def test_rf_covtype_train_full_fvec(self):
h2o.beta_features = True
csvFilename = 'covtype.data'
csvPathname = 'standard/' + csvFilename
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, schema='put', hex_key=csvFilename + ".hex",
timeoutSecs=180)
for trial in range(1):
# params is mutable. This is default.
kwargs = paramDict
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = kwargs['ntrees'] * 60
start = time.time()
print "Note train.csv is used for both train and validation"
rfv = h2o_cmd.runRF(parseResult=parseResult, timeoutSecs=timeoutSecs, noPoll=True, **kwargs)
h2o_jobs.pollStatsWhileBusy(timeoutSecs=timeoutSecs, retryDelaySecs=5)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
job_key = rfv['job_key']
model_key = rfv['destination_key']
rfv = h2o_cmd.runRFView(data_key=parseResult['destination_key'],
model_key=model_key,
timeoutSecs=timeoutSecs, retryDelaySecs=1, print_params=True)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv)
self.assertLess(classification_error, 3, "train.csv should have full classification error: %s < 3" % classification_error)
print "Trial #", trial, "completed"
def test_rf_covtype_train_full_fvec(self):
h2o.beta_features = True
csvFilename = 'covtype.data'
csvPathname = 'standard/' + csvFilename
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, schema='put', hex_key=csvFilename + ".hex",
timeoutSecs=180)
for trial in range(1):
# params is mutable. This is default.
kwargs = paramDict
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = kwargs['ntrees'] * 60
start = time.time()
print "Note train.csv is used for both train and validation"
rfv = h2o_cmd.runRF(parseResult=parseResult, timeoutSecs=timeoutSecs, noPoll=True, **kwargs)
h2o_jobs.pollStatsWhileBusy(timeoutSecs=timeoutSecs, retryDelaySecs=5)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
job_key = rfv['job_key']
model_key = rfv['destination_key']
rfv = h2o_cmd.runRFView(data_key=parseResult['destination_key'],
model_key=model_key, timeoutSecs=timeoutSecs, retryDelaySecs=1)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv)
# hmm..just using defaults above in RF?
self.assertLess(classification_error, 4.8, "train.csv should have full classification error: %s < 4.8" % classification_error)
print "Trial #", trial, "completed"
def test_rf_big1_nopoll_fvec(self):
h2o.beta_features = True
csvFilename = 'hhp_107_01.data.gz'
hex_key = csvFilename + ".hex"
print "\n" + csvFilename
parseResult = h2i.import_parse(bucket='smalldata', path=csvFilename,
hex_key=hex_key, timeoutSecs=30, schema='put')
rfViewInitial = []
# dispatch multiple jobs back to back
for jobDispatch in range(3):
start = time.time()
kwargs = {}
if OVERWRITE_RF_MODEL:
print "Since we're overwriting here, we have to wait for each to complete noPoll=False"
model_key = 'RF_model'
else:
model_key = 'RF_model' + str(jobDispatch)
kwargs['ntrees'] = 1
if OVERWRITE_RF_MODEL:
print "Change the number of trees, while keeping the rf model key name the same"
print "Checks that we correctly overwrite previous rf model"
kwargs['ntrees'] += 1
kwargs['seed'] = random.randint(0, sys.maxint)
# FIX! what model keys do these get?
randomNode = h2o.nodes[random.randint(0,len(h2o.nodes)-1)]
h2o_cmd.runRF(node=randomNode, parseResult=parseResult, destination_key=model_key,
timeoutSecs=300, noPoll=False if OVERWRITE_RF_MODEL else True, **kwargs)
print "rf job dispatch end on ", csvFilename, 'took', time.time() - start, 'seconds'
print "\njobDispatch #", jobDispatch
h2o_jobs.pollWaitJobs(pattern='RF_model', timeoutSecs=300, pollTimeoutSecs=10, retryDelaySecs=5)
# we saved the initial response?
# if we do another poll they should be done now, and better to get it that
# way rather than the inspect (to match what simpleCheckGLM is expected
first = None
print "rfViewInitial", rfViewInitial
for rfView in rfViewInitial:
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['_dataKey']
model_key = rfView['_key']
ntree = rfView['ntree']
print "Temporary hack: need to do two rf views minimum, to complete a RF (confusion matrix creation)"
# allow it to poll to complete
rfViewResult = h2o_cmd.runRFView(None, data_key, model_key, ntree=ntree, timeoutSecs=60, noPoll=False)
if first is None: # we'll use this to compare the others
first = rfViewResult.copy()
firstModelKey = model_key
print "first", h2o.dump_json(first)
else:
print "Comparing", model_key, "to", firstModelKey
df = h2o_util.JsonDiff(rfViewResult, first, vice_versa=True, with_values=True)
print "df.difference:", h2o.dump_json(df.difference)
def scoreRF(scoreParseResult,
trainResult,
vactual=None,
timeoutSecs=120,
**kwargs):
# Run validation on dataset
parseKey = scoreParseResult['destination_key']
if h2o.beta_features:
# this is how we're supposed to do scorin?
rfModelKey = trainResult['drf_model']['_key']
predictKey = 'Predict.hex'
start = time.time()
predictResult = h2o_cmd.runPredict(data_key=parseKey,
model_key=rfModelKey,
destination_key=predictKey,
timeoutSecs=timeoutSecs,
**kwargs)
h2o_cmd.runInspect(key='Predict.hex', verbose=True)
predictCMResult = h2o.nodes[0].predict_confusion_matrix(
actual=parseKey,
vactual=vactual,
predict=predictKey,
vpredict='predict',
timeoutSecs=timeoutSecs,
**kwargs)
rftime = time.time() - start
cm = predictCMResult['cm']
# These will move into the h2o_gbm.py
pctWrong = h2o_gbm.pp_cm_summary(cm)
print "\nTest\n==========\n"
print h2o_gbm.pp_cm(cm)
scoreResult = predictCMResult
else:
ntree = trainResult['ntree']
rfModelKey = trainResult['model_key']
start = time.time()
# NOTE: response_variable is required, and passed from kwargs here
# out_of_bag_error_estimate=0 is required for scoring. H2O will assert if 1 and different data set
# compared to training
kwargs['out_of_bag_error_estimate'] = 0
scoreResult = h2o_cmd.runRFView(None,
parseKey,
rfModelKey,
ntree=ntree,
timeoutSecs=timeoutSecs,
**kwargs)
rftime = time.time() - start
h2o.verboseprint("RF score results: ", scoreResult)
h2o.verboseprint("RF computation took {0} sec".format(rftime))
scoreResult['python_call_timer'] = rftime
return scoreResult
def test_rf_covtype_fvec(self):
importFolderPath = "/home/0xdiag/datasets/standard"
csvFilename = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on the normal covtype.data"
parseKey = h2i.parseImportFolderFile(None, csvFilename, importFolderPath, key2=key2,
header=0, timeoutSecs=180)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
rfViewInitial = []
for jobDispatch in range(1):
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(jobDispatch)
# don't poll for fvec
rfResult = h2o_cmd.runRFOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, noPoll=True, rfView=False, **kwargs)
elapsed = time.time() - start
print "RF dispatch end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
print h2o.dump_json(rfResult)
# FIX! are these already in there?
rfView = {}
rfView['data_key'] = key2
rfView['model_key'] = kwargs['model_key']
rfView['ntree'] = kwargs['ntree']
rfViewInitial.append(rfView)
print "rf job dispatch end on ", csvPathname, 'took', time.time() - start, 'seconds'
print "\njobDispatch #", jobDispatch
h2o_jobs.pollWaitJobs(pattern='RF_model', timeoutSecs=180, pollTimeoutSecs=120, retryDelaySecs=5)
# we saved the initial response?
# if we do another poll they should be done now, and better to get it that
# way rather than the inspect (to match what simpleCheckGLM is expected
print "rfViewInitial", rfViewInitial
for rfView in rfViewInitial:
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['data_key']
model_key = rfView['model_key']
ntree = rfView['ntree']
# allow it to poll to complete
rfViewResult = h2o_cmd.runRFView(None, data_key, model_key, ntree=ntree, timeoutSecs=60, noPoll=False)
def scoreRF(scoreParseKey, trainResult, **kwargs):
# Run validation on dataset
rfModelKey = trainResult['model_key']
ntree = trainResult['ntree']
start = time.time()
scoreResult = h2o_cmd.runRFView(modelKey=rfModelKey, parseKey=scoreParseKey, ntree=ntree, **kwargs)
rftime = time.time()-start
h2o.verboseprint("RF score results: ", scoreResult)
h2o.verboseprint("RF computation took {0} sec".format(rftime))
scoreResult['python_call_timer'] = rftime
return scoreResult
def scoreRF(scoreParseResult, trainResult, vactual=None, timeoutSecs=120, **kwargs):
# Run validation on dataset
parseKey = scoreParseResult['destination_key']
if h2o.beta_features:
# this is how we're supposed to do scorin?
rfModelKey = trainResult['drf_model']['_key']
predictKey = 'Predict.hex'
start = time.time()
predictResult = h2o_cmd.runPredict(
data_key=parseKey,
model_key=rfModelKey,
destination_key=predictKey,
timeoutSecs=timeoutSecs, **kwargs)
h2o_cmd.runInspect(key='Predict.hex', verbose=True)
predictCMResult = h2o.nodes[0].predict_confusion_matrix(
actual=parseKey,
vactual=vactual,
predict=predictKey,
vpredict='predict',
timeoutSecs=timeoutSecs, **kwargs)
rftime = time.time()-start
cm = predictCMResult['cm']
# These will move into the h2o_gbm.py
pctWrong = h2o_gbm.pp_cm_summary(cm);
print "\nTest\n==========\n"
print h2o_gbm.pp_cm(cm)
scoreResult = predictCMResult
else:
ntree = trainResult['ntree']
rfModelKey = trainResult['model_key']
start = time.time()
# NOTE: response_variable is required, and passed from kwargs here
# out_of_bag_error_estimate=0 is required for scoring. H2O will assert if 1 and different data set
# compared to training
kwargs['out_of_bag_error_estimate'] = 0
scoreResult = h2o_cmd.runRFView(None, parseKey, rfModelKey, ntree=ntree, timeoutSecs=timeoutSecs, **kwargs)
rftime = time.time()-start
h2o.verboseprint("RF score results: ", scoreResult)
h2o.verboseprint("RF computation took {0} sec".format(rftime))
scoreResult['python_call_timer'] = rftime
return scoreResult
def scoreRF(scoreParseKey, trainResult, **kwargs):
# Run validation on dataset
rfModelKey = trainResult["model_key"]
ntree = trainResult["ntree"]
start = time.time()
data_key = scoreParseKey["destination_key"]
scoreResult = h2o_cmd.runRFView(None, data_key, rfModelKey, ntree, **kwargs)
rftime = time.time() - start
h2o.verboseprint("RF score results: ", scoreResult)
h2o.verboseprint("RF computation took {0} sec".format(rftime))
scoreResult["python_call_timer"] = rftime
return scoreResult
def scoreRF(scoreParseKey, trainResult, **kwargs):
# Run validation on dataset
rfModelKey = trainResult['model_key']
ntree = trainResult['ntree']
start = time.time()
data_key = scoreParseKey['destination_key']
# NOTE: response_variable is required, and passed from kwargs here
scoreResult = h2o_cmd.runRFView(None, data_key, rfModelKey, ntree, **kwargs)
rftime = time.time()-start
h2o.verboseprint("RF score results: ", scoreResult)
h2o.verboseprint("RF computation took {0} sec".format(rftime))
scoreResult['python_call_timer'] = rftime
return scoreResult
def scoreRF(scoreParseKey, trainResult, **kwargs):
# Run validation on dataset
rfModelKey = trainResult['model_key']
ntree = trainResult['ntree']
start = time.time()
data_key = scoreParseKey['destination_key']
scoreResult = h2o_cmd.runRFView(None, data_key, rfModelKey, ntree,
**kwargs)
rftime = time.time() - start
h2o.verboseprint("RF score results: ", scoreResult)
h2o.verboseprint("RF computation took {0} sec".format(rftime))
scoreResult['python_call_timer'] = rftime
return scoreResult
def test_rf_covtype_train_full_fvec(self):
h2o.beta_features = True
csvFilename = "covtype.data"
csvPathname = "standard/" + csvFilename
parseResult = h2i.import_parse(
bucket="home-0xdiag-datasets", path=csvPathname, schema="put", hex_key=csvFilename + ".hex", timeoutSecs=180
)
for trial in range(1):
# params is mutable. This is default.
kwargs = paramDict
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = kwargs["ntrees"] * 60
start = time.time()
print "Note train.csv is used for both train and validation"
rfv = h2o_cmd.runRF(parseResult=parseResult, timeoutSecs=timeoutSecs, noPoll=True, **kwargs)
h2o_jobs.pollStatsWhileBusy(timeoutSecs=timeoutSecs, retryDelaySecs=5)
elapsed = time.time() - start
print "RF end on ", csvPathname, "took", elapsed, "seconds.", "%d pct. of timeout" % (
(elapsed / timeoutSecs) * 100
)
job_key = rfv["job_key"]
model_key = rfv["destination_key"]
rfv = h2o_cmd.runRFView(
data_key=parseResult["destination_key"], model_key=model_key, timeoutSecs=timeoutSecs, retryDelaySecs=1
)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv)
# hmm..just using defaults above in RF?
self.assertLess(
classification_error,
4.8,
"train.csv should have full classification error: %s < 4.8" % classification_error,
)
print "Trial #", trial, "completed"
def test_rf_big1_nopoll(self):
csvFilename = 'hhp_107_01.data.gz'
csvPathname = h2o.find_file("smalldata/" + csvFilename)
key2 = csvFilename + ".hex"
print "\n" + csvPathname
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname,
key2=key2,
timeoutSecs=15)
rfViewInitial = []
# dispatch multiple jobs back to back
for jobDispatch in range(3):
start = time.time()
kwargs = {}
if OVERWRITE_RF_MODEL:
print "Since we're overwriting here, we have to wait for each to complete nopoll=False"
model_key = 'RF_model'
else:
model_key = 'RF_model' + str(jobDispatch)
kwargs['ntree'] = 7
if OVERWRITE_RF_MODEL:
print "Change the number of trees, while keeping the rf model key name the same"
print "Checks that we correctly overwrite previous rf model"
kwargs['ntree'] += 1
kwargs['seed'] = random.randint(0, sys.maxint)
# FIX! what model keys do these get?
randomNode = h2o.nodes[random.randint(0, len(h2o.nodes) - 1)]
h2o_cmd.runRFOnly(node=randomNode,
parseKey=parseKey,
model_key=model_key,
timeoutSecs=300,
noPoll=False if OVERWRITE_RF_MODEL else True,
**kwargs)
# FIX! are these already in there?
rfView = {}
rfView['data_key'] = key2
rfView['model_key'] = model_key
rfView['ntree'] = kwargs['ntree']
rfViewInitial.append(rfView)
print "rf job dispatch end on ", csvPathname, 'took', time.time(
) - start, 'seconds'
print "\njobDispatch #", jobDispatch
h2o_jobs.pollWaitJobs(pattern='RF_model',
timeoutSecs=300,
pollTimeoutSecs=10,
retryDelaySecs=5)
# we saved the initial response?
# if we do another poll they should be done now, and better to get it that
# way rather than the inspect (to match what simpleCheckGLM is expected
first = None
print "rfViewInitial", rfViewInitial
for rfView in rfViewInitial:
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['data_key']
model_key = rfView['model_key']
ntree = rfView['ntree']
# a = h2o.nodes[0].random_forest_view(data_key, model_key, noPoll=True)
print "Temporary hack: need to do two rf views minimum, to complete a RF (confusion matrix creation)"
# allow it to poll to complete
rfViewResult = h2o_cmd.runRFView(None,
data_key,
model_key,
ntree=ntree,
timeoutSecs=60,
noPoll=False)
if first is None: # we'll use this to compare the others
first = rfViewResult.copy()
firstModelKey = model_key
print "first", h2o.dump_json(first)
else:
print "Comparing", model_key, "to", firstModelKey
df = h2o_util.JsonDiff(rfViewResult,
first,
vice_versa=True,
with_values=True)
print "df.difference:", h2o.dump_json(df.difference)
def test_rf_covtype20x_fvec(self):
h2o.beta_features = True
importFolderPath = 'standard'
if DO_SMALL:
csvFilenameTrain = 'covtype.data'
hex_key = 'covtype1x.data.A.hex'
else:
csvFilenameTrain = 'covtype20x.data'
hex_key = 'covtype20x.data.A.hex'
csvPathname = importFolderPath + "/" + csvFilenameTrain
parseResultTrain = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, hex_key=hex_key, timeoutSecs=500)
inspect = h2o_cmd.runInspect(key=parseResultTrain['destination_key'])
dataKeyTrain = parseResultTrain['destination_key']
print "Parse end", dataKeyTrain
# have to re import since source key is gone
# we could just copy the key, but sometimes we change the test/train data to covtype.data
if DO_SMALL:
csvFilenameTest = 'covtype.data'
hex_key = 'covtype1x.data.B.hex'
dataKeyTest2 = 'covtype1x.data.C.hex'
else:
csvFilenameTest = 'covtype20x.data'
hex_key = 'covtype20x.data.B.hex'
dataKeyTest2 = 'covtype20x.data.C.hex'
csvPathname = importFolderPath + "/" + csvFilenameTest
parseResultTest = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, hex_key=hex_key, timeoutSecs=500)
print "Parse result['destination_key']:", parseResultTest['destination_key']
inspect = h2o_cmd.runInspect(key=parseResultTest['destination_key'])
dataKeyTest = parseResultTest['destination_key']
print "Parse end", dataKeyTest
# make a 3rd key so the predict is uncached too!
execExpr = dataKeyTest2 + "=" + dataKeyTest
if h2o.beta_features:
kwargs = {'str': execExpr, 'timeoutSecs': 15}
else:
kwargs = {'expression': execExpr, 'timeoutSecs': 15}
resultExec = h2o_cmd.runExec(**kwargs)
# train
# this does RFView to understand when RF completes, so the time reported for RFView here, should be
# considered the "first RFView" times..subsequent have some caching?.
# unless the no_confusion_matrix works
# params is mutable. This is default.
if h2o.beta_features:
paramDict = drf2ParamDict
params = {
'ntrees': 20,
'destination_key': 'RF_model'
}
else:
paramDict = drf1ParamDict
params = {
'ntree': 20,
'out_of_bag_error_estimate': 1,
'model_key': 'RF_model'
}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
if h2o.beta_features:
timeoutSecs = 30 + kwargs['ntrees'] * 60
else:
timeoutSecs = 30 + kwargs['ntree'] * 60
start = time.time()
rf = h2o_cmd.runRF(parseResult=parseResultTrain,
timeoutSecs=timeoutSecs, retryDelaySecs=1, **kwargs)
print "rf job end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
print "\nRFView start after job completion"
if h2o.beta_features:
model_key = kwargs['destination_key']
ntree = kwargs['ntrees']
else:
model_key = kwargs['model_key']
ntree = kwargs['ntree']
start = time.time()
# this does the RFModel view for v2. but only model_key is used. Data doesn't matter? (nor ntree)
h2o_cmd.runRFView(None, dataKeyTrain, model_key, ntree=ntree, timeoutSecs=timeoutSecs)
print "First rfview end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
for trial in range(1):
# scoring
start = time.time()
rfView = h2o_cmd.runRFView(None, dataKeyTest,
model_key, ntree=ntree, timeoutSecs=timeoutSecs, out_of_bag_error_estimate=0, retryDelaySecs=1)
print "rfview", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
self.assertAlmostEqual(classification_error, 50, delta=50,
msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest2)
print "predict", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
parseKey = parseResultTrain['destination_key']
rfModelKey = rfView['drf_model']['_key']
predictKey = 'Predict.hex'
start = time.time()
predictResult = h2o_cmd.runPredict(
data_key=parseKey,
model_key=rfModelKey,
destination_key=predictKey,
timeoutSecs=timeoutSecs)
predictCMResult = h2o.nodes[0].predict_confusion_matrix(
actual=parseKey,
vactual='C54',
predict=predictKey,
vpredict='predict',
)
cm = predictCMResult['cm']
# These will move into the h2o_gbm.py
pctWrong = h2o_gbm.pp_cm_summary(cm);
print "\nTest\n==========\n"
print h2o_gbm.pp_cm(cm)
print "Trial #", trial, "completed"
def test_rf_covtype_train_oobe3(self):
print "\nUse randomFilter to sample the dataset randomly. then slice it"
importFolderPath = "/home/0xdiag/datasets/standard"
csvFilename = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on the normal covtype.data"
parseKey = h2i.parseImportFolderFile(None, csvFilename, importFolderPath, key2=key2,
header=0, timeoutSecs=100)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = num_rows
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
dataKeyTrain = "rTrain"
# FIX! too many digits (10) in the 2nd param seems to cause stack trace
execExpr = dataKeyTest + "=randomFilter(" + key2 + "," + str(pct10) + ",12345)"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTest, timeoutSecs=10)
execExpr = dataKeyTrain + "=randomFilter(" + key2 + "," + str(rowsForPct[9]) + ",12345)"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTrain, timeoutSecs=10)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
for trial in range(1,10):
# always slice from the beginning
rowsToUse = rowsForPct[trial%10]
resultKey = "r" + str(trial)
execExpr = resultKey + "=slice(" + dataKeyTrain + ",1," + str(rowsToUse) + ")"
# execExpr = resultKey + "=slice(" + dataKeyTrain + ",1)"
h2o_exec.exec_expr(None, execExpr, resultKey=resultKey, timeoutSecs=10)
parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(trial)
rfv = h2o_cmd.runRFOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]), delta=0.2)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
kwargs['iterative_cm'] = 1
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
fullScorePctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]), delta=0.2)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/num_rows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
def test_RF_mnist_both(self):
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist*gz'),
# to see results a 2nd time
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist*gz'),
]
# IMPORT**********************************************
# since H2O deletes the source key, we should re-import every iteration if we re-use the src in the list
(importFolderResult, importPattern) = h2i.import_only(bucket='home-0xdiag-datasets', path=importFolderPath + "/*")
### print "importHDFSResult:", h2o.dump_json(importFolderResult)
if 'files' in importFolderResult:
succeededList = importFolderResult['files']
else:
succeededList = importFolderResult['succeeded']
### print "succeededList:", h2o.dump_json(succeededList)
self.assertGreater(len(succeededList),1,"Should see more than 1 files in the import?")
# why does this hang? can't look at storeview after import?
print "\nTrying StoreView after the import folder"
h2o_cmd.runStoreView(timeoutSecs=30)
trial = 0
allDelta = []
for (trainCsvFilename, testCsvFilename, timeoutSecs, rfSeed, parsePattern) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath+"/"+testCsvFilename,
hex_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
print "Use multi-file parse to grab both the mnist_testing.csv.gz and mnist_training.csv.gz for training"
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath+"/"+parsePattern,
hex_key=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
# print "This is the 'ignore=' we'll use"
# no longer use. depend on h2o to get it right.
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, forRF=True)
ntree = 25
params = {
'response_variable': 0,
# 'ignore': ignore_x,
'ntree': ntree,
'iterative_cm': 1,
'out_of_bag_error_estimate': 1,
# 'data_key='mnist_training.csv.hex'
'features': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'exclusive_split_limit': None,
'depth': 2147483647,
'stat_type': 'ENTROPY',
'sampling_strategy': 'RANDOM',
'sample': 67,
# 'model_key': '__RFModel_7055e6cf-a0de-44db-b165-f5994730ac77',
'model_key': 'RF_model',
'bin_limit': 1024,
# 'seed': 784834182943470027,
'use_non_local_data': 1,
# 'class_weights': '0=1.0,1=1.0,2=1.0,3=1.0,4=1.0,5=1.0,6=1.0,7=1.0,8=1.0,9=1.0',
}
if rfSeed is None:
params['seed'] = random.randint(0,sys.maxint)
else:
params['seed'] = rfSeed
print "RF seed:", params['seed']
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRF(parseResult=parseResult, rfView=True,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfView = h2o_cmd.runRFView(data_key=testKey2, model_key=modelKey, ntree=ntree, out_of_bag_error_estimate=0,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
print "classification error is expected to be low because we included the test data in with the training!"
self.assertAlmostEqual(classification_error, 0.0003, delta=0.0003, msg="Classification error %s differs too much" % classification_error)
leaves = rfView['trees']['leaves']
# Expected values are from this case:
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
leavesExpected = {'min': 4996, 'mean': 5064.1, 'max': 5148}
for l in leaves:
# self.assertAlmostEqual(leaves[l], leavesExpected[l], delta=10, msg="leaves %s %s %s differs too much" % (l, leaves[l], leavesExpected[l]))
delta = ((leaves[l] - leavesExpected[l])/leaves[l]) * 100
d = "seed: %s %s leaves: %s expected: %s pct. different %s" % (params['seed'], l, leaves[l], leavesExpected[l], delta)
print d
allDelta.append(d)
depth = rfView['trees']['depth']
depthExpected = {'min': 21, 'mean': 23.8, 'max': 25}
for l in depth:
# self.assertAlmostEqual(depth[l], depthExpected[l], delta=1, msg="depth %s %s %s differs too much" % (l, depth[l], depthExpected[l]))
delta = ((depth[l] - depthExpected[l])/leaves[l]) * 100
d = "seed: %s %s depth: %s expected: %s pct. different %s" % (params['seed'], l, depth[l], depthExpected[l], delta)
print d
allDelta.append(d)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=modelKey, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# Done *******************************************************
print "\nShowing the results again from all the trials, to see variance"
for d in allDelta:
print d
def test_rf_covtype_train_oobe(self):
print "\nMichal will hate me for another file needed: covtype.shuffled.data"
importFolderPath = "/home/0xdiag/datasets"
csvFilename = 'covtype.shuffled.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on the normal covtype.data"
parseKey = h2i.parseImportFolderFile(None, csvFilename, importFolderPath, key2=key2,
header=0, timeoutSecs=180)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = last10
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# this was with 10 trees
# expectTrainPctRightList = [0, 85.27, 88.45, 89.99, 91.11, 91.96, 92.51, 93.03, 93.45, 93.78]
# expectScorePctRightList = [0, 89.10, 91,90, 93.26, 94.25, 94.74, 95.10, 95.42, 95.72, 95.92]
# 0 isn't used
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
# start at 90% rows + 1
execExpr = dataKeyTest + " = slice(" + key2 + "," + str(rowsForPct[9]+1) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTest, timeoutSecs=10)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
for trial in range(1,10):
# always slice from the beginning
rowsToUse = rowsForPct[trial%10]
resultKey = "r" + str(trial)
execExpr = resultKey + " = slice(" + key2 + ",1," + str(rowsToUse) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=resultKey, timeoutSecs=10)
parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(trial)
rfv = h2o_cmd.runRFOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]),
delta=0.2)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
# no_confusion_matrix=1&
# clear_confusion_matrix=1
### dataKeyTest = data_key
kwargs['clear_confusion_matrix'] = 1
kwargs['no_confusion_matrix'] = 0
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
fullScorePctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]),
delta=0.2)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/num_rows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print"expectTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "expectScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
def test_RF_mnist_reals_fvec(self):
h2o.beta_features = True
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
("mnist_reals_training.csv.gz", "mnist_reals_testing.csv.gz", 600),
]
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + testCsvFilename,
hex_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + trainCsvFilename,
hex_key=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, forRF=True)
ntrees = 10
params = {
'response': 'C1',
'ignored_cols_by_name': ignore_x,
'ntrees': ntrees,
'mtries': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'max_depth': 15,
'sample_rate': 0.67,
'destination_key': 'RF_model',
'nbins': 1024,
'seed': 784834182943470027,
'importance': 0,
'balance_classes': 0,
}
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfv = h2o_cmd.runRF(parseResult=parseResult, rfView=True,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfv, **params)
rf_model = rfv['drf_model']
used_trees = rf_model['N']
data_key = rf_model['_dataKey']
model_key = rf_model['_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfv = h2o_cmd.runRFView(data_key=testKey2, model_key=model_key, ntrees=ntrees, out_of_bag_error_estimate=0,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfv, **params)
self.assertAlmostEqual(classification_error, 9, delta=1.0, msg="Classification error %s differs too much" % classification_error)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
def test_rf_log(self):
SYNDATASETS_DIR = h2o.make_syn_dir()
tryList = [
(10000, 10, 'cA', 300),
]
for (rowCount, colCount, hex_key, timeoutSecs) in tryList:
SEEDPERFILE = random.randint(0, sys.maxint)
# CREATE test dataset******************************************************
csvFilename = 'syn_test_' + str(rowCount) + 'x' + str(colCount) + '.csv'
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "Creating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
testParseResult = h2i.import_parse(path=csvPathname, hex_key=hex_key, schema='put', timeoutSecs=10)
print "Test Parse result['destination_key']:", testParseResult['destination_key']
dataKeyTest = testParseResult['destination_key']
# CREATE train dataset******************************************************
csvFilename = 'syn_train_' + str(rowCount) + 'x' + str(colCount) + '.csv'
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "Creating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
trainParseResult = h2i.import_parse(path=csvPathname, hex_key=hex_key, schema='put', timeoutSecs=10)
print "Train Parse result['destination_key']:", trainParseResult['destination_key']
dataKeyTrain = trainParseResult['destination_key']
# RF train******************************************************
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['response_variable'] = colCount
rfv = h2o_cmd.runRF(parseResult=trainParseResult, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
expectTrainPctRight = 98
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRight,\
msg="OOBE: pct. right for training not close enough %6.2f %6.2f"% (oobeTrainPctRight, expectTrainPctRight), delta=1)
# RF score******************************************************
print "Now score with the 2nd random dataset"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
kwargs['iterative_cm'] = 1
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, ntree=ntree)
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
fullScorePctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
expectScorePctRight = 99
self.assertAlmostEqual(fullScorePctRight,expectScorePctRight,
msg="Full: pct. right for scoring not close enough %6.2f %6.2f"% (fullScorePctRight, expectScorePctRight), delta=1)
def test_rfview_score(self):
csvPathnameTrain = h2o.find_dataset(
'UCI/UCI-large/covtype/covtype.data')
print "Train with:", csvPathnameTrain
parseKeyTrain = h2o_cmd.parseFile(csvPathname=csvPathnameTrain,
key2="covtype.hex",
timeoutSecs=15)
dataKeyTrain = parseKeyTrain['destination_key']
csvPathnameTest = h2o.find_dataset(
'UCI/UCI-large/covtype/covtype.data')
print "Test with:", csvPathnameTest
parseKeyTest = h2o_cmd.parseFile(csvPathname=csvPathnameTrain,
key2="covtype.hex",
timeoutSecs=15)
dataKeyTest = parseKeyTest['destination_key']
for trial in range(5):
# params is mutable. This is default.
params = {
'ntree': 13,
'parallel': 1,
'out_of_bag_error_estimate': 0
}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + kwargs['ntree'] * 10 * (kwargs['parallel'] and 1
or 5)
rfv = h2o_cmd.runRFOnly(parseKey=parseKeyTrain,
timeoutSecs=timeoutSecs,
retryDelaySecs=1,
**kwargs)
### print "rf response:", h2o.dump_json(rfv)
model_key = rfv['model_key']
# pop the stuff from kwargs that were passing as params
kwargs.pop('model_key', None)
data_key = rfv['data_key']
kwargs.pop('data_key', None)
ntree = rfv['ntree']
kwargs.pop('ntree', None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
# new web page for predict? throw it in here for now
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 0
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 1
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 1
kwargs['class_weights'] = '1=1,2=2,3=3,4=4,5=5,6=6,7=7'
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
print "Trial #", trial, "completed"
def test_rf_big1_nopoll(self):
csvFilename = 'hhp_107_01.data.gz'
csvPathname = h2o.find_file("smalldata/" + csvFilename)
key2 = csvFilename + ".hex"
print "\n" + csvPathname
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=key2, timeoutSecs=15)
firstRfView = None
# dispatch multiple jobs back to back
for jobDispatch in range(3):
start = time.time()
kwargs = {}
if OVERWRITE_RF_MODEL:
print "Since we're overwriting here, we have to wait for each to complete nopoll=False"
model_key = 'RF_model'
else:
model_key = 'RF_model' + str(jobDispatch)
print "Change the number of trees, while keeping the rf model key name the same"
print "Checks that we correctly overwrite previous rf model"
if OVERWRITE_RF_MODEL:
kwargs['ntree'] = 7 + jobDispatch
else:
kwargs['ntree'] = 7
# don't change the seed if we're overwriting the model. It should get
# different results just from changing the tree count
kwargs['seed'] = random.randint(0, sys.maxint)
# FIX! what model keys do these get?
randomNode = h2o.nodes[random.randint(0,len(h2o.nodes)-1)]
h2o_cmd.runRFOnly(node=randomNode, parseKey=parseKey, model_key=model_key, timeoutSecs=300,
noPoll=True, **kwargs)
# FIX! are these already in there?
rfView = {}
rfView['data_key'] = key2
rfView['model_key'] = model_key
rfView['ntree'] = kwargs['ntree']
print "rf job dispatch end on ", csvPathname, 'took', time.time() - start, 'seconds'
print "\njobDispatch #", jobDispatch
# we're going to compare rf results to previous as we go along (so we save rf view results
h2o_jobs.pollWaitJobs(pattern='RF_model', timeoutSecs=300, pollTimeoutSecs=10, retryDelaySecs=5)
# In this test we're waiting after each one, so we can save the RFView results for comparison to future
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['data_key']
model_key = rfView['model_key']
ntree = rfView['ntree']
# a = h2o.nodes[0].random_forest_view(data_key, model_key, noPoll=True)
print "Temporary hack: need to do two rf views minimum, to complete a RF (confusion matrix creation)"
# allow it to poll to complete
rfViewResult = h2o_cmd.runRFView(None, data_key, model_key, ntree=ntree, timeoutSecs=60, noPoll=False)
if firstRfView is None: # we'll use this to compare the others
firstRfView = rfViewResult.copy()
firstModelKey = model_key
print "firstRfView", h2o.dump_json(firstRfView)
else:
print "Comparing", model_key, "to", firstModelKey
df = h2o_util.JsonDiff(rfViewResult, firstRfView, vice_versa=True, with_values=True)
print "df.difference:", h2o.dump_json(df.difference)
self.assertGreater(len(df.difference), 29,
msg="Want >=30 , not %d differences between the two rfView json responses. %s" % \
(len(df.difference), h2o.dump_json(df.difference)))
def test_rf_covtype_train_oobe3(self):
print "\nUse randomFilter to sample the dataset randomly. then slice it"
importFolderPath = "standard"
csvFilename = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilename
hex_key = csvFilename + ".hex"
print "\nUsing header=0 on the normal covtype.data"
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, hex_key=hex_key,
header=0, timeoutSecs=100)
inspect = h2o_cmd.runInspect(None, parseResult['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = num_rows
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
dataKeyTrain = "rTrain"
# FIX! too many digits (10) in the 2nd param seems to cause stack trace
execExpr = dataKeyTest + "=randomFilter(" + hex_key + "," + str(pct10) + ",12345)"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTest, timeoutSecs=10)
execExpr = dataKeyTrain + "=randomFilter(" + hex_key + "," + str(rowsForPct[9]) + ",12345)"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTrain, timeoutSecs=10)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
for trial in range(1,10):
# always slice from the beginning
rowsToUse = rowsForPct[trial%10]
resultKey = "r" + str(trial)
execExpr = resultKey + "=slice(" + dataKeyTrain + ",1," + str(rowsToUse) + ")"
# execExpr = resultKey + "=slice(" + dataKeyTrain + ",1)"
h2o_exec.exec_expr(None, execExpr, resultKey=resultKey, timeoutSecs=10)
parseResult['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(trial)
rfv = h2o_cmd.runRF(parseResult=parseResult, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]), delta=0.2)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
kwargs['iterative_cm'] = 1
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, ntree=ntree)
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
fullScorePctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]), delta=0.2)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/num_rows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
def test_rf_covtype20x(self):
importFolderPath = '/home/0xdiag/datasets/standard'
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
csvFilenameTrain = 'covtype20x.data'
key2 = 'covtype20x.data.A.hex'
parseKeyTrain = h2i.parseImportFolderFile(None, csvFilenameTrain, importFolderPath, key2=key2, timeoutSecs=500)
print csvFilenameTrain, 'parse time:', parseKeyTrain['response']['time']
inspect = h2o_cmd.runInspect(key=parseKeyTrain['destination_key'])
dataKeyTrain = parseKeyTrain['destination_key']
print "Parse end", dataKeyTrain
# have to re import since source key is gone
# we could just copy the key, but sometimes we change the test/train data to covtype.data
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
csvFilenameTest = 'covtype20x.data'
key2 = 'covtype20x.data.B.hex'
parseKeyTest = h2i.parseImportFolderFile(None, csvFilenameTest, importFolderPath, key2=key2, timeoutSecs=500)
print csvFilenameTest, 'parse time:', parseKeyTest['response']['time']
print "Parse result['destination_key']:", parseKeyTest['destination_key']
inspect = h2o_cmd.runInspect(key=parseKeyTest['destination_key'])
dataKeyTest = parseKeyTest['destination_key']
dataKeyTest2 = 'covtype20x.data.C.hex'
print "Parse end", dataKeyTest
# make a 3rd key so the predict is uncached too!
execExpr = dataKeyTest2 + "=" + dataKeyTest
resultExec = h2o_cmd.runExecOnly(expression=execExpr, timeoutSecs=15)
# train
# this does RFView to understand when RF completes, so the time reported for RFView here, should be
# considered the "first RFView" times..subsequent have some caching?.
# unless the no_confusion_matrix works
# params is mutable. This is default.
print "RF with no_confusion_matrix=1, so we can 'time' the RFView separately after job completion?"
params = {
'ntree': 6,
'parallel': 1,
'out_of_bag_error_estimate': 0,
'no_confusion_matrix': 1,
'model_key': 'RF_model'
}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + kwargs['ntree'] * 60 * (kwargs['parallel'] and 1 or 5)
start = time.time()
rfv = h2o_cmd.runRFOnly(parseKey=parseKeyTrain,
timeoutSecs=timeoutSecs, retryDelaySecs=1, noPoll=True, **kwargs)
print "rf job dispatch end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
### print "rf response:", h2o.dump_json(rfv)
start = time.time()
h2o_jobs.pollWaitJobs(pattern='RF_model', timeoutSecs=180, pollTimeoutSecs=500, retryDelaySecs=5)
print "rf job end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
print "\nRFView start after job completion"
model_key = kwargs['model_key']
ntree = kwargs['ntree']
start = time.time()
h2o_cmd.runRFView(None, dataKeyTrain, model_key, ntree, timeoutSecs)
print "First rfview end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
for trial in range(3):
# scoring
start = time.time()
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree, timeoutSecs, out_of_bag_error_estimate=0, retryDelaySecs=1)
print "rfview", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest2)
print "predict", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
print "Trial #", trial, "completed"
def rf_covtype_train_oobe(self, csvFilename, checkExpectedResults=True):
# the expected results are only for the shuffled version
# since getting 10% samples etc of the smallish dataset will vary between
# shuffled and non-shuffled datasets
importFolderPath = "standard"
csvPathname = importFolderPath + "/" + csvFilename
hex_key = csvFilename + ".hex"
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathname,
hex_key=hex_key,
timeoutSecs=180)
inspect = h2o_cmd.runInspect(key=parseResult['destination_key'])
print "\n" + csvPathname, \
" numRows:", "{:,}".format(inspect['numRows']), \
" numCols:", "{:,}".format(inspect['numCols'])
numCols = inspect['numCols']
numRows = inspect['numRows']
pct10 = int(numRows * .1)
rowsForPct = [i * pct10 for i in range(0, 11)]
# this can be slightly less than 10%
last10 = numRows - rowsForPct[9]
rowsForPct[10] = numRows
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# 0 isn't used
expectTrainPctRightList = [
0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79
]
expectScorePctRightList = [
0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78
]
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
trial = 0
for rowPct in [0.9]:
trial += 1
# Not using this now (did use it for slicing)
rowsToUse = rowsForPct[trial % 10]
resultKey = "r_" + csvFilename + "_" + str(trial)
# just do random split for now
dataKeyTrain = 'rTrain.hex'
dataKeyTest = 'rTest.hex'
h2o_cmd.createTestTrain(hex_key,
dataKeyTrain,
dataKeyTest,
trainPercent=90,
outputClass=4,
outputCol=numCols - 1,
changeToBinomial=not DO_MULTINOMIAL)
sliceResult = {'destination_key': dataKeyTrain}
# adjust timeoutSecs with the number of trees
kwargs = paramDict.copy()
kwargs['destination_key'] = "model_" + csvFilename + "_" + str(
trial)
timeoutSecs = 30 + kwargs['ntrees'] * 20
start = time.time()
rfv = h2o_cmd.runRF(parseResult=sliceResult,
timeoutSecs=timeoutSecs,
**kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
(error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, **kwargs)
# oobeTrainPctRight = 100 * (1.0 - error)
oobeTrainPctRight = 100 - error
if checkExpectedResults:
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]), delta=ALLOWED_DELTA)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%. Note this is silly if we trained on 100% of the data"
print "Or sorted by output class, so that the last 10% is the last few classes"
rf_model = rfv['drf_model']
used_trees = rf_model['N']
data_key = rf_model['_dataKey']
model_key = rf_model['_key']
rfvScoring = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
used_trees,
timeoutSecs,
retryDelaySecs=1,
**kwargs)
(error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfvScoring, **kwargs)
fullScorePctRight = 100 - error
h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
if checkExpectedResults:
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]), delta=ALLOWED_DELTA)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (
rowsToUse * 100.0 / numRows), "pct. of all rows"
actualDelta = [
abs(a - b)
for a, b in zip(expectTrainPctRightList, actualTrainPctRightList)
]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [
abs(a - b)
for a, b in zip(expectScorePctRightList, actualScorePctRightList)
]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
return rfvScoring
def test_rfview_score(self):
# for determinism, I guess we should spit out the seed?
# random.seed(SEED)
SEED = random.randint(0, sys.maxint)
# if you have to force to redo a test
# SEED =
random.seed(SEED)
print "\nUsing random seed:", SEED
csvPathnameTrain = h2o.find_file('smalldata/covtype/covtype.20k.data')
print "Train with:", csvPathnameTrain
parseKeyTrain = h2o_cmd.parseFile(csvPathname=csvPathnameTrain,
key2="covtype.20k.hex",
timeoutSecs=10)
dataKeyTrain = parseKeyTrain['destination_key']
csvPathnameTest = h2o.find_dataset(
'UCI/UCI-large/covtype/covtype.data')
print "Test with:", csvPathnameTest
parseKeyTest = h2o_cmd.parseFile(csvPathname=csvPathnameTrain,
key2="covtype.hex",
timeoutSecs=10)
dataKeyTest = parseKeyTest['destination_key']
for trial in range(5):
# params is mutable. This is default.
params = {
'ntree': 13,
'parallel': 1,
'out_of_bag_error_estimate': 0
}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + 15 * (kwargs['parallel'] and 5 or 10)
rfv = h2o_cmd.runRFOnly(parseKey=parseKeyTrain,
timeoutSecs=timeoutSecs,
retryDelaySecs=1,
**kwargs)
### print "rf response:", h2o.dump_json(rfv)
model_key = rfv['model_key']
# pop the stuff from kwargs that were passing as params
kwargs.pop('model_key', None)
data_key = rfv['data_key']
kwargs.pop('data_key', None)
ntree = rfv['ntree']
kwargs.pop('ntree', None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
# no_confusion_matrix=1&
# clear_confusion_matrix=1
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
kwargs['no_confusion_matrix'] = 0
kwargs['clear_confusion_matrix'] = 0
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
kwargs['no_confusion_matrix'] = 0
kwargs['clear_confusion_matrix'] = 1
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
kwargs['no_confusion_matrix'] = 1
kwargs['clear_confusion_matrix'] = 0
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
kwargs['no_confusion_matrix'] = 1
kwargs['clear_confusion_matrix'] = 1
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
kwargs['no_confusion_matrix'] = 0
kwargs['clear_confusion_matrix'] = 0
kwargs['class_weights'] = '1=1,2=2,3=3,4=4,5=5,6=6,7=7'
h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
print "Trial #", trial, "completed"
def test_rf_change_data_key_fvec(self):
importFolderPath = 'standard'
csvFilenameTrain = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilenameTrain
parseResultTrain = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathname,
timeoutSecs=500)
inspect = h2o_cmd.runInspect(key=parseResultTrain['destination_key'])
dataKeyTrain = parseResultTrain['destination_key']
print "Parse end", dataKeyTrain
# we could train on covtype, and then use covtype20x for test? or vice versa
# parseResult = parseResult
# dataKeyTest = dataKeyTrain
csvFilenameTest = 'covtype20x.data'
csvPathname = importFolderPath + "/" + csvFilenameTest
parseResultTest = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathname,
timeoutSecs=500)
print "Parse result['destination_key']:", parseResultTest[
'destination_key']
inspect = h2o_cmd.runInspect(key=parseResultTest['destination_key'])
dataKeyTest = parseResultTest['destination_key']
print "Parse end", dataKeyTest
# train
# this does RFView to understand when RF completes, so the time reported for RFView here, should be
# considered the "first RFView" times..subsequent have some caching?.
# unless the no_confusion_matrix works
# params is mutable. This is default.
params = {'ntrees': 2, 'destination_key': 'RF_model'}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 100
start = time.time()
rfv = h2o_cmd.runRF(parseResult=parseResultTrain,
timeoutSecs=timeoutSecs,
retryDelaySecs=1,
noPoll=True,
**kwargs)
print "rf job dispatch end on ", dataKeyTrain, 'took', time.time(
) - start, 'seconds'
### print "rf response:", h2o.dump_json(rfv)
start = time.time()
h2o_jobs.pollWaitJobs(pattern='RF_model',
timeoutSecs=360,
pollTimeoutSecs=120,
retryDelaySecs=5)
print "rf job end on ", dataKeyTrain, 'took', time.time(
) - start, 'seconds'
print "\nRFView start after job completion"
model_key = kwargs['destination_key']
ntrees = kwargs['ntrees']
start = time.time()
h2o_cmd.runRFView(None, dataKeyTrain, model_key, ntrees, timeoutSecs)
print "First rfview end on ", dataKeyTrain, 'took', time.time(
) - start, 'seconds'
for trial in range(3):
# scoring
start = time.time()
rfView = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntrees,
timeoutSecs,
out_of_bag_error_estimate=1,
retryDelaySecs=1)
print "rfview", trial, "end on ", dataKeyTest, 'took', time.time(
) - start, 'seconds.'
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntrees)
# FIX! should update this expected classification error
# self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
print "predict", trial, "end on ", dataKeyTest, 'took', time.time(
) - start, 'seconds.'
print "Trial #", trial, "completed"
def test_rf_covtype_train_oobe(self):
print "\nMichal will hate me for another file needed: covtype.shuffled.data"
importFolderPath = "/home/0xdiag/datasets/standard"
csvFilename = 'covtype.shuffled.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on the normal covtype.data"
parseKey = h2i.parseImportFolderFile(None, csvFilename, importFolderPath, key2=key2,
header=0, timeoutSecs=180)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = last10
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# 0 isn't used
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
# start at 90% rows + 1
execExpr = dataKeyTest + " = slice(" + key2 + "," + str(rowsForPct[9]+1) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTest, timeoutSecs=10)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
for trial in range(1,10):
# always slice from the beginning
rowsToUse = rowsForPct[trial%10]
resultKey = "r" + str(trial)
execExpr = resultKey + " = slice(" + key2 + ",1," + str(rowsToUse) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=resultKey, timeoutSecs=10)
parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(trial)
rfv = h2o_cmd.runRFOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]),
delta=0.2)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
# does this still exist?
# no_confusion_matrix=1
kwargs['iterative_cm'] = 1
kwargs['no_confusion_matrix'] = 1
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
h2o.nodes[0].random_forest_predict(model_key=model_key, key=dataKeyTest)
fullScorePctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]),
delta=0.2)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/num_rows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print"expectTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "expectScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
def rf_covtype_train_oobe(self, csvFilename, checkExpectedResults=True):
# the expected results are only for the shuffled version
# since getting 10% samples etc of the smallish dataset will vary between
# shuffled and non-shuffled datasets
importFolderPath = "standard"
csvPathname = importFolderPath + "/" + csvFilename
hex_key = csvFilename + ".hex"
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname,
hex_key=hex_key, timeoutSecs=180)
inspect = h2o_cmd.runInspect(key=parseResult['destination_key'])
print "\n" + csvPathname, \
" numRows:", "{:,}".format(inspect['numRows']), \
" numCols:", "{:,}".format(inspect['numCols'])
numCols = inspect['numCols']
numRows = inspect['numRows']
pct10 = int(numRows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = numRows - rowsForPct[9]
rowsForPct[10] = numRows
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# 0 isn't used
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
trial = 0
for rowPct in [0.9]:
trial += 1
# Not using this now (did use it for slicing)
rowsToUse = rowsForPct[trial%10]
resultKey = "r_" + csvFilename + "_" + str(trial)
# just do random split for now
dataKeyTrain = 'rTrain.hex'
dataKeyTest = 'rTest.hex'
createTestTrain(hex_key, dataKeyTrain, dataKeyTest, percent=0.90, outputClass=4, numCols=numCols)
sliceResult = {'destination_key': dataKeyTrain}
# adjust timeoutSecs with the number of trees
kwargs = paramDict.copy()
kwargs['destination_key'] = "model_" + csvFilename + "_" + str(trial)
timeoutSecs = 30 + kwargs['ntrees'] * 20
start = time.time()
rfv = h2o_cmd.runRF(parseResult=sliceResult, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
(error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv)
# oobeTrainPctRight = 100 * (1.0 - error)
oobeTrainPctRight = 100 - error
if checkExpectedResults:
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]), delta=ALLOWED_DELTA)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%. Note this is silly if we trained on 100% of the data"
print "Or sorted by output class, so that the last 10% is the last few classes"
rf_model = rfv['drf_model']
used_trees = rf_model['N']
data_key = rf_model['_dataKey']
model_key = rf_model['_selfKey']
rfvScoring = h2o_cmd.runRFView(None, dataKeyTest, model_key, used_trees,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
(error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfvScoring)
fullScorePctRight = 100 - error
h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
if checkExpectedResults:
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]), delta=ALLOWED_DELTA)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/numRows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
# return the last rfv done during training
return rfv
def test_mnist8m_RF_bench(self):
overallWallStart = time.time()
importFolderPath = '/home/0xdiag/datasets/mnist/mnist8m'
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
output = None
if not os.path.exists('rfbench.csv'):
output = open('rfbench.csv', 'w')
output.write(','.join(csv_header) + '\n')
else:
output = open('rfbench.csv', 'a')
csvWrt = csv.DictWriter(output,
fieldnames=csv_header,
restval=None,
dialect='excel',
extrasaction='ignore',
delimiter=',')
try:
java_heap_GB = h2o.nodes[0].java_heap_GB
#Train File Parsing#
trainParseWallStart = time.time()
print "Training file is: ", files['train']
csvPathname = files['train']
destKey = files['train'] + '.hex'
parseKey = h2i.parseImportFolderFile(None,
csvPathname,
importFolderPath,
key2=destKey,
timeoutSecs=3600,
retryDelaySecs=5,
pollTimeoutSecs=120)
trainParseWallTime = time.time() - trainParseWallStart
#End Train File Parse#
inspect = h2o.nodes[0].inspect(parseKey['destination_key'])
row = {
'java_heap_GB': java_heap_GB,
'dataset': 'mnist8m',
'nTrainRows': inspect['num_rows'],
'nCols': inspect['num_cols'],
#'nIgnoredCols':nIgnoredCols,'ignoredCols':ignoredCols,
'trainParseWallTime': trainParseWallTime
}
#RF+RFView (train)#
kwargs = configs.copy()
trainRFStart = time.time()
rfView = h2o_cmd.runRFOnly(parseKey=parseKey,
rfView=True,
timeoutSecs=3600,
pollTimeoutSecs=60,
retryDelaySecs=2,
**kwargs)
trainViewTime = time.time() - trainRFStart
#End RF+RFView (train)#
row.update({'trainViewTime': trainViewTime})
h2o_rf.simpleCheckRFView(None, rfView, **kwargs)
modelKey = rfView['model_key']
#Test File Parsing#
testParseWallStart = time.time()
print "Testing file is: ", files['test']
csvPathname = files['test']
destKey = files['test'] + '.hex'
parseKey = h2i.parseImportFolderFile(None,
csvPathname,
importFolderPath,
key2=destKey,
timeoutSecs=3600,
retryDelaySecs=5,
pollTimeoutSecs=120)
testParseWallTime = time.time() - testParseWallStart
#End Test File Parse#
inspect = h2o.nodes[0].inspect(parseKey['destination_key'])
row.update({'nTestRows': inspect['num_rows']})
row.update({'testParseWallTime': testParseWallTime})
modelKey = rfView['model_key']
#RFView (score on test)#
kwargs = configs.copy()
testRFStart = time.time()
kwargs.update({
'model_key': modelKey,
'ntree': 10,
'out_of_bag_error_estimate': 1
})
rfView = h2o_cmd.runRFView(data_key=destKey,
timeoutSecs=3600,
doSimpleCheck=False,
**kwargs)
testViewTime = time.time() - testRFStart
#End RFView (score on test)#
pprint(rfView)
errRate = rfView['confusion_matrix']['classification_error']
row.update({'testViewTime': testViewTime})
overallWallTime = time.time() - overallWallStart
row.update({'overallWallTime': overallWallTime})
row.update({'errRate': errRate})
print row
csvWrt.writerow(row)
#h2o.nodes[0].remove_key(k)
finally:
output.close()
def test_rf_covtype_fvec(self):
importFolderPath = "/home/0xdiag/datasets/standard"
csvFilename = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on the normal covtype.data"
parseKey = h2i.parseImportFolderFile(None,
csvFilename,
importFolderPath,
key2=key2,
header=0,
timeoutSecs=180)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
rfViewInitial = []
for jobDispatch in range(1):
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(jobDispatch)
# don't poll for fvec
rfResult = h2o_cmd.runRFOnly(parseKey=parseKey,
timeoutSecs=timeoutSecs,
noPoll=True,
rfView=False,
**kwargs)
elapsed = time.time() - start
print "RF dispatch end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
print h2o.dump_json(rfResult)
# FIX! are these already in there?
rfView = {}
rfView['data_key'] = key2
rfView['model_key'] = kwargs['model_key']
rfView['ntree'] = kwargs['ntree']
rfViewInitial.append(rfView)
print "rf job dispatch end on ", csvPathname, 'took', time.time(
) - start, 'seconds'
print "\njobDispatch #", jobDispatch
h2o_jobs.pollWaitJobs(pattern='RF_model',
timeoutSecs=180,
pollTimeoutSecs=120,
retryDelaySecs=5)
# we saved the initial response?
# if we do another poll they should be done now, and better to get it that
# way rather than the inspect (to match what simpleCheckGLM is expected
print "rfViewInitial", rfViewInitial
for rfView in rfViewInitial:
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['data_key']
model_key = rfView['model_key']
ntree = rfView['ntree']
# allow it to poll to complete
rfViewResult = h2o_cmd.runRFView(None,
data_key,
model_key,
ntree=ntree,
timeoutSecs=60,
noPoll=False)
def test_RF_mnist_both(self):
csvFilelist = [
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
]
# IMPORT**********************************************
trial = 0
allDelta = []
importFolderPath = "mnist"
for (trainCsvFilename, testCsvFilename, timeoutSecs, rfSeed, parsePattern) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
csvPathname = importFolderPath + "/" + testCsvFilename
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, schema='local',
hex_key=testKey, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
print "Use multi-file parse to grab both the mnist_testing.csv.gz and mnist_training.csv.gz for training"
trainKey = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
csvPathname = importFolderPath + "/" + parsePattern
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=parsePattern, schema='local', timeoutSecs=300)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
y = 0 # first column is pixel value
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, returnIgnoreX=True)
ntree = 100
params = {
'response': y,
'ignored_cols': ignore_x,
'ntrees': ntree,
# 'data_key='mnist_training.csv.hex'
'mtries': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'max_depth': 500,
'destination_key': 'RF_model',
'nbins': 1024,
}
if rfSeed is None:
params['seed'] = random.randint(0,sys.maxint)
else:
params['seed'] = rfSeed
print "RF seed:", params['seed']
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRF(parseResult=parseResult,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
kwargs = {'response': y}
rfView = h2o_cmd.runRFView(data_key=testKey, model_key=modelKey,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
print "classification error is expected to be low because we included the test data in with the training!"
self.assertAlmostEqual(classification_error, 0.028, delta=0.01, msg="Classification error %s differs too much" % classification_error)
treeStats = rfView['drf_model']['treesStats']
# Expected values are from this case:
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
expected = {'minLeaves': 4996, 'meanLeaves': 5064.1, 'maxLeaves': 5148}
expected += {'minDepth': 21, 'meanDepth': 23.8, 'maxDepth': 25}
for key in expected:
delta = ((expected[key]- actual[key])/expected[key]) * 100
d = "seed: %s %s %s %s %s pct. different %s" % (params['seed'], key, actual[key], expected[key], delta)
print d
allDelta.append(d)
# FIX! should change this to an assert?
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=modelKey, data_key=testKey, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# Done *******************************************************
print "\nShowing the results again from all the trials, to see variance"
for d in allDelta:
print d
def run_rf(files,configs):
overallWallStart = time.time()
output = None
#if not os.path.exists('rfbench.csv'):
# output = open('rfbench.csv','w')
# output.write(','.join(csv_header)+'\n')
#else:
# output = open('rfbench.csv','a')
#csvWrt = csv.DictWriter(output, fieldnames=csv_header, restval=None,
# dialect='excel', extrasaction='ignore',delimiter=',')
#csvWrt.writeheader()
try:
java_heap_GB = h2o.nodes[0].java_heap_GB
#Train File Parsing#
trainParseWallStart = time.time()
print "Training file is: ", files['train']
importFolderPath = "mnist/mnist8m"
csvPathname = importFolderPath + "/" + files['train']
hex_key = files['train'] + '.hex'
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, hex_key=hex_key
timeoutSecs=3600,retryDelaySecs=5,pollTimeoutSecs=120)
trainParseWallTime = time.time() - trainParseWallStart
#End Train File Parse#
inspect = h2o.nodes[0].inspect(parseResult['destination_key'])
row = {'java_heap_GB':java_heap_GB,'dataset':'mnist8m',
'nTrainRows': inspect['num_rows'],'nCols':inspect['num_cols'],
#'nIgnoredCols':nIgnoredCols,'ignoredCols':ignoredCols,
'trainParseWallTime':trainParseWallTime}
#RF+RFView (train)#
kwargs = configs.copy()
trainRFStart = time.time()
rfView = h2o_cmd.runRFOnly(parseResult=parseResult,rfView=True,
timeoutSecs= 3600,pollTimeoutSecs= 60,retryDelaySecs = 2, **kwargs)
trainViewTime = time.time() - trainRFStart
#End RF+RFView (train)#
row.update({'trainViewTime':trainViewTime})
h2o_rf.simpleCheckRFView(None, rfView, **kwargs)
modelKey = rfView['model_key']
#Test File Parsing#
testParseWallStart = time.time()
print "Testing file is: ", files['test']
importFolderPath = "mnist/mnist8m"
csvPathname = importFolderPath + "/" + files['test']
hex_key = files['test'] + '.hex'
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, hex_key=hex_key
timeoutSecs=3600,retryDelaySecs=5,pollTimeoutSecs=120)
testParseWallTime = time.time() - testParseWallStart
#End Test File Parse#
inspect = h2o.nodes[0].inspect(parseResult['destination_key'])
row.update({'nTestRows':inspect['num_rows']})
row.update({'testParseWallTime':testParseWallTime})
modelKey = rfView['model_key']
#RFView (score on test)#
kwargs = configs.copy()
testRFStart = time.time()
kwargs.update({'model_key':modelKey,'ntree':10})
rfView = h2o_cmd.runRFView(data_key=hex_key,timeoutSecs=180,
doSimpleCheck=False,**kwargs)
testViewTime = time.time() - testRFStart
#End RFView (score on test)#
pprint(rfView)
errRate = rfView['confusion_matrix']['classification_error']
row.update({'testViewTime':testViewTime})
overallWallTime = time.time() - overallWallStart
row.update({'overallWallTime':overallWallTime})
row.update({'errRate':errRate})
print row
#csvWrt.writerow(row)
#h2o.nodes[0].remove_key(k)
finally:
output.close()
def test_RF_mnist(self):
importFolderPath = "/home/0xdiag/datasets/mnist"
csvFilelist = [
# ("mnist_testing.csv.gz", "mnist_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_testing.csv.gz", "mnist_testing.csv.gz", 600),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600),
]
# IMPORT**********************************************
# since H2O deletes the source key, we should re-import every iteration if we re-use the src in the list
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
### print "importHDFSResult:", h2o.dump_json(importFolderResult)
if 'files' in importFolderResult:
succeededList = importFolderResult['files']
else:
succeededList = importFolderResult['succeeded']
### print "succeededList:", h2o.dump_json(succeededList)
self.assertGreater(len(succeededList),1,"Should see more than 1 files in the import?")
# why does this hang? can't look at storeview after import?
print "\nTrying StoreView after the import folder"
h2o_cmd.runStoreView(timeoutSecs=30)
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseKey = h2i.parseImportFolderFile(None, testCsvFilename, importFolderPath,
key2=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseKey['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseKey['destination_key'], timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseKey = h2i.parseImportFolderFile(None, trainCsvFilename, importFolderPath,
key2=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseKey['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseKey['destination_key'], timeoutSecs=300, forRF=True)
ntree = 10
params = {
'response_variable': 0,
'ignore': ignore_x,
'ntree': ntree,
'iterative_cm': 1,
'out_of_bag_error_estimate': 1,
# 'data_key='mnist_training.csv.hex'
'features': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'exclusive_split_limit': None,
'depth': 2147483647,
'stat_type': 'ENTROPY',
'sampling_strategy': 'RANDOM',
'sample': 67,
# 'model_key': '__RFModel_7055e6cf-a0de-44db-b165-f5994730ac77',
'model_key': 'RF_model',
'bin_limit': 1024,
'seed': 784834182943470027,
'parallel': 1,
'use_non_local_data': 0,
'class_weights': '0=1.0,1=1.0,2=1.0,3=1.0,4=1.0,5=1.0,6=1.0,7=1.0,8=1.0,9=1.0',
}
kwargs = params.copy()
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRFOnly(parseKey=parseKey, rfView=True,
timeoutSecs=timeoutSecs, pollTimeoutsecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfView = h2o_cmd.runRFView(data_key=testKey2, model_key=modelKey, ntree=ntree, out_of_bag_error_estimate=0,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=modelKey, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
def test_rf_change_data_key_fvec(self):
h2o.beta_features = True
importFolderPath = 'standard'
csvFilenameTrain = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilenameTrain
parseResultTrain = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, timeoutSecs=500)
inspect = h2o_cmd.runInspect(key=parseResultTrain['destination_key'])
dataKeyTrain = parseResultTrain['destination_key']
print "Parse end", dataKeyTrain
# we could train on covtype, and then use covtype20x for test? or vice versa
# parseResult = parseResult
# dataKeyTest = dataKeyTrain
csvFilenameTest = 'covtype20x.data'
csvPathname = importFolderPath + "/" + csvFilenameTest
parseResultTest = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, timeoutSecs=500)
print "Parse result['destination_key']:", parseResultTest['destination_key']
inspect = h2o_cmd.runInspect(key=parseResultTest['destination_key'])
dataKeyTest = parseResultTest['destination_key']
print "Parse end", dataKeyTest
# train
# this does RFView to understand when RF completes, so the time reported for RFView here, should be
# considered the "first RFView" times..subsequent have some caching?.
# unless the no_confusion_matrix works
# params is mutable. This is default.
params = {
'ntrees': 6,
'destination_key': 'RF_model'
}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + kwargs['ntrees'] * 60
start = time.time()
rfv = h2o_cmd.runRF(parseResult=parseResultTrain,
timeoutSecs=timeoutSecs, retryDelaySecs=1, noPoll=True, **kwargs)
print "rf job dispatch end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
### print "rf response:", h2o.dump_json(rfv)
start = time.time()
h2o_jobs.pollWaitJobs(pattern='RF_model', timeoutSecs=180, pollTimeoutSecs=120, retryDelaySecs=5)
print "rf job end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
print "\nRFView start after job completion"
model_key = kwargs['destination_key']
ntrees = kwargs['ntrees']
start = time.time()
h2o_cmd.runRFView(None, dataKeyTrain, model_key, ntrees, timeoutSecs)
print "First rfview end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
for trial in range(3):
# scoring
start = time.time()
rfView = h2o_cmd.runRFView(None, dataKeyTest,
model_key, ntrees, timeoutSecs, out_of_bag_error_estimate=1, retryDelaySecs=1)
print "rfview", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntrees)
# FIX! should update this expected classification error
# self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
print "predict", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
print "Trial #", trial, "completed"
def test_rf_mnist_both_fvec(self):
h2o.beta_features = True
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist*gz'),
# to see results a 2nd time
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist*gz'),
]
# IMPORT**********************************************
# since H2O deletes the source key, we should re-import every iteration if we re-use the src in the list
(importFolderResult, importPattern) = h2i.import_only(bucket='home-0xdiag-datasets', path=importFolderPath + "/*")
### print "importHDFSResult:", h2o.dump_json(importFolderResult)
if 'files' in importFolderResult:
succeededList = importFolderResult['files']
else:
succeededList = importFolderResult['succeeded']
### print "succeededList:", h2o.dump_json(succeededList)
self.assertGreater(len(succeededList),1,"Should see more than 1 files in the import?")
# why does this hang? can't look at storeview after import?
print "\nTrying StoreView after the import folder"
h2o_cmd.runStoreView(timeoutSecs=30)
trial = 0
allDelta = []
for (trainCsvFilename, testCsvFilename, timeoutSecs, rfSeed, parsePattern) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath+"/"+testCsvFilename,
hex_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
print "Use multi-file parse to grab both the mnist_testing.csv.gz and mnist_training.csv.gz for training"
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath+"/"+parsePattern,
hex_key=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
print "Not using ignore from this..have to adjust cols?"
h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, forRF=True)
ntree = 2
params = {
'response': 'C1',
# 'ignored_cols_by_name': ignore_x,
'ntrees': ntree,
'mtries': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'max_depth': 20,
'sample_rate': 0.67,
'destination_key': 'RF_model',
'nbins': 100,
'importance': 0,
'balance_classes': 0,
}
if rfSeed is None:
params['seed'] = random.randint(0,sys.maxint)
else:
params['seed'] = rfSeed
print "RF seed:", params['seed']
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRF(parseResult=parseResult, rfView=True,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# print 'rfView:', h2o.dump_json(rfView)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['drf_model']['_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response': y}
rfView = h2o_cmd.runRFView(data_key=testKey2, model_key=modelKey, ntree=ntree, out_of_bag_error_estimate=0,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
# training and test data are unique, so error won't be low?
# self.assertAlmostEqual(classification_error, 0.0003, delta=0.0003, msg="Classification error %s differs too much" % classification_error)
leaves = {
'min': rfView['drf_model']['treeStats']['minLeaves'],
'mean': rfView['drf_model']['treeStats']['meanLeaves'],
'max': rfView['drf_model']['treeStats']['maxLeaves'],
}
# Expected values are from this case:
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
leavesExpected = {'min': 537, 'mean': 1118.05, 'max': 1701}
for l in leaves:
# self.assertAlmostEqual(leaves[l], leavesExpected[l], delta=10, msg="leaves %s %s %s differs too much" % (l, leaves[l], leavesExpected[l]))
delta = ((leaves[l] - leavesExpected[l])/leaves[l]) * 100
d = "seed: %s leaves %s %s %s pct. different %s" % (params['seed'], l, leaves[l], leavesExpected[l], delta)
print d
allDelta.append(d)
depth = {
'min': rfView['drf_model']['treeStats']['minDepth'],
'mean': rfView['drf_model']['treeStats']['meanDepth'],
'max': rfView['drf_model']['treeStats']['maxDepth'],
}
depthExpected = {'min': 20, 'mean': 20, 'max': 20}
for l in depth:
# self.assertAlmostEqual(depth[l], depthExpected[l], delta=1, msg="depth %s %s %s differs too much" % (l, depth[l], depthExpected[l]))
delta = ((depth[l] - depthExpected[l])/leaves[l]) * 100
d = "seed: %s depth %s %s %s pct. different %s" % (params['seed'], l, depth[l], depthExpected[l], delta)
print d
allDelta.append(d)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=modelKey, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# Done *******************************************************
print "\nShowing the results again from all the trials, to see variance"
for d in allDelta:
print d
def test_rf_log_fvec(self):
h2o.beta_features = True
SYNDATASETS_DIR = h2o.make_syn_dir()
tryList = [
(10000, 100, 'cA', 300),
]
for (rowCount, colCount, hex_key, timeoutSecs) in tryList:
SEEDPERFILE = random.randint(0, sys.maxint)
# CREATE test dataset******************************************************
csvFilename = 'syn_test_' + str(rowCount) + 'x' + str(colCount) + '.csv'
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "Creating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
testParseResult = h2i.import_parse(path=csvPathname, hex_key=hex_key, schema='put', timeoutSecs=10)
print "Test Parse result['destination_key']:", testParseResult['destination_key']
dataKeyTest = testParseResult['destination_key']
# CREATE train dataset******************************************************
csvFilename = 'syn_train_' + str(rowCount) + 'x' + str(colCount) + '.csv'
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "Creating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
trainParseResult = h2i.import_parse(path=csvPathname, hex_key=hex_key, schema='put', timeoutSecs=10)
print "Train Parse result['destination_key']:", trainParseResult['destination_key']
dataKeyTrain = trainParseResult['destination_key']
# RF train******************************************************
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntrees'] * 20
start = time.time()
# do oobe
kwargs['response'] = "C" + str(colCount+1)
rfv = h2o_cmd.runRF(parseResult=trainParseResult, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
rf_model = rfv['drf_model']
used_trees = rf_model['N']
data_key = rf_model['_dataKey']
model_key = rf_model['_key']
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, ntree=used_trees)
oobeTrainPctRight = 100.0 - classification_error
expectTrainPctRight = 94
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRight,\
msg="OOBE: pct. right for training not close enough %6.2f %6.2f"% (oobeTrainPctRight, expectTrainPctRight), delta=5)
# RF score******************************************************
print "Now score with the 2nd random dataset"
rfv = h2o_cmd.runRFView(data_key=dataKeyTest, model_key=model_key,
timeoutSecs=timeoutSecs, retryDelaySecs=1)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, ntree=used_trees)
self.assertTrue(classification_error<=5.0, msg="Classification error %s too big" % classification_error)
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
fullScorePctRight = 100.0 - classification_error
expectScorePctRight = 94
self.assertTrue(fullScorePctRight >= expectScorePctRight,
msg="Full: pct. right for scoring not close enough %6.2f %6.2f"% (fullScorePctRight, expectScorePctRight), delta=5)
def rf_covtype_train_oobe(self, csvFilename, checkExpectedResults=True):
# the expected results are only for the shuffled version
# since getting 10% samples etc of the smallish dataset will vary between
# shuffled and non-shuffled datasets
importFolderPath = "/home/0xdiag/datasets/standard"
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on", csvFilename
parseKey = h2i.parseImportFolderFile(None,
csvFilename,
importFolderPath,
key2=key2,
header=0,
timeoutSecs=180)
inspect = h2o_cmd.runInspect(key=parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0, 11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = num_rows
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# 0 isn't used
expectTrainPctRightList = [
0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79
]
expectScorePctRightList = [
0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78
]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
# start at 90% rows + 1
execExpr = dataKeyTest + " = slice(" + key2 + "," + str(rowsForPct[9] +
1) + ")"
h2o_exec.exec_expr(None,
execExpr,
resultKey=dataKeyTest,
timeoutSecs=10)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
# don't use the smaller samples..bad error rates, plus for sorted covtype, you can get just one class!
for trial in range(8, 9):
# always slice from the beginning
rowsToUse = rowsForPct[trial % 10]
resultKey = "r_" + csvFilename + "_" + str(trial)
execExpr = resultKey + " = slice(" + key2 + ",1," + str(
rowsToUse) + ")"
h2o_exec.exec_expr(None,
execExpr,
resultKey=resultKey,
timeoutSecs=10)
# hack so the RF will use the sliced result
# FIX! don't use the sliced bit..use the whole data for rf training below
### parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + csvFilename + "_" + str(trial)
# kwargs['model_key'] = "model"
# double check the rows/cols
inspect = h2o_cmd.runInspect(key=parseKey['destination_key'])
h2o_cmd.infoFromInspect(inspect, "going into RF")
start = time.time()
rfv = h2o_cmd.runRFOnly(parseKey=parseKey,
timeoutSecs=timeoutSecs,
**kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (
1.0 - rfv['confusion_matrix']['classification_error'])
if checkExpectedResults:
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]), delta=ALLOWED_DELTA)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%. Note this is silly if we trained on 100% of the data"
print "Or sorted by output class, so that the last 10% is the last few classes"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key', None)
data_key = rfv['data_key']
kwargs.pop('data_key', None)
ntree = rfv['ntree']
kwargs.pop('ntree', None)
kwargs['iterative_cm'] = 1
kwargs['no_confusion_matrix'] = 0
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
# double check the rows/cols
inspect = h2o_cmd.runInspect(key=dataKeyTest)
h2o_cmd.infoFromInspect(inspect, "dataKeyTest")
rfvScoring = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
fullScorePctRight = 100 * (
1.0 - rfvScoring['confusion_matrix']['classification_error'])
if checkExpectedResults:
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]), delta=ALLOWED_DELTA)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (
rowsToUse * 100.0 / num_rows), "pct. of all rows"
actualDelta = [
abs(a - b)
for a, b in zip(expectTrainPctRightList, actualTrainPctRightList)
]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [
abs(a - b)
for a, b in zip(expectScorePctRightList, actualScorePctRightList)
]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
# return the last rfv done during training
return rfv
def test_RF_mnist_both(self):
importFolderPath = "/home/0xdiag/datasets/mnist_repl"
csvFilelist = [
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
("mnist_training.csv.gz", "mnist_testing_0.csv.gz", 600, None,
'*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing_0.csv.gz", 600, None,
'*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing_0.csv.gz", 600, None,
'*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing_0.csv.gz", 600, None,
'*mnist_training*gz'),
]
# IMPORT**********************************************
# since H2O deletes the source key, we should re-import every iteration if we re-use the src in the list
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
### print "importHDFSResult:", h2o.dump_json(importFolderResult)
succeededList = importFolderResult['files']
### print "succeededList:", h2o.dump_json(succeededList)
self.assertGreater(len(succeededList), 1,
"Should see more than 1 files in the import?")
# why does this hang? can't look at storeview after import?
print "\nTrying StoreView after the import folder"
h2o_cmd.runStoreView(timeoutSecs=30)
trial = 0
allDelta = []
for (trainCsvFilename, testCsvFilename, timeoutSecs, rfSeed,
parsePattern) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseKey = h2i.parseImportFolderFile(None,
testCsvFilename,
importFolderPath,
key2=testKey2,
timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseKey['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y,
key=parseKey['destination_key'],
timeoutSecs=300)
# PARSE train****************************************
print "Use multi-file parse to grab both the mnist_testing.csv.gz and mnist_training.csv.gz for training"
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseKey = h2i.parseImportFolderFile(None,
parsePattern,
importFolderPath,
key2=trainKey2,
timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseKey['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(
y,
key=parseKey['destination_key'],
timeoutSecs=300,
forRF=True)
ntree = 100
params = {
'response_variable':
0,
'ignore':
ignore_x,
'ntree':
ntree,
'iterative_cm':
1,
'out_of_bag_error_estimate':
1,
# 'data_key='mnist_training.csv.hex'
'features':
28, # fix because we ignore some cols, which will change the srt(cols) calc?
'exclusive_split_limit':
None,
'depth':
2147483647,
'stat_type':
'ENTROPY',
'sampling_strategy':
'RANDOM',
'sample':
67,
# 'model_key': '__RFModel_7055e6cf-a0de-44db-b165-f5994730ac77',
'model_key':
'RF_model',
'bin_limit':
1024,
# 'seed': 784834182943470027,
'parallel':
1,
'use_non_local_data':
0,
'class_weights':
'0=1.0,1=1.0,2=1.0,3=1.0,4=1.0,5=1.0,6=1.0,7=1.0,8=1.0,9=1.0',
}
if rfSeed is None:
params['seed'] = random.randint(0, sys.maxint)
else:
params['seed'] = rfSeed
print "RF seed:", params['seed']
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRFOnly(parseKey=parseKey,
rfView=False,
timeoutSecs=timeoutSecs,
pollTimeoutsecs=60,
retryDelaySecs=2,
**kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfView = h2o_cmd.runRFView(data_key=testKey2,
model_key=modelKey,
ntree=ntree,
out_of_bag_error_estimate=0,
timeoutSecs=60,
pollTimeoutSecs=60,
noSimpleCheck=False,
**kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
print "classification error is expected to be low because we included the test data in with the training!"
self.assertAlmostEqual(
classification_error,
0.028,
delta=0.01,
msg="Classification error %s differs too much" %
classification_error)
leaves = rfView['trees']['leaves']
# Expected values are from this case:
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
leavesExpected = {'min': 4996, 'mean': 5064.1, 'max': 5148}
for l in leaves:
# self.assertAlmostEqual(leaves[l], leavesExpected[l], delta=10, msg="leaves %s %s %s differs too much" % (l, leaves[l], leavesExpected[l]))
delta = ((leaves[l] - leavesExpected[l]) / leaves[l]) * 100
d = "seed: %s leaves %s %s %s pct. different %s" % (
params['seed'], l, leaves[l], leavesExpected[l], delta)
print d
allDelta.append(d)
depth = rfView['trees']['depth']
depthExpected = {'min': 21, 'mean': 23.8, 'max': 25}
for l in depth:
# self.assertAlmostEqual(depth[l], depthExpected[l], delta=1, msg="depth %s %s %s differs too much" % (l, depth[l], depthExpected[l]))
delta = ((depth[l] - depthExpected[l]) / leaves[l]) * 100
d = "seed: %s depth %s %s %s pct. different %s" % (
params['seed'], l, depth[l], depthExpected[l], delta)
print d
allDelta.append(d)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(
model_key=modelKey, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# Done *******************************************************
print "\nShowing the results again from all the trials, to see variance"
for d in allDelta:
print d
def test_rf_log_fvec(self):
h2o.beta_features = True
SYNDATASETS_DIR = h2o.make_syn_dir()
tryList = [
(10000, 100, 'cA', 300),
]
for (rowCount, colCount, hex_key, timeoutSecs) in tryList:
SEEDPERFILE = random.randint(0, sys.maxint)
# CREATE test dataset******************************************************
csvFilename = 'syn_test_' + str(rowCount) + 'x' + str(
colCount) + '.csv'
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "Creating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
testParseResult = h2i.import_parse(path=csvPathname,
hex_key=hex_key,
schema='put',
timeoutSecs=10)
print "Test Parse result['destination_key']:", testParseResult[
'destination_key']
dataKeyTest = testParseResult['destination_key']
# CREATE train dataset******************************************************
csvFilename = 'syn_train_' + str(rowCount) + 'x' + str(
colCount) + '.csv'
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "Creating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
trainParseResult = h2i.import_parse(path=csvPathname,
hex_key=hex_key,
schema='put',
timeoutSecs=10)
print "Train Parse result['destination_key']:", trainParseResult[
'destination_key']
dataKeyTrain = trainParseResult['destination_key']
# RF train******************************************************
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntrees'] * 20
start = time.time()
# do oobe
kwargs['response'] = "C" + str(colCount + 1)
rfv = h2o_cmd.runRF(parseResult=trainParseResult,
timeoutSecs=timeoutSecs,
**kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
rf_model = rfv['drf_model']
used_trees = rf_model['N']
data_key = rf_model['_dataKey']
model_key = rf_model['_key']
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, ntree=used_trees)
oobeTrainPctRight = 100.0 - classification_error
expectTrainPctRight = 94
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRight,\
msg="OOBE: pct. right for training not close enough %6.2f %6.2f"% (oobeTrainPctRight, expectTrainPctRight), delta=5)
# RF score******************************************************
print "Now score with the 2nd random dataset"
rfv = h2o_cmd.runRFView(data_key=dataKeyTest,
model_key=model_key,
timeoutSecs=timeoutSecs,
retryDelaySecs=1)
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfv, ntree=used_trees)
self.assertTrue(classification_error <= 5.0,
msg="Classification error %s too big" %
classification_error)
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
fullScorePctRight = 100.0 - classification_error
expectScorePctRight = 94
self.assertTrue(
fullScorePctRight >= expectScorePctRight,
msg="Full: pct. right for scoring not close enough %6.2f %6.2f"
% (fullScorePctRight, expectScorePctRight),
delta=5)
def test_rf_covtype_train_oobe(self):
print "\nUse randomBitVector and filter to separate the dataset randomly"
importFolderPath = "/home/0xdiag/datasets"
csvFilename = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
print "\nUsing header=0 on the normal covtype.data"
# don't import it, just so we don't have all the key names cluttering the view all
# in the browser
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=key2, header=0, timeoutSecs=100)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = last10
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# this was with 10 trees
# expectTrainPctRightList = [0, 85.27, 88.45, 89.99, 91.11, 91.96, 92.51, 93.03, 93.45, 93.78]
# expectScorePctRightList = [0, 89.10, 91,90, 93.26, 94.25, 94.74, 95.10, 95.42, 95.72, 95.92]
# 0 isn't used
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
dataKeyTrain = "rTrain"
# start at 90% rows + 1
# randomBitVector(size,selected)
# randomFilter(srcFrame,rows,seed)
# filter(srcFrame,bitVect)
### h2b.browseTheCloud()
# odd. output is byte, all other exec outputs are 8 byte? (at least the ones below?)
execExpr = "rbv=randomBitVector(" + str(num_rows) + "," + str(last10) + ",12345)"
h2o_exec.exec_expr(None, execExpr, resultKey="rbv", timeoutSecs=10)
# complement the bit vector
execExpr = "not_rbv=colSwap(rbv,0,rbv[0]==0?1:0)"
h2o_exec.exec_expr(None, execExpr, resultKey="not_rbv", timeoutSecs=10)
execExpr = dataKeyTest + "=filter(" + key2 + ",rbv)"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTest, timeoutSecs=10)
execExpr = dataKeyTrain + "=filter(" + key2 + ",not_rbv)"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTrain, timeoutSecs=10)
### time.sleep(3600)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
for trial in range(1,10):
# always slice from the beginning
rowsToUse = rowsForPct[trial%10]
resultKey = "r" + str(trial)
execExpr = resultKey + "=slice(" + dataKeyTrain + ",1," + str(rowsToUse) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=resultKey, timeoutSecs=10)
parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(trial)
rfv = h2o_cmd.runRFOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]),
delta=0.2)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
# no_confusion_matrix=1&
# clear_confusion_matrix=1
### dataKeyTest = data_key
kwargs['clear_confusion_matrix'] = 1
kwargs['no_confusion_matrix'] = 0
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
fullScorePctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]),
delta=0.2)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/num_rows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print"expectTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "expectScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
def run_rf(files,configs):
overallWallStart = time.time()
output = None
#if not os.path.exists('rfbench.csv'):
# output = open('rfbench.csv','w')
# output.write(','.join(csv_header)+'\n')
#else:
# output = open('rfbench.csv','a')
#csvWrt = csv.DictWriter(output, fieldnames=csv_header, restval=None,
# dialect='excel', extrasaction='ignore',delimiter=',')
#csvWrt.writeheader()
try:
java_heap_GB = h2o.nodes[0].java_heap_GB
#Train File Parsing#
trainParseWallStart = time.time()
print "Training file is: ", files['train']
importFolderPath = "mnist/mnist8m"
csvPathname = importFolderPath + "/" + files['train']
hex_key = files['train'] + '.hex'
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, schema='local', hex_key=hex_key
timeoutSecs=3600,retryDelaySecs=5,pollTimeoutSecs=120)
trainParseWallTime = time.time() - trainParseWallStart
#End Train File Parse#
inspect = h2o.nodes[0].inspect(parseResult['destination_key'])
row = {'java_heap_GB':java_heap_GB,'dataset':'mnist8m',
'nTrainRows': inspect['numRows'],'nCols':inspect['numCols'],
#'nIgnoredCols':nIgnoredCols,'ignoredCols':ignoredCols,
'trainParseWallTime':trainParseWallTime}
#RF+RFView (train)#
kwargs = configs.copy()
trainRFStart = time.time()
rfView = h2o_cmd.runRF(parseResult=parseResult,rfView=True,
timeoutSecs= 3600,pollTimeoutSecs= 60,retryDelaySecs = 2, **kwargs)
trainViewTime = time.time() - trainRFStart
#End RF+RFView (train)#
row.update({'trainViewTime':trainViewTime})
h2o_rf.simpleCheckRFView(None, rfView, **kwargs)
modelKey = rfView['model_key']
#Test File Parsing#
testParseWallStart = time.time()
print "Testing file is: ", files['test']
importFolderPath = "mnist/mnist8m"
csvPathname = importFolderPath + "/" + files['test']
hex_key = files['test'] + '.hex'
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, schema='local', hex_key=hex_key
timeoutSecs=3600,retryDelaySecs=5,pollTimeoutSecs=120)
testParseWallTime = time.time() - testParseWallStart
#End Test File Parse#
inspect = h2o.nodes[0].inspect(parseResult['destination_key'])
row.update({'nTestRows':inspect['numRows']})
row.update({'testParseWallTime':testParseWallTime})
modelKey = rfView['model_key']
#RFView (score on test)#
kwargs = configs.copy()
testRFStart = time.time()
kwargs.update({'model_key':modelKey,'ntree':10})
rfView = h2o_cmd.runRFView(data_key=hex_key,timeoutSecs=180,
doSimpleCheck=False,**kwargs)
testViewTime = time.time() - testRFStart
#End RFView (score on test)#
pprint(rfView)
errRate = rfView['confusion_matrix']['classification_error']
row.update({'testViewTime':testViewTime})
overallWallTime = time.time() - overallWallStart
row.update({'overallWallTime':overallWallTime})
row.update({'errRate':errRate})
print row
#csvWrt.writerow(row)
#h2o.nodes[0].remove_key(k)
finally:
output.close()
def test_rf_change_data_key(self):
importFolderPath = '/home/0xdiag/datasets/standard'
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
csvFilenameTrain = 'covtype.data'
parseKeyTrain = h2i.parseImportFolderFile(None, csvFilenameTrain, importFolderPath, timeoutSecs=500)
print csvFilenameTrain, 'parse time:', parseKeyTrain['response']['time']
inspect = h2o_cmd.runInspect(key=parseKeyTrain['destination_key'])
dataKeyTrain = parseKeyTrain['destination_key']
print "Parse end", dataKeyTrain
# we could train on covtype, and then use covtype20x for test? or vice versa
# parseKey = parseKey
# dataKeyTest = dataKeyTrain
csvFilenameTest = 'covtype20x.data'
parseKeyTest = h2i.parseImportFolderFile(None, csvFilenameTest, importFolderPath, timeoutSecs=500)
print csvFilenameTest, 'parse time:', parseKeyTest['response']['time']
print "Parse result['destination_key']:", parseKeyTest['destination_key']
inspect = h2o_cmd.runInspect(key=parseKeyTest['destination_key'])
dataKeyTest = parseKeyTest['destination_key']
print "Parse end", dataKeyTest
# train
# this does RFView to understand when RF completes, so the time reported for RFView here, should be
# considered the "first RFView" times..subsequent have some caching?.
# unless the no_confusion_matrix works
# params is mutable. This is default.
print "RF with no_confusion_matrix=1, so we can 'time' the RFView separately after job completion?"
params = {
'ntree': 6,
'parallel': 1,
'out_of_bag_error_estimate': 0,
'no_confusion_matrix': 1,
'model_key': 'RF_model'
}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + kwargs['ntree'] * 60 * (kwargs['parallel'] and 1 or 5)
start = time.time()
rfv = h2o_cmd.runRFOnly(parseKey=parseKeyTrain,
timeoutSecs=timeoutSecs, retryDelaySecs=1, noPoll=True, **kwargs)
print "rf job dispatch end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
### print "rf response:", h2o.dump_json(rfv)
start = time.time()
h2o_jobs.pollWaitJobs(pattern='RF_model', timeoutSecs=180, pollTimeoutSecs=120, retryDelaySecs=5)
print "rf job end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
print "\nRFView start after job completion"
model_key = kwargs['model_key']
ntree = kwargs['ntree']
start = time.time()
h2o_cmd.runRFView(None, dataKeyTrain, model_key, ntree, timeoutSecs)
print "First rfview end on ", dataKeyTrain, 'took', time.time() - start, 'seconds'
for trial in range(3):
# scoring
start = time.time()
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree, timeoutSecs, out_of_bag_error_estimate=1, retryDelaySecs=1)
print "rfview", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
print "predict", trial, "end on ", dataKeyTest, 'took', time.time() - start, 'seconds.'
print "Trial #", trial, "completed"
def test_rf_covtype_train_oobe2(self):
print "\nUse randomBitVector and filter to separate the dataset randomly"
importFolderPath = "/home/0xdiag/datasets/standard"
csvFilename = 'covtype.data'
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on the normal covtype.data"
parseKey = h2i.parseImportFolderFile(None,
csvFilename,
importFolderPath,
key2=key2,
header=0,
timeoutSecs=100)
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0, 11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = last10
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
expectTrainPctRightList = [
0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79
]
expectScorePctRightList = [
0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78
]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
dataKeyTrain = "rTrain"
# start at 90% rows + 1
# randomBitVector(size,selected)
# randomFilter(srcFrame,rows,seed)
# filter(srcFrame,bitVect)
# odd. output is byte, all other exec outputs are 8 byte? (at least the ones below?)
execExpr = "rbv=randomBitVector(" + str(num_rows) + "," + str(
last10) + ",12345)"
h2o_exec.exec_expr(None, execExpr, resultKey="rbv", timeoutSecs=10)
# complement the bit vector
execExpr = "not_rbv=colSwap(rbv,0,rbv[0]==0?1:0)"
h2o_exec.exec_expr(None, execExpr, resultKey="not_rbv", timeoutSecs=10)
execExpr = dataKeyTest + "=filter(" + key2 + ",rbv)"
h2o_exec.exec_expr(None,
execExpr,
resultKey=dataKeyTest,
timeoutSecs=10)
execExpr = dataKeyTrain + "=filter(" + key2 + ",not_rbv)"
h2o_exec.exec_expr(None,
execExpr,
resultKey=dataKeyTrain,
timeoutSecs=10)
### time.sleep(3600)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
for trial in range(1, 10):
# always slice from the beginning
rowsToUse = rowsForPct[trial % 10]
resultKey = "r" + str(trial)
execExpr = resultKey + "=slice(" + dataKeyTrain + ",1," + str(
rowsToUse) + ")"
# execExpr = resultKey + "=slice(" + dataKeyTrain + ",1)"
h2o_exec.exec_expr(None,
execExpr,
resultKey=resultKey,
timeoutSecs=10)
parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
start = time.time()
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + str(trial)
rfv = h2o_cmd.runRFOnly(parseKey=parseKey,
timeoutSecs=timeoutSecs,
**kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (
1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]),
delta=0.2)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key', None)
data_key = rfv['data_key']
kwargs.pop('data_key', None)
ntree = rfv['ntree']
kwargs.pop('ntree', None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
# http://192.168.1.28:54321/GeneratePredictionsPage.html?model_key=__RFModel_0e2531bc-2552-4f65-8a4a-843031b0cb99&key=iris
# http://192.168.1.28:54321/RFView.html?data_key=iris.hex&model_key=__RFModel_0e2531bc-2552-4f65-8a4a-843031b0cb99&ntree=50&response_variable=4&class_weights=Iris-setosa%3D1.0%2CIris-versicolor%3D1.0%2CIris-virginica%3D1.0&out_of_bag_error_estimate=true&iterative_cm=true
kwargs['iterative_cm'] = 1
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
rfv = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
fullScorePctRight = 100 * (
1.0 - rfv['confusion_matrix']['classification_error'])
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]),
delta=0.2)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (
rowsToUse * 100.0 / num_rows), "pct. of all rows"
actualDelta = [
abs(a - b)
for a, b in zip(expectTrainPctRightList, actualTrainPctRightList)
]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [
abs(a - b)
for a, b in zip(expectScorePctRightList, actualScorePctRightList)
]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
def rf_covtype_train_oobe(self, csvFilename, checkExpectedResults=True):
# the expected results are only for the shuffled version
# since getting 10% samples etc of the smallish dataset will vary between
# shuffled and non-shuffled datasets
importFolderPath = "/home/0xdiag/datasets/standard"
csvPathname = importFolderPath + "/" + csvFilename
key2 = csvFilename + ".hex"
h2i.setupImportFolder(None, importFolderPath)
print "\nUsing header=0 on", csvFilename
parseKey = h2i.parseImportFolderFile(None, csvFilename, importFolderPath, key2=key2,
header=0, timeoutSecs=180)
inspect = h2o_cmd.runInspect(key=parseKey['destination_key'])
print "\n" + csvPathname, \
" num_rows:", "{:,}".format(inspect['num_rows']), \
" num_cols:", "{:,}".format(inspect['num_cols'])
# how many rows for each pct?
num_rows = inspect['num_rows']
pct10 = int(num_rows * .1)
rowsForPct = [i * pct10 for i in range(0,11)]
# this can be slightly less than 10%
last10 = num_rows - rowsForPct[9]
rowsForPct[10] = num_rows
# use mod below for picking "rows-to-do" in case we do more than 9 trials
# use 10 if 0 just to see (we copied 10 to 0 above)
rowsForPct[0] = rowsForPct[10]
# 0 isn't used
expectTrainPctRightList = [0, 85.16, 88.45, 90.24, 91.27, 92.03, 92.64, 93.11, 93.48, 93.79]
expectScorePctRightList = [0, 88.81, 91.72, 93.06, 94.02, 94.52, 95.09, 95.41, 95.77, 95.78]
print "Creating the key of the last 10% data, for scoring"
dataKeyTest = "rTest"
# start at 90% rows + 1
execExpr = dataKeyTest + " = slice(" + key2 + "," + str(rowsForPct[9]+1) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=dataKeyTest, timeoutSecs=10)
# keep the 0 entry empty
actualTrainPctRightList = [0]
actualScorePctRightList = [0]
# don't use the smaller samples..bad error rates, plus for sorted covtype, you can get just one class!
for trial in range(8,9):
# always slice from the beginning
rowsToUse = rowsForPct[trial%10]
resultKey = "r_" + csvFilename + "_" + str(trial)
execExpr = resultKey + " = slice(" + key2 + ",1," + str(rowsToUse) + ")"
h2o_exec.exec_expr(None, execExpr, resultKey=resultKey, timeoutSecs=10)
# hack so the RF will use the sliced result
# FIX! don't use the sliced bit..use the whole data for rf training below
### parseKey['destination_key'] = resultKey
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
kwargs = paramDict.copy()
timeoutSecs = 30 + kwargs['ntree'] * 20
# do oobe
kwargs['out_of_bag_error_estimate'] = 1
kwargs['model_key'] = "model_" + csvFilename + "_" + str(trial)
# kwargs['model_key'] = "model"
# double check the rows/cols
inspect = h2o_cmd.runInspect(key=parseKey['destination_key'])
h2o_cmd.infoFromInspect(inspect, "going into RF")
start = time.time()
rfv = h2o_cmd.runRFOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
elapsed = time.time() - start
print "RF end on ", csvPathname, 'took', elapsed, 'seconds.', \
"%d pct. of timeout" % ((elapsed/timeoutSecs) * 100)
oobeTrainPctRight = 100 * (1.0 - rfv['confusion_matrix']['classification_error'])
if checkExpectedResults:
self.assertAlmostEqual(oobeTrainPctRight, expectTrainPctRightList[trial],
msg="OOBE: pct. right for %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), oobeTrainPctRight, expectTrainPctRightList[trial]), delta=ALLOWED_DELTA)
actualTrainPctRightList.append(oobeTrainPctRight)
print "Now score on the last 10%. Note this is silly if we trained on 100% of the data"
print "Or sorted by output class, so that the last 10% is the last few classes"
# pop the stuff from kwargs that were passing as params
model_key = rfv['model_key']
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
kwargs['iterative_cm'] = 1
kwargs['no_confusion_matrix'] = 0
# do full scoring
kwargs['out_of_bag_error_estimate'] = 0
# double check the rows/cols
inspect = h2o_cmd.runInspect(key=dataKeyTest)
h2o_cmd.infoFromInspect(inspect, "dataKeyTest")
rfvScoring = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
fullScorePctRight = 100 * (1.0 - rfvScoring['confusion_matrix']['classification_error'])
if checkExpectedResults:
self.assertAlmostEqual(fullScorePctRight,expectScorePctRightList[trial],
msg="Full: pct. right for scoring after %s pct. training not close enough %6.2f %6.2f"% \
((trial*10), fullScorePctRight, expectScorePctRightList[trial]), delta=ALLOWED_DELTA)
actualScorePctRightList.append(fullScorePctRight)
print "Trial #", trial, "completed", "using %6.2f" % (rowsToUse*100.0/num_rows), "pct. of all rows"
actualDelta = [abs(a-b) for a,b in zip(expectTrainPctRightList, actualTrainPctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualTrainPctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualTrainPctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
actualDelta = [abs(a-b) for a,b in zip(expectScorePctRightList, actualScorePctRightList)]
niceFp = ["{0:0.2f}".format(i) for i in actualScorePctRightList]
print "maybe should update with actual. Remove single quotes"
print "actualScorePctRightList =", niceFp
niceFp = ["{0:0.2f}".format(i) for i in actualDelta]
print "actualDelta =", niceFp
# return the last rfv done during training
return rfv
def test_rf_covtype_fvec(self):
h2o.beta_features = True # fvec
importFolderPath = "standard"
# Parse Train ******************************************************
csvTrainFilename = 'covtype.shuffled.90pct.data'
csvTrainPathname = importFolderPath + "/" + csvTrainFilename
hex_key = csvTrainFilename + ".hex"
parseTrainResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvTrainPathname, hex_key=hex_key,
timeoutSecs=180, doSummary=False)
inspect = h2o_cmd.runInspect(None, parseTrainResult['destination_key'])
# Parse Test ******************************************************
csvTestFilename = 'covtype.shuffled.10pct.data'
csvTestPathname = importFolderPath + "/" + csvTestFilename
hex_key = csvTestFilename + ".hex"
parseTestResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvTestPathname, hex_key=hex_key,
timeoutSecs=180)
inspect = h2o_cmd.runInspect(None, parseTestResult['destination_key'])
rfViewInitial = []
xList = []
eList = []
fList = []
trial = 0
depthList = [10, 20, 30, 40]
ntreesList = [5, 10, 20, 30]
# ntreesList = [2]
nbinsList = [10, 100, 1000]
if TRY == 'max_depth':
tryList = depthList
elif TRY == 'ntrees':
tryList = ntreesList
elif TRY == 'nbins':
tryList = nbinsList
else:
raise Exception("huh? %s" % TRY)
for d in tryList:
if TRY == 'max_depth':
paramDict['max_depth'] = d
elif TRY == 'ntrees':
paramDict['ntrees'] = d
elif TRY == 'nbins':
paramDict['nbins'] = d
else:
raise Exception("huh? %s" % TRY)
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
if DO_OOBE:
paramDict['validation'] = None
else:
paramDict['validation'] = parseTestResult['destination_key']
timeoutSecs = 30 + paramDict['ntrees'] * 200
# do ten starts, to see the bad id problem?
TRIES = 10
for i in range(TRIES):
lastOne = i == (TRIES-1)
# have unique model names
trial += 1
kwargs = paramDict.copy()
# kwargs['destination_key'] = 'RFModel_' + str(trial)
# let h2o name it
kwargs['destination_key'] = None
start = time.time()
rfResult = h2o_cmd.runRF(parseResult=parseTrainResult, timeoutSecs=timeoutSecs,
noPoll=False, rfView=False, **kwargs)
trainElapsed = time.time() - start
print 'rf train end', i, 'on', csvTrainPathname, 'took', trainElapsed, 'seconds'
# don't cancel the last one
if not lastOne:
time.sleep(1)
h2o_jobs.cancelAllJobs(timeoutSecs=2)
### print "rfView", h2o.dump_json(rfView)
model_key = rfResult['drf_model']['_selfKey']
data_key = rfResult['drf_model']['_dataKey']
ntrees = kwargs['ntrees']
rfView = h2o_cmd.runRFView(None, model_key=model_key, timeoutSecs=60, retryDelaySecs=5, doSimpleCheck=False)
## print "rfView:", h2o.dump_json(rfView)
rf_model = rfView['drf_model']
cm = rf_model['cm']
### print "cm:", h2o.dump_json(cm)
ntrees = rf_model['N']
errs = rf_model['errs']
N = rf_model['N']
varimp = rf_model['varimp']
treeStats = rf_model['treeStats']
print "maxDepth:", treeStats['maxDepth']
print "maxLeaves:", treeStats['maxLeaves']
print "minDepth:", treeStats['minDepth']
print "minLeaves:", treeStats['minLeaves']
print "meanLeaves:", treeStats['meanLeaves']
print "meanDepth:", treeStats['meanDepth']
print "errs[0]:", errs[0]
print "errs[-1]:", errs[-1]
print "errs:", errs
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView)
# we iterate over params, so can't really do this check
# self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
# FIX! should update this expected classification error
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=data_key)
eList.append(classErrorPctList[4])
fList.append(trainElapsed)
if DO_PLOT:
if TRY == 'max_depth':
xLabel = 'max_depth'
elif TRY == 'ntrees':
xLabel = 'ntrees'
elif TRY == 'nbins':
xLabel = 'nbins'
else:
raise Exception("huh? %s" % TRY)
xList.append(paramDict[xLabel])
if DO_PLOT:
eLabel = 'class 4 pctWrong'
fLabel = 'trainElapsed'
eListTitle = ""
fListTitle = ""
h2o_gbm.plotLists(xList, xLabel, eListTitle, eList, eLabel, fListTitle, fList, fLabel)
def test_rf_covtype20x_fvec(self):
h2o.beta_features = True
importFolderPath = 'standard'
if DO_SMALL:
csvFilenameTrain = 'covtype.data'
hex_key = 'covtype1x.data.A.hex'
else:
csvFilenameTrain = 'covtype20x.data'
hex_key = 'covtype20x.data.A.hex'
csvPathname = importFolderPath + "/" + csvFilenameTrain
parseResultTrain = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathname,
hex_key=hex_key,
timeoutSecs=500)
inspect = h2o_cmd.runInspect(key=parseResultTrain['destination_key'])
dataKeyTrain = parseResultTrain['destination_key']
print "Parse end", dataKeyTrain
# have to re import since source key is gone
# we could just copy the key, but sometimes we change the test/train data to covtype.data
if DO_SMALL:
csvFilenameTest = 'covtype.data'
hex_key = 'covtype1x.data.B.hex'
dataKeyTest2 = 'covtype1x.data.C.hex'
else:
csvFilenameTest = 'covtype20x.data'
hex_key = 'covtype20x.data.B.hex'
dataKeyTest2 = 'covtype20x.data.C.hex'
csvPathname = importFolderPath + "/" + csvFilenameTest
parseResultTest = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathname,
hex_key=hex_key,
timeoutSecs=500)
print "Parse result['destination_key']:", parseResultTest[
'destination_key']
inspect = h2o_cmd.runInspect(key=parseResultTest['destination_key'])
dataKeyTest = parseResultTest['destination_key']
print "Parse end", dataKeyTest
# make a 3rd key so the predict is uncached too!
execExpr = dataKeyTest2 + "=" + dataKeyTest
kwargs = {'str': execExpr, 'timeoutSecs': 15}
resultExec = h2o_cmd.runExec(**kwargs)
# train
# this does RFView to understand when RF completes, so the time reported for RFView here, should be
# considered the "first RFView" times..subsequent have some caching?.
# unless the no_confusion_matrix works
# params is mutable. This is default.
paramDict = drf2ParamDict
params = {'ntrees': 20, 'destination_key': 'RF_model'}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
timeoutSecs = 30 + kwargs['ntrees'] * 60
start = time.time()
rf = h2o_cmd.runRF(parseResult=parseResultTrain,
timeoutSecs=timeoutSecs,
retryDelaySecs=1,
**kwargs)
print "rf job end on ", dataKeyTrain, 'took', time.time(
) - start, 'seconds'
print "\nRFView start after job completion"
model_key = kwargs['destination_key']
ntree = kwargs['ntrees']
start = time.time()
# this does the RFModel view for v2. but only model_key is used. Data doesn't matter? (nor ntree)
h2o_cmd.runRFView(None,
dataKeyTrain,
model_key,
ntree=ntree,
timeoutSecs=timeoutSecs)
print "First rfview end on ", dataKeyTrain, 'took', time.time(
) - start, 'seconds'
for trial in range(1):
# scoring
start = time.time()
rfView = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree=ntree,
timeoutSecs=timeoutSecs,
out_of_bag_error_estimate=0,
retryDelaySecs=1)
print "rfview", trial, "end on ", dataKeyTest, 'took', time.time(
) - start, 'seconds.'
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
self.assertAlmostEqual(
classification_error,
50,
delta=50,
msg="Classification error %s differs too much" %
classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest2)
print "predict", trial, "end on ", dataKeyTest, 'took', time.time(
) - start, 'seconds.'
parseKey = parseResultTrain['destination_key']
rfModelKey = rfView['drf_model']['_key']
predictKey = 'Predict.hex'
start = time.time()
predictResult = h2o_cmd.runPredict(data_key=parseKey,
model_key=rfModelKey,
destination_key=predictKey,
timeoutSecs=timeoutSecs)
predictCMResult = h2o.nodes[0].predict_confusion_matrix(
actual=parseKey,
vactual='C55',
predict=predictKey,
vpredict='predict',
)
cm = predictCMResult['cm']
# These will move into the h2o_gbm.py
pctWrong = h2o_gbm.pp_cm_summary(cm)
print "\nTest\n==========\n"
print h2o_gbm.pp_cm(cm)
print "Trial #", trial, "completed"
def test_RF_mnist_both(self):
csvFilelist = [
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, None, '*mnist_training*gz'),
]
# IMPORT**********************************************
trial = 0
allDelta = []
importFolderPath = "mnist"
for (trainCsvFilename, testCsvFilename, timeoutSecs, rfSeed, parsePattern) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
csvPathname = importFolderPath + "/" + testCsvFilename
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=csvPathname, schema='local',
hex_key=testKey, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
print "Use multi-file parse to grab both the mnist_testing.csv.gz and mnist_training.csv.gz for training"
trainKey = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
csvPathname = importFolderPath + "/" + parsePattern
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=parsePattern, schema='local', timeoutSecs=300)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, returnIgnoreX=True)
ntree = 100
params = {
'response': 0,
'ignored_cols': ignore_x,
'ntrees': ntree,
# 'data_key='mnist_training.csv.hex'
'mtries': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'max_depth': 500,
'destination_key': 'RF_model',
'nbins': 1024,
}
if rfSeed is None:
params['seed'] = random.randint(0,sys.maxint)
else:
params['seed'] = rfSeed
print "RF seed:", params['seed']
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRF(parseResult=parseResult, rfView=False,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
kwargs = {'response': y}
rfView = h2o_cmd.runRFView(data_key=testKey, model_key=modelKey,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
print "classification error is expected to be low because we included the test data in with the training!"
self.assertAlmostEqual(classification_error, 0.028, delta=0.01, msg="Classification error %s differs too much" % classification_error)
treeStats = rfView['drf_model']['treesStats']
# Expected values are from this case:
# ("mnist_training.csv.gz", "mnist_testing.csv.gz", 600, 784834182943470027),
expected = {'minLeaves': 4996, 'meanLeaves': 5064.1, 'maxLeaves': 5148}
expected += {'minDepth': 21, 'meanDepth': 23.8, 'maxDepth': 25}
for key in expected:
delta = ((expected[key]- actual[key])/expected[key]) * 100
d = "seed: %s %s %s %s %s pct. different %s" % (params['seed'], key, actual[key], expected[key], delta)
print d
allDelta.append(d)
# FIX! should change this to an assert?
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=modelKey, data_key=testKey, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# Done *******************************************************
print "\nShowing the results again from all the trials, to see variance"
for d in allDelta:
print d
def test_rf_big1_overwrite_model_fvec(self):
h2o.beta_features = True
csvFilename = 'hhp_107_01.data.gz'
hex_key = csvFilename + ".hex"
print "\n" + csvFilename
parseResult = h2i.import_parse(bucket='smalldata',
path=csvFilename,
hex_key=hex_key,
timeoutSecs=15,
schema='put')
firstRfView = None
# dispatch multiple jobs back to back
for jobDispatch in range(3):
start = time.time()
kwargs = {}
if OVERWRITE_RF_MODEL:
print "Since we're overwriting here, we have to wait for each to complete noPoll=False"
model_key = 'RF_model'
else:
model_key = 'RF_model' + str(jobDispatch)
print "Change the number of trees, while keeping the rf model key name the same"
print "Checks that we correctly overwrite previous rf model"
if OVERWRITE_RF_MODEL:
kwargs['ntrees'] = 1 + jobDispatch
else:
kwargs['ntrees'] = 1
# don't change the seed if we're overwriting the model. It should get
# different results just from changing the tree count
kwargs['seed'] = random.randint(0, sys.maxint)
# FIX! what model keys do these get?
randomNode = h2o.nodes[random.randint(0, len(h2o.nodes) - 1)]
h2o_cmd.runRF(node=randomNode,
parseResult=parseResult,
destination_key=model_key,
timeoutSecs=300,
noPoll=True,
**kwargs)
# FIX! are these already in there?
rfView = {}
rfView['_dataKey'] = hex_key
rfView['_key'] = model_key
print "rf job dispatch end on ", csvFilename, 'took', time.time(
) - start, 'seconds'
print "\njobDispatch #", jobDispatch
# we're going to compare rf results to previous as we go along (so we save rf view results
h2o_jobs.pollWaitJobs(pattern='RF_model',
timeoutSecs=300,
pollTimeoutSecs=10,
retryDelaySecs=5)
# In this test we're waiting after each one, so we can save the RFView results for comparison to future
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['_dataKey']
model_key = rfView['_key']
print "Temporary hack: need to do two rf views minimum, to complete a RF (confusion matrix creation)"
# allow it to poll to complete
rfViewResult = h2o_cmd.runRFView(None,
data_key,
model_key,
timeoutSecs=60,
noPoll=False)
if firstRfView is None: # we'll use this to compare the others
firstRfView = rfViewResult.copy()
firstModelKey = model_key
print "firstRfView", h2o.dump_json(firstRfView)
else:
print "Comparing", model_key, "to", firstModelKey
df = h2o_util.JsonDiff(rfViewResult,
firstRfView,
vice_versa=True,
with_values=True)
print "df.difference:", h2o.dump_json(df.difference)
self.assertGreater(len(df.difference), 29,
msg="Want >=30 , not %d differences between the two rfView json responses. %s" % \
(len(df.difference), h2o.dump_json(df.difference)))
def test_rfview_score(self):
csvPathnameTrain = 'standard/covtype.data'
print "Train with:", csvPathnameTrain
parseResultTrain = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathnameTrain,
schema='put',
hex_key="covtype.hex",
timeoutSecs=15)
dataKeyTrain = parseResultTrain['destination_key']
csvPathnameTest = 'standard/covtype.data'
print "Test with:", csvPathnameTest
parseResultTest = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvPathnameTest,
schema='put',
hex_key="covtype.hex",
timeoutSecs=15)
dataKeyTest = parseResultTest['destination_key']
for trial in range(5):
# params is mutable. This is default.
params = {'ntree': 13, 'out_of_bag_error_estimate': 0}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + kwargs['ntree'] * 10
rfv = h2o_cmd.runRF(parseResult=parseResultTrain,
timeoutSecs=timeoutSecs,
retryDelaySecs=1,
**kwargs)
### print "rf response:", h2o.dump_json(rfv)
model_key = rfv['model_key']
# pop the stuff from kwargs that were passing as params
kwargs.pop('model_key', None)
data_key = rfv['data_key']
kwargs.pop('data_key', None)
ntree = rfv['ntree']
kwargs.pop('ntree', None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
rfView = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
**kwargs)
# new web page for predict? throw it in here for now
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if 'sampling_strategy' in kwargs and kwargs[
'sampling_strategy'] != 'STRATIFIED_LOCAL':
check_err = True
else:
check_err = False
if check_err:
self.assertAlmostEqual(
classification_error,
0.03,
delta=0.5,
msg="Classification error %s differs too much" %
classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 0
rfView = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if check_err:
self.assertAlmostEqual(
classification_error,
0.03,
delta=0.5,
msg="Classification error %s differs too much" %
classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 1
rfView = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if check_err:
self.assertAlmostEqual(
classification_error,
0.03,
delta=0.5,
msg="Classification error %s differs too much" %
classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 1
kwargs['class_weights'] = '1=1,2=2,3=3,4=4,5=5,6=6,7=7'
rfView = h2o_cmd.runRFView(None,
dataKeyTest,
model_key,
ntree,
timeoutSecs,
retryDelaySecs=1,
print_params=True,
**kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if check_err:
self.assertAlmostEqual(
classification_error,
0.03,
delta=0.5,
msg="Classification error %s differs too much" %
classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
print "Trial #", trial, "completed"
def test_RF_mnist_fvec(self):
h2o.beta_features = True
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_testing.csv.gz", "mnist_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_testing.csv.gz", "mnist_testing.csv.gz", 600),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600),
]
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets',
path=importFolderPath + "/" +
testCsvFilename,
hex_key=testKey2,
timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
# x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets',
path=importFolderPath + "/" +
trainCsvFilename,
schema='local',
hex_key=trainKey2,
timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(
y,
key=parseResult['destination_key'],
timeoutSecs=300,
forRF=True)
params = {
'response': 'C' + str(y),
'cols': None,
'ignored_cols_by_name': ignore_x,
'classification': 1,
'validation': None,
'ntrees': 10,
'max_depth': 20,
'min_rows': None,
'nbins': 1000,
'mtries': None,
'sample_rate': 0.66,
'seed': None,
}
rfViewInitial = []
for jobDispatch in range(1):
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
params['destination_key'] = 'RFModel_' + str('jobDispatch')
kwargs = params.copy()
timeoutSecs = 1200
start = time.time()
rfResult = h2o_cmd.runRF(parseResult=parseResult,
timeoutSecs=timeoutSecs,
noPoll=not DO_POLL,
rfView=DO_POLL,
**kwargs)
elapsed = time.time() - start
# print h2o.dump_json(rfResult)
print "rf job dispatch end on ", trainCsvFilename, 'took', time.time(
) - start, 'seconds'
print "\njobDispatch #", jobDispatch
# FIX! are these already in there?
rfView = {}
rfView['data_key'] = trainKey2
rfView['model_key'] = kwargs['destination_key']
rfView['ntrees'] = kwargs['ntrees']
rfViewInitial.append(rfView)
if not DO_POLL:
h2o_jobs.pollStatsWhileBusy(timeoutSecs=1200,
pollTimeoutSecs=120,
retryDelaySecs=5)
# FIX! need to add the rfview and predict stuff
# we saved the initial response?
# if we do another poll they should be done now, and better to get it that
# way rather than the inspect (to match what simpleCheckGLM is expected
print "rfViewInitial", rfViewInitial
for rfView in rfViewInitial:
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['data_key']
model_key = rfView['model_key']
ntrees = rfView['ntrees']
rfView = h2o_cmd.runRFView(None,
model_key=model_key,
timeoutSecs=60,
noPoll=not DO_POLL,
doSimpleCheck=False)
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView)
self.assertAlmostEqual(
classification_error,
10,
delta=2,
msg="Classification error %s differs too much" %
classification_error)
if not DO_POLL:
h2o_jobs.pollStatsWhileBusy(timeoutSecs=300,
pollTimeoutSecs=120,
retryDelaySecs=5)
# rfView = h2o_cmd.runRFView(None, data_key, model_key, timeoutSecs=60, noPoll=True, doSimpleCheck=False)
# print "rfView:", h2o.dump_json(rfView)
# "N":1,
# "errs":[0.25,0.1682814508676529],
# "testKey":"syn_binary_10000x10.hex",
# "cm":[[3621,1399],[1515,3465]]}}
rf_model = rfView['drf_model']
cms = rf_model['cms']
ntrees = rf_model['N']
errs = rf_model['errs']
N = rf_model['N']
# FIX! should update this expected classification error
## (classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntrees)
## self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=data_key)
def test_RF_mnist_fvec(self):
h2o.beta_features = True
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_testing.csv.gz", "mnist_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_testing.csv.gz", "mnist_testing.csv.gz", 600),
("mnist_training.csv.gz", "mnist_testing.csv.gz", 600),
]
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + testCsvFilename,
hex_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
# x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + trainCsvFilename, schema='local',
hex_key=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, forRF=True)
params = {
'response': 'C' + str(y+1),
'cols': None,
'ignored_cols_by_name': ignore_x,
'classification': 1,
'validation': None,
'ntrees': 2,
'max_depth': 20,
'min_rows': None,
'nbins': 1000,
'mtries': None,
'sample_rate': 0.66,
'seed': None,
}
rfViewInitial = []
for jobDispatch in range(1):
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
params['destination_key'] = 'RFModel_' + str('jobDispatch')
kwargs = params.copy()
timeoutSecs = 1200
start = time.time()
rfResult = h2o_cmd.runRF(parseResult=parseResult, timeoutSecs=timeoutSecs, noPoll=not DO_POLL, rfView=DO_POLL, **kwargs)
elapsed = time.time() - start
# print h2o.dump_json(rfResult)
print "rf job dispatch end on ", trainCsvFilename, 'took', time.time() - start, 'seconds'
print "\njobDispatch #", jobDispatch
# FIX! are these already in there?
rfView = {}
rfView['data_key'] = trainKey2
rfView['model_key'] = kwargs['destination_key']
rfView['ntrees'] = kwargs['ntrees']
rfViewInitial.append(rfView)
if not DO_POLL:
h2o_jobs.pollStatsWhileBusy(timeoutSecs=1200, pollTimeoutSecs=120, retryDelaySecs=5)
# FIX! need to add the rfview and predict stuff
# we saved the initial response?
# if we do another poll they should be done now, and better to get it that
# way rather than the inspect (to match what simpleCheckGLM is expected
print "rfViewInitial", rfViewInitial
for rfView in rfViewInitial:
print "Checking completed job:", rfView
print "rfView", h2o.dump_json(rfView)
data_key = rfView['data_key']
model_key = rfView['model_key']
ntrees = rfView['ntrees']
rfView = h2o_cmd.runRFView(None, model_key=model_key, timeoutSecs=60, noPoll=not DO_POLL, doSimpleCheck=False)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView)
self.assertAlmostEqual(classification_error, 10, delta=2, msg="Classification error %s differs too much" % classification_error)
if not DO_POLL:
h2o_jobs.pollStatsWhileBusy(timeoutSecs=300, pollTimeoutSecs=120, retryDelaySecs=5)
# rfView = h2o_cmd.runRFView(None, data_key, model_key, timeoutSecs=60, noPoll=True, doSimpleCheck=False)
# print "rfView:", h2o.dump_json(rfView)
# "N":1,
# "errs":[0.25,0.1682814508676529],
# "testKey":"syn_binary_10000x10.hex",
# "cm":[[3621,1399],[1515,3465]]}}
rf_model = rfView['drf_model']
cms = rf_model['cms']
errs = rf_model['errs']
# FIX! should update this expected classification error
## (classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntrees)
## self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=data_key)
def test_rfview_score(self):
csvPathnameTrain = 'UCI/UCI-large/covtype/covtype.data'
print "Train with:", csvPathnameTrain
parseResultTrain = h2i.import_parse(bucket='datasets', path=csvPathnameTrain, schema='put',
hex_key="covtype.hex", timeoutSecs=15)
dataKeyTrain = parseResultTrain['destination_key']
csvPathnameTest = 'UCI/UCI-large/covtype/covtype.data'
print "Test with:", csvPathnameTest
parseResultTest = h2i.import_parse(bucket='datasets', path=csvPathnameTest, schema='put',
hex_key="covtype.hex", timeoutSecs=15)
dataKeyTest = parseResultTest['destination_key']
for trial in range(5):
# params is mutable. This is default.
params = {'ntree': 13, 'parallel': 1, 'out_of_bag_error_estimate': 0}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + kwargs['ntree'] * 10 * (kwargs['parallel'] and 1 or 5)
rfv = h2o_cmd.runRF(parseResult=parseResultTrain, timeoutSecs=timeoutSecs, retryDelaySecs=1, **kwargs)
### print "rf response:", h2o.dump_json(rfv)
model_key = rfv['model_key']
# pop the stuff from kwargs that were passing as params
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
rfView = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
# new web page for predict? throw it in here for now
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if 'sampling_strategy' in kwargs and kwargs['sampling_strategy'] != 'STRATIFIED_LOCAL':
check_err = True
else:
check_err = False
if check_err:
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 0
rfView = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if check_err:
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 1
rfView = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if check_err:
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
kwargs['iterative_cm'] = 1
kwargs['class_weights'] = '1=1,2=2,3=3,4=4,5=5,6=6,7=7'
rfView = h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
# FIX! should update this expected classification error
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView, ntree=ntree)
# don't check error if stratified
if check_err:
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=dataKeyTest)
elapsed = time.time() - start
print "predict end on ", dataKeyTest, 'took', elapsed, 'seconds.'
print "Trial #", trial, "completed"
def test_rf_params_rand2(self):
# for determinism, I guess we should spit out the seed?
# random.seed(SEED)
SEED = random.randint(0, sys.maxint)
# if you have to force to redo a test
# SEED =
random.seed(SEED)
print "\nUsing random seed:", SEED
csvPathnameTrain = h2o.find_file('smalldata/covtype/covtype.20k.data')
print "Train with:", csvPathnameTrain
parseKeyTrain = h2o_cmd.parseFile(csvPathname=csvPathnameTrain, key2="covtype.20k.hex", timeoutSecs=10)
dataKeyTrain = parseKeyTrain['destination_key']
csvPathnameTest = h2o.find_dataset('UCI/UCI-large/covtype/covtype.data')
print "Test with:", csvPathnameTest
parseKeyTest = h2o_cmd.parseFile(csvPathname=csvPathnameTrain, key2="covtype.hex", timeoutSecs=10)
dataKeyTest = parseKeyTest['destination_key']
for trial in range(5):
# params is mutable. This is default.
params = {'ntree': 13, 'parallel': 1, 'out_of_bag_error_estimate': 0}
colX = h2o_rf.pickRandRfParams(paramDict, params)
kwargs = params.copy()
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
timeoutSecs = 30 + 15 * (kwargs['parallel'] and 5 or 10)
rfv = h2o_cmd.runRFOnly(parseKey=parseKeyTrain, timeoutSecs=timeoutSecs, retryDelaySecs=1, **kwargs)
### print "rf response:", h2o.dump_json(rfv)
model_key = rfv['model_key']
# pop the stuff from kwargs that were passing as params
kwargs.pop('model_key',None)
data_key = rfv['data_key']
kwargs.pop('data_key',None)
ntree = rfv['ntree']
kwargs.pop('ntree',None)
# scoring
# RFView.html?
# dataKeyTest=a5m.hex&
# model_key=__RFModel_81c5063c-e724-4ebe-bfc1-3ac6838bc628&
# response_variable=1&
# ntree=50&
# class_weights=-1%3D1.0%2C0%3D1.0%2C1%3D1.0&
# out_of_bag_error_estimate=1&
# no_confusion_matrix=1&
# clear_confusion_matrix=1
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
kwargs['no_confusion_matrix'] = 0
kwargs['clear_confusion_matrix'] = 0
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
kwargs['no_confusion_matrix'] = 0
kwargs['clear_confusion_matrix'] = 1
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
kwargs['no_confusion_matrix'] = 1
kwargs['clear_confusion_matrix'] = 0
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
kwargs['no_confusion_matrix'] = 1
kwargs['clear_confusion_matrix'] = 1
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
kwargs['no_confusion_matrix'] = 0
kwargs['clear_confusion_matrix'] = 0
kwargs['class_weights'] = '1=1,2=2,3=3,4=4,5=5,6=6,7=7'
h2o_cmd.runRFView(None, dataKeyTest, model_key, ntree,
timeoutSecs, retryDelaySecs=1, print_params=True, **kwargs)
print "Trial #", trial, "completed"
def test_RF_mnist_reals(self):
importFolderPath = "/home/0xdiag/datasets/mnist"
csvFilelist = [
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
("mnist_reals_training.csv.gz", "mnist_reals_testing.csv.gz", 600),
]
# IMPORT**********************************************
# since H2O deletes the source key, we should re-import every iteration if we re-use the src in the list
importFolderResult = h2i.setupImportFolder(None, importFolderPath)
### print "importHDFSResult:", h2o.dump_json(importFolderResult)
succeededList = importFolderResult['files']
### print "succeededList:", h2o.dump_json(succeededList)
self.assertGreater(len(succeededList), 1,
"Should see more than 1 files in the import?")
# why does this hang? can't look at storeview after import?
print "\nTrying StoreView after the import folder"
h2o_cmd.runStoreView(timeoutSecs=30)
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseKey = h2i.parseImportFolderFile(None,
testCsvFilename,
importFolderPath,
key2=testKey2,
timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseKey['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y,
key=parseKey['destination_key'],
timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseKey = h2i.parseImportFolderFile(None,
trainCsvFilename,
importFolderPath,
key2=trainKey2,
timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseKey['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(
y,
key=parseKey['destination_key'],
timeoutSecs=300,
forRF=True)
ntree = 100
params = {
'response_variable':
0,
'ignore':
ignore_x,
'ntree':
ntree,
'iterative_cm':
1,
'out_of_bag_error_estimate':
1,
# 'data_key='mnist_reals_training.csv.hex'
'features':
28, # fix because we ignore some cols, which will change the srt(cols) calc?
'exclusive_split_limit':
None,
'depth':
2147483647,
'stat_type':
'ENTROPY',
'sampling_strategy':
'RANDOM',
'sample':
67,
# 'model_key': '__RFModel_7055e6cf-a0de-44db-b165-f5994730ac77',
'model_key':
'RF_model',
'bin_limit':
1024,
'seed':
784834182943470027,
'parallel':
1,
'use_non_local_data':
0,
'class_weights':
'0=1.0,1=1.0,2=1.0,3=1.0,4=1.0,5=1.0,6=1.0,7=1.0,8=1.0,9=1.0',
}
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRFOnly(parseKey=parseKey,
rfView=False,
timeoutSecs=timeoutSecs,
pollTimeoutsecs=60,
retryDelaySecs=2,
**kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfView = h2o_cmd.runRFView(data_key=testKey2,
model_key=modelKey,
ntree=ntree,
out_of_bag_error_estimate=0,
timeoutSecs=60,
pollTimeoutSecs=60,
noSimpleCheck=False,
**kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
self.assertAlmostEqual(
classification_error,
0.03,
delta=0.5,
msg="Classification error %s differs too much" %
classification_error)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(
model_key=modelKey, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
def test_RF_mnist_reals(self):
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
("mnist_reals_training.csv.gz", "mnist_reals_testing.csv.gz", 600),
]
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + testCsvFilename,
hex_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + trainCsvFilename,
hex_key=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
print "This is the 'ignore=' we'll use"
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, forRF=True)
ntree = 10
params = {
'response_variable': 0,
'ignore': ignore_x,
'ntree': ntree,
'iterative_cm': 1,
'out_of_bag_error_estimate': 1,
# 'data_key='mnist_reals_training.csv.hex'
'features': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'exclusive_split_limit': None,
'depth': 2147483647,
'stat_type': 'ENTROPY',
'sampling_strategy': 'RANDOM',
'sample': 67,
# 'model_key': '__RFModel_7055e6cf-a0de-44db-b165-f5994730ac77',
'model_key': 'RF_model',
'bin_limit': 1024,
'seed': 784834182943470027,
'parallel': 1,
'use_non_local_data': 0,
'class_weights': '0=1.0,1=1.0,2=1.0,3=1.0,4=1.0,5=1.0,6=1.0,7=1.0,8=1.0,9=1.0',
}
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfView = h2o_cmd.runRF(parseResult=parseResult, rfView=True,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfView, **params)
modelKey = rfView['model_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfView = h2o_cmd.runRFView(data_key=testKey2, model_key=modelKey, ntree=ntree, out_of_bag_error_estimate=0,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfView, **params)
self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=modelKey, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
def test_rf_covtype_fvec(self):
h2o.beta_features = True # fvec
importFolderPath = "standard"
# Parse Train ******************************************************
csvTrainFilename = 'covtype.shuffled.90pct.data'
csvTrainPathname = importFolderPath + "/" + csvTrainFilename
hex_key = csvTrainFilename + ".hex"
parseTrainResult = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvTrainPathname,
hex_key=hex_key,
timeoutSecs=180,
doSummary=False)
inspect = h2o_cmd.runInspect(None, parseTrainResult['destination_key'])
# Parse Test ******************************************************
csvTestFilename = 'covtype.shuffled.10pct.data'
csvTestPathname = importFolderPath + "/" + csvTestFilename
hex_key = csvTestFilename + ".hex"
parseTestResult = h2i.import_parse(bucket='home-0xdiag-datasets',
path=csvTestPathname,
hex_key=hex_key,
timeoutSecs=180)
inspect = h2o_cmd.runInspect(None, parseTestResult['destination_key'])
rfViewInitial = []
xList = []
eList = []
fList = []
trial = 0
depthList = [10, 20, 30, 40]
ntreesList = [5, 10, 20, 30]
# ntreesList = [2]
nbinsList = [10, 100, 1000]
if TRY == 'max_depth':
tryList = depthList
elif TRY == 'ntrees':
tryList = ntreesList
elif TRY == 'nbins':
tryList = nbinsList
else:
raise Exception("huh? %s" % TRY)
for d in tryList:
if TRY == 'max_depth':
paramDict['max_depth'] = d
elif TRY == 'ntrees':
paramDict['ntrees'] = d
elif TRY == 'nbins':
paramDict['nbins'] = d
else:
raise Exception("huh? %s" % TRY)
# adjust timeoutSecs with the number of trees
# seems ec2 can be really slow
if DO_OOBE:
paramDict['validation'] = None
else:
paramDict['validation'] = parseTestResult['destination_key']
timeoutSecs = 30 + paramDict['ntrees'] * 200
# do ten starts, to see the bad id problem?
TRIES = 5
for i in range(TRIES):
lastOne = i == (TRIES - 1)
# have unique model names
trial += 1
kwargs = paramDict.copy()
model_key = 'RFModel_' + str(trial)
kwargs['destination_key'] = model_key
data_key = parseTrainResult['destination_key']
start = time.time()
rfResult = h2o_cmd.runRF(parseResult=parseTrainResult,
timeoutSecs=timeoutSecs,
noPoll=True,
rfView=False,
**kwargs)
trainElapsed = time.time() - start
print 'rf train end', i, 'on', csvTrainPathname, 'took', trainElapsed, 'seconds'
# don't cancel the last one
if not lastOne:
time.sleep(1)
h2o_jobs.cancelAllJobs(timeoutSecs=2)
### print "rfView", h2o.dump_json(rfView)
print "We have a result from the RF above, completed but didn't do RFView yet"
# could the RF indicate 'done' too soon?
# if rfResult['state']=='RUNNING':
# raise Exception("Why is this RF still in RUNNING state? %s" % h2o.dump_json(rfResult))
# if 'drf_model' not in rfResult:
# raise Exception("How come there's no drf_model in this RF result? %s" % h2o.dump_json(rfResult))
h2o_jobs.pollWaitJobs(timeoutSecs=300)
rfView = h2o_cmd.runRFView(None,
model_key=model_key,
timeoutSecs=60,
retryDelaySecs=5,
doSimpleCheck=False)
print "rfView:", h2o.dump_json(rfView)
rf_model = rfView['drf_model']
cms = rf_model['cms']
### print "cm:", h2o.dump_json(cm)
ntrees = rf_model['N']
errs = rf_model['errs']
N = rf_model['N']
varimp = rf_model['varimp']
treeStats = rf_model['treeStats']
print "maxDepth:", treeStats['maxDepth']
print "maxLeaves:", treeStats['maxLeaves']
print "minDepth:", treeStats['minDepth']
print "minLeaves:", treeStats['minLeaves']
print "meanLeaves:", treeStats['meanLeaves']
print "meanDepth:", treeStats['meanDepth']
print "errs[0]:", errs[0]
print "errs[-1]:", errs[-1]
print "errs:", errs
(classification_error, classErrorPctList,
totalScores) = h2o_rf.simpleCheckRFView(rfv=rfView)
# we iterate over params, so can't really do this check
# self.assertAlmostEqual(classification_error, 0.03, delta=0.5, msg="Classification error %s differs too much" % classification_error)
print "classErrorPctList:", classErrorPctList
self.assertEqual(
len(classErrorPctList), 7,
"Should be 7 output classes, so should have 7 class error percentages from a reasonable predict"
)
# FIX! should update this expected classification error
predict = h2o.nodes[0].generate_predictions(model_key=model_key,
data_key=data_key)
eList.append(classErrorPctList[4])
fList.append(trainElapsed)
if DO_PLOT:
if TRY == 'max_depth':
xLabel = 'max_depth'
elif TRY == 'ntrees':
xLabel = 'ntrees'
elif TRY == 'nbins':
xLabel = 'nbins'
else:
raise Exception("huh? %s" % TRY)
xList.append(paramDict[xLabel])
if DO_PLOT:
eLabel = 'class 4 pctWrong'
fLabel = 'trainElapsed'
eListTitle = ""
fListTitle = ""
h2o_gbm.plotLists(xList, xLabel, eListTitle, eList, eLabel,
fListTitle, fList, fLabel)
def test_RF_mnist_reals_fvec(self):
importFolderPath = "mnist"
csvFilelist = [
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
# ("a.csv", "b.csv", 60),
# ("mnist_reals_testing.csv.gz", "mnist_reals_testing.csv.gz", 600),
("mnist_reals_training.csv.gz", "mnist_reals_testing.csv.gz", 600),
]
trial = 0
for (trainCsvFilename, testCsvFilename, timeoutSecs) in csvFilelist:
trialStart = time.time()
# PARSE test****************************************
testKey2 = testCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + testCsvFilename,
hex_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", testCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
print "We won't use this pruning of x on test data. See if it prunes the same as the training"
y = 0 # first column is pixel value
print "y:"
x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300)
# PARSE train****************************************
trainKey2 = trainCsvFilename + "_" + str(trial) + ".hex"
start = time.time()
parseResult = h2i.import_parse(bucket='home-0xdiag-datasets', path=importFolderPath + "/" + trainCsvFilename,
hex_key=trainKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "parse end on ", trainCsvFilename, 'took', elapsed, 'seconds',\
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
print "parse result:", parseResult['destination_key']
# RF+RFView (train)****************************************
ignore_x = h2o_glm.goodXFromColumnInfo(y, key=parseResult['destination_key'], timeoutSecs=300, returnIgnoreX=True)
ntrees = 10
params = {
'response': 'C1',
'ignored_cols_by_name': ignore_x,
'ntrees': ntrees,
'mtries': 28, # fix because we ignore some cols, which will change the srt(cols) calc?
'max_depth': 15,
'sample_rate': 0.67,
'destination_key': 'RF_model',
'nbins': 1024,
'seed': 784834182943470027,
'importance': 0,
'balance_classes': 0,
}
kwargs = params.copy()
print "Trying rf"
timeoutSecs = 1800
start = time.time()
rfv = h2o_cmd.runRF(parseResult=parseResult, rfView=True,
timeoutSecs=timeoutSecs, pollTimeoutSecs=60, retryDelaySecs=2, **kwargs)
elapsed = time.time() - start
print "RF completed in", elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
h2o_rf.simpleCheckRFView(None, rfv, **params)
rf_model = rfv['drf_model']
used_trees = rf_model['N']
data_key = rf_model['_dataKey']
model_key = rf_model['_key']
# RFView (score on test)****************************************
start = time.time()
# FIX! 1 on oobe causes stack trace?
kwargs = {'response_variable': y}
rfv = h2o_cmd.runRFView(data_key=testKey2, model_key=model_key, ntrees=ntrees, out_of_bag_error_estimate=0,
timeoutSecs=60, pollTimeoutSecs=60, noSimpleCheck=False, **kwargs)
elapsed = time.time() - start
print "RFView in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
(classification_error, classErrorPctList, totalScores) = h2o_rf.simpleCheckRFView(None, rfv, **params)
self.assertAlmostEqual(classification_error, 9, delta=1.0, msg="Classification error %s differs too much" % classification_error)
# Predict (on test)****************************************
start = time.time()
predict = h2o.nodes[0].generate_predictions(model_key=model_key, data_key=testKey2, timeoutSecs=timeoutSecs)
elapsed = time.time() - start
print "generate_predictions in", elapsed, "secs", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
