def write_negative_instances(num_ps_rows, num_qs_rows, prods_db_c,
num_random_qs_per_prod, quests_db_c, outc, outdb):
for rowid in (range(1, num_ps_rows + 1)):
p_id = MyUtils_dbs.search_in_alltables_db(
prods_db_c, "SELECT id FROM", "WHERE rowid = " + str(rowid))[0][0]
neg_qs_ids = []
neg_qs_indices = np.random.choice(a=range(1, num_qs_rows),
size=num_random_qs_per_prod,
replace=False)
for neg_qs_index in neg_qs_indices:
neg_qs_ids.append(
MyUtils_dbs.search_in_alltables_db(
quests_db_c, "SELECT id FROM",
"WHERE `index` = " + str(neg_qs_index))[0][0])
if product_has_allfeatures(prods_db_c,
p_id) and allquestions_have_allfeatures(
quests_db_c, str(neg_qs_ids)):
insertion_sequence = [(p_id, q_id, 0) for q_id in neg_qs_ids]
outc.executemany("INSERT INTO instances VALUES (?,?,?)",
insertion_sequence)
else:
logging.info(
"Product %s excluded from the instances due to not having all the features",
p_id)
if rowid % (num_ps_rows // 10) == 0:
logging.info("Working on category: +10%%...")
outdb.commit()
def sort_candidates(candidates_db_path, ranked_candidates_outdb_path, prod_reps_dbpath, quest_reps_dbpath):
MyUtils.init_logging("Rank_candidates_nn.log")
### Connecting to the databases: candidates, test products, test questions
candidates_nn_db = sqlite3.connect(candidates_db_path)
cands_db_c = candidates_nn_db.cursor()
testprods_rep_c = sqlite3.connect(prod_reps_dbpath).cursor()
testquests_rep_c = sqlite3.connect(quest_reps_dbpath).cursor()
f = open(ranked_candidates_outdb_path, "w"); f.close()
outdb = sqlite3.connect(ranked_candidates_outdb_path)
outdb_c = outdb.cursor()
outdb_c.execute('''CREATE TABLE candidates( p_id varchar(63),
q_id varchar(63),
distance int
)''')
###
test_products_ids = cands_db_c.execute("SELECT DISTINCT p_id FROM candidates").fetchall()
logging.info(test_products_ids[0])
#logging.debug(test_products_ids)
for tpl_pid in test_products_ids:
pid = tpl_pid[0]
product_representation = MyUtils_dbs.search_in_alltables_db(testprods_rep_c, "SELECT * FROM ",
"WHERE id = '" + str(pid) + "'")[0]
product_tuple = MyUtils.prodls_tonamedtuple(product_representation, offset=1)
quests_ids = list(map ( lambda results_tpl : results_tpl[0], cands_db_c.execute("SELECT q_id FROM candidates WHERE p_id = ?", tpl_pid).fetchall()))
logging.debug(quests_ids)
product_qs_sorted = sort_product_candidates(product_tuple, quests_ids, testprods_rep_c, testquests_rep_c)
outdb.executemany("INSERT INTO candidates VALUES (?,?,?)", product_qs_sorted)
outdb.commit()
def shuffle_db_table(db_path):
db_conn = sqlite3.connect(db_path)
c = db_conn.cursor()
temp_db = db_path + "_temp"
f = open(temp_db, "w")
f.close() #clean outdb
outdb_conn = sqlite3.connect(temp_db)
outc = outdb_conn.cursor()
outc.execute('''CREATE TABLE instances(p varchar(63),
q varchar(63),
y tinyint) ''')
outdb_conn.commit()
tot_num_of_rows = MyUtils_dbs.get_tot_num_rows_db(c)
rand_indices = np.random.choice(range(1, tot_num_of_rows + 1),
tot_num_of_rows,
replace=False)
for ind in rand_indices:
picked_row = c.execute("SELECT * FROM instances WHERE rowid = " +
str(ind)).fetchone()
p = picked_row[0]
q = picked_row[1]
y = picked_row[2]
outc.execute('''INSERT INTO instances VALUES (?,?,?);''',
(str(p), str(q), str(y)))
outdb_conn.commit()
logging.info("Instances have been shuffled.")
os.rename(src=temp_db, dst=db_path)
def obtain_category_instances(category_dirpath, categ_products_db,
categ_questions_db, max_neg_cardinality):
logging.info("Extracting instances for the category: %s",
os.path.basename(category_dirpath))
prods_db_c = categ_products_db.cursor()
quests_db_c = categ_questions_db.cursor()
outdbname = MyUtils_flags.FLAG_INSTANCEIDS + ".db"
f = open(os.path.join(category_dirpath, outdbname), "w")
f.close()
outdb = sqlite3.connect(os.path.join(category_dirpath, outdbname))
outc = outdb.cursor()
outc.execute('''CREATE TABLE instances(p varchar(63),
q varchar(63),
y tinyint) ''')
outdb.commit()
### Get the number of Ps and Qs. Generally, |Ps| << |Qs| (eg. 119,43608)
num_ps_rows = MyUtils_dbs.get_tot_num_rows_db(prods_db_c)
logging.info("Number of products in category: %s", num_ps_rows)
num_qs_rows = MyUtils_dbs.get_tot_num_rows_db(quests_db_c)
logging.info("Number of questions in category: %s", num_qs_rows)
num_possible_instances = num_ps_rows * num_qs_rows
logging.info("Potential total number of instances from the category: %s",
num_possible_instances)
cardinality = min(num_possible_instances, max_neg_cardinality)
logging.info(
"Considering the upper boundary, the number of negative instances to include in the category dataset is: %s ",
cardinality)
num_random_qs_per_prod = cardinality // num_ps_rows
logging.info("Number of random negative examples per product: %s",
num_random_qs_per_prod)
write_positive_instances(num_ps_rows, prods_db_c, quests_db_c, outc)
outdb.commit()
write_negative_instances(num_ps_rows, num_qs_rows, prods_db_c,
num_random_qs_per_prod, quests_db_c, outc, outdb)
outdb.commit()
shuffle_db_table(os.path.join(category_dirpath, outdbname))
categ_products_db.close()
categ_questions_db.close()
outdb.close()
def write_positive_instances(num_ps_rows, prods_db_c, quests_db_c, outc):
###Iterate over the products:
for rowid in (range(1, num_ps_rows + 1)):
p_id = MyUtils_dbs.search_in_alltables_db(
prods_db_c, "SELECT id FROM", "WHERE rowid = " + str(rowid))[0][0]
###Get all the Qs asked for the selected Ps; they will always be part of the dataset,since there are so few Ps
q_ids_results = MyUtils_dbs.search_in_alltables_db(
quests_db_c, "SELECT id FROM",
"WHERE id LIKE '" + str(p_id) + "%'")
q_ids_ls = [tpl[0] for tpl in q_ids_results]
#filter: p and qs must have all features
if product_has_allfeatures(prods_db_c,
p_id) and allquestions_have_allfeatures(
quests_db_c, str(q_ids_ls)):
insertion_sequence = [(p_id, q_id, 1) for q_id in q_ids_ls]
outc.executemany("INSERT INTO instances VALUES (?,?,?)",
insertion_sequence)
else:
logging.info(
"Product %s excluded from the instances due to not having all the features",
p_id)
def generator_of_batches(batch_size, dataset_type):
if dataset_type == MyUtils_flags.FLAG_TRAIN:
dataset_length = MyUtils_dbs.get_nn_dataset_length(
MyUtils_flags.FLAG_TRAIN)
db_conn = sqlite3.connect(F.NN_TRAIN_INSTANCES_DB)
if dataset_type == MyUtils_flags.FLAG_VALID:
dataset_length = MyUtils_dbs.get_nn_dataset_length(
MyUtils_flags.FLAG_VALID)
db_conn = sqlite3.connect(F.NN_VALID_INSTANCES_DB)
if dataset_type == MyUtils_flags.FLAG_TEST:
dataset_length = MyUtils_dbs.get_nn_dataset_length(
MyUtils_flags.FLAG_TEST)
db_conn = sqlite3.connect(F.NN_TEST_INSTANCES_DB)
c = db_conn.cursor()
num_of_batches = dataset_length // batch_size + 1
half_mark_offset = dataset_length // 2
for i in range(0, num_of_batches):
start_index_pos = i * (batch_size // 2)
end_index_pos = min((i + 1) * (batch_size // 2), half_mark_offset)
start_index_neg = half_mark_offset + i * (batch_size // 2)
end_index_neg = half_mark_offset + min(
(i + 1) * (batch_size // 2), dataset_length)
c.execute("SELECT p_id, q_id, x,y FROM instances WHERE rowid IN " +
str(tuple(range(start_index_pos, end_index_pos))))
rows = c.fetchall()
c.execute("SELECT p_id, q_id, x,y FROM instances WHERE rowid IN " +
str(tuple(range(start_index_neg, end_index_neg))))
rows_neg = c.fetchall()
rows.extend(rows_neg)
batch = list(
map(
lambda elem: (str(elem[0]), str(elem[1]), json.loads(elem[2]),
int(elem[3])), rows))
yield batch
def sort_product_candidates(product_tuple, quests_ids, testprods_rep_c, testquests_rep_c):
distance_list = [] #list of tuples, (p_id, q_id, distance), sorted on tpl[2]
for quest_id in quests_ids:
question_representation = MyUtils_dbs.search_in_alltables_db(testquests_rep_c, "SELECT * FROM ",
"WHERE id = '" + str(quest_id) + "'")[0]
logging.debug("Question representation: %s", question_representation)
question_tuple = MyUtils.quest_lstonamedtuple(question_representation, offset=1)
pq_dist = CD.compute_dist_pq(product_tuple, question_tuple)
distance_list.append((product_tuple.id, quest_id, pq_dist))
distance_list_sorted = sorted(distance_list, key=lambda tpl : tpl[2])
return distance_list_sorted
def get_instance_encoded_dictionary(prod_id, question_id, ps_db_c, qs_db_c,
d2v_model):
product_row = MyUtils_dbs.search_in_alltables_db(
ps_db_c, "SELECT * FROM ", "WHERE id = '" + prod_id + "'")
question_row = MyUtils_dbs.search_in_alltables_db(
qs_db_c, "SELECT * FROM ", "WHERE id = '" + str(question_id) + "'")
prod_tuple = MyUtils.prodls_tonamedtuple(product_row[0])
q_tuple = MyUtils.quest_lstonamedtuple(question_row[0])
instance_x = {}
instance_x["p_descvec"] = MyUtils_strings.fromstring_toarray(
prod_tuple.descvec)
instance_x["p_titlevec"] = MyUtils_strings.fromstring_toarray(
prod_tuple.titlevec)
instance_x["p_kwsVectors"] = MyUtils_strings.fromlls_toarrays(
prod_tuple.kwsVectors)
#logging.debug("instance_x['p_kwsVectors'].shape : %s", np.array(instance_x["p_kwsVectors"]).shape)
instance_x["p_mdcategories"] = MyUtils_strings.categories_to_vecs_lls(
MyUtils_strings.fromlls_toarrays(prod_tuple.mdcategories), d2v_model)
if len(np.array(instance_x["p_mdcategories"]).shape) >= 3:
logging.debug("instance_x['p_mdcategories'].shape : %s",
np.array(instance_x["p_mdcategories"]).shape)
instance_x["p_mdcategories"] = instance_x["p_mdcategories"][0]
instance_x["q_questionVec"] = MyUtils_strings.fromstring_toarray(
q_tuple.questionVec)
instance_x["q_questionType"] = q_tuple.questionType
instance_x["q_kwsVectors"] = MyUtils_strings.fromlls_toarrays(
q_tuple.kwsVectors)
instance_y = 1 if q_tuple.id[0:10] in prod_id else 0
instance = namedtuple('instance', 'x y')
inst = instance(x=instance_x, y=instance_y)
return inst
def train_NN(learning_rate = 0.01, max_epochs=1000, batch_size=32, dropout_rate=0, hiddenlayers_ls=None):
MyUtils.init_logging("train_NN.log")
hiddenlayers_ls_str = [str(num_elems) for num_elems in hiddenlayers_ls]
tensorboard_dir_path = os.path.join(F.TENSORBOARD_ANN_DIR, 'trainingset_' + str(
MyUtils_dbs.get_nn_dataset_length("train")),
"bs_" + str(batch_size),
"hls_" + "-".join(hiddenlayers_ls_str),
"lr_" + str(learning_rate),
"drop_" + str(dropout_rate) + "eps_" + str(max_epochs))
if not os.path.exists(tensorboard_dir_path):
os.makedirs(tensorboard_dir_path)
MyUtils_filesystem.clean_directory(tensorboard_dir_path)
tf.reset_default_graph()
session = tf.Session()
logging.info("Creating the placeholders for input and labels...")
(input_placeholder, labels_placeholder) = NN.get_model_placeholders(batch_size)
placeholders = (input_placeholder, labels_placeholder)
logging.info("Connecting the loss computation and forward structure...")
train_loss = NN.nn_loss_computation(logits=NN.nn_inference(input=input_placeholder, layers_hidden_units_ls=hiddenlayers_ls,
dropout_rate=dropout_rate),
labels=labels_placeholder)
lrate_tensor = tf.placeholder(shape=[], dtype=tf.float32, name="lrate_tensor")
####### Defining the optimizer
if str(learning_rate).lower() == MyUtils_flags.FLAG_ADAM:
starting_lrate = MyUtils_flags.FLAG_ADAM
optimizer = tf.train.AdamOptimizer()
else:
if str(learning_rate).lower() == MyUtils_flags.FLAG_RMSPROP:
starting_lrate = MyUtils_flags.FLAG_RMSPROP
optimizer = tf.train.RMSPropOptimizer(0.001)
else:
starting_lrate = learning_rate
optimizer = tf.train.GradientDescentOptimizer(lrate_tensor)
if str(learning_rate).lower() == MyUtils_flags.FLAG_CLR:
_best_lr, min_lr, max_lr = CLR.find_cyclical_lrate_loop(placeholders, batch_size, hiddenlayers_ls,
dropout_rate)
# Summaries, and gathering information:
train_loss_summary = tf.summary.scalar('Cross-entropy', train_loss)
predictions = tf.argmax(tf.nn.softmax(logits=NN.nn_inference(input_placeholder, hiddenlayers_ls, dropout_rate)),
axis=1, name="predictions")
tf_metric, tf_metric_update = tf.metrics.accuracy(labels=labels_placeholder, predictions=predictions,
name="accuracy")
accuracy_summary = tf.summary.scalar('Accuracy', tf_metric_update)
logging.info("Defining the optimizer's minimization task on the loss function...")
minimizer_task = optimizer.minimize(train_loss)
#Global variables are initialized after the graph structure
tf.global_variables_initializer().run(session=session)
#defining the tasks that will be run inside the training loop
training_tasks = [minimizer_task, train_loss, predictions, tf_metric_update]
validation_tasks = [tf_metric_update, predictions]
validation_writing_tasks = [accuracy_summary]
train_writing_tasks = [train_loss_summary, accuracy_summary]
tasks_dictionary = {MyUtils_flags.FLAG_TRAIN_TASKS: training_tasks,
MyUtils_flags.FLAG_WRITING_TRAIN_TASKS: train_writing_tasks,
MyUtils_flags.FLAG_VALIDATION_TASKS: validation_tasks,
MyUtils_flags.FLAG_WRITING_VALIDATION_TASKS: validation_writing_tasks}
#connection to the validation dataset
valid_db_conn = sqlite3.connect(F.NN_VALID_INSTANCES_DB)
valid_db_cursor = valid_db_conn.cursor()
if str(learning_rate).lower() == MyUtils_flags.FLAG_CLR:
CLR.training_loop_clr(tasks_dictionary, placeholders, batch_size,
max_epochs, min_lr, max_lr, valid_db_cursor, tensorboard_dir_path)
else:
training_loop(tasks_dictionary, placeholders, starting_lrate,
batch_size, max_epochs, valid_db_cursor, tensorboard_dir_path, session)
def define_negative_examples(doc2vec_model, dataset_typeflag):
MyUtils.init_logging("NN_Dataset_Instances-define_negative_examples.log",
logging.INFO)
f = open(F.PRODS_WITH_NOTASKEDQUESTS_IDS, "w")
f.close()
prodsnegativeqs_outfile = open(F.PRODS_WITH_NOTASKEDQUESTS_IDS, "a")
prodsnegativeqs_outfile.write("id_questionsNotAsked\n")
### Connect with the database to read from: candidate negative examples
db_conn = sqlite3.connect(F.CANDIDATE_NEGQS_DB)
c = db_conn.cursor()
### IF we are working to create the training dataset,
### then we before allowing a question Q asked for P2 to be a negative example for P1,
### we check the similarity between P1 and P2 (it must not be too high)
if dataset_typeflag == MyUtils_flags.FLAG_TRAIN:
### Determining the maximum allowed similarity between products. Creates the similarity db if it does not exist
if os.path.exists(F.SIMILARITY_PRODUCTS_DB) == True:
p_sim_breakpoint = ES.get_products_similarity_breakpoint(
fraction=0.97)
else:
p_sim_breakpoint = ES.explore_products_similarity(N=500,
fraction=0.97)
### Connect with the databases of product and questions representations, to be able to pick the products P1 and P2
product_reps_dbconn = sqlite3.connect(F.PRODUCTS_FINAL_TRAIN_DB)
product_reps_c = product_reps_dbconn.cursor()
segment_size = 10**4
for input_segment in pd.read_csv(F.PRODSWITHQUESTS_IDS,
sep="_",
chunksize=segment_size):
for id_askedqs_t in input_segment.itertuples():
prod_id = id_askedqs_t.id
#logging.debug("Reading from F.PRODSWITHQUESTS_IDS, the product.id is: %s", prod_id)
asked_qs = ast.literal_eval(id_askedqs_t.questionsAsked)
t = (prod_id, )
c.execute('SELECT * FROM prodnegatives WHERE prod_id=?', t)
row = c.fetchone()
if row is None: #i.e. if the product in the file PRODSWITHQUESTS_IDS was excluded from the previous random subsampling
continue
candidatenegativeqs_rawstring = row[1]
candidatenegativeqs_string = "[" + candidatenegativeqs_rawstring[:
-1] + "]"
candidatenegativeqs_ls = ast.literal_eval(
candidatenegativeqs_string)
candidatenegativeqs_ls1 = [
q_id for q_id in candidatenegativeqs_ls if q_id not in asked_qs
]
if dataset_typeflag == MyUtils_flags.FLAG_TRAIN:
p1_row = MyUtils_dbs.search_in_alltables_db(
dbcursor=product_reps_c,
query_pretext="SELECT * FROM",
query_aftertext=" WHERE id='" + str(prod_id) + "'")[0]
candidatenegativeqs_asins = list(
map(lambda q_id: q_id[0:10], candidatenegativeqs_ls1))
p2_rows = MyUtils_dbs.search_in_alltables_db(
dbcursor=product_reps_c,
query_pretext="SELECT * FROM",
query_aftertext="WHERE id IN " +
str(tuple(candidatenegativeqs_asins)))
qids_and_p2rows = list(zip(candidatenegativeqs_ls1, p2_rows))
for q_id, p2_row in qids_and_p2rows:
#logging.debug("p1_row : %s", p1_row)
if p2_row is not None and len(p2_row) > 0:
#there are questions without corresponding products, in which case no similarity check is to be done
p1_tuple = MyUtils.prodls_tonamedtuple(p1_row) #[1:]?
p2_tuple = MyUtils.prodls_tonamedtuple(p2_row)
p1_p2_sim, _simparts = PS.compute_2products_similarity_singleprocess(
prod1_tuple=p1_tuple,
prod2_tuple=p2_tuple,
d2v_model=doc2vec_model)
if p1_p2_sim > p_sim_breakpoint:
candidatenegativeqs_ls1.remove(q_id)
logging.info(
"Removing question from the candidate negative examples, "
+
"because the similarity between %s and %s is > %s",
prod_id, p2_tuple.id, p_sim_breakpoint)
logging.info(
"Choosing negative examples: P-to-p similarity checks done for product: %s",
prod_id)
random_indices = sorted(
np.random.choice(a=range(len(candidatenegativeqs_ls1)),
size=min(len(candidatenegativeqs_ls1),
len(asked_qs)),
replace=False,
p=None))
#logging.info(candidatenegativeqs_ls1)
negativeqs_ls = [
candidatenegativeqs_ls1[i] for i in random_indices
]
#logging.info(negativeqs_ls)
prodsnegativeqs_outfile.write(prod_id + "_" + str(negativeqs_ls) +
"\n")
prodsnegativeqs_outfile.close()
def get_num_training_iterations(batch_size):
trainset_length = MyUtils_dbs.get_nn_dataset_length(
MyUtils_flags.FLAG_TRAIN)
max_iter = trainset_length // batch_size # in 1 epoch, you can not have more iterations than batches
logging.info("Number of iterations per epoch: %s", max_iter)
return max_iter
def run_hedge(actions_ls=None,
eta=None,
max_T=None,
balanced_instances=True,
restart_candidates=True):
MyUtils.init_logging("Hedge-run_hedge.log")
### initialization: either we use the single global balanced dataset, or the imbalanced category datasets
if balanced_instances:
dbs_paths = [(F.ONLINE_INSTANCEIDS_GLOBAL_DB,
F.PRODUCTS_FINAL_TRAIN_DB, F.QUESTIONS_FINAL_TRAIN_DB)]
else:
category_dirpaths = MyUtils_filesystem.get_category_dirpaths()
dbs_paths = [
] #list of tuples, with 3 elements: instancedb, products_db, qs_db
for c_dir_path in category_dirpaths:
for fname in os.listdir(c_dir_path):
if "db" in fname:
if MyUtils_flags.FLAG_INSTANCEIDS in fname:
categ_instances_dbpath = os.path.join(
c_dir_path, fname)
elif MyUtils_flags.FLAG_PRODUCTS in fname:
categ_prods_dbpath = os.path.join(c_dir_path, fname)
else:
categ_qs_dbpath = os.path.join(c_dir_path, fname)
dbs_paths.append(
(categ_instances_dbpath, categ_prods_dbpath, categ_qs_dbpath))
### connecting with the database containing the candidates
if balanced_instances:
output_candidates_dbpath = F.CANDIDATES_ONLINE_BALANCED_DB
else:
output_candidates_dbpath = F.CANDIDATES_ONLINE_UNBALANCED_DB
if restart_candidates:
f = open(output_candidates_dbpath, "w")
f.close()
output_candidates_db = sqlite3.connect(output_candidates_dbpath)
output_candidates_c = output_candidates_db.cursor()
if restart_candidates:
output_candidates_c.execute("""CREATE TABLE candidates (
p_id varchar(63),
q_id varchar(63) )""")
#For each dataset: connect to databases of instances, Ps, and Qs
for (instances_dbpath, prods_dbpath, quests_dbpath) in dbs_paths:
instances_db = sqlite3.connect(instances_dbpath)
instances_ids_c = instances_db.cursor()
prods_db = sqlite3.connect(prods_dbpath)
ps_c = prods_db.cursor()
quests_db = sqlite3.connect(quests_dbpath)
qs_c = quests_db.cursor()
chosen_dataset_name = os.path.basename(
os.path.dirname(instances_dbpath))
logging.info("Online Learning: operating on dataset: %s",
chosen_dataset_name)
#### define the number of rounds
if max_T is None:
max_T = MyUtils_dbs.get_tot_num_rows_db(instances_ids_c)
logging.info(
"Total number of rounds (i.e. instances in the training set): %s",
max_T)
#### define the actions
if actions_ls is None:
if balanced_instances == False:
actions_ls = ACD.get_actionsforcategories()
else:
actions_ls = AGD.get_actionsforbalanced()
#### define the "learning rate"
if eta is None:
eta = np.sqrt((2 * np.log(len(actions_ls))) / max_T)
#### output directory for Tensorboard logging
results_dirpath = os.path.join(
'OnlineLearning', 'Experiments_results', str(chosen_dataset_name),
'numactions_' + str(len(actions_ls)), 'instances_' +
str(max_T)) #datetime.datetime.today().strftime('%Y-%m-%d')
if not os.path.exists(results_dirpath):
os.makedirs(results_dirpath)
MyUtils_filesystem.clean_directory(results_dirpath)
#### the actual core of the algorithm
hedge_loop(eta, max_T, actions_ls, instances_ids_c, ps_c, qs_c,
output_candidates_db, balanced_instances,
results_dirpath)
else:
logging.info("Online Learning results already computed for : %s",
results_dirpath)
def attach_text_to_candidates(ranked_candidates_dbpath, prods_initial_dbpath, quests_initial_dbpath, prod_reps_dbpath, quest_reps_dbpath, final_outdb_path):
MyUtils.init_logging("Attach_text_to_candidates.log")
candidates_nn_db = sqlite3.connect(ranked_candidates_dbpath)
cands_db_c = candidates_nn_db.cursor()
f = open(F.RANKING_TEMP_DB, 'w'); f.close()
temp_db = sqlite3.connect(F.RANKING_TEMP_DB)
temp_db_c = temp_db.cursor()
testprods_initial_c = sqlite3.connect(prods_initial_dbpath).cursor()
testquests_initial_c = sqlite3.connect(quests_initial_dbpath).cursor()
testprods_rep_c = sqlite3.connect(prod_reps_dbpath).cursor()
testquests_rep_c = sqlite3.connect(quest_reps_dbpath).cursor()
temp_db_c.execute('''CREATE TABLE candidates( p_id varchar(63),
q_id varchar(63),
distance int,
p_titletext varchar(1023),
p_descriptiontext varchar(8191),
p_categorytext varchar (4095),
q_text varchar (8191)
)''')
num_of_candidates = MyUtils_dbs.get_tot_num_rows_db(cands_db_c)
logging.info(num_of_candidates)
counter_questionsameid = 0
last_prod_id = 'x'
for rowindex in range(1, num_of_candidates + 1):
row = cands_db_c.execute("SELECT * FROM candidates WHERE rowid = ?", (rowindex,)).fetchone()
#logging.info("info: %s", row)
prod_id = row[0]
quest_id = row[1]
distance = row[2]
if last_prod_id != prod_id:
product_titleinfo,product_descinfo, product_categinfo = \
MyUtils_dbs.search_in_alltables_db(testprods_initial_c, "SELECT title, description, categories FROM",
"WHERE asin = '" + str(prod_id) + "'")[0]
product_representation = MyUtils_dbs.search_in_alltables_db(testprods_rep_c, "SELECT * FROM ",
"WHERE id = '" + str(prod_id) + "'")[0]
prod_tpl = MyUtils.prodls_tonamedtuple(product_representation, offset=1)
counter_questionsameid = 0
###get question's unixTime
if len(quest_id)< 21: #format : @nan0
base_endpoint = 14
question_unixTime = str(quest_id[11:base_endpoint])
else:
base_endpoint = 23
question_unixTime = str(quest_id[11:base_endpoint])
logging.debug("Question unixTime: %s", question_unixTime)
if base_endpoint == 23: #if we have a valid unixTime specification
possible_questions_text = MyUtils_dbs.search_in_alltables_db(testquests_initial_c, "SELECT question FROM",
"WHERE asin = '" + str(quest_id[0:10]) + "'"
+ " AND unixTime LIKE '" + question_unixTime + "%'")
else: #if we have NULL in the unixTime field
possible_questions_text = MyUtils_dbs.search_in_alltables_db(testquests_initial_c, "SELECT question FROM",
"WHERE asin = '" + str(quest_id[0:10]) + "'"
+ " AND unixTime IS NULL")
base_q_id = str(quest_id[0:23])
possible_questions_reps = MyUtils_dbs.search_in_alltables_db(testquests_rep_c, "SELECT * FROM ",
"WHERE id LIKE '" + str(base_q_id) + "%'")
logging.debug("possible_questions_reps: %s", possible_questions_reps)
logging.debug("possible_questions_text:%s", possible_questions_text)
if len(possible_questions_text) > 1:
possible_questions_tuples = list(map ( lambda q_ls : MyUtils.quest_lstonamedtuple(q_ls, offset=1), possible_questions_reps))
possible_questions_distances = list(map (lambda q_tpl : CD.compute_dist_pq(prod_tpl, q_tpl) , possible_questions_tuples))
qs_dist_lts = list(zip(possible_questions_tuples, possible_questions_distances))
qs_dist_lts_sorted = sorted( qs_dist_lts, key=lambda tpl : tpl[1])
#logging.info("sorted question tuples: %s", qs_dist_lts_sorted)
question_textinfo = possible_questions_text[counter_questionsameid][0]
counter_questionsameid= counter_questionsameid+1
else:
question_textinfo = possible_questions_text[0][0]
logging.debug("question_textinfo: %s", question_textinfo)
temp_db_c.execute("INSERT INTO candidates VALUES (?,?,?,?,?,?,?)", (prod_id, quest_id, distance,
product_titleinfo, product_descinfo, product_categinfo, question_textinfo))
logging.debug("***")
temp_db.commit()
os.rename(F.RANKING_TEMP_DB , final_outdb_path)
def find_cyclical_lrate_loop(placeholders,
batch_size,
hiddenlayers_ls,
drop_rate,
lrate_start=10**(-7),
lrate_end=0.2):
trainset_length = MyUtils_dbs.get_nn_dataset_length(
MyUtils_flags.FLAG_TRAIN)
max_iter = trainset_length // batch_size # in 1 epoch, you can not have more iterations than batches
hiddenlayers_ls_str = [str(num_elems) for num_elems in hiddenlayers_ls]
tensorboard_dir_path = os.path.join(
F.TENSORBOARD_ANN_DIR,
'trainingset_' + str(MyUtils_dbs.get_nn_dataset_length("train")),
"bs_" + str(batch_size), "hls_" + "-".join(hiddenlayers_ls_str),
"lr_clr", "drop_" + str(drop_rate) + "_explore")
if not os.path.exists(tensorboard_dir_path):
os.makedirs(tensorboard_dir_path)
session = tf.Session(
) # separate session for trying to find the optimal l.r. for the Cyclical Learning Rate
logging.info("*** Session: Cyclical Learning Rate")
(input_placeholder, labels_placeholder) = placeholders
train_loss = NN.nn_loss_computation(logits=NN.nn_inference(
input=input_placeholder,
layers_hidden_units_ls=hiddenlayers_ls,
dropout_rate=drop_rate),
labels=labels_placeholder)
#train_loss_summary = tf.summary.scalar('Cross-entropy', train_loss)
predictions = tf.argmax(tf.nn.softmax(
logits=NN.nn_inference(input_placeholder, hiddenlayers_ls, drop_rate)),
axis=1)
#tf_metric, tf_metric_update = tf.metrics.accuracy(labels=labels_placeholder, predictions=predictions,
# name="CLR_train_accuracy")
running_vars = tf.get_collection(tf.GraphKeys.LOCAL_VARIABLES,
scope="CLR_train_accuracy")
#train_accuracy_summary = tf.summary.scalar('Accuracy', tf_metric_update)
lrate_tensor = tf.placeholder(shape=[],
dtype=tf.float32,
name="lrate_tensor")
optimizer = tf.train.GradientDescentOptimizer(lrate_tensor)
minimizer_task = optimizer.minimize(train_loss)
# Global variables are initialized after the graph structure
tf.global_variables_initializer().run(session=session)
running_vars_initializer = tf.variables_initializer(var_list=running_vars)
session.run(running_vars_initializer)
logging.info(
"Number of iterations per epoch (and linear steps in the search for the learning rate): %s",
max_iter)
lrate_increase = (lrate_end - lrate_start) / max_iter
logging.info(
"Step increase of the learning rate in the exploration epochs: %s",
round(lrate_increase, 7))
trial_epochs = 5
loss_matrix = np.zeros((trial_epochs, max_iter))
accuracy_matrix = np.zeros((trial_epochs, max_iter))
for i in range(1, trial_epochs + 1):
start_epoch_time = time()
logging.info(
"Search for the base learning rate; Starting training epoch n. %s",
i)
batch_generator = NN.generator_of_batches(batch_size,
MyUtils_flags.FLAG_TRAIN)
# Train, in the current epoch
for j in range(0, max_iter):
session.run(running_vars_initializer
) #new batch: re-initializing the accuracy computation
batch = \
batch_generator.__next__()
current_iteration_feed_dict = NN.fill_feed_dict(
batch, input_placeholder, labels_placeholder)
learning_rate = lrate_start + lrate_increase * j
current_iteration_feed_dict.update({lrate_tensor: learning_rate})
if j % (max_iter // 20) == 0:
logging.info("Iteration: %s on %s .", j, max_iter)
_, current_loss, b_predictions, b_labels = session.run(
[minimizer_task, train_loss, predictions, labels_placeholder],
feed_dict=current_iteration_feed_dict)
loss_matrix[i - 1][j] = current_loss
accuracy_matrix[i - 1][j] = get_batch_accuracy(
b_predictions, b_labels)
end_epoch_time = time()
logging.info(
"Searching for the base values for the cyclical learning rate. " +
"Training on epoch %s executed. Time elapsed: %s", i,
round(end_epoch_time - start_epoch_time, 3))
best_lr, min_lr, max_lr = pick_lr_boundaries(loss_matrix, lrate_start,
lrate_increase)
session.close()
#write the lr to a logfile
lrfile = open(os.path.join(tensorboard_dir_path, "found_lr.log"), "w")
lrfile.write(
"Cyclical Learning rate: applying the LR test on " +
str(trial_epochs) + "epochs ;\n " +
"the average learning rate granting the steepest descent of the loss function is: "
+ str(best_lr))
lrfile.close()
return best_lr, min_lr, max_lr