Conditional Density Estimation notebook #40
Conversation
| @@ -43,6 +43,7 @@ def motorcycle_data(): | |||
| df = pd.read_csv("./data/motor.csv", index_col=0) | |||
| X, Y = df["times"].values.reshape(-1, 1), df["accel"].values.reshape(-1, 1) | |||
| Y = (Y - Y.mean()) / Y.std() | |||
| X /= X.max() | |||
There was a problem hiding this comment.
How is this relevant to the problem?
There was a problem hiding this comment.
Rescales inputs between [0,1] instead of the [0,60], which makes the length-scales initialisation to 1 more sensible.
|
Can you summarise what did you add to fix the issue? |
|
@vdutor was it intentional that you did not save it as jupytext .py instead of .ipynb format? |
|
Hi @st-- , yes this notebook takes a little bit too long to generate on the fly. I was thinking that it makes sense to have some precompiled notebooks as well in our documentation to explain some more complicated models rather than keeping it to toy examples. What do you think? |
docs/notebooks/deep_cde.ipynb
Outdated
| "source": [ | ||
| "## Deep Gaussian process with latent variables\n", | ||
| "\n", | ||
| "To tackle the problem we suggest a Deep Gaussian process with a latent variable in the first layer. The latent variable will be able to capture the \n", |
There was a problem hiding this comment.
A bit of word shuffling: "We suggest a Deep Gaussian process with a latent variable in the first layer to improve the error bars on the given dataset. The latent variable allows to model the heteroscedasticity, while an extra layer makes the model more expressive to catch sharp transitions."
| "source": [ | ||
| "### Latent Variable Layer\n", | ||
| "\n", | ||
| "This layer concatenates the inputs with a latent variable. See Dutordoir, Salimbeni et al. Conditional Density with Gaussian processes (2018) <cite data-cite=\"dutordoir2018cde\"/> for full details. We choose a one-dimensional input and a full parameterisation for the latent variables. This means that we do not need to train a recognition network, which is useful for fitting but can only be done in the case of small datasets, as is the case here." |
There was a problem hiding this comment.
Make a ref [...] instead of the long text?
docs/notebooks/deep_cde.ipynb
Outdated
| "source": [ | ||
| "### Fit\n", | ||
| "\n", | ||
| "We can now fit the model. Because of the `DirectlyParameterizedEncoder` it is important to set the batch size to the number of datapoints and turn off shuffle. This is so that we use the associated latent variable for each datapoint. If we would use an Amortized Encoder network this would not be necessary." |
There was a problem hiding this comment.
"It is important to set the batch size to the number of datapoints and turn off shuffle, checkout DirectlyParameterizedEncoder for details"? This needs a bit of explanation, what DirectlyParameterizedEncoder is.
…lux into vincent/notebook/cde
Notebook building and fitting a deep (two layer) latent variable model using VI. No changes to the core of GPflux are required but careful setting of fitting options is necessary. For example, it is important to set
shuffletoFalseandbatch_sizeto the number of datapoints to have correct optimisation of the latent variables.