[theano-users] Unknown import error

[Ramana Subramanyam]

Hi,
This is the traceback I'm getting when I tried to compute ReLU with bigger
values(As it was reported in OpenAI Gym that ReLU from tensor.nnet.relu
isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected

Regards,
Ramana

[Frédéric Bastien]

It would be great to know why they don't like that implementation.

I don't know why you get this error. Can you delete your Theano cache and
try again?

Fred

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with bigger
values(As it was reported in OpenAI Gym that ReLU from tensor.nnet.relu
isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Ramana Subramanyam]

Hi,
Somewhere else, I saw a comment that says it doesn't perform well with
bigger values. I am not able to recollect where I saw that. I will try to
reproduce with some big random values and cross check with a numpy
implementation. If it doesn't match,then I will ask Salimans.
Deleting cache solves this, but this happens quite often and today it
affected the training process. For example, 20 epochs run as expected, on
21st I see this error. I trained the network on a notebook and haven't
faced when I executed my code as a python file.

Regards,
Ramana

Post by Frédéric Bastien
It would be great to know why they don't like that implementation.
I don't know why you get this error. Can you delete your Theano cache and
try again?
Fred

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with
bigger values(As it was reported in OpenAI Gym that ReLU from
tensor.nnet.relu isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Ramana Subramanyam]

Hi,

Post by Frédéric Bastien
It would be great to know why they don't like that implementation.

While training an outdated GAN implementation, I used the relu from theano
in the generator and at ~500 epoch the generator loss became 100% and
saturated. Can it be said that encountering a dying relu problem is more
while using our implementation?

Post by Frédéric Bastien
I don't know why you get this error. Can you delete your Theano cache and
try again?
Fred

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with
bigger values(As it was reported in OpenAI Gym that ReLU from
tensor.nnet.relu isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Frédéric Bastien]

The error happen during the compilation of a Theano function.

Do you compile new Theano function at each epoch?

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with bigger
values(As it was reported in OpenAI Gym that ReLU from tensor.nnet.relu
isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Ramana Subramanyam]

Hi,
Nope, I compile outside the loop and the call to the function was perfectly
fine for 20 times and the 21st time this happened. Does theano wipe off
cache on it's own? If that was the case, I think it is possible for that to
have happened.
Also, I resumed my training and it has reached around 30 epochs and there
is no problem in the current run.

Ramana

Post by Frédéric Bastien
The error happen during the compilation of a Theano function.
Do you compile new Theano function at each epoch?

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with
bigger values(As it was reported in OpenAI Gym that ReLU from
tensor.nnet.relu isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Frédéric Bastien]

There is some cleanup, but only when the process shutdown and only on old
modules. We stop using the module for 1 weeks before deleting them. So if
you have an experiments running for more then 1 weeks, in theory, it could
happen.

Tell us if it happen again.

Fred

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with bigger
values(As it was reported in OpenAI Gym that ReLU from tensor.nnet.relu
isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Frédéric Bastien]

Also, note that this happen during the compilation of a Theano function. It
seem related to this code:

<ipython-input-50-125142517723> in <module>() 2
print(np.random.uniform(10000, 10001, 96).shape) 3 test_var =
theano.shared(np.random.uniform(10000, 10001, 96).reshape(2, 3, 4,
4))----> 4 theano.function([], theano.tensor.nnet.relu(test_var))()

Fred

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with bigger
values(As it was reported in OpenAI Gym that ReLU from tensor.nnet.relu
isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

[Ramana Subramanyam]

Hi,
Sorry for the delayed reply. My notebook is up for two weeks, but my
training isn't that long.

Post by Frédéric Bastien
Also, note that this happen during the compilation of a Theano function.
<ipython-input-50-125142517723> in <module>() 2 print(np.random.uniform(10000, 10001, 96).shape) 3 test_var = theano.shared(np.random.uniform(10000, 10001, 96).reshape(2, 3, 4, 4))----> 4 theano.function([], theano.tensor.nnet.relu(test_var))()
Fred
This was just an example that I showed. But not the traceback of the code

which stopped at 20th epoch. I actually don't understand that one week
window which you said. Does it apply to imports as well (importing theano)
or only compiled theano functions ?

Post by Frédéric Bastien

Post by Ramana Subramanyam
Hi,
This is the traceback I'm getting when I tried to compute ReLU with
bigger values(As it was reported in OpenAI Gym that ReLU from
tensor.nnet.relu isn't stable,
https://github.com/openai/improved-gan/blob/master/mnist_svhn_cifar10/nn.py#L12
) : http://dpaste.com/28DM3WX
I tried on CPU and it works as expected
Regards,
Ramana
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.

Regards,
Ramana