zronaghi

modfile = 'https://gmao.gsfc.nasa.gov/gmaoftp/geoscf/COVID_NO2/examples/model.csv
model = cudf.read_csv(modfile,parse_dates=['ISO8601'],date_parser=lambda x: dt.datetime.strptime(x, '%Y-%m-%dT%H:%M:%SZ'))

zronaghi

Xall = obsl.merge(model,how='inner',on='ISO8601')
Yall = Xall['value'] - Xall['NO2']
...
train = xgb.DMatrix(data=Xtrain,label=Ytrain)
params = {'booster':'gbtree','tree_method':'gpu_hist'}
bst = xgb.train(params,train)

zronaghi

bst.set_param({"predictor": "gpu_predictor"}) 
shap_array = np.abs(bst.predict(X,pred_contribs=True))
shap_values = pd.DataFrame(data=shap_array[:,:-1],columns=list(bst.feature_names))

zronaghi

import cudf
obsfile = 'https://gmao.gsfc.nasa.gov/gmaoftp/geoscf/COVID_NO2/examples/obs.csv'
obsdat = cudf.read_csv(obsfile,parse_dates=['ISO8601'],date_parser=lambda x: dt.datetime.strptime(x, '%Y-%m-%dT%H:%M:%SZ'))
allobs = obsdat.loc[(obsdat['obstype']=='no2')]

zronaghi

num_round = 20
param = {'booster' : 'gbtree' }
param['tree_method'] = 'gpu_hist'
start_time = time.perf_counter()
bst = xgb.train(param,train,num_round)
end_time = time.perf_counter()
print('This took {0:.3f} seconds'.format(end_time-start_time))

zronaghi

import requests
import xgboost as xgb
import time

SVMURL = 'https://gmao.gsfc.nasa.gov/gmaoftp/geoscf/gc-xgb/svm'
ifile = 'gcxgb_example_train.svm'
urlfile = '/'.join([SVMURL,ifile])
r = requests.get(urlfile)
open(ifile, 'wb').write(r.content)
train = xgb.DMatrix(ifile)

zronaghi

import cupy as cp
 
def get_dist_mat_gpu(coords):
# Calculate distance matrix for all elements in coords. 
# ‘coords’ is a cupy array of [1D image coordinates, wavelengths] within a group of 7 files (7 wavelengths), in GPU memory. The 2D images coordinates are 1st converted from the 1D values “on the fly” and the Euclidian distance matrix is returned. 
 
    coords_x = coords % 4096
    coords_y = coords // 4096
    coords_xb = coords_x[:, cp.newaxis]
    coords_yb = coords_y[:, cp.newaxis]