def job(model_type, f, train_set, test_set, output_fold, batch_size, pc, ps):
cm = SDBN(pc, ps)
for r in range(1, 2):
# np.random.seed(r)
ranker = PDGDLinearRanker(FEATURE_SIZE, Learning_rate)
print("PDGD fold{} {} run{} start!".format(f, model_type, r))
ndcg_scores, cndcg_scores, cmrr_scores, final_weights = run(
train_set, test_set, ranker, NUM_INTERACTION, cm, batch_size)
os.makedirs(os.path.dirname("{}/fold{}/".format(output_fold, f)),
exist_ok=True) # create directory if not exist
with open(
"{}/fold{}/{}_run{}_ndcg.txt".format(output_fold, f,
model_type, r),
"wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/fold{}/{}_run{}_cndcg.txt".format(output_fold, f,
model_type, r),
"wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"{}/fold{}/{}_run{}_cmrr.txt".format(output_fold, f,
model_type, r),
"wb") as fp:
pickle.dump(cmrr_scores, fp)
with open(
"{}/fold{}/{}_run{}_weights.txt".format(
output_fold, f, model_type, r), "wb") as fp:
pickle.dump(final_weights, fp)
print("PDGD fold{} {} run{} finished!".format(f, model_type, r))
def job(model_type, f, train_set, test_set, tau, sigma, gamma, num_rankers, learning_rate_decay, output_fold):
if model_type == "perfect":
pc = [0.0, 0.2, 0.4, 0.8, 1.0]
ps = [0.0, 0.0, 0.0, 0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
elif model_type == "informational":
pc = [0.4, 0.6, 0.7, 0.8, 0.9]
ps = [0.1, 0.2, 0.3, 0.4, 0.5]
cm = SDBN(pc, ps)
for r in range(1, 26):
# np.random.seed(r)
ranker = ESLinearRanker(FEATURE_SIZE, Learning_rate, sigma, tau, gamma, learning_rate_decay=learning_rate_decay)
print("ES fold{} {} run{} start!".format(f, model_type, r))
ndcg_scores, cndcg_scores, final_weight = run(train_set, test_set, ranker, NUM_INTERACTION, cm, num_rankers)
with open(
"{}/fold{}/{}_sigma{}_run{}_ndcg.txt".format(output_fold, f, model_type, sigma, r),
"wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/fold{}/{}_sigma{}_run{}_cndcg.txt".format(output_fold, f, model_type, sigma, r),
"wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"{}/fold{}/{}_sigma{}_run{}_final_weight.txt".format(output_fold, f, model_type, sigma, r),
"wb") as fp:
pickle.dump(final_weight, fp)
print("ES sigma{} fold{} {} run{} finished!".format(output_fold, sigma, f, model_type, r))
def job(model_type, f, train_set, test_set, tau, r):
if model_type == "perfect":
pc = [0.0, 0.5, 1.0]
ps = [0.0, 0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.5, 0.95]
ps = [0.2, 0.5, 0.9]
elif model_type == "informational":
pc = [0.4, 0.7, 0.9]
ps = [0.1, 0.3, 0.5]
cm = SDBN(train_set, pc, ps)
# np.random.seed(r)
ranker1 = PDGDLinearRanker(FEATURE_SIZE, Learning_rate, tau)
ranker2 = PDGDLinearRanker(FEATURE_SIZE, Learning_rate, tau)
print("PDGD tau{} fold{} {} run{} start!".format(tau, f, model_type, r))
final_weight1, final_weight2 = run(train_set, test_set, ranker1, ranker2,
NUM_INTERACTION, cm)
with open(
"./results/multiple_ranker/mq2007/PDGD/fold{}/{}_tau{}_run{}_ranker1_weights.txt"
.format(f, model_type, tau, r), "wb") as fp:
pickle.dump(final_weight1, fp)
with open(
"./results/multiple_ranker/mq2007/PDGD/fold{}/{}_tau{}_run{}_ranker2_weights.txt"
.format(f, model_type, tau, r), "wb") as fp:
pickle.dump(final_weight2, fp)
print("PDGD tau{} fold{} {} run{} finished!".format(tau, f, model_type, r))
def job(model_type, f, train_set, test_set, tau):
if model_type == "perfect":
pc = [0.0, 0.5, 1.0]
ps = [0.0, 0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.5, 0.95]
ps = [0.2, 0.5, 0.9]
elif model_type == "informational":
pc = [0.4, 0.7, 0.9]
ps = [0.1, 0.3, 0.5]
cm = SDBN(pc, ps)
for r in range(1, 26):
# np.random.seed(r)
ranker = PDGDLinearRanker(FEATURE_SIZE, Learning_rate, tau)
print("PDGD tau{} fold{} {} run{} start!".format(
tau, f, model_type, r))
ndcg_scores, cndcg_scores, final_weight = run(train_set, test_set,
ranker, NUM_INTERACTION,
cm)
with open(
"../results/exploration/mq2007/PDGD/fold{}/{}_tau{}_run{}_ndcg.txt"
.format(f, model_type, tau, r), "wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"../results/exploration/mq2007/PDGD/fold{}/{}_tau{}_run{}_cndcg.txt"
.format(f, model_type, tau, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"../results/exploration/mq2007/PDGD/fold{}/{}_tau{}_run{}_final_weight.txt"
.format(f, model_type, tau, r), "wb") as fp:
pickle.dump(final_weight, fp)
print("PDGD tau{} fold{} {} run{} finished!".format(
tau, f, model_type, r))
def job(model_type, Learning_rate, NUM_INTERACTION, f, train_set, intent_paths,
output_fold, num_groups, group_sequence):
if model_type == "perfect":
pc = [0.0, 1.0]
ps = [0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.95]
ps = [0.2, 0.9]
elif model_type == "informational":
pc = [0.3, 0.7]
ps = [0.1, 0.5]
elif model_type == "noisy":
pc = [0.4, 0.6]
ps = [0.0, 0.0]
# cm = PBM(pc, 1)
cm = SDBN(pc, ps)
for r in range(1, 26):
random.seed(r)
np.random.seed(r)
datasets = get_groups_dataset(train_set,
intent_paths,
num_groups=num_groups)
ranker = PDGDLinearRanker(FEATURE_SIZE, Learning_rate)
print("PDGD intent change {} fold{} run{} start!".format(
model_type, f, r))
ndcg_scores, cndcg_scores = run(datasets, ranker, NUM_INTERACTION, cm,
group_sequence)
# create directory if not exist
os.makedirs(os.path.dirname("{}/current_intent/fold{}/".format(
output_fold, f)),
exist_ok=True)
with open(
"{}/current_intent/fold{}/{}_run{}_cndcg.txt".format(
output_fold, f, model_type, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"{}/current_intent/fold{}/{}_run{}_ndcg.txt".format(
output_fold, f, model_type, r), "wb") as fp:
pickle.dump(ndcg_scores[0], fp)
for i in range(len(ndcg_scores) - 1): # the intent ndcg start from 1.
os.makedirs(os.path.dirname("{}/intent{}/fold{}/".format(
output_fold, i + 1, f)),
exist_ok=True) # create directory if not exist\
with open(
"{}/intent{}/fold{}/{}_run{}_ndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(ndcg_scores[i + 1], fp)
print("PDGD intent change {} run{} finish!".format(model_type, r))
print()
def job(model_type, f, train_set, test_set, tau, step_size, gamma, num_rankers,
learning_rate_decay, output_fold):
if model_type == "perfect":
pc = [0.0, 0.5, 1.0]
ps = [0.0, 0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.5, 0.95]
ps = [0.2, 0.5, 0.9]
elif model_type == "informational":
pc = [0.4, 0.7, 0.9]
ps = [0.1, 0.3, 0.5]
# if model_type == "perfect":
# pc = [0.0, 0.2, 0.4, 0.8, 1.0]
# ps = [0.0, 0.0, 0.0, 0.0, 0.0]
# elif model_type == "navigational":
# pc = [0.05, 0.3, 0.5, 0.7, 0.95]
# ps = [0.2, 0.3, 0.5, 0.7, 0.9]
# elif model_type == "informational":
# pc = [0.4, 0.6, 0.7, 0.8, 0.9]
# ps = [0.1, 0.2, 0.3, 0.4, 0.5]
cm = SDBN(pc, ps)
for r in range(1, 26):
# np.random.seed(r)
ranker = COLTRLinearRanker(FEATURE_SIZE,
Learning_rate,
step_size,
tau,
gamma,
learning_rate_decay=learning_rate_decay)
print("COTLR {} tau{} fold{} {} run{} start!".format(
output_fold, tau, f, model_type, r))
ndcg_scores, cndcg_scores, final_weight = run(train_set, test_set,
ranker, NUM_INTERACTION,
cm, num_rankers)
with open(
"{}/fold{}/{}_tau{}_run{}_ndcg.txt".format(
output_fold, f, model_type, tau, r), "wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/fold{}/{}_tau{}_run{}_cndcg.txt".format(
output_fold, f, model_type, tau, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"{}/fold{}/{}_tau{}_run{}_final_weight.txt".format(
output_fold, f, model_type, tau, r), "wb") as fp:
pickle.dump(final_weight, fp)
print("COTLR {} tau{} fold{} {} run{} finished!".format(
output_fold, tau, f, model_type, r))
utility.send_progress("@arvin {}".format(model_type), r, 25,
"final ndcg {}".format(ndcg_scores[-1]))
def job(model_type, Learning_rate, NUM_INTERACTION, f, train_set, intent_paths,
output_fold, num_groups):
if model_type == "perfect":
pc = [0.0, 1.0]
ps = [0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.95]
ps = [0.2, 0.9]
elif model_type == "informational":
pc = [0.3, 0.7]
ps = [0.1, 0.5]
elif model_type == "noisy":
pc = [0.4, 0.6]
ps = [0.0, 0.0]
# cm = PBM(pc, 1)
cm = SDBN(pc, ps)
for r in range(1, 26):
random.seed(r)
np.random.seed(r)
datasets = get_groups_dataset(train_set,
intent_paths,
num_groups=num_groups)
# for i in range(len(datasets)):
for i in [1, 0]:
ranker = PDGDLinearRanker(FEATURE_SIZE, Learning_rate)
print("PDGD intent fixed {} intent {} run{} start!".format(
model_type, i, r))
ndcg_scores, cndcg_scores = run(datasets[i], ranker,
NUM_INTERACTION, cm)
os.makedirs(os.path.dirname("{}/group{}/fold{}/".format(
output_fold, i + 1, f)),
exist_ok=True)
with open(
"{}/group{}/fold{}/{}_run{}_ndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/group{}/fold{}/{}_run{}_cndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
print("PDGD intent fixed {} intent {} run{} finished!".format(
model_type, i, r))
print()
def job(model_type, f, train_set, intent_paths, tau, step_size, gamma,
num_rankers, learning_rate_decay, output_fold):
if model_type == "perfect":
pc = [0.0, 1.0]
ps = [0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.95]
ps = [0.2, 0.9]
elif model_type == "informational":
pc = [0.3, 0.7]
ps = [0.1, 0.5]
# cm = PBM(pc, 1)
cm = SDBN(pc, ps)
for r in range(1, 26):
random.seed(r)
np.random.seed(r)
datasets = get_groups_dataset(train_set, intent_paths)
for i in range(len(datasets)):
ranker = COLTRLinearRanker(FEATURE_SIZE,
Learning_rate,
step_size,
tau,
gamma,
learning_rate_decay=learning_rate_decay)
print("COLTR fixed intent {} fold{} run{} start!".format(
model_type, f, r))
ndcg_scores, cndcg_scores = run(datasets[i], ranker,
NUM_INTERACTION, cm, num_rankers)
# create directory if not exist
os.makedirs(os.path.dirname("{}/group{}/fold{}/".format(
output_fold, i + 1, f)),
exist_ok=True)
with open(
"{}/group{}/fold{}/{}_run{}_ndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/group{}/fold{}/{}_run{}_cndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
print("COLTR fixed intent {} run{} finish!".format(model_type, r))
print()
def job(model_type, f, train_set, intent_paths, delta, alpha, FEATURE_SIZE,
num_rankers, output_fold):
if model_type == "perfect":
pc = [0.0, 1.0]
ps = [0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.95]
ps = [0.2, 0.9]
elif model_type == "informational":
pc = [0.3, 0.7]
ps = [0.1, 0.5]
# cm = PBM(pc, 1)
cm = SDBN(pc, ps)
for r in range(1, 26):
random.seed(r)
np.random.seed(r)
datasets = get_groups_dataset(train_set, intent_paths)
# create directory if not exist
for i in range(len(datasets)):
ranker = ProbabilisticRanker(delta, alpha, FEATURE_SIZE)
print("PDGD intent fixed {} intent {} run{} start!".format(
model_type, i, r))
ndcg_scores, cndcg_scores = run(datasets[i], ranker,
NUM_INTERACTION, cm, num_rankers)
os.makedirs(os.path.dirname("{}/group{}/fold{}/".format(
output_fold, i + 1, f)),
exist_ok=True)
with open(
"{}/group{}/fold{}/{}_run{}_ndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/group{}/fold{}/{}_run{}_cndcg.txt".format(
output_fold, i + 1, f, model_type, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
print("PDGD intent fixed {} intent {} run{} finished!".format(
model_type, i, r))
print()
def job(model_type, f, train_set, test_set, output_fold):
if model_type == "perfect":
# pc = [0.0, 0.2, 0.4, 0.8, 1.0]
# pc = [0.0, 0.5, 1.0]
pc = [0.0, 1.0]
# ps = [0.0, 0.0, 0.0, 0.0, 0.0]
# ps = [0.0, 0.0, 0.0]
ps = [0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
# pc = [0.05, 0.5, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
# ps = [0.2, 0.5, 0.9]
elif model_type == "informational":
# pc = [0.4, 0.6, 0.7, 0.8, 0.9]
# pc = [0.4, 0.7, 0.9]
pc = [0.1, 0.9]
# ps = [0.1, 0.2, 0.3, 0.4, 0.5]
# ps = [0.1, 0.3, 0.5]
ps = [0.1, 0.5]
cm = SDBN(pc, ps)
for r in range(1, 16):
# np.random.seed(r)
FEATURE_SIZE = 136
ranker = PDGDLinearRanker(FEATURE_SIZE, Learning_rate)
print("PDGD fold{} {} run{} start!".format(f, model_type, r))
ndcg_scores, cndcg_scores, final_weights = run(train_set, test_set, ranker, NUM_INTERACTION, cm)
os.makedirs(os.path.dirname("{}/fold{}/".format(output_fold, f)),
exist_ok=True) # create directory if not exist
with open(
"{}/fold{}/{}_run{}_ndcg.txt".format(output_fold, f, model_type, r),
"wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"{}/fold{}/{}_run{}_cndcg.txt".format(output_fold, f, model_type, r),
"wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"{}/fold{}/{}_run{}_weights.txt".format(output_fold, f, model_type, r),
"wb") as fp:
pickle.dump(final_weights, fp)
print("PDGD fold{} {} run{} finished!".format(f, model_type, r))
def job(model_type, f, train_set, test_set, tau, step_size, gamma, num_rankers,
learning_rate_decay):
if model_type == "perfect":
pc = [0.0, 0.5, 1.0]
ps = [0.0, 0.0, 0.0]
elif model_type == "navigational":
pc = [0.05, 0.5, 0.95]
ps = [0.2, 0.5, 0.9]
elif model_type == "informational":
pc = [0.4, 0.7, 0.9]
ps = [0.1, 0.3, 0.5]
cm = SDBN(pc, ps)
for r in range(1, 26):
# np.random.seed(r)
ranker = CMAESLinearRanker(FEATURE_SIZE,
Learning_rate,
step_size,
tau,
gamma,
learning_rate_decay=learning_rate_decay)
print("COTLR start!")
ndcg_scores, cndcg_scores, final_weight = run(train_set, test_set,
ranker, NUM_INTERACTION,
cm, num_rankers)
with open(
"../results/COLTR/mq2007/fold{}/{}_tau{}_run{}_ndcg.txt".
format(f, model_type, tau, r), "wb") as fp:
pickle.dump(ndcg_scores, fp)
with open(
"../results/COLTR/mq2007/fold{}/{}_tau{}_run{}_cndcg.txt".
format(f, model_type, tau, r), "wb") as fp:
pickle.dump(cndcg_scores, fp)
with open(
"../results/COLTR/mq2007/fold{}/{}_tau{}_run{}_final_weight.txt"
.format(f, model_type, tau, r), "wb") as fp:
pickle.dump(final_weight, fp)
print("COTLR tau{} fold{} {} run{} finished!".format(
tau, f, model_type, r))
from clickModel.RCTR import RCTR
from clickModel.Mixed import Mixed
from utils import read_file as rf
from utils import utility
from dataset import LetorDataset
# import matplotlib.pyplot as plt
import numpy as np
import multiprocessing as mp
train_path = "../datasets/ltrc_yahoo/set1.train.txt"
print("loading training set.......")
train_set = LetorDataset(train_path, 700)
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
mixed_models = [DCTR(pc), SDBN(pc, ps), UBM(pc)]
datasets_simulator = [
('SDBN', SDBN(pc, ps)),
# ('SDCM', SDCM(pc)),
# ('CM', CM(pc)),
('DCTR', DCTR(pc)),
('UBM', UBM(pc)),
('SDBN_reverse', SDBN_reverse(pc, ps))
]
click_model = RCTR()
for dataset, simulator in datasets_simulator:
for id in range(1, 16):
click_log_path = "../click_logs/{}/train_set{}.txt".format(dataset, id)
click_log = rf.read_click_log(click_log_path)
click_model.train(click_log)
# plt.show()
# %%
if __name__ == "__main__":
# %%
train_path = "../datasets/ltrc_yahoo/set1.LetorDataset.txt"
print("loading training set.......")
with open(train_path, "rb") as fp:
train_set = pickle.load(fp)
# %%
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
datasets_simulator = [
('SDBN', SDBN(pc, ps)),
# ('SDCM', SDCM(pc)),
# ('CM', CM(pc)),
('DCTR', DCTR(pc)),
('UBM', UBM(pc)),
# ('SDBN_reverse', SDBN_reverse(pc, ps))
]
progress = 0
for dataset, simulator in datasets_simulator:
for id in range(2, 16):
click_log_path = "../click_logs/{}/train_set{}.txt".format(
dataset, id)
test_click_log_path = "../click_logs/{}/seen_set{}.txt".format(
dataset, id)
query_frequency_path = "../click_logs/{}/query_frequency{}.txt".format(
print("loading testing set.......")
test_set = LetorDataset(test_path, 700)
# %%
# pc = [0.4, 0.6, 0.7, 0.8, 0.9]
# ps = [0.1, 0.2, 0.3, 0.4, 0.5]
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
for id in range(1, 16):
p1 = mp.Process(target=generate_dataset,
args=(train_set, test_set, DCTR(pc),
"../feature_click_datasets/DCTR/", id))
p2 = mp.Process(target=generate_dataset,
args=(train_set, test_set, CM(pc),
"../feature_click_datasets/CM/", id))
p3 = mp.Process(target=generate_dataset,
args=(train_set, test_set, SDBN(pc, ps),
"../feature_click_datasets/SDBN/", id))
p4 = mp.Process(target=generate_dataset,
args=(train_set, test_set, SDCM(pc),
"../feature_click_datasets/SDCM/", id))
p1.start()
p2.start()
p3.start()
p4.start()
p1.join()
p2.join()
p3.join()
p4.join()
print(" ")
import pickle
import bz2
from dataset import LetorDataset
from clickModel.SDBN import SDBN
import sys
import tensorflow as tf
train_path = "../datasets/ltrc_yahoo/set1.train.txt"
print("loading training set.......")
dataset = LetorDataset(train_path, 700)
model = NCM(64, 1024, 10240)
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
simulator = SDBN(pc, ps)
click_log_path = "../feature_click_datasets/{}/train_set_test.txt".format(
"SDBN", "1")
# click_log_path = "../click_logs/{}/train_set{}_small.txt".format("SDBN", "1")
click_log = rf.read_click_log(click_log_path)
model.initial_representation(click_log)
# model.save_training_set_numpy(click_log, "test", "SDBN")
# model.save_training_set(click_log, "../click_logs/{}/train_set{}_small_NCM.tfrecord".format("SDBN", "1"), "SDBN")
model.save_training_tfrecord(click_log, "NCM_test.tfrecord", "SDBN")
# model.save_training_set_numpy(click_log, "../click_logs/{}/train_set{}_NCM".format("SDBN", "1"), "SDBN")
# model.save_training_set_numpy(click_log, "test", "SDBN")
# plt.show()
# %%
if __name__ == "__main__":
# %%
train_path = "../datasets/ltrc_yahoo/set1.train.txt"
test_path = "../datasets/ltrc_yahoo/set1.test.txt"
print("loading training set.......")
train_set = LetorDataset(train_path, 700)
# %%
# print("loading testing set.......")
# test_set = LetorDataset(test_path, 700)
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
datasets_simulator = [('SDBN', SDBN(pc, ps)), ('SDCM', SDCM(pc)),
('CM', CM(pc)), ('DCTR', DCTR(pc))]
# datasets = ['CM']
for dataset, simulator in datasets_simulator:
for id in range(1, 2):
click_log_path = "../feature_click_datasets/{}/train_set{}.txt".format(
dataset, id)
test_click_log_path = "../feature_click_datasets/{}/seen_set{}.txt".format(
dataset, id)
query_frequency_path = "../feature_click_datasets/{}/query_frequency{}.txt".format(
dataset, id)
click_log = rf.read_click_log(click_log_path)
test_click_log = rf.read_click_log(test_click_log_path)
query_frequency = rf.read_query_frequency(query_frequency_path)
click_models = [SDBN(), SDCM(), CM(), DCTR()]
for freq in frequencies:
perplexities = click_model.get_perplexity(np.array(test_logs[freq]))
MSEs = click_model.get_MSE(np.array(test_logs[freq]), dataset, simulator)
perplexity_line = "Frequency " + freq + " perplexities:"
MSEs_line = "Frequency " + freq + " MSE:"
for perp in perplexities:
perplexity_line += " " + str(perp)
for MSE in MSEs:
MSEs_line += " " + str(MSE)
f.write(perplexity_line + "\n")
f.write(MSEs_line + "\n")
f.close()
if __name__ == "__main__":
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
Mixed_models = [DCTR(pc), SDBN(pc, ps), UBM(pc)]
# simulators = [SDBN(pc, ps), Mixed(Mixed_models), DCTR(pc), UBM(pc)]
simulators = [SDBN(pc, ps), DCTR(pc), UBM(pc)]
dataset_path = "../datasets/ltrc_yahoo/set1.train.txt"
print("loading training set.......")
dataset = LetorDataset(dataset_path, 700)
for r in range(1, 2):
for simulator in simulators:
run(simulator, dataset, r)
generator = "Mixed"
click_log_path = "../feature_click_datasets/{}/train_set1.txt".format(generator)
test_click_log_path = "../feature_click_datasets/{}/seen_set1.txt".format(generator)
click_log = rf.read_click_log(click_log_path)
test_click_log = rf.read_click_log(test_click_log_path)
# #
dataset = tf.data.TFRecordDataset(filenames='../feature_click_datasets/{}/train_set1.tfrecord'.format(generator))
# # # test_dataset = tf.data.TFRecordDataset(filenames='../feature_click_datasets/SDBN/seen_set1.tfrecord')
# # #%%
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
Mixed_models = [DCTR(pc), CM(pc), SDBN(pc, ps), SDCM(pc), UBM(pc)]
simulator = Mixed(Mixed_models)
print(click_log.shape)
print(test_click_log.shape)
#
click_model = LSTMv2(700, 1024, train_set, batch_size=128, epoch=5)
print(click_model.get_MSE(test_click_log[np.random.choice(test_click_log.shape[0], 1000)], train_set, simulator))
click_model.train(dataset)
print(click_model.get_MSE(test_click_log[np.random.choice(test_click_log.shape[0], 1000)], train_set, simulator))
click_model.model.save("../click_model_results/LSTM_models/{}_train_set1.h5".format(generator))
# test model
def run(train_set, test_set, ranker1, ranker2, num_interation, click_model):
click_predictor = SDBN()
ndcg_scores1 = []
cndcg_scores1 = []
ndcg_scores2 = []
cndcg_scores2 = []
query_set = train_set.get_all_querys()
np.random.shuffle(query_set)
index = np.random.randint(query_set.shape[0], size=num_interation)
pdf = np.random.normal(size=query_set.shape[0])
e_x = np.exp((pdf - np.max(pdf)) / 0.2)
probs = e_x / e_x.sum(axis=0)
querys = np.random.choice(query_set,
replace=True,
p=probs,
size=num_interation)
num_interaction = 0
correct = 0
wrong = 0
test1 = 0
test2 = 0
for qid in querys:
num_interaction += 1
# qid = query_set[i]
result_list1, scores1 = ranker1.get_query_result_list(train_set, qid)
result_list2, scores2 = ranker2.get_query_result_list(train_set, qid)
clicked_doc1, click_label1, _ = click_model.simulate(
qid, result_list1, train_set)
clicked_doc2, click_label2, _ = click_model.simulate(
qid, result_list2, train_set)
#
last_exam = None
# if len(clicked_doc2) > 0:
# last_exam = np.where(click_label2 == 1)[0][-1] + 1
#
# click_predictor.online_training(qid, result_list2, click_label2)
# reduce, reduced_index = click_predictor.click_noise_reduce(qid, result_list2, click_label2, 0.5, 20)
#
# if reduce:
# for rank in reduced_index:
# # print(train_set.get_relevance_label_by_query_and_docid(qid, result_list2[rank]))
# if train_set.get_relevance_label_by_query_and_docid(qid, result_list2[rank]) == 0:
# correct += 1
# else:
# wrong += 1
# # print(correct, wrong)
clicked_doc_index = 0
for j in np.where(click_label2 == 1)[0]:
rel = train_set.get_relevance_label_by_query_and_docid(
qid, result_list2[clicked_doc_index])
if rel == 0:
click_label2[j] = 0
clicked_doc_index += 1
ranker1.update_to_clicks(click_label1, result_list1, scores1,
train_set.get_all_features_by_query(qid))
ranker2.update_to_clicks(click_label2, result_list2, scores2,
train_set.get_all_features_by_query(qid),
last_exam)
all_result1 = ranker1.get_all_query_result_list(test_set)
ndcg1 = evl_tool.average_ndcg_at_k(test_set, all_result1, 10)
cndcg1 = evl_tool.query_ndcg_at_k(train_set, result_list1, qid, 10)
all_result2 = ranker2.get_all_query_result_list(test_set)
ndcg2 = evl_tool.average_ndcg_at_k(test_set, all_result2, 10)
cndcg2 = evl_tool.query_ndcg_at_k(train_set, result_list2, qid, 10)
ndcg_scores1.append(ndcg1)
cndcg_scores1.append(cndcg1)
ndcg_scores2.append(ndcg2)
cndcg_scores2.append(cndcg2)
final_weight1 = ranker1.get_current_weights()
final_weight2 = ranker2.get_current_weights()
test1 += ndcg1
test2 += ndcg2
print(test1, test2)
print(np.mean(ndcg_scores1), np.mean(ndcg_scores2))
return ndcg_scores1, cndcg_scores1, final_weight1, ndcg_scores2, cndcg_scores2, final_weight2
from clickModel.NCM_TF import NCM
from utils import read_file as rf
import numpy as np
import pickle
from dataset import LetorDataset
from clickModel.SDBN import SDBN
model = NCM(774, 100, 10240+1024+1, 2)
pc = [0.05, 0.3, 0.5, 0.7, 0.95]
ps = [0.2, 0.3, 0.5, 0.7, 0.9]
simulator = SDBN(pc, ps)
click_log_path = "../feature_click_datasets/{}/train_set{}.txt".format("SDBN", "_test")
click_log = rf.read_click_log(click_log_path)
model.initial_representation(click_log)
# session = np.array(['1112', '16', '3', '45', '37', '31', '22', '5', '34', '17', '21', '0', '0', '1', '0', '0', '0', '0', '0', '0', '0' ])
#
model.save_training_set(click_log, "")
# with open("X.txt", "rb") as fp:
# X = pickle.load(fp)
#
# with open("Y.txt", "rb") as fp:
# Y = pickle.load(fp)
#
# train_path = "../datasets/ltrc_yahoo/set1.train.txt"
#
