def task_seg_ml_sklearn_crossval():
"""
Run ML Scikit learn methods for Sen et al. and Hagen et al. dataset task
segmentation
"""
classifiers = methods.multiplelearners()
print('\n\n--- Experiment with Sen et al. dataset ')
sen_aol = datasets.lucchese_aol(
representation=datasets.representation().glove)
sen_aol.load_sequential_queries()
classifiers.run(sen_aol)
dataset = datasets.lucchese_aol()
dataset.load_sequential_pair()
methods.task_rules(dataset=dataset).test()
print('\n\n--- Experiment with Hagen et al. dataset ')
hagen_aol = datasets.hagen_aol(
representation=datasets.representation().glove)
hagen_aol.load_sequential_queries()
classifiers.run(hagen_aol)
dataset = datasets.hagen_aol()
dataset.load_sequential_pair()
methods.task_rules(dataset=dataset).test()
def task_ext(representation=datasets.representation(width=8).glove,
save_dir='models/irdcs_model'):
lambda_loss = 0.1
cell = 'GRU'
ui, lr, batch, mi = 1, 1e-5, 128, 150
# Datasets
qdatasets = []
qdataset = datasets.sen_aol(representation=representation)
qdataset.load_augmented(textdata=False)
qdatasets.append(qdataset)
qdataset = datasets.volske_trek(representation=representation)
qdataset.load_augmented_filter_user(textdata=False)
qdatasets.append(qdataset)
qdatasets_names = ['Sen et al.', 'Volske et al. Trec ']
# Clustering
for i in range(len(qdatasets)):
sen_aol = qdatasets[i]
ds_name = qdatasets_names[i]
print('\n\n--- Experiment RDC ' + ds_name + ' dataset, sequence pretrain')
hagen_aol = datasets.hagen_aol(representation=representation)
hagen_aol.load_random_pair_dual()
pe = 10; plr = 1e-5
methods.dc_rnn(pretrain_dataset=hagen_aol, dataset=sen_aol,
save_dir=save_dir, rnn=deep_clustering.IRDCS,
pretrain_epochs=pe, batch_size=batch, maxiter=mi, update_interval=ui,
learning_rate=lr, pretrain=True, pretrain_lr=plr, cell=cell)
print('\n\n--- Experiment RDC ' + ds_name + ' dataset, segmentation pretrain')
hagen_aol = datasets.hagen_aol(representation=representation)
hagen_aol.load_sequential_pair_dual()
pe = 30; plr = 1e-4
methods.dc_rnn(pretrain_dataset=hagen_aol, dataset=sen_aol,
save_dir=save_dir, rnn=deep_clustering.IRDCS,
pretrain_epochs=pe, batch_size=batch, maxiter=mi, update_interval=ui,
learning_rate=lr, pretrain=True, pretrain_lr=plr, cell=cell)
print('\n\n--- Experiment RDC ' + ds_name + ' dataset, no pretrain')
methods.dc_rnn(pretrain_dataset=hagen_aol, dataset=sen_aol,
save_dir=save_dir, rnn=deep_clustering.IRDCS,
pretrain_epochs=pe, batch_size=batch, maxiter=mi, update_interval=ui,
learning_rate=lr, pretrain=False, pretrain_lr=plr, cell=cell, lambda_loss=lambda_loss)
def task_seg_context():
"""
Use adjacent queries to add context to the query pair. Use Hagen et al., 2013
dataset and the BRNN architecture with GRU cells
"""
for m in range(0, 6):
hagen_aol = datasets.hagen_aol(
representation=datasets.representation().glove)
hagen_aol.load_sequential_queries(m=m, n=m+1)
hagen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = 'models/rnn_model_context'
brnn.ITERATIONS = 60000
brnn.CELL = 'GRU'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Hagen et al. dataset. m,n= ' + str(m) +',' + str(m+1)
print(description, file=sys.stderr)
print(description)
brnn.crossval(hagen_aol)
del hagen_aol
if m == 0:
continue
hagen_aol = datasets.hagen_aol(
representation=datasets.representation().glove)
hagen_aol.load_sequential_queries(m=m, n=1)
hagen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = 'models/rnn_model_context'
brnn.ITERATIONS = 60000
brnn.CELL = 'GRU'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Hagen et al. dataset. m,n= ' + str(m) +',' + str(1)
print(description, file=sys.stderr)
print(description)
brnn.crossval(hagen_aol)
del hagen_aol
def task_seg_crossval_brnn():
"""
Run a experiment with the bidirectional RNN to detect session changes in
the AOL dataset from (Sen et al., 2018) and (Hagen et al., 2013)
"""
m = 0
hagen_aol=datasets.hagen_aol(representation=datasets.representation().glove)
hagen_aol.load_sequential_queries(m=m, n=m+1)
hagen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = 'models/rnn_model_transfer'
brnn.ITERATIONS = 60000
brnn.CELL = 'LSTM'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Hagen et al. dataset. m,n= ' + str(m) +',' + str(m+1)
print(description, file=sys.stderr)
print(description)
brnn.crossval(hagen_aol)
del hagen_aol
m = 0
sen_aol=datasets.lucchese_aol(representation=datasets.representation().glove)
sen_aol.load_sequential_queries(m=m, n=m+1)
sen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = 'models/rnn_model_transfer'
brnn.ITERATIONS = 20000
brnn.CELL = 'LSTM'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Sen et al. dataset. m,n= ' + str(m) +',' + str(m+1)
print(description, file=sys.stderr)
print(description)
brnn.crossval(sen_aol, train_eval_runs=1)
del sen_aol
m = 0
hagen_aol=datasets.hagen_aol(representation=datasets.representation().glove)
hagen_aol.load_sequential_queries(m=m, n=m+1)
hagen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = 'models/rnn_model_transfer'
brnn.ITERATIONS = 60000
brnn.CELL = 'GRU'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Hagen et al. dataset. m,n= ' + str(m) +',' + str(m+1)
print(description, file=sys.stderr)
print(description)
brnn.crossval(hagen_aol)
del hagen_aol
m = 0
sen_aol=datasets.lucchese_aol(representation=datasets.representation().glove)
sen_aol.load_sequential_queries(m=m, n=m+1)
sen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = 'models/rnn_model_transfer'
brnn.ITERATIONS = 20000
brnn.CELL = 'GRU'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Sen et al. dataset. m,n= ' + str(m) +',' + str(m+1)
print(description, file=sys.stderr)
print(description)
brnn.crossval(sen_aol, train_eval_runs=1)
del sen_aol
def task_seg_transfer_learning():
"""
Pretrain BRNN on Hagen et al., 2013 dataset. Then, fine tune for Sen et al., 2018 session segmentation
"""
model_dir = 'models/rnn_model'
pretrain_model_dir = 'models/rnn_model_transfer_pretrain'
hagen_aol = datasets.hagen_aol(
representation=datasets.representation().glove)
hagen_aol.load_sequential_queries()
brnn = methods.brnn()
brnn.MODEL_DIR = pretrain_model_dir
brnn.ITERATIONS = 40000
brnn.CELL = 'GRU'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + \
'. Sen et al. dataset. Pretraining with Hagen et al., test set 10%'
print(description, file=sys.stderr)
print(description)
brnn.train_test(hagen_aol, test_size=0.1)
del hagen_aol
sen_aol = datasets.lucchese_aol(
representation=datasets.representation().glove)
sen_aol.load_sequential_queries()
sen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = model_dir
brnn.ITERATIONS = 40000 + 20000
brnn.CELL = 'GRU'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Sen et al. dataset, crossval, fine tuning'
print(description, file=sys.stderr)
print(description)
brnn.crossval(sen_aol, transfer=True, pretrain_dir=pretrain_model_dir)
del sen_aol
hagen_aol = datasets.hagen_aol(
representation=datasets.representation().glove)
hagen_aol.load_sequential_queries()
brnn = methods.brnn()
brnn.MODEL_DIR = pretrain_model_dir
brnn.ITERATIONS = 40000
brnn.CELL = 'LSTM'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + \
'. Sen et al. dataset. Pretraining with Hagen et al., test set 10%'
print(description, file=sys.stderr)
print(description)
brnn.train_test(hagen_aol, test_size=0.1)
del hagen_aol
sen_aol = datasets.lucchese_aol(
representation=datasets.representation().glove)
sen_aol.load_sequential_queries()
sen_aol.kfold()
brnn = methods.brnn()
brnn.MODEL_DIR = model_dir
brnn.ITERATIONS = 40000 + 20000
brnn.CELL = 'LSTM'
brnn.BATCH_SIZE = 256
description = '\n\n--- Experiment with ' + str(brnn.HIDDEN_UNITS) + \
brnn.CELL + '. Sen et al. dataset, crossval, fine tuning'
print(description, file=sys.stderr)
print(description)
brnn.crossval(sen_aol, transfer=True, pretrain_dir=pretrain_model_dir)
del sen_aol