Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ripsnet #587

Open
wants to merge 31 commits into
base: master
Choose a base branch
from
Open

ripsnet #587

wants to merge 31 commits into from

Conversation

hensel-f
Copy link

@hensel-f hensel-f commented Mar 1, 2022

A pull request for adding RipsNet to GUDHI. An architecture for fast and robust estimation of persistence diagram vectorizations of point clouds.

Notice that, for the moment, RipsNet documentation is only available from installation/tensorflow. We will deal later with this when perslay and PD gradient (other tensorflow features) will be merged

Copy link
Contributor

@VincentRouvreau VincentRouvreau left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

To be complete, you also missed some required changes in src/python/CMakeLists.txt as Mathieu did on PD gradient PR, but it's ok, this file is a nightmare ;-)

src/python/doc/ripsnet.inc Show resolved Hide resolved
src/python/doc/ripsnet.inc Outdated Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
src/python/doc/ripsnet.inc Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
src/python/doc/ripsnet.rst Outdated Show resolved Hide resolved
Copy link
Contributor

@VincentRouvreau VincentRouvreau left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In order to generate the documentation, I am proposing to you a quick 'fix'. In src/python/doc/installation.rst, in the Tesorflow section (~ l.398), add:

:doc:`RipsNet <ripsnet>` module requires `TensorFlow <https://https://www.tensorflow.org/install/>`_.

You will be able to see that your bibliography is correct (you can remove the links). I can show you the result on CI when it will be built.

src/python/test/test_ripsnet.py Outdated Show resolved Hide resolved
@hensel-f hensel-f marked this pull request as ready for review March 11, 2022 13:46
@VincentRouvreau

This comment was marked as resolved.

Copy link
Member

@mglisse mglisse left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A first batch of comments.
(not all comments require code changes, some are questions)

Comment on lines 68 to 69
tf_data_test = tf.ragged.constant([
[list(c) for c in list(data_test[i])] for i in range(len(data_test))], ragged_rank=1)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Uh, so we build data_test using numpy (from lists), only to convert it again to lists here, and finally build a tensorflow object from those lists? Would it be possible to skip some of those conversions? We probably don't need to import numpy at all.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, you're right, thanks. I updated it.

Comment on lines 73 to 74
Once RN is properly trained (which we skip in this documentation) it can be used to make predictions.
A possible output is:
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It isn't obvious what the example is computing. Maybe adding a comment or 2 would help (or an image). Is data_test a list of 3 point sets in 2D, and is the output some kind of vectorized persistence diagram? It becomes clear once we read the detailed doc, but I think some minimal comments in the example would still make sense.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good point, I added a sentence to describe it.

return outputs


class TFBlock(tf.keras.layers.Layer):
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This seems very general. Is it related to tf.keras.Sequential, or is there some other utility already providing this composition?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is possible that there is an existing utility for this composition, but I couldn't find anything explanation of how to make it work with ragged inputs.

"""
super().__init__(dynamic=True, **kwargs)
self._supports_ragged_inputs = True
self.pop = perm_op
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Using the name pop is confusing (I was wondering from which list you we removing an element), could we stick to perm_op or anything that isn't already an English word with an unrelated meaning?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, I've updated it to avoid confusion, thanks.

Comment on lines 113 to 114
def build(self, input_shape):
super().build(input_shape)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Isn't that exactly what happens if you don't define this function?
I am also trying to understand the difference between this and what you did for DenseRaggedBlock and RipsNet.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm also not exactly sure what the effect of this is, I modeled it after an example I saw somewhere. But I've changed it to mach the case for RipsNet and DenseRagged. Is that fine, or what would you suggest?

return outputs


class DenseRaggedBlock(tf.keras.layers.Layer):
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks identical to TFBlock except for _supports_ragged_inputs? Would TFBlock(ragged=True) make sense? Or could it even be implicit, ragged iff the first layer is?

Copy link
Author

@hensel-f hensel-f Apr 26, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Indeed, I have changed TFBlock such that it supports ragged inputs if the first layer is an instance of DenseRagged. So, indeed I think that DenseRaggedBlock is no longer needed. I have commented it for now but if the change is confirmed it can be deleted.

elif self.pop == 'sum':
pop_ragged = PermopRagged(tf.math.reduce_sum)
else:
raise ValueError(f'Permutation invariant operation: {self.pop} is not allowed, must be "mean" or "sum".')
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If perm_op is not one of those 2 strings, should we assume that it is a function, so users can pass tf.math.reduce_max if they want? Or is that useless?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Well, the only concern I have is that if we leave this entirely up to the user, then their input function may be such that our theoretical guarantees for RipsNet may perhaps no longer be satisfied. I'm not sure what the best option is in this case. What do you think?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it's good to let the user do whatever they want (as long as it is properly documented). Our theoretical results are of the form "if ..., then" ; but nothing prevents to use RipsNet with some other perm_op and a user may be interested in doing so.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Okay, sounds good, I've removed the requirement that the permutation invariant function has to be 'mean' or 'sum', so that users can specify their own functions.

Comment on lines +55 to +58
outputs = tf.ragged.map_flat_values(tf.matmul, inputs, self.kernel)
if self.use_bias:
outputs = tf.ragged.map_flat_values(tf.nn.bias_add, outputs, self.bias)
outputs = tf.ragged.map_flat_values(self.activation, outputs)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks like an ad hoc reimplementation of a dense layer, applied to each element. Would it make sense first to define a small network that takes an input of size 2 (if we work with 2d points) and that may be composed of several layers, and only then apply (map) it to all the points in the tensor, so there is a single call to a map function for the whole phi1? It seems conceptually simpler, but it might be slower if tensorflow doesn't realize that it is equivalent.

Copy link
Author

@hensel-f hensel-f Apr 26, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry, I'm not exactly sure what you mean, or if it makes it faster. But if you have a concrete change in mind please just adapt it directly or let me know.

Copy link
Contributor

@tlacombe tlacombe left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Short reading of the doc.


.. testcode::

from gudhi.tensorflow import *
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It may be better to be explicit, i.e. naming the imported functions or to do import gudhi.tensorflow as gtf just to make clear to the user which functions below are indeed coming from the gudhi.tensorflow package.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree, that makes it a bit clearer.

activation_fct = 'gelu'
output_activation = 'sigmoid'
dropout = 0
kernel_regularization = 0
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps it may be worth to comment (not in detail) what are these hyper-parameters ; in particular how one is supposed to chose ragged_layers_size and dense_layers_size.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure, I added a comment saying they should be tuned according to the specific dataset in order to reach a better performance.


RN.predict(tf_data_test)

Once RN is properly trained (which we skip in this documentation) it can be used to make predictions.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it be cumbersome to provide a sort of minimal working example to train RipsNet ?

A user may not be familiar with tensorflow and have no clue on how to train the RipsNet model at this stage. Perhaps just setting the typical optimizer = ..., loss_function = ... , and do a single step of gradient descent here, would help to and not discourage the user not familiar with tf?

Another option is to write a Tutorial ( https://github.com/GUDHI/TDA-tutorial ) to reproduce, say, the synthetic experiment of the paper (multiple_circles) and to refer to it in this doc (I understand that we don't want this doc to be too long).

A final option (that requires more development) would be to provide a method train to RipsNet that does the job with some default parameters, so that one could get starting by simply going for something like

RN = ripsnet.RipsNet(...)
RN.train(train_data)
RN.predict(test_data)

Of course these are just suggestions.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think the best option is to link to the notebook containing the synthetic examples. This provides a very nice example where one can see the workflow.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've adapted Mathieu's original tutorial on the synthetic data so that it illustrates the use (including setup and training) of a RipsNet architecture. So, as @tlacombe suggested, I think it may be nice to include and link to this tutorial somewhere.

Copy link
Author

@hensel-f hensel-f Jun 7, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've opened a PR (here: GUDHI/TDA-tutorial#59) to include the tutorial notebook so that we can then link to it.

RN.predict(tf_data_test)

Once RN is properly trained (which we skip in this documentation) it can be used to make predictions.
In this example RipsNet estimates persistence vectorizations (of output size 25) of a list of point clouds (of 3 points) in 2D.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Isn't it simply "a list of 3 point clouds (of 3 points)" ?
(Here, I understand it as "each point cloud is made of 3 points", but perhaps my English is just broken.)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could also add "yielding 3 vectorizations hence an output with shape nb_training_data x output_units " for clarity?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure, I will add a sentence for clarification, thanks for the suggestion. (In this example each of the pointclouds only has 3 points actually, so I guess you understood it as intended.)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants