raminmh / cfc Goto Github PK

I am using on the Temporal Fusion Transformers model, and planning to use the CFC.
is it possible to use the LTC layer in the Temporal Fusion Transformers ?

Is it possible to train cfc neural network with other network structures such as transformer?
Are there specific examples?

Hello, I would like to ask, if cfc is used in time series, is the sample the entire sequence, or a fixed-length window of the time series? For the case of one piece of data per day, does ts need to be set to all 1? How does cfc capture long-term dependencies? If so Windows, it is obvious that there is no recorded time relationship between windows

On the page https://ncps.readthedocs.io/en/latest/examples/torch_first_steps.html

Under heading The LTC model with NCP wiring
wiring = AutoNCP(16, out_features) # 16 units, 1 motor neuron
ltc_model = LTC(in_features, wiring, batch_first=True)
issue - out_features and in_features not defined
number of units appears to be 8 and out and in features are 1

under Training the model
issue - model undefined

Hi
Thank you so much, I wasn’t expecting such a quick response and I’m very sorry for not replying to this earlier as my computer has unfortunately been on the blink for quite a while.
I was just wondering if you could release a version of this (the traffic model) where we could use our own optimizer because I’d like to further test it out with my favourite one:)

If you could mark in the code the optimum position to put the optimisation algorithm then I'd be further grateful for that too :)

Thanks, Sam.

A regularizing kernel could be used to penalize the neural network weights. This can help to avoid overfitting the model and improve its overall performance.

This could be done by adding the following line of code after the creation of each dense layer in the "backbone" section of the build method:

copy code
self.backbone.append(
tf.keras.layers.Dense(
self.hparams["backbone_units"],
backbone_activation,
kernel_regularizer=tf.keras.regularizers.L2(
self.hparams["weight_decay"]
),
)
)
This would add an L2 regularizer kernel to each dense layer in the "backbone" section, using the "weight_decay" parameter specified in the hyperparameters to set the level of regularization.

Hello,

The dataset seems to be unavailable (getting "Unable to establish SSL connection." for https://pub.ist.ac.at/~mlechner/datasets/walker.zip).

Hello, I am very interested in your model. When I run your traffic_with_cfc.py, I found ''--use_mixed_ltc'', but I did not find the MixedLTCCell function in the code. I really want to know what it is, thank you

With pytorch_lightning 1.7 : ModuleNotFoundError: No module named 'pytorch_lightning.metrics'

Solution : change imports in python train_*.py from import pytorch_lightning.metrics.functional... into ```import torchmetrics.functional `` and install torchmetrics (pip install torchmetrics for example) .

I'm trying to adopt CfC network from your PyTorch example (person_atcivity) to make a timeseries regression. Given one dimensional array of floating-point numbers (lets consider them equally spaced in time) how can I feed them into the network? As I understand your current PyTorch implementation works as a classifier, that is the Labels array should be finite number of classes. But in case of regression we have floating point values as output. Also how should we structure the inputs? Lets say we have 10 last values of the timeseries as input and we want to predict 11th value, how do we do that? If you can provide a short pseudocode example it would be much appreciated. Thanks.

Hi,

I am trying to estimate ECG metrics with a combination of Conv-layers and CFCs, taking inspiration from the Atari tutorial. Unlike Atari, where convolutions extract features from images, in my case, convolutions work on the time dimension of the ECG signal itself. Therefore, the time dimension is modified by the convolutions. Will this have an impact on the performance of the CFC?

Also, if you can comment on the implementation of the CFC, that will be very helpful for understanding what each input to the CFC function does. As of now, tutorials are very helpful in showcasing the potential of LTCs and CFCs, but stronger documentation is needed for development.

Can you also share a sample code you used for generating gradient maps for the CFCs, as highlighted in the paper?

Thank you very much for your time!

ValueError: Unable to restore custom object of type _tf_keras_rnn_layer currently. Please make sure that the layer implements get_configand from_config when saving. In addition, please use the custom_objects arg when calling load_model().

Hi there, I have been fascinated by your models and have deeply enjoyed testing them out.

Longshot but would you possibly be able to train the CfC models on the Traffic volume dataset (like you did with the LTC networks) please and post it on here so I could try it out? 

Thanks, Sam.