kjhall01 / xcast Goto Github PK

Hi, I just read through the Xcast paper. Wanted to check if xcast also has tools for spatiotemporal downscaling of gridded datasets; if not, request for the feature to be added to the module.

limits prediction probabilities to (23%, 57%) because minimum transformed value is 1 / (1 + 2*e) and maximum value is e / (e + 2)

seems like transposing is taking up a lot of time during cross validation. hopefully this fixes it

Need to add Probabilistic Methods

X, Y, T = xc.load_sample_data() # load test data

gamma = xc.rGammaRegression()
gamma.fit(X, Y)
pred = gamma.predict(X)

ValueError: Some value(s) of y are out of the valid range of the loss 'HalfGammaLoss'.

line 79 in base_verification.py needs to be removed - artifact from old version

I installed xcast following instruction on GitHub page but got error when importing the package in Spyder.

import xcast as xc
Traceback (most recent call last):

Cell In[1], line 1
import xcast as xc

File /opt/anaconda3/lib/python3.11/site-packages/xcast-0.6.9-py3.9.egg/xcast/init.py:7
from .visualization import view_probabilistic, view_reliability, reliability_diagram, view_taylor, view_roc, view

File /opt/anaconda3/lib/python3.11/site-packages/xcast-0.6.9-py3.9.egg/xcast/visualization/init.py:4
from .roc import view_roc

File /opt/anaconda3/lib/python3.11/site-packages/xcast-0.6.9-py3.9.egg/xcast/visualization/roc.py:3
from scipy import interp

ImportError: cannot import name 'interp' from 'scipy' (/opt/anaconda3/lib/python3.11/site-packages/scipy/init.py)

How can I solve this?

ELR here uses thresholds in the range [0, 1] when it should, probably, be [mindata, maxdata].

The threshold and the input should be scaled to the same range.

forcing into multi-class for now. Need to re-build.

The functions fail to generate correct plots for a domain that crosses the dateline (eg. 155E to 185W and 20S to 3N)

Since RankedTerciles uses xr.quantile, which uses dask.apply_gu_func, which does not accommodate multiple chunks across S/M, XCast needs a new implementation of percentile-based one-hot encoding.

1. Update documentation - In Progress
2. Smart dimension-name guessing - ETA 1/1/2022
3. keep attributes, XR.dataarray store in estimator objects ? - ETA 1/1/2022
4. Remove strict chunking rules, switch to manual dask mapping over two dimensions only - ETA 1/1/2022
5. Fix Conda-Install; needs to work on Python 3.9 & 3.10

I love the idea of xcast. If I want to predict multiple lead_times in a single iteration, would I just stack lead_time and X or Y?

I think more concrete examples in the documentation would help users.
i.e. something like applying xcast to s2s-ai-challenge mock data like I did in https://gist.github.com/aaronspring/36e112e992e36fba935f73404dbbd3cd and related issue (likely your approach is much more performant because the vectorize=True essentially loops the xr.apply_ufunc call, but would be great to check)

Select everything within polygon
Etc

I was wondering how the scope of xcast differs from xbatcher?

Hi @kjhall01 !

I am interested in using xcast for a few projects, but still trying to get my arms around the api and generally xarray I/O with xcast. Attached is a notebook (zipped) where I ran into some issues. Happy to set up a meeting too if that is best.

-Thomas
xcast_demo.ipynb.zip

• some methods .fit on continous observations data, and are capable of producing probabilistic forecasts.
• predict_proba takes the number of outputs from the number of outputs on the training data
• methods trained on continuous data expect only a single output for predict_proba - but that is an issue for tercile forecasts

Seems that calling

X, Y, T = xc.NMME_IMD_ISMR() # load test data

does not work.

undefined variable "unassign_labels" should say "unassigned_labels". Bug will not affect many since it only covers the case where all coords except one can be assigned, and then the remaining dim is assigned to the remaining unassigned label.

it makes cross validation way easier when the One-Hot encoding is part of the classifier

A Y-Mask is calculated in .fit, but not actually used to mask out erroneous missing data.

it should be. also apply an x-mask , and institute assertions for erroneous missing data

check these

the amount of warnings printed probably significnatly impacts performance

every slice in the cross-validation process should return the same chunking scheme as the initial arguments, X & Y. Currently, It sometimes re-chunks across the S dimension (likely a by-product of xr.concat in the function definition).

Either have XCast accomodate multiple chunks along S/M (ideal) , or have cross validator maintain number of chunks along S/M.