def text_experiment(config):
np.random.seed(seed=27)
dataset = 'wikipedia'
data_path = '../data/' + dataset + '/'
trfile = data_path + '/splits/train%d.pkl' % config['fold']
trainset = pickle.load(open(trfile, 'rb'))
print('Overall %d reviews' % len(trainset))
# Step 2: Fit a emb model
print(config_to_name(config))
if config['cont_train']:
dummy_config = config.copy()
dummy_config['cont_train'] = False
dummy_config['use_sideinfo'] = False
dummy_config['exposure'] = False
dummy_config['max_iter'] = 300000
mfile = config_to_name(dummy_config) + '.pkl'
print(mfile)
train_noside = pickle.load(open(data_path + 'splits/' + mfile, "rb"))
init_model = train_noside['model']
else:
init_model = None
tag_dict = pickle.load(open(data_path + 'simple_tag_dict.pkl', 'rb'))
tags = [''] * 12
for tag in tag_dict.keys():
tags[tag_dict[tag]] = tag
voc_dict = pickle.load(open(data_path + 'voc_dict.pkl', 'rb'))
dictionary = voc_dict['dic']
reverse_dictionary = voc_dict['rev_dic']
sample_doc = trainset[16][0:20, :]
print('Sample data', sample_doc[:, 0],
[reverse_dictionary[i] for i in sample_doc[:, 0]])
print('Prior tags', sample_doc[:, 1], [tags[i] for i in sample_doc[:, 1]])
emb_model, logg = fit_emb(trainset, config, init_model)
print('Training done!')
tsfile = data_path + '/splits/test%d.pkl' % config['fold']
testset = pickle.load(open(tsfile, 'rb'))
print('Overall %d reviews' % len(testset))
loss_array, pos_loss_array = evaluate_emb(testset, emb_model, config)
# Step 4: Save result and Visualize the embeddings.
mfile = config_to_name(config) + '.pkl'
pickle.dump(
dict(model=emb_model,
logg=logg,
loss_array=loss_array,
pos_loss_array=pos_loss_array),
open(data_path + 'splits/' + mfile, "wb"))
def restaurant_experiment(config):
np.random.seed(seed=27)
dataset = 'restaurant'
data_path = '/rigel/dsi/users/ll3105/sa-data/' + dataset + '/'
trfile = data_path + '/splits/train0.pkl'
trainset = pickle.load(open(trfile, 'rb'))
print('Overall %d reviews' % len(trainset))
voc_dict = pickle.load(open(data_path + 'voc_dict.pkl', 'rb'))
dictionary = voc_dict['dic']
reverse_dictionary = voc_dict['rev_dic']
print('Sample data', trainset[0]['text'][:10],
[reverse_dictionary[i] for i in trainset[0]['text'][:10]])
# Step 2: Function to generate a training batch for the skip-gram model.
sidevec, batch, labels = generate_batch(trainset[0], config=config)
#print(trainset[0])
for i in range(8):
msg = ''
for j in xrange(2 * config['half_window']):
msg = msg + str(
batch[i, j]) + ',' + reverse_dictionary[batch[i, j]] + ';'
print(msg, ' *', sidevec[0:3], '->', labels[i],
reverse_dictionary[labels[i]])
# Step 3: Fit a emb model
print(config_to_name(config))
if config['cont_train']:
dummy_config = config.copy()
dummy_config['cont_train'] = False
dummy_config['use_sideinfo'] = False
dummy_config['exposure'] = False
dummy_config['max_iter'] = 300000
mfile = config_to_name(dummy_config) + '.pkl'
print(mfile)
train_noside = pickle.load(open(data_path + 'splits/' + mfile, "rb"))
init_model = train_noside['model']
else:
init_model = None
emb_model, logg = fit_emb(trainset, config, voc_dict, init_model)
# Step 4: Save result and Visualize the embeddings.
mfile = config_to_name(config) + '.pkl'
pickle.dump(dict(model=emb_model, logg=logg),
open(data_path + 'splits/' + mfile, "wb"))
print('Training done!')
def movie_experiment(config):
np.random.seed(seed=27)
dataset = 'movie'
data_path = '/rigel/dsi/users/ll3105/sa-data/' + dataset + '/'
trfile = data_path + '/splits/train%d.pkl' % config['fold']
trainset = pickle.load(open(trfile, 'rb'))
print('Overall %d reviews' % len(trainset))
# Step 2: Fit a emb model
print(config_to_name(config))
if config['cont_train']:
dummy_config = config.copy()
dummy_config['cont_train'] = False
dummy_config['use_sideinfo'] = False
dummy_config['exposure'] = False
dummy_config['max_iter'] = 300000
mfile = config_to_name(dummy_config) + '.pkl'
print(mfile)
train_noside = pickle.load(open(data_path + 'splits/' + mfile, "rb"))
init_model = train_noside['model']
else:
init_model = None
emb_model, logg = fit_emb(trainset, config, init_model)
tsfile = data_path + '/splits/test%d.pkl' % config['fold']
testset = pickle.load(open(tsfile, 'rb'))
print('Overall %d reviews' % len(testset))
loss_array, pos_loss_array = evaluate_emb(testset, emb_model, config)
# Step 4: Save result and Visualize the embeddings.
mfile = config_to_name(config) + '.pkl'
pickle.dump(dict(model=emb_model, logg=logg, loss_array=loss_array, pos_loss_array=pos_loss_array), open(data_path + 'splits/' + mfile, "wb"))
print('Training done!')
# load dictionary
voc_dict = pickle.load(open(data_path + 'voc_dict.pkl', 'rb'))
dictionary = voc_dict['dic']
reverse_dictionary = voc_dict['rev_dic']
vocabulary_size = len(dictionary)
print('Vocabulary size is %d' % vocabulary_size)
# load data
tsfile = data_path + '/splits/test0.pkl'
testset = pickle.load(open(tsfile, 'rb'))
print('Overall %d reviews' % len(testset))
# load model
config = dict(use_sideinfo=True, K=128, max_iter=200000, context_type='skip_gram', skip_window=1, num_skips=2, voc_size=len(dictionary), reg=1.0)
mfile = config_to_name(config) + '.pkl'
emb_model = pickle.load(open(data_path + 'splits/' + mfile, "rb"))
# batch_size becomes a variable
num_sampled = 64 # Number of negative examples to sample.
reviews = testset
num_review = len(reviews)
print('Sample data', reviews[0]['text'][:10], [reverse_dictionary[i] for i in reviews[0]['text'][:10]])
# Step 3: Function to generate a training batch for the skip-gram model.
graph = tf.Graph()
with graph.as_default():
# Input data.
def load_model(config, path):
mfile = config_to_name(config) + '.pkl'
loader = pickle.load(open(path + mfile, "rb"))
model = loader['model']
return model['alpha'], model['rho'], model['weight'], model['invmu']
reg_weight=10,
exposure=False,
cont_train=False,
sample_ratio=0.1)
config1 = config0.copy()
config1.update({'exposure': True})
config2 = config1.copy()
config2.update({'use_sideinfo': True})
data_path = '/rigel/dsi/users/ll3105/sa-data/' + dataset + '/'
reviews = pickle.load(open(data_path + 'splits/test%d.pkl' % fold, 'rb'))
losses = []
for conf in [config0, config1, config2]:
mfile = config_to_name(conf)
#loss_array = pickle.load(open(data_path + 'splits/test/loss_' + mfile + '.pkl', "rb"))
loader = pickle.load(open(data_path + 'splits/' + mfile + '.pkl',
"rb"))
loss_array = loader['loss_array']
#rlosses = np.array([np.mean(rloss) for rloss in loss_array])
for ir in xrange(len(loss_array)):
_, ind, rate = sparse.find(reviews['scores'][ir, :])
loss_array[ir] = loss_array[ir][ind]
rlosses = np.concatenate(loss_array)
losses.append(rlosses)
print('\\multirow{2}{*}{K = %d}' % K, end='')