Hi,

First of all, thank you for opening your great script!

I have a question about latent encoder.

In your code, posterior and prior use same latent encoder. Exactly using same (x,y) pair encoder is no problem, but the linear layers to sample the latent variable would be different functions for prior and posterior I think.

If my understanding about the code is wrong, then could I know which point I missed?

Thanks.

Best Regards, Jaesik Yoon.

Hi,

I was wondering what the reasoning behind the specific parameterizations of the standard deviation in the latent encoder is, i.e. why is it bounded? I usually just use a softplus in such settings.

Also, they are different between the encoder and decoder, not sure if that is intentional. I'd assume both should be with tf.sigmoid?

(Latent) Encoder: Bounds SD between 0.1 and 1

    # Compute sigma
    sigma = 0.1 + 0.9 * tf.sigmoid(log_sigma)

Decoder: Bounds SD to be higher than 0.1 and...?

    # Bound the variance
    sigma = 0.1 + 0.9 * tf.nn.softplus(log_sigma)

And thank you for that well-documented repository :).

1. For the latest version of matplotlib, set_facecolor rather than set_axis_bgcolor should be used in plot_functions?
2. We should move TensorFlow Distributions to TensorFlow Probability?
3. It might be better to add information about package versions in the documentation?

When I run the 'attentive_neural_process.ipynb', I got the mistakes as follows:

ValueError: Variable layer_3/kernel already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope? Originally defined at:

File "", line 28, in batch_mlp output, output_sizes[-1], name="layer_{}".format(i + 1)) File "", line 31, in call hidden = batch_mlp(encoder_input, self._output_sizes, "latent_encoder") File "", line 58, in call prior = self._latent_encoder(context_x, context_y)

I think NP implementation in ANP is a little bit different with NP paper. right?

In original NP, there is the only latent encoder and latent encoder uses context and target data, while implementation has concated features of latent and deterministic and latent encoder uses only context data.

By default, Colab will use TensorFlow 2.x as of now, this results in failed execution of cells. By specifying the %tensorflow_version magic, Colab uses the specified TensorFlow version.

So here are two results (two columns) of ANP after 50000+ iterations. The first row is the default plot and the plots in the second row contains 16 samples of the curve (given same context x & y). And the red curve is the mean of these 16 curves. As we can see, there has very limited variation.

This is true also for NP. However, if I set use_deterministic_path=False, the variation start to emerge:

My guess is that during training the decoder only prefer the deterministic path and just ignore the latent path. Then no matter what latent code is delivered the output won't change much. What's your opinion?

Comments on sampling from either prior (during testing) or the posterior (during training) are swapped.

"Note that targets contain contexts by design" -> is this is referring to how the latent encoder had previously been run on the contexts, rather than the targets themselves actually containing contexts?

Thanks!

Hello, I don't know if this is the right forum to ask.....

Is there a available demo on the one-shot classification using the Omniglot dataset presented in section 4.3 of the paper. I am unable to understand how you combined the different channels of convolutional output to compute the statistics and combine it with test image to perform classification?