def __init__(self):
self.mc = 0
self.rtcl = 3
self.nrtcl = 5
self.scl = 4
self.server = 2
self.current = Node(2, 0, 4)
self.rt_queue = deque([])
self.nrt_queue = deque([])
self.preempt = -1
self.rt_iat = self.nrt_iat = self.rt_st = self.nrt_st = 0
# Results
self.m = 0
self.b = 0
self.nrt = 0
self.nnrt = 0
self.rt_times = []
self.nrt_times = []
self.rt_mean = []
self.nrt_mean = []
self.rt_percentile = []
self.nrt_percentile = []
self.results = Result()
class Simulation:
def __init__(self):
self.mc = 0
self.rtcl = 3
self.nrtcl = 5
self.scl = 4
self.server = 2
self.current = Node(2, 0, 4)
self.rt_queue = deque([])
self.nrt_queue = deque([])
self.preempt = -1
self.rt_iat = self.nrt_iat = self.rt_st = self.nrt_st = 0
# Results
self.m = 0
self.b = 0
self.nrt = 0
self.nnrt = 0
self.rt_times = []
self.nrt_times = []
self.rt_mean = []
self.nrt_mean = []
self.rt_percentile = []
self.nrt_percentile = []
self.results = Result()
def get_inputs(self):
# Get inputs
print("\nEnter following details: ")
self.rt_iat = float(
input("Enter mean inter-arrival time of RT messages: "))
self.nrt_iat = float(
input("Enter mean inter-arrival time of nonRT messages: "))
self.rt_st = float(input("Enter mean service time of RT message: "))
self.nrt_st = float(
input("Enter mean service time of non RT message: "))
self.m = int(input("Enter number of batches: "))
self.b = int(input("Enter batch size: "))
def calculate_time(self, t):
# Calculate psuedo-random time
time = (-1 * t) * log(random.random())
time = round(time, 4)
return time
def rt_arrival(self):
# Arrival of RT Event
self.mc = self.rtcl
self.rt_queue.append(Node(1, self.mc, self.calculate_time(self.rt_st)))
self.rtcl = round(self.mc + self.calculate_time(self.rt_iat), 4)
if len(
self.rt_queue
) > 0 and self.server == 0: # RT Queue not empty and server idle
self.current = self.rt_queue.popleft()
self.scl = round(self.mc + self.current.time, 4)
self.server = 1
elif len(
self.rt_queue
) > 0 and self.server == 2: # RT Queue not empty and server executing nRT job
remaining = self.scl - self.mc
self.current.time = round(remaining, 4)
if remaining != 0: # Preempt the current nRT job
self.preempt = self.current.time
self.nrt_queue.appendleft(self.current)
self.current = self.rt_queue.popleft()
self.scl = round(self.mc + self.current.time, 4)
self.server = 1
def nrt_arrival(self):
# Arrival of nRT Event
self.mc = self.nrtcl
self.nrt_queue.append(
Node(2, self.mc, self.calculate_time(self.nrt_st)))
self.nrtcl = round(self.mc + self.calculate_time(self.nrt_iat), 4)
if len(self.nrt_queue
) > 0 and self.server == 0: # Schedule the job is server idle
self.current = self.nrt_queue.popleft()
self.scl = round(self.mc + self.current.time, 4)
self.server = 2
def service_complete(self):
# Job completion event
self.mc = self.scl
if self.current.type == 1 and self.nrt != self.b: # Calculate total time for job
self.nrt += 1
self.rt_times.append(self.mc - self.current.arrival)
elif self.current.type == 2 and self.nnrt != self.b:
self.nnrt += 1
self.nrt_times.append(self.mc - self.current.arrival)
if len(self.rt_queue) > 0: # RT Queue is not empty
self.current = self.rt_queue.popleft()
self.scl = round(self.mc + self.current.time, 4)
self.server = 1
elif len(self.nrt_queue
) > 0 and self.preempt != -1: # Schedule preempted event
self.current = self.nrt_queue.popleft()
self.scl = round(self.mc + self.current.time, 4)
self.server = 2
self.preempt = -1
elif len(self.nrt_queue) > 0: # nRT queue is not empty
self.current = self.nrt_queue.popleft()
self.scl = round(self.mc + self.current.time, 4)
self.server = 2
else:
self.current = None
self.server = 0
def print_status(self):
# Print current system status
if self.server == 0:
data = "{}| {}| {}| {}| {}| {}| {}|".format(
str(self.mc).ljust(10),
str(self.rtcl).ljust(10),
str(self.nrtcl).ljust(10),
str(len(self.rt_queue)).ljust(10),
str(len(self.nrt_queue)).ljust(10), '-'.ljust(10),
str(self.server).ljust(15))
else:
data = "{}| {}| {}| {}| {}| {}| {}|".format(
str(self.mc).ljust(10),
str(self.rtcl).ljust(10),
str(self.nrtcl).ljust(10),
str(len(self.rt_queue)).ljust(10),
str(len(self.nrt_queue)).ljust(10),
str(self.scl).ljust(10),
str(self.server).ljust(15))
if (self.preempt != -1):
data += " s={}".format(str(self.preempt))
print("".ljust(102, '-'))
print(data)
def calculate_ci(self):
# Pop batch 0
self.rt_mean.pop(0)
self.nrt_mean.pop(0)
self.rt_percentile.pop(0)
self.nrt_percentile.pop(0)
# Response time for RT
rt_mean = np.mean(self.rt_mean)
rt_std = np.std(self.rt_mean)
rtci_lb = rt_mean - (2.0086 * (rt_std / sqrt(len(self.rt_mean))))
rtci_ub = rt_mean + (2.0086 * (rt_std / sqrt(len(self.rt_mean))))
# Percentile for RT
rt_percentile_mean = np.mean(self.rt_percentile)
rt_percentile_std = np.std(self.rt_percentile)
rtci_lb2 = rt_percentile_mean - (
2.0086 * (rt_percentile_std / sqrt(len(self.rt_mean))))
rtci_ub2 = rt_percentile_mean + (
2.0086 * (rt_percentile_std / sqrt(len(self.rt_mean))))
# Response time for nRT
nrt_mean = np.mean(self.nrt_mean)
nrt_std = np.std(self.nrt_mean)
nrtci_lb = nrt_mean - (2.0086 * (nrt_std / sqrt(len(self.nrt_mean))))
nrtci_ub = nrt_mean + (2.0086 * (nrt_std / sqrt(len(self.nrt_mean))))
# Percentile for nRT
nrt_percentile_mean = np.mean(self.nrt_percentile)
nrt_percentile_std = np.std(self.nrt_percentile)
nrtci_lb2 = nrt_percentile_mean - (
2.0086 * (nrt_percentile_std / sqrt(len(self.rt_mean))))
nrtci_ub2 = nrt_percentile_mean + (
2.0086 * (nrt_percentile_std / sqrt(len(self.rt_mean))))
# Calculate results for plotting graph
self.results.nrt_miat.append(self.nrt_iat)
self.results.rt_mean.append(round(rt_mean, 4))
self.results.rt_percentile.append(round(rt_percentile_mean, 4))
self.results.rtci.append((round(rtci_lb, 4), round(rtci_ub, 4)))
self.results.percentile_rtci.append((round(rtci_lb2,
4), round(rtci_ub2, 4)))
self.results.nrt_mean.append(round(nrt_mean, 4))
self.results.nrt_percentile.append(round(nrt_percentile_mean, 4))
self.results.nrtci.append((round(nrtci_lb, 4), round(nrtci_ub, 4)))
self.results.percentile_nrtci.append((round(nrtci_lb2,
4), round(nrtci_ub2, 4)))
# Print batch results
print("\n\nMean Inter-arrival time of nRT message: {}".format(
self.nrt_iat))
print("Mean of RT response time: {:0.4f}".format(rt_mean))
print("95th percentile of RT response time: {:0.4f}".format(
rt_percentile_mean))
print("Mean of nRT response time: {:0.4f}".format(nrt_mean))
print("95th percentile of nRT response time: {:0.4f}".format(
nrt_percentile_mean))
print("RT Mean Confidence Interval: {:0.4f} - {:0.4f}".format(
rtci_lb, rtci_ub))
print("nRT Mean Confidence Interval: {:0.4f} - {:0.4f}".format(
nrtci_lb, nrtci_ub))
print("RT Percentile Confidence Interval: {:0.4f} - {:0.4f}".format(
rtci_lb2, rtci_ub2))
print("nRT Percentile Confidence Interval: {:0.4f} - {:0.4f}".format(
nrtci_lb2, nrtci_ub2))
def process(self):
# Run the simulation
while (True):
if self.nrt == self.b and self.nnrt == self.b: # Batch complete
rt_mean = np.mean(self.rt_times)
nrt_mean = np.mean(self.nrt_times)
rt_percentile = np.percentile(self.rt_times, 95)
nrt_percentile = np.percentile(self.nrt_times, 95)
self.rt_mean.append(round(rt_mean, 4))
self.nrt_mean.append(round(nrt_mean, 4))
self.rt_percentile.append(round(rt_percentile, 4))
self.nrt_percentile.append(round(nrt_percentile, 4))
# Reset the RT and nRT buffers
self.rt_times = []
self.nrt_times = []
self.nrt = self.nnrt = 0
self.m -= 1
if self.m == 0:
break
clocks = []
# Find which event occurs next
if self.server == 0:
clocks = [self.rtcl, self.nrtcl]
else:
clocks = [self.rtcl, self.nrtcl, self.scl]
if self.rtcl == min(clocks):
self.rt_arrival()
elif self.nrtcl == min(clocks):
self.nrt_arrival()
else:
self.service_complete()
# Calculate confidence interval
self.calculate_ci()
def reset(self, no_batches):
# Reset for new batch
self.mc = 0
self.rtcl = 3
self.nrtcl = 5
self.scl = 4
self.server = 2
self.current = Node(2, 0, 4)
self.rt_queue = deque([])
self.nrt_queue = deque([])
self.preempt = -1
self.nrt = 0
self.nnrt = 0
self.rt_times = []
self.nrt_times = []
self.rt_mean = []
self.nrt_mean = []
self.rt_percentile = []
self.nrt_percentile = []
self.m = no_batches
def run(self):
# Start Simulation
self.get_inputs()
no_batches = self.m
while (self.nrt_iat <= max_nrt_miat):
self.reset(no_batches)
self.process()
self.nrt_iat += nrt_miat_increments
self.results.display()
from time import time
from experiments.common import DCIclassify
from model.dci import DCI
from model.pivotselection import pivot_selection
import numpy as np
from util.results import Result
optimize = True
mds_home= '../datasets/MDS'
dataset_home='../datasets/Webis-CLS-10'
nfolds=5
outfile = './DCI.varpivot.dat'
if exists(outfile):
rperf = Result.load(outfile, False)
else:
rperf = Result(['dataset', 'task', 'method', 'fold', 'npivots', 'acc', 'dci_time', 'svm_time'])
pivot_range = [10,25,50,100,250,500,1000,1500,2000,2500,5000]
for source, target, fold, taskname in MDS_task_generator(abspath(mds_home), nfolds=nfolds):
s_pivots, t_pivots = pivot_selection(max(pivot_range), source.X, source.y, source.U, target.U,
source.V, target.V,
phi=1, cross=True)
for npivots in pivot_range:
for dcf in ['cosine','linear']:
dci = DCI(dcf=dcf, unify=False, post='normal')
acc, dci_time, svm_time, _ = DCIclassify(source, target, s_pivots[:npivots], t_pivots[:npivots], dci, optimize=True)
import numpy as np
from util.file import *
from quantification.helpers import *
from util.plot_correction import plot_corr
from util.results import Result
VECTORSPATH = '../vectors'
sample_size = 200
samples_by_prevalence = 100
results_table = Result()
def add_results(results, dataset, method, approach):
for metric, score in results.items():
results_table.add(dataset=dataset,
method=method,
approach=approach,
metric=metric,
score=score)
for vectorset_name in list_dirs(VECTORSPATH):
print(vectorset_name)
vectorset = join(VECTORSPATH, vectorset_name)
train_x = np.load(join(vectorset, 'train.vec.npy'))
train_y = np.load(join(vectorset, 'train.y.npy'))
train_y_pred = np.load(join(vectorset, 'train.y_pred.npy'))
train_y_prob = np.load(join(vectorset, 'train.y_prob.npy'))
from data.tasks_topics import Topic_task_generator
from model.dci import DCI
from experiments.common import DCIinduction, pivot_selection_timed
from util.results import Result
optimize = True
npivots = 1000
dcf = 'cosine'
twentynews_home = '../datasets/20news'
sraa_home = '../datasets/SRAA'
reuters_home = '../datasets/Reuters21578'
rperf = Result([
'dataset', 'task', 'method', 'acc', 'pivot_t', 'dci_t', 'svm_t', 'test_t'
])
for source, target, task, dataset in \
Topic_task_generator(reuters_home=reuters_home, sraa_home=sraa_home, twenty_home=twentynews_home):
s_pivots, t_pivots, pivot_time = pivot_selection_timed(npivots,
source.X,
source.y,
source.U,
target.U,
source.V,
target.V,
phi=1,
cross=True,
show=min(
10, npivots))
from data.tasks import WebisCLS10_task_generator, WebisCLS10_crossdomain_crosslingual_task_generator
from experiments.common import DCIclassify
from model.dci import DCI
from model.pivotselection import pivot_selection
import os
from time import time
from util.results import Result
dcf = 'cosine'
npivots = 450
optimize = True
dataset_home = '../datasets/Webis-CLS-10'
rperf = Result([
'dataset', 'task', 'method', 'acc', 'pivot_t', 'dci_t', 'svm_t', 'test_t'
])
for source, target, oracle, taskname in WebisCLS10_crossdomain_crosslingual_task_generator(
os.path.abspath(dataset_home)):
# pivot selection
tinit = time()
s_pivots, t_pivots = pivot_selection(npivots,
source.X,
source.y,
source.U,
target.U,
source.V,
target.V,
oracle=oracle,
phi=30,
from sklearn.model_selection import cross_val_score, GridSearchCV
from sklearn.svm import LinearSVC
from data.tasks import WebisCLS10_task_generator
from data.domain import unify_feat_space
import os
from util.results import Result
parameters = {'C': [10 ** i for i in range(-5, 5)]}
dataset_home='../datasets/Webis-CLS-10'
results = Result(['dataset', 'task', 'method', 'acc'])
for source, target, oracle, taskname in WebisCLS10_task_generator(os.path.abspath(dataset_home)):
# upper
svm = GridSearchCV(LinearSVC(), parameters, n_jobs=-1, verbose=1, cv=5)
source, target = unify_feat_space(source, target)
acc = cross_val_score(svm, target.X, target.y, cv=5).mean()
results.add(dataset='Webis-CLS-10', task=taskname, method='Upper', acc=acc)
# lower
svm = GridSearchCV(LinearSVC(), parameters, n_jobs=-1, verbose=1, cv=5)
svm.fit(source.X, source.y)
yte_ = svm.predict(target.X)
acc = (target.y == yte_).mean()
results.add(dataset='Webis-CLS-10', task=taskname, method='Lower', acc=acc)
results.pivot(grand_totals=True)
from sklearn.model_selection import cross_val_score, GridSearchCV
from sklearn.svm import LinearSVC
from data.domain import unify_feat_space
from data.tasks import MDS_task_generator, UpperMDS_task_generator
from os.path import abspath
from util.results import Result
dataset_home='../datasets/MDS'
nfolds=5
upper={}
parameters = {'C': [10 ** i for i in range(-5, 5)]}
results = Result(['dataset', 'task', 'method', 'fold', 'acc'])
for domain in UpperMDS_task_generator(abspath(dataset_home)):
svm = GridSearchCV(LinearSVC(), parameters, n_jobs=-1, verbose=1, cv=5)
upper[domain.domain] = cross_val_score(svm, domain.X, domain.y, cv=5).mean()
for source, target, fold, taskname in MDS_task_generator(abspath(dataset_home), nfolds=nfolds):
svm = GridSearchCV(LinearSVC(), parameters, n_jobs=-1, verbose=1, cv=5)
target_domain_name = target.domain
source, target = unify_feat_space(source, target)
svm.fit(source.X, source.y)
yte_ = svm.predict(target.X)
acc = (yte_ == target.y).mean()
results.add(dataset='MDS', task=taskname, method='Lower', fold=fold, acc=acc)
results.add(dataset='MDS', task=taskname, method='Upper', fold=fold, acc=upper[target_domain_name])
results.pivot(grand_totals=True)
from sklearn.svm import LinearSVC
from data.domain import *
from sklearn.model_selection import cross_val_score
from data.tasks_topics import Topic_task_generator
from util.results import Result
twentynews_home = '../datasets/20news'
sraa_home = '../datasets/SRAA'
reuters_home = '../datasets/Reuters21578'
results = Result(['dataset', 'task', 'method', 'acc'])
for source, target, task, dataset in \
Topic_task_generator(reuters_home=reuters_home, sraa_home=sraa_home, twenty_home=twentynews_home):
source, target = unify_feat_space(source, target)
svm = LinearSVC()
svm.fit(source.X, source.y)
yte_ = svm.predict(target.X)
acc = (yte_ == target.y).mean()
results.add(dataset=dataset, task=task, method='Lower', acc=acc)
upper = cross_val_score(svm, target.X, target.y, cv=5).mean()
results.add(dataset=dataset, task=task, method='Upper', acc=upper)
results.pivot(grand_totals=True)
from os.path import abspath
from time import time
from experiments.common import DCIclassify
from model.dci import DCI
from model.pivotselection import pivot_selection
from util.results import Result
dcf = 'cosine'
npivots = 1000
optimize = True
mds_home = '../datasets/MDS'
nfolds = 5
rperf = Result([
'dataset', 'task', 'method', 'fold', 'acc', 'pivot_t', 'dci_t', 'svm_t',
'test_t'
])
for source, target, fold, taskname in MDS_task_generator(abspath(mds_home),
nfolds=nfolds):
tinit = time()
s_pivots, t_pivots = pivot_selection(npivots,
source.X,
source.y,
source.U,
target.U,
source.V,
target.V,
phi=1,
show=min(10, npivots),
from sklearn.model_selection import train_test_split
import pickle
from tqdm import tqdm
from sklearn.model_selection import StratifiedKFold
from os.path import join
import scipy
from util.file import *
cuda = torch.device('cuda')
VECTORSPATH = '../vectors'
sample_size = 200
samples_by_prevalence = 100
results_table = Result()
def add_results(results, dataset, method, approach):
for metric, score in results.items():
results_table.add(dataset=dataset,
method=method,
approach=approach,
metric=metric,
score=score)
for iter in range(10):
hidden = 64
drop_p = 0.0
layers = 2
from experiments.common import pivot_selection_timed, DCItransduction
from model.dci import DCI
import os
from util.results import Result
optimize = False
transductive = True
dcf='cosine'
npivots = 450
svmlight_home='../../svm_light'
dataset_home='../datasets/Webis-CLS-10'
methodname = ('T' if transductive else 'I') + f'DCI'
rperf = Result(['dataset', 'task', 'method', 'acc', 'pivot_t', 'dci_t', 'svm_t', 'test_t'])
for source, target, oracle, taskname in WebisCLS10_task_generator(os.path.abspath(dataset_home)):
s_pivots, t_pivots, pivot_time = pivot_selection_timed(
npivots, source.X, source.y, source.U, target.U, source.V, target.V,
oracle=oracle, phi=30, show=min(20, npivots), cross=False
)
dci = DCI(dcf=dcf, unify=True, post='normal')
acc, dci_time, svm_time, test_time = DCItransduction(
source, target, s_pivots, t_pivots, dci, svmlight_home, optimize=optimize, transductive=transductive
)
rperf.add(dataset='Webis-CLS-10', task=taskname, method=methodname,
acc=acc,
pivot_t=pivot_time, dci_t=dci_time, svm_t=svm_time, test_t=test_time)
from sklearn.model_selection import cross_val_score
from data.tasks import MDS_task_generator, UpperMDS_task_generator
from data.domain import *
from os.path import abspath
from classification.svmlight import SVMlight
from util.results import Result
dataset_home='../datasets/MDS'
svmlight_home='../../svm_light'
nfolds=5
results = Result(['dataset', 'task', 'method', 'acc'])
for domain in UpperMDS_task_generator(abspath(dataset_home)):
isvm = SVMlight(svmlightbase=svmlight_home, verbose=0, transduction=None)
score = cross_val_score(isvm, domain.X, domain.y, cv=nfolds).mean()
results.add(dataset='MDS', task=domain.domain, method='UPPER', acc=score)
results.pivot(grand_totals=True)
results = Result(['dataset', 'task', 'method', 'fold', 'acc'])
for source, target, fold, task in MDS_task_generator(abspath(dataset_home), nfolds=nfolds):
source_name = source.domain
target_name = target.domain
source, target = unify_feat_space(source, target)
isvm = SVMlight(svmlightbase=svmlight_home, verbose=0, transduction=None).fit(source.X, source.y)
tsvm = SVMlight(svmlightbase=svmlight_home, verbose=0, transduction=target.X).fit(source.X, source.y)
yte_ = isvm.predict(target.X)
tyte_ = tsvm.transduced_labels