sequences (samples x time)
X_train shape: (20000, 100)
X_test shape: (5000, 100)
Build model...
Traceback (most recent call last):
File "/home/sun/seya/examples/imdb_brnn.py", line 59, in
model.add(brnn) # try using another Bidirectional RNN inside the Bidirectional RNN. Inception meets callback hell.
File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 139, in add
output_tensor = layer(self.outputs[0])
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 458, in call
self.build(input_shapes[0])
TypeError: build() takes exactly 1 argument (2 given)

Process finished with exit code 1

Answer 1 · 2016-04-25T18:05:54.000Z

seya is not ready for keras-1 yet, sorry!

Answer 2 · 2016-04-29T20:04:01.000Z

is this git still going to be maintained for newest versions of Keras?

Answer 3 · 2016-04-29T22:34:35.000Z

yeah... I'm just finishing something on reinforcement learning and then I'll come back to this.
But François told me adapting layers from keras 0.3.3 to keras 1 should be simple.
If you guys want to try to adapt something and make a PR I'll be glad to review the code

Answer 4 · 2016-04-30T13:44:13.000Z

would definitely love to get attention.py working on K1.0.1 :-/
i'm poking at it but no love yet -

Answer 5 · 2016-04-30T15:59:55.000Z

it seems that the main idea is to move the get_output to __call__. I hope this is not super urgent for you T_T sorry!

Answer 6 · 2016-04-30T16:55:44.000Z

I would also like to use the spatial transformer in a project of mine. Since this is your code Eder I think it would be faster if you update. We'll wait for you :)

Answer 7 · 2016-05-19T20:04:49.000Z

btw, I started translating seya to keras1. SpatialTransformer worked for me (with Theano): https://github.com/EderSantana/seya/blob/keras1/seya/layers/attention.py

Answer 8 · 2016-08-05T12:10:11.000Z

Bidirectional RNN also gives me the same error, Is the seya ready for keras-1 now? Thanks

Answer 9 · 2016-08-05T16:41:17.000Z

@xypan1232 not really we are translating one layer at a time when necessity arrives...
But I'd check keras itself for bidirectional RNNs. All the RNN layers now have go_backwards option

Answer 10 · 2016-10-31T02:33:04.000Z

@EderSantana What's the process you're going through to convert? I'm not familiar with the old Keras (started using it at 0.9 / 1.0 I believe) but want to use some of the non-implemented layers in this library. Can I make the changes and commit a PR?

Answer 11 · 2016-10-31T02:41:55.000Z

@StuartFarmer sure, make a PR to the keras1 branch

Answer 12 · 2016-10-31T10:41:21.000Z

@EderSantana Is there a specific commit I should look at for diffs to gain an understanding of what I need to do? I can probably poke around and figure it out, but if you could provide guidance, I could submit a better PR, I believe.

Answer 13 · 2016-11-02T16:40:34.000Z

sorry, I don't get what you mean... are you writing a new BRNN that works on keras1, right? as long as your new code works on keras-1 and runs the example script, it should be good.

Answer 14 · 2016-11-05T21:56:57.000Z

Right. I'm wondering if you know the specific changes between the Keras 0.3.3 API and the Keras 1.0 API that I should be looking for.

Answer 15 · 2016-11-08T05:56:29.000Z

oh, yeah! checkout the new functional API changes: https://github.com/fchollet/keras/blob/master/docs/templates/getting-started/functional-api-guide.md

Answer 16 · 2016-11-08T13:35:59.000Z

Awesome! Seems easy enough :) I'll submit some PRs soon.