Re: [theano-users] diff is not differentiable

be converted to a > scan op but I don’t know if that’ll help with grad or just slow down things. > > On Jul 21, 2016, at 2:09 PM, Frédéric Bastien <frederic.bast...@gmail.com> > wrote: > > We haven't implemetned the grad for the diff op if the input isn't a > vector. > >

[theano-users] Software Developer Positions at MILA

://rh-carriere-dmz.synchro.umontreal.ca/psc/rhprpr9_car/EMPLOYEE/HRMS/c/HRS_HRAM.HRS_CE.GBL You need to fill your application there. Note, the web site is only in French. If you need help, contact me privately. Frédéric Bastien -- --- You received this message because you are subscribed

Re: [theano-users] Optimization Warning for g++

/Theano/Theano.git > > Any advice? > > Thanks, > > Craig > > On Wed, Jul 20, 2016 at 1:23 PM, Frédéric Bastien < > frederic.bast...@gmail.com> wrote: > >> Thanks. >> >> We updated that logic in the development version of Theano: >> >

Re: [theano-users] Sum vector according to indices

If the number of split is constant, it can be done. You split in 3 vector: v1, v2, v3 = vector[0:splits[0]], vector[splits[0]:splits[1]], ... s1=v1.sum() s2=v3.sum() theano.tensor.concatenate(s1,s2,s3) Fred On Wed, Jul 20, 2016 at 5:34 PM, Aditya Gudimella < aditya.gudime...@gmail.com> wrote:

Re: [theano-users] SVM with LeNet

I can't help you. But I remember that Pylearn2 allowed to train SVM on top of a Theano model. Maybe you can look there. Note, Pylearn2 don't have developer anymore. Fred On Fri, Jul 22, 2016 at 12:11 PM, Beatriz G. wrote: > Hi Everyone. I am trying to use SVM classifier in

Re: [theano-users] ERROR (theano.gof.opt): Optimization failure due to: constant_folding

We also observed some slowness on Windows compared to Linux when compiling C code. But not after that. Do you still see slowdown after the compilation of C code is done? This is cached, so if you launch the same job multiple time, the 2nd and following time won't compile c code. There was changes

Re: [theano-users] Return 0 with divide by zero

You can use the value you want. I still need to understand the result of your previous code. Fred Le 30 juil. 2016 01:00, "Varglur" a écrit : > c_without_inf = theano.tensor.set_subtensor(c[inf_idx], -999) > > > may I ask why should we set "inf" to -999 ?? > > > 在

Re: [theano-users] Limit in CNMeM (CUDA) parameter configuration

I suppose that your GPU is shared with the monitor? If so, this is normal that you can't raise the value too high. This % is the % of the total GPU memory. Not of the free memory. So if we take too much, then it don't let enought space for cudnn internal stuff. So mostly, don't use a too high

Re: [theano-users] Re: U-Net on theano?

Just a quick comment, don't pad. You can use the gradient of the convolution instead. It will be faster. Lasagne does it correctly (as probably other framework that have some special layer for this). Fred On Sat, Jul 16, 2016 at 3:43 PM, Robb Brown wrote: > Yes, you can

Re: [theano-users] Re: very slow compile time for GRU with attention

Update to Theano development version should speed up the compilation: http://www.deeplearning.net/software/theano/install.html#bleeding-edge-install-instructions Also, the first time you compile on a computer, it will take more time. This is normal as we compile C code. But we cache it. So the

Re: [theano-users] Error using floatX = float16 to save memory

We have a PR that upgrade some stuff about float16: https://github.com/Theano/Theano/pull/4764/files It probably fix your problem. Can you try it to confirm that you don't have a different problem? thanks Frédéric On Fri, Jul 15, 2016 at 4:55 AM, wrote: > ok I

Re: [theano-users] U-Net on theano?

I'm pretty sure it can be done in Theano. Fred On Wed, Jul 13, 2016 at 4:27 PM, wrote: > Hello Theano users, > > Would U-Net architecture be implementable in Theano, or are some specific > layer models missing? >

Re: [theano-users] Re: ANN: cuDNN Batch Normalization, dilated convolutions, CPU parallel convolution and pooling

yes. On Wed, Jul 13, 2016 at 3:27 PM, Doug wrote: > Does dnn_batch_normalization_train handle maintaining the population > statistics for mean/std? Since Epsilon is a parameter for the function I'm > hoping so. > > > On Wednesday, July 13, 2016 at 12:30:03 PM UTC-4, nouiz

Re: [theano-users] storing results from multiple calls to scan function without evaluation

Hi, There seem to be a problem in the way you use Theano. You should compile a Theano function by calling theano.function() and reuse that many times. If you don't do that, then it will be too slow. The time theano.function() take is significant and probably longer then just one call to the

Re: [theano-users] Equivalent of zip(..) in Theano

I think you can do it with advance indexing. I didn't try it, but I think something like this: x = tensor3() range = T.arange(3) x[range, range] If this one don't work try other variation advanced subtensor. If that fail you could do an alloc of zeros with call to set_subtensor with the

Re: [theano-users] Function Compiling Fail when several processes are running

You don't need to delete them. You could just add the device number in them and so they get reused. Reusing them help speed up the compilation. I just made a PR that would help having different compiledir for different device: https://github.com/Theano/Theano/pull/4831 If you use it with this

[theano-users] Bump Theano min version of NumPy, SciPy and Python

Hi, We want to bump NumPy minimal version to 1.9.1(from 1.7.1) and SciPy 0.14(from 0.11) and Python to 2.7 now in the dev version. Any strong objection? I nothing by Monday, we will do it at the start of next week. Fred -- --- You received this message because you are subscribed to the

Re: [theano-users] Invalid value ("'high'") for configuration variable "exception_verbosity". Valid options are ('low', 'high')

Don't quote string in the config file. This is the problem. Fred Le 16 août 2016 21:45, "Steve Leach" a écrit : > > Whilst chasing another mysterious error, (input needed to compute the > shape of the input is missing??) which recommended setting > exception_verbosity

Re: [theano-users] Convolution with sparse input image

What is the % of sparcity? On the GPU if it isn't high, it won't speed up the computation to do sparse computation. We have some block sparse operation in that module, maybe it can help you:

Re: [theano-users] Re: Error by evaluatyion the function for some parameters

Try test values. Check Theano doc, I'm offline and can't give the link. Fred Le 1 févr. 2017 10:08, "roman.foell via theano-users" < theano-users@googlegroups.com> a écrit : > When I debugged through File "theano\compile\function_module.py'' in line > 822 the error message appears in self.fn(),

Re: [theano-users] Re: Error by evaluation the function for some parameters

There is an error on your graph. Try the flag optimizer=None. If the error don't help you, the test values. Check Theano doc, I'm offline. Le mer. 1 févr. 2017 10:08, roman.foell via theano-users < theano-users@googlegroups.com> a écrit : > When I debugged through File

Re: [theano-users] Re: How to quantize theano variable to arbitrary resolution?

For float 16, 32 add 64, just car twice. From memory, Theano won't remove the first cast. If it does, you could disable and optimization. For the other, can you show me the Python code? Le jeu. 2 févr. 2017 12:39, Michael Klachko a écrit : > I'm also interested in how

Re: [theano-users] Keras new version cannot 'load_weights' of old model weights, help!!

This is a question for keras, not Theano. The problem code from change in Keras, not Theano so we can't help. Keras have a tendency to break compatibility. I don't know if there is a fix except using the old keras, write to them, maybe they can help you. Fred On Tue, Jan 31, 2017 at 9:03 AM,

Re: [theano-users] Hessian through a gamma function

THis is what I wrote there. I think theano-dev would be a better place to continue this discussion: The grad of gammaln use psi. But we didn't implement the grad of psi. You can implement the method grad() of the class Psi in the file theano/scalar/basic_scipy.py. This would make this case work.

Re: [theano-users] Re: Using Cuda8 with theano

On Mon, Feb 6, 2017 at 10:37 AM Frédéric Bastien <frederic.bast...@gmail.com> wrote: > Delete your Theano cache. You probably have it populated with module that > request cuda 7.5. Run: > > theano-cache purge > > otherwise, by default it is under ~/.theano > > Fred >

Re: [theano-users] Re: Info about types on new GPU backend

Hi, for the gpu, int are only supported in the new gpu back-end (device=cuda*). In the old back-end, they would end up being on CPU. This is why at many places it is told to not use int on the GPU. But it isn't true with the new back-end. For the reduction being slow, we didn't parallelize it on

Re: [theano-users] Slow theano.tensor.blas.Gemv on CPU

Theano is not about to link directly to BLAS. The simplest way to fix this is to use anaconda and install mkl-service. Make sure to use Theano dev version. We tested this recently and it work on Mac Linux and Windows. Fred Le 30 janv. 2017 20:56, "Raphael Shu" a écrit : Hi,

Re: [theano-users] GPU Acceleration

Try to modify the environment variable LIBRARY_PATH as the LD version. Fred Le mer. 1 févr. 2017 14:28, Lucian Willoughby a écrit : > I am trying to enable GPU acceleration on my Mac OSX: > > I have install CUDA toolkit > I have downloaded cuDNN and place the include

Re: [theano-users] What's the complexity of bincount or inc_subtensor in Theano on the GPU?

I don't recall the implementation of binvount and I'm offline. For inc_subtensor, we only modify the data the same number of time as the new data, y in the code bellow inc_subtensor(x[idx], y) On the GPU, we use the GPU feature AtomicAdd Le lun. 30 janv. 2017 07:51, Kiuhnm Mnhuik

Re: [theano-users] Re: Using Cuda8 with theano

Delete your Theano cache. You probably have it populated with module that request cuda 7.5. Run: theano-cache purge otherwise, by default it is under ~/.theano Fred On Sun, Feb 5, 2017 at 10:50 PM, Jayendra Parmar wrote: > defintely I can run the cuda samples > > $

Re: [theano-users] Re: Numpy has no attribute 'distutils' when `import theano`

THis PR fix it: https://github.com/Theano/Theano/pull/5497 We can't merge it now due to CI error, but it should be done in the next few hours. You can use that PR in the mean time. Fred On Mon, Feb 6, 2017 at 8:31 AM, Shir Gur wrote: > Same same > > looks like an old bug

Re: [theano-users] Info about types on new GPU backend

The new back-end support float32 and float64 (as uint* int* too). But gamers GPUs have very slow float64. It will work, but you won't have any speed up from the GPU unless you use float32. Fred On Mon, Feb 6, 2017 at 7:20 AM, Kiuhnm Mnhuik wrote: > I'm using Theano

Re: [theano-users] tensor.choose does not have gradient method

I replied to theano-dev. please, don't ask on both mailing list. It take more time from us. Fred On Mon, Jan 23, 2017 at 5:45 PM, Helmut Strey wrote: > I am working on implementing hidden-markov-models in pymc3 that is using > theano to implement the probabilistic

Re: [theano-users] Re: g++ undefined instruction error?

The "-march=native" is a good fix depending of the cause. There is 2 causes for this problems: 1) a bug in g++. Then this is a good work around. 2) your home directory is shared on many computer with different CPU. In some cases, our cache should recompile some of the files and sadly don't do it.

Re: [theano-users] Dimshuffle add

Hi, Theano will automatically add the dimshuffle in that case for you. We have a new tutorial on broadcasting that should help you: http://deeplearning.net/software/theano_versions/dev/tutorial/broadcasting.html If you still have questions after, don't hesitate. Fred On Thu, Jan 26, 2017 at

Re: [theano-users] Re: Could not initialize pygpu, support disabled

Hi, Theano OpenCL support is incomplete. Unless you want to help us make it work, I won't recommend that you use it. Fred Le 29 janv. 2017 19:09, a écrit : > were you able to solve this?? > I am having the same problem, please somebody help > > -- > > --- >

Re: [theano-users] Theano, dilation convolutions in 3D

That is the problem. You use the old interface. Use the new one and filter_dilation will work. We should not change the interface again, so this problem shouldn't happen again. The old interface was specific to one implementation, while the new one isn't. Fred On Wed, Jan 25, 2017 at 9:21 AM,

Re: [theano-users] How to Define A New Type that Allows Differentation Just as With Tensors?

I replied on stack overflow On Wed, Jan 25, 2017 at 3:25 PM, wrote: > Hi all, > > I followed the example of creating new operations and types in Python and > I am trying to create a Type that allows differentiation (building upon the > very basic "Double" type from the

Re: [theano-users] saving statistics for later inspection

The sharedvariable is the right way. The OP should take it as input and return a new value for it. At first, it should return new data. Not modify inplace the old data. Doing it inplace can be done with more work. You need to tell Theano about this. Keep this for when all work and of you see a

Re: [theano-users] pygpu.gpuarray.GpuArrayException

Hi, To be sure, you use the development version of Theano, not Theano 0.8.2? And you also installed the development version of libgpuarray? What are your Theano flag? Fred On Fri, Jan 27, 2017 at 6:21 AM, Tecx Nitrom wrote: > > I am trying to set theano to use gpu but

Re: [theano-users] How can I change subsets value in theano tensor impeded into loop?

That way of working will be super slow. Try to make your code "vectorized" by using advanced indexing when possible: https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#advanced-indexing On Tue, Jan 24, 2017 at 4:17 PM, Feras Almasri wrote: > layer_Fmaps of

Re: [theano-users] Why float is 12 bytes in torch?

The title name "torch" Are you on the right mailing list? This is Theano mailing list. a float always take only 4 bytes. But depending of how you count, there is overhead for the structure of the numpy.ndarray that surround it. There is also question of memory alignment. So depending of how you

Re: [theano-users] How can I change subsets value in theano tensor impeded into loop?

I don't undertand, what Theano don't support? We support advanced indexing. Fred On Wed, Jan 25, 2017 at 2:53 PM, Feras Almasri wrote: > thanks for mentioning this but theano function itself doesn't accept this > implementation. how can I implement it ? > > On Wednesday,

Re: [theano-users] New error after bleeding edge update.

Pygpu test don't test cudnn. Theano is not able to find cudnn, this is the problem. How did you install it? Check the doc to make sore to install it correctly. Fred Le lun. 20 févr. 2017 20:27, Ragav Venkatesan a écrit : > I installed 0.9rc with the new libgpu array and

Re: [theano-users] Why environment flags not working?

That line must be executed before the import in Theano. Otherwise it will be ignored.i suppose your change make Theano imported before this line. Fred Le mar. 21 févr. 2017 03:46, David Leon a écrit : > Previously I set Theano configuration variables by OS environment as

Re: [theano-users] Custom allocator

No. On the GPU, we have our own allocator on top of cuda malloc. On the CPU, we use NumPy allocator. So if you find a way to change numpy allocator, then you will be good. There was discussion of letting this possible. I don't know if this was finished or not. We allow to keep all allocated

Re: [theano-users] The new batchnorm correct usage.

There was a bug on the new GPU back-end related to BN. It was fixed probably last Friday. Make sure to update Theano again to the dev version to be sure to have the fix. Fred On Mon, Feb 20, 2017 at 3:21 AM Ragav Venkatesan wrote: > Is this a correct usage of

Re: [theano-users] theano-nose seems to be broken

The problem would be to udpate doctest. But changing to an different software stack could also fix it. Fred On Wed, Feb 15, 2017 at 5:20 AM Ramana Subramanyam wrote: > Hi, > My cron job has updated the master to the upstream master and this > morning, the *theano-nose

Re: [theano-users] Memory issue about theano

On the GPU, you need memory for the parameters, but also for temporary memory. But it seem Theano isn't even able to set the parameters on the GPU. Is tensorflow running at the same time? If so, it reserve all the GPU memory, so Theano won't have any left. Can you give the full error? It

Re: [theano-users] PyGPU installation error...

What is your NVIDIA driver? There is some driver that are buggy that make us raise an error. It will show up as failure in the tests. see: https://github.com/Theano/Theano/issues/5530 On Sat, Feb 18, 2017 at 3:45 PM André L wrote: > (myVE) andrelopes@X

Re: [theano-users] pygpu.gpuarray.GpuArrayException

You need to install clblas or clblast to have a blas for opencl. Without that, it is missing many operation. Note, in Theano, we are missing many/some operation in OpenCL. We don'T have time to work in the short term on that. So if your goal is to use opencl with Theano, you will probably need

Re: [theano-users] Re: Speed difference between backends

The new back-end do compilation during the first call to the Theano function. Can you redo your timing while ignoring the first call to the Theano function? If that don't fix the problem, can you enable Theano profiler and send us both profile? Use those Theano flags: profile=True

Re: [theano-users] IPC device pointer to shared variable

The shared variable won't be moved off the GPU. But we could have it contain a new object. We do not garanty to reuse that memory. Maybe it was replaced by a view to some other memory and that this ptr isn't one returned by cuda malloc. I don't know if that is a requirement or not. There is this

Re: [theano-users] New error after bleeding edge update.

This flag should not change where the computation happen. So if it don't work on GPU, this is a different problem. I'll put are merged for the 0.9rc. So it should be done today. We should get new conda package, wait for them. They should work well and have all dependencies. Fred Le dim. 19 févr.

Re: [theano-users] Re: Speed difference between backends

I don't understand what you mean by "after the training starts". You must not include the first call to the Theano function in your timing. Did you do that? Fred On Fri, Feb 17, 2017 at 7:53 AM Adam Becker wrote: > I noticed this problem late 2016 while investigating

Re: [theano-users] Speed difference between backends

Adam, I ran your script and I still see the slow down of scan with the new back-end. But I think we can consider this example as having few work to do at each iteration. Ozan, can you share the code of what is done at each iteration of the scan? With the size. This is to know if Adam code show

Re: [theano-users] NotImplementedError: In() instances and tuple inputs trigger the old semantics...

What is the running code? Did you update theano as the last message suggest? What is your Theano version? Make sure to use Theano 0.9beta1 or 0.9rc1. If the code come from the Deep learning tutorial, update it. Old combination of DLT and Theano could give this. Current Theano version don't raise

Re: [theano-users] GPU spends most of the time in GpuAlloc

as broadcastable. This would allow you to do what you want efficently without having alloc in the graph. This is a very good use of broadcasting. Frédéric On Wed, Feb 15, 2017 at 12:16 PM Frédéric Bastien < frederic.bast...@gmail.com> wrote: > tile generate alloc. To help you about the bro

Re: [theano-users] Disabling optimizations

There is no easy way to disable opt on part of the graph. I think you can use x*(x<=a).astype('float32') The cast will get merged by Theano with the comparisons and as all inputs and outputs of that fused op are in float32, Theano should move that to the GPU. Fred Le sam. 11 févr. 2017 06:51,

Re: [theano-users] THEANO LAPACK/BLAS

On Windows we fixed something yesterday that could help the error you point. So update Theano to the dev version again. Fred On Wed, Feb 15, 2017 at 7:54 PM 'Ahmad Albab' via theano-users < theano-users@googlegroups.com> wrote: > The following error happened while compiling the node',

Re: [theano-users] New error after bleeding edge update.

Strange. Can you do this to make sure one dependency is installed? conda install mkl-service If that don't fix it, try this PR: https://github.com/Theano/Theano/pull/5544 A temporary work around is to use this THeano flag: blas.ldflags= Fred On Thu, Feb 16, 2017 at 2:11 AM Ragav Venkatesan

Re: [theano-users] Re: Multiple matrix product in theano

of other suggestions that I could try ? > If for-loop approach is doomed because of the max recursion issue, is > there perhaps a way I could rewrite my code such that scan becomes > reasonably faster ? > > > > On 16 February 2017 at 22:22, Frédéric Bastien <frederic...@gma

Re: [theano-users] Re: Multiple matrix product in theano

Why do you tell it should only work well if len(A) is small? batched_dot should work well for a big len(A). Fred On Wed, Feb 15, 2017 at 10:50 AM Adam Becker wrote: > Besides that, use subtensor and batched_dot can further vectorize things. > > > A =

Re: [theano-users] Re: Info about types on new GPU backend

I have no idea if what you propose would work well. You can make a new OP that use pycuda for the computation. We do that for our fft op in the new back-end: https://github.com/Theano/Theano/blob/master/theano/gpuarray/fft.py On Sat, Feb 11, 2017 at 6:45 AM Kiuhnm Mnhuik

Re: [theano-users] Storage map footprint of a node

Theano profile could help you. Theano flag profile=True and profile_memory=True will print at process exit time memory information. If you call debugprint, on a theano function while the profiling was enabled and the function was called before calling debugprint. It will print timing

Re: [theano-users] theano-nose seems to be broken

Which version of Python do you use? I have the impression tha your version of doctest is too old for the version of Python you use. On Wed, Feb 15, 2017 at 5:20 AM Ramana Subramanyam wrote: > Hi, > My cron job has updated the master to the upstream master and this >

Re: [theano-users] THEANO LAPACK/BLAS

Hi, Install anaconda and make sure to use Theano dev version. It will tell you to install the conda package mkl-service. This will make Theano happy and fast. Le mer. 15 févr. 2017 11:59, 'Ahmad Albab' via theano-users < theano-users@googlegroups.com> a écrit : > i'm new to python, currently im

Re: [theano-users] Downgrade Theano from 0.9.0dev5 to 0.9.0dev4

What isthe error that you have with the new back-end? It it better to fix it then revert back to previous version. with dev5 version, you can still use the old back-end if you wish. On Tue, Jan 24, 2017 at 12:18 AM, nabarun goswami wrote: > Hi, > > I am unable to use

Re: [theano-users] Downgrade Theano from 0.9.0dev5 to 0.9.0dev4

; Nabarun > > On Tue, 24 Jan 2017 at 11:04 PM, Frédéric Bastien < > frederic.bast...@gmail.com> wrote: > >> What isthe error that you have with the new back-end? >> >> It it better to fix it then revert back to previous version. >> >> with dev5 versi

Re: [theano-users] nvidia TX1 Embedded Board not support

Which version of Theano do you use? Can you confirm you use Theano development version? 0.9.devN? Fred On Mon, Jan 16, 2017 at 3:11 AM, sang hyun Yoon wrote: > can i help.. > Thank you.. > > WARNING (theano.gof.cmodule): OPTIMIZATION WARNING: Theano was not able to > find the

Re: [theano-users] Scan: Return tuple of length 1?

Thanks for the idea. At a few places, we allow to pass return_list=True to force the return a of list. I added that option to scan in that PR: https://github.com/Theano/Theano/pull/5420 Why we don't always return a list? It is because the best interface depend of what you want. If you use

Re: [theano-users] Return a tuple of length-1 from scan?

I answered in the other email on this subject. On Mon, Jan 16, 2017 at 9:57 AM, Peter O'Connor wrote: > Hi all. > > There's some really unintuitive/inconsistent behaviour with theano's scan > and I'm wondering how I should work around it. > > Take the following test,

Re: [theano-users] Getting NaN as cost in the pretraining stage of a dbn

Quick reply offline. You need to assist the hyper parameter. The second later cost is very close to 0 . So maybe you don't need a third layer? If you want to investigate more, in Theano web site, there is a page on finding the cause of nan. Le 19 janv. 2017 05:05, "Sayantan Ghosh"

Re: [theano-users] Re: Nan After updating to bleeding edge

forum 2016-08-21 9:45 GMT-03:00 Frédéric Bastien <frederic.bast...@gmail.com>: > If you used git to get Theano, try this command > > git revlog > > Or > > git reflog > > If you used pip, I don't know. Try the devN one after each other > > Le 20 août 2016 21

Re: [theano-users] Re: Nan After updating to bleeding edge

a high learning rate. Was working fine before i update, but now i dont know which version of theano i was using. Probably dev1. Do you know if i can get it? Im desperate, my research has halted because of it. 2016-08-20 22:26 GMT-03:00 Frédéric Bastien <frederic.bast...@gmail.com>: > Try

Re: [theano-users] Shifting the location of tensor3 elements based on an offset vector

You could use advanced indexing where you create the index you want to use and they will be copied over. I think this would be more performant then using scan. On Wed, Aug 17, 2016 at 11:44 AM, Franck Dernoncourt < franck.dernonco...@gmail.com> wrote: > I couldn't find any dedicated function

Re: [theano-users] theano-test failing on nvidia-docker image

Some test fail as your GPU isn't supported by cudnn. But others are failing as one dependency is missing: ImportError: No module named nose_parameterized Is this the docker image you are talking about? https://github.com/nouiz/Theano-Docker Which docker file did you used?

Re: [theano-users] Why does GpuDnnSoftmax needs 'bc01'?

Use the CPU softmax. Theano will move it to the GPU and cudnn softmax automatically if Theano can find cudnn: http://deeplearning.net/software/theano/library/tensor/nnet/nnet.html?highlight=softmax#theano.tensor.nnet.nnet.softmax Fred On Mon, Aug 22, 2016 at 3:40 PM, Shir Gur

Re: [theano-users] Re: Theano 0.9.0rc1 on Windows 10 x64 and Ubuntu 16.04 have same pygpu errors

It is optional. But to get the fastest convolution, it is better to have it. If you don't use convolution then you don't need it. Free Le lun. 27 févr. 2017 04:36, 侠贵族 a écrit : > I did not install cuDnn. Dose the new cuda backend must need it? > > -- > > --- > You

Re: [theano-users] Re: theano on Mac OS -- clang7.3.0: cannot compile functions (null pointer comparison)

To force on the CPU, you need two Theano flags. device=CPU And force_device=True This should make Theano not use the GPU at all. Fred Le dim. 26 févr. 2017 20:37, Thomas Ploetz a écrit : > Here is the full output: > > >gcc --version > Configured with:

Re: [theano-users] ReLUs in Restricted Boltzmann Machines

Check the deep learning tutorial. They have rbm examples. Le dim. 26 févr. 2017 20:42, Russell Tsuchida a écrit : > Hi all, > > Using the code at the tutorial > as a starting point, I've > been able to implement an

Re: [theano-users] Re: I killed a Theano process using CNMeM, but the GPU RAM was released. Why, and how to release it?

`, it did not help (GPU 0 > is the one where the GPU memory didn't get released). > > On 23 August 2016 at 09:26, Frédéric Bastien <frederic.bast...@gmail.com> > wrote: > >> With nvidia-smi, I think you can reset a GPU. I never did it, but I know >> some cluster

Re: [theano-users] Re: I killed a Theano process using CNMeM, but the GPU RAM was released. Why, and how to release it?

If python is dead then Theano can't use GPU memory anymore. The small memory you see used is used by nvidia driver, not Theano. Fred Le 20 août 2016 08:56, "Franck Dernoncourt" a écrit : > Thanks, I ran `pkill -9 python`, there is no Python process running (I >

Re: [theano-users] Error using floatX = float16 to save memory

thanks. On Wed, Aug 24, 2016 at 5:16 AM, wrote: > Fred, > many thanks for your help. > > I reinstalled and updated anaconda. > I reinstalled theano and gpuarray/pygpu: > Theano==0.9.0.dev2 > pygpu==0.2.1 > > Then I tested a small 3D convnet using first: > floatX =

Re: [theano-users] Freeing memory after memory error

There is one easy think to try, call a_theano_function.free(): http://deeplearning.net/software/theano/library/compile/function.html?highlight=function#theano.compile.function_module.Function.free We don't free all the memory in that case. This should do that. But as Pascal said, we didn't test

Re: [theano-users] Avoiding memory duplication when using advanced indexing with repeated indices

interface with it from Theano. > > Regards, > Arvid Fahlström Myrman > > On Thursday, 1 September 2016 13:00:55 CEST Frédéric Bastien wrote: > > Currently it is not possible. But batchedGemm for Douglas allow this I > > think. If you use the GPU you could call batchedGemm directly a

Re: [theano-users] What exactly does theano's clone do to shared variables?

On Wed, Aug 31, 2016 at 1:47 PM, AA wrote: > Alright, so when calculating the derivative with respect to a shared > variable that has been replicated, it won't consider appearances of > the cloned shared variable when calculating the derivative, but the value > of the

Re: [theano-users] Scan not working

"numpy.dtype has the wrong size, try recompiling" is in the error. This is a problem related to numpy version and something Theano cache. Installing a different Theano version, deleting the cache. Changing numpy version can fix this. >From an comment on an issue, he reinstalled and this fixed

Re: [theano-users] Re: Does theano have a function to resize a tensor type (just like the function 'resize' function in cv2)

The simplest would be to wrap cv2 implementation in a Theano Op: http://deeplearning.net/software/theano/extending/ If you have questions, don't hesitate. I think it would be great to have this in Theano. Fred On Mon, Aug 29, 2016 at 10:24 PM, wrote: > OK,So sad ...

Re: [theano-users] What exactly does theano's clone do to shared variables?

Hi, When we clone a shared variable, it will create a new object. So if you compare them equal, it will be false. But they will represent the same value under the hood. This mean that if you change the value of one of cloned or the original shared variable. This change will be reflected in the

Re: [theano-users] Avoiding memory duplication when using advanced indexing with repeated indices

Currently it is not possible. But batchedGemm for Douglas allow this I think. If you use the GPU you could call batchedGemm directly and modify it to take as input y. You need to be a C programmer, no need for cuda programming. Are you interested? We don't have time to do it shortly. Fred Le 25

Re: [theano-users] Re: Issue with GTX 1080

I don't know how much slower it would be. Maybe nvcc.flags=-arch=sm_50 work. This would would be better then 30 if it work. I have never timed the slow down. Fred On Tue, Aug 30, 2016 at 11:45 PM, Florin Andrei wrote: > It is the new 2016 Nvidia Titan X Pascal > >

Re: [theano-users] Theano import error, via keras

Remove the comment on your Theano config file. Le 3 sept. 2016 12:06, "Nikky U" a écrit : > Hi All, > > can someone help me diagnose where I am going wrong. > > I installed anaconda, tensorflow, cuda, keras and theano. I can import > standard python packages like numpy,

Re: [theano-users] Re: CNMeM is disabled, CuDNN not available

If the copied destination is already in LD_LIBRARY_PATH, then you don't need that step. On Fri, Sep 9, 2016 at 10:51 AM, Naomi Fridman wrote: > > Hi Poornachandra Sandur, > > Thank you, great answer and super practical. > > But why to add the extracted directory to

Re: [theano-users] output of conv layer using stride

Theano don't have a good user interface for that. But you could use some internal function that help for that. Like calling op.infer_shape with the right parameter. Otherwise, I think it would be easier to use Lasagne or keras. Lasagne is a thin layer on top of Theano that provide that type of

Re: [theano-users] Error using floatX = float16 to save memory

On Fri, Oct 7, 2016 at 11:31 AM, Pascal Lamblin wrote: > On Fri, Oct 07, 2016, luca.wagner.0...@gmail.com wrote: > > Hi Fred, > > I did a test using: > > > > theano.tensor.nnet.conv3d2d import conv3d > > That's the old conv3d2d code, that should not be needed with

Re: [theano-users] Unsolvable Theano import error "No module named cPickle" and "No module named graph"

Uninstall and reinstall Theano. Uninstall many times. You could have multiple Theano version that conflict together Le 6 oct. 2016 17:11, "Rav" a écrit : > I'm not sure what you refer to. > I just run theano from where ever i want (jupyter, terminal, ide) > Some tips would be

Re: [theano-users] Error using floatX = float16 to save memory

For float16, always use device=cuda Not device=gpu. This could be your problem. Can you test that? thanks Fred On Tue, Oct 4, 2016 at 10:21 AM, wrote: > Hi Fred, > I tested the convnet using > > floatX= float32, > device=gpu > theano.tensor.nnet.conv3d2d.conv3d

Re: [theano-users] 'module' object has no attribute 'Subtensor'

n Theano récent. > > cheers, > Thomas > > > On 21 September 2016 at 16:53, Frédéric Bastien < > frederic.bast...@gmail.com> wrote: > >> If you use the same Theano version, it should work. Can you remove all >> Theano version. Try to import it and make su

Re: [theano-users] 'FromFunctionOp' object has no attribute 'grad'

I think scan need to compute all the grad internally, even if they don't get used. You could implement a grad method to the as_op that will add dummy op to the graph that could be removed later if they aren't needed. Check those 2 URLS:

  1   2   3   4   >