[
https://issues.apache.org/jira/browse/TIKA-2672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16534966#comment-16534966
]
Thejan Wijesinghe commented on TIKA-2672:
-----------------------------------------
[[email protected]] sorry for the delay, so the dl4j update works fine for
vgg, but not with inception v3. Spent almost two days trying to fix this.
First, I updated the keras 1-inception v3 model with keras 2-inception v3, but
couldn't get this working, then I assumed this should be a problem related to
the dynamic input shapes(None shapes) of the model, so changed the entire model
to static shapes starting from the input layer's change to (None, None, 3) to
(299, 299, 3) but still no luck, I didn't see any breaking changes of their api
either, so filed an issue today
[https://github.com/deeplearning4j/deeplearning4j/issues/5831] , they have
ignored the inception v3 test in their tests too. Anyway if they are targeting
to support full backward capability, the model import api should work with a
keras-1 model as well. Let's wait for their instructions.
> Upgrade dl4j to 1.0.0-beta
> --------------------------
>
> Key: TIKA-2672
> URL: https://issues.apache.org/jira/browse/TIKA-2672
> Project: Tika
> Issue Type: Task
> Reporter: Tim Allison
> Priority: Major
> Attachments: TIKA-2672.patch
>
>
> Let's try to upgrade dl4j. I think I got us most of the way there, but I got
> this error when reading the json config file. Can someone with more
> knowledge of layer specs help ([~thammegowda], perhaps :))?
> {noformat}
> org.deeplearning4j.exception.DL4JInvalidConfigException: Invalid
> configuration for layer (idx=-1, name=convolution2d_2, type=ConvolutionLayer)
> for width dimension: Invalid input configuration for kernel width. Require 0
> < kW <= inWidth + 2*padW; got (kW=3, inWidth=1, padW=0)
> Input type = InputTypeConvolutional(h=149,w=1,c=32), kernel = [3, 3], strides
> = [1, 1], padding = [0, 0], layer size (output channels) = 32, convolution
> mode = Truncate
> {noformat}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
