dbbert / dfn Goto Github PK

Hi! Thanks for sharing your amazing work!
I am confused about the role of filter bias in models/model_stereoPrediction.py (line 112), Could you explain how the filter bias help this task?

## add a bias
output = ConcatLayer([input, filters_biases])
output = FeaturePoolLayer(output, pool_size=2, pool_function=theano.tensor.sum)

Thank you very much!

First of all, nice work done.
I am much inspired by the 'Dynamic Filter Networks' paper, and I am planning to implement some ideas with dynamic filter layers in Tensorflow.

But first, I am trying to re-produce your results in my own way. Steerable filter was an easy single-channel example, and now I am trying to make a generalized version of dynamic filter layer...

I have several questions (and suggestions) for the existing Tensorflow version layer.

1. Why is there assertion checking assert num_filters_in <= 3?
2. According to the current version, the same filters are used for each channel? (code here). I think a generalized version, which apply different filters on different channels, can be made by setting filter_size_prod=kernel_h*kernel_w*num_filters_in and reshape the reshape_filter as size (kernel_w, kernel_h, num_filters_in, filter_size_prod).
3. Are you planning to release the Tensorflow version soon? If so, when will you release the code?

Correct me if I am wrong or misunderstood your code.

In dataset_bouncingMnistOriginal.py
After GetBatch(), the dim of the data is (16L, 1L, 64L, 64L, 20L), so you should use
data = data[case_id, :].reshape(self.image_size_, self.image_size_, -1)
in line149, instead of
data = data[case_id, :].reshape(-1, self.image_size_, self.image_size_)
:)

The url for datasets highwayDriving and movingObjects cannot be visited. Could you share the datasets in public?

I read your excellent paper and run the code. I find there is not bouncingMnist_originalTest.py. I can get the bouncingMnist_originalTest.py file easily by changing one line code "self.data_ = f['train'].value.reshape(-1, 28, 28)" to "self.data_ = f['test'].value.reshape(-1, 28, 28)". Do I understand it right? Thank you.

Hi! First of all, thanks for sharing your work!

I do not seem to be able to find the proper versions of Theano and Lasagne to make the code work. There are major changes among the different releases of these libraries that are causing me too much trouble for finding those that make the code work. Currently I am stuck with:

• Ubuntu 16.04
• Theano 0.9
• Lasagne 0.2.dev1
• PyGPU 0.6.9 (which seems to be the proper version to work with Theano 0.9)
• CuDNN 5.1 (as more recent versions do not work properly with PyGPU 0.6.9)
• Cuda 8.0

It is the closest I feel I've been of pulling this out, however I get a Theano error where it is trying to check the type of a numpy.ndarray with ".type" which numpy arrays don't have as an attribute.

Other combinations of Theano-Lasagne versions gave a wide range of errors where it cannot make the GPU work (related to the back-end Theano uses in more recent versions) or not being able to find specific functions (related, I guess, to some incompatibilities among specific versions of Theano and Lasagne).

It would be of great help if you could provide the used versions for making the code work. I am using Docker, so changing to different OS or Cuda/CuDNN versions wouldn't be a problem either.

Thanks in advance!

Hi, the datasets of highwayDriving and movingObjects cannot be visited.

Could you please share them again? Thanks.

I really appreciate your work. That's amazing.

I'm really interested in your model and try to realize it in tensorflow. However, the loss doesn't decrease. I feel very frustrated. I couldn't figure it out. Could you mind release a tensorflow version of mnist experiment?

Thank you very much.

Best Regards,
Jacqueline

elif options['loss'] == 'binary_crossentropy':
        # Clipping to avoid NaN's in binary crossentropy: https://github.com/Lasagne/Lasagne/issues/436
        output_frame = T.clip(output_frame, np.finfo(np.float32).eps, 1-np.finfo(np.float32).eps)
        frame_loss = binary_crossentropy(output_frame, target[:,[i],:,:])

if you use binary_crossentropy , should you promise the target is 0 or 1 ? but the target is float in [0,1].
and i think the output should using sigmoid function to promise [0,1]

What package version of sonnet are you using please? When I am running on the latest version of that package I am seeing the error:

AttributeError                            Traceback (most recent call last)
Cell In[5], [line 2](vscode-notebook-cell:?execution_count=5&line=2)
      [1](vscode-notebook-cell:?execution_count=5&line=1) """Defining the model (using a Sonnet module)"""
----> [2](vscode-notebook-cell:?execution_count=5&line=2) class Model(snt.AbstractModule):
      [3](vscode-notebook-cell:?execution_count=5&line=3)     def __init__(self, filter_size=(9,9), name="model"):
      [4](vscode-notebook-cell:?execution_count=5&line=4)         super(Model, self).__init__(name=name)

AttributeError: module 'sonnet' has no attribute 'AbstractModule'

for the file experiment_bouncingMnistOriginal_tensorflow.ipynb