Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
  • Sep 19 12:14

    raver119 on r119_sd

    less spam Signed-off-by: raver… (compare)

  • Sep 19 11:02
    AlexDBlack labeled #8248
  • Sep 19 11:02
    AlexDBlack labeled #8248
  • Sep 19 10:52
    ChambersLiu commented #8162
  • Sep 19 10:46

    raver119 on r119_sd

    one more test fixed Signed-off… (compare)

  • Sep 19 10:27

    raver119 on r119_sd

    one test fixed Signed-off-by: … (compare)

  • Sep 19 10:18

    raver119 on r119_sd

    Context iArgs now long Signed-… (compare)

  • Sep 19 09:21
    orausch opened #8248
  • Sep 19 08:50
    dyoann commented #8162
  • Sep 19 08:35
    ChambersLiu commented #8162
  • Sep 19 05:56
    liuzhengcast edited #8244
  • Sep 19 05:56
    liuzhengcast closed #8244
  • Sep 19 05:55
    liuzhengcast edited #8244
  • Sep 19 01:52

    AlexDBlack on ab_migrate_samediff

    (compare)

  • Sep 19 01:52

    AlexDBlack on master

    SameDiff ops (#8247) * update … (compare)

  • Sep 19 01:52
    AlexDBlack closed #8247
  • Sep 19 01:52
    AlexDBlack opened #8247
  • Sep 19 01:38

    AlexDBlack on ab_migrate_samediff

    update javadocs and a few metho… add PRelu op Signed-off-by: Ry… test and fixes Signed-off-by: … and 3 more (compare)

  • Sep 19 01:28

    AlexDBlack on master

    RL4J - AsyncTrainingListener (#… (compare)

  • Sep 19 01:28
    AlexDBlack closed #8072
dollarHome
@dollarHome

to get resnet on fp16 right now you'll probably have to cast params to fp16, and assign these params to your model

Thanks @raver119 Is there an example which shows how to do that? I'm interested in inference only using a pretrained model.

raver119
@raver119
sorry, nope.
but params of your nn is just INDArray
but you'd better file an issue
and we'll provide convenient method to do that
we'll need to do that for quantized types anyway
dollarHome
@dollarHome
ok will do. Thanks @raver119
so basically, you are saying, right now, after building the ComputationGraph, and initPretrained(), I would getParams() which is a INDArray, cast it to FP16, and reassign that to the ComputationGraph?
I was more thinking along the lines of TF where you can load a pretrained FP16 model and run inference on that.
dollarHome
@dollarHome
@raver119 @AlexDBlack filed deeplearning4j/deeplearning4j#7520
dollarHome
@dollarHome
@AlexDBlack another issue your way :) deeplearning4j/deeplearning4j#7521 I can help here with some guidance
Alex Black
@AlexDBlack
@dollarHome Thanks, both good ideas, I've replied to both
probably won't get to them in the short-term, unfortunately
though the former is a bit simpler...
dollarHome
@dollarHome
@AlexDBlack thanks, I saw the replies. Is there a way I can help here for #7521 ? If I can better understand what needs to be done, I can probably pitch in.
Alex Black
@AlexDBlack
@dollarHome I appreciate the offer, but unfortunately we're a bit blocked until this CUDA branch is merged: deeplearning4j/deeplearning4j#7095
if we start allowing different weight formats in conv layer, we need to either reimplement the conv layer in java (more trouble than it's worth, and a waste of time in the long run) or we need to switch to conv2d op now (which we already plan to do in the future). However, if we switch to conv2d op before CUDA branch is merged, that means it'll be CPU only (no CUDA, unless using cuDNN)
dollarHome
@dollarHome
@AlexDBlack just wondering, when is the tentative plan for CUDA branch merge? When you said "it'll be CPU only" do you mean, the optimization will be applicable to only CPU and CUDA wont be able to take advantage of it unless using cuDNN? or CUDA path wont work at all if this feature is implemented before merging the CUDA branch?
@saudet saw your comment on the github. Thanks for pointing out the code. this will help me get more familiar with the code path.
Alex Black
@AlexDBlack
@dollarHome it'll be merged when it's ready - but not before the next release. It's nearing completion, but could still be a few weeks yet, we'll see.
and when I mean CPU only, I mean even when using CUDA backend it'll execute on CPU (unless using cuDNN)
dollarHome
@dollarHome
@AlexDBlack Thanks for the clarification.
Lukasz Jastrzebski
@elyast
quick question I moved from 1.0.0-beta3 to 1.0.0-beta4 and I'm seeing following exception when doing distributed training using Spark: java.lang.IllegalStateException: Cannot convert to list: feature set rank must be in range 2 to 5 inclusive. Got shape: [784]
at org.nd4j.linalg.dataset.DataSet.getHelper(DataSet.java:826)
at org.nd4j.linalg.dataset.DataSet.asList(DataSet.java:796)
at org.deeplearning4j.spark.data.BatchAndExportDataSetsFunction.processList(BatchAndExportDataSetsFunction.java:133)
raver119
@raver119
@elyast show your source code as !gist please
gitterBot
@raver120
To use gist: paste your code/exception/large output log into https://gist.github.com, click 'Create Secret Gist' and paste URL link here
Alex Black
@AlexDBlack
hm... hypothetically that's something we missed in the rank 1/2 changes... should be [1,784] usually in getHelper (though that doesn't exclude possibility of wrong input rank in original dataset creation)
dollarHome
@dollarHome
Hi,I was looking for data on which companies use DL4J and for what usecases. All I found was https://skymind.ai/wiki/use-cases which gives generalized info on usecases and industries. Do we have any public or sharable info on which enterprise/company is using DL4J and for what usecases? Thanks!
ChrisN
@chrisvnicholson
@dollarHome Feel free to DM me about that. Some case studies we can talk about publicly and others we can't.
dionisole
@dionisole
@dionisole
Hi guys, just a quick question: For CnnSentenceDataSetIterator if i set: RemoveWord the program works fine but if i set: UseUnknownVector (beacause i want to use the unknown vector) the program exits with an error code saying that it cannot load word vectors
Any idea on this?
raver119
@raver119
there's probably just no unk word/vector in your model?
dionisole
@dionisole
There are many unknown words for sure as i perturb the words with unicode characters so there are plenty of them
@raver119
return new CnnSentenceDataSetIterator.Builder()
.sentenceProvider(sentenceProvider)
.wordVectors(wordVectors)
.minibatchSize(minibatchSize)
.maxSentenceLength(maxSentenceLength)
.useNormalizedWordVectors(false)
.unknownWordHandling(UseUnknownVector)
.build();
shouldn't this work?
public static final CnnSentenceDataSetIterator.UnknownWordHandling UseUnknownVector = CnnSentenceDataSetIterator.UnknownWordHandling.UseUnknownVector;
raver119
@raver119
once again.
there's probably just no unk word/vector in your model?
it means: there's no UNK word defined in your w2v model
not in your corpus
but in w2v model
dionisole
@dionisole
Ahh i thought that if a word is not found in W2v it just returns the closest vector found.
@raver119
so i have to define a no UNK word vector in my model so every time a word is not found the defined vector will be returned
Am i right?
raver119
@raver119
yep
dionisole
@dionisole
But what if i want the closest word vector found in the corpuus?
@raver119
raver119
@raver119
to get "closest" vector you should have original vector first :)
which isn't possible if you face out-of-vocabulary word
that's how w2v work
you might want to use fasttext instead, which is able to handle oov words
however, fasttext isn't ideal as well
Richard Corbishley
@rcorbish
Can someone give me a pointer to how I link a signed eclipse agreement to a pull? I have a pull request which fails checking.
Samuel Audet
@saudet
@rcorbish make sure your commits are signed with "git -s" as well