def evaluate_emb(reviews, model, config):
graph = tf.Graph()
with graph.as_default():
tf.set_random_seed(27)
# construct model graph
print('Building graph...')
builder = GraphBuilder()
inputs, outputs, model_param = builder.construct_model_graph(
reviews, config, model, training=False)
init = tf.global_variables_initializer()
with tf.Session(graph=graph) as session:
# We must initialize all variables before we use them.
print('Initializing...')
init.run()
llh_array = []
pos_llh_array = []
mul_llh_array = []
review_size = reviews['scores'].shape[0]
print('Calculating llh of instances...')
for step in xrange(review_size):
att, index, label = generate_batch(reviews, step)
if index.size <= 1: # neglect views with only one entry
continue
feed_dict = {
inputs['input_att']: att,
inputs['input_ind']: index,
inputs['input_label']: label
}
ins_llh_val, pos_llh_val, mul_llh_val = session.run(
(outputs['ins_llh'], outputs['pos_llh'], outputs['mul_llh']),
feed_dict=feed_dict)
#if step == 0:
# predicts = session.run(outputs['debugv'], feed_dict=feed_dict)
# print('%d movies ' % predicts.shape[0])
# print(predicts)
llh_array.append(ins_llh_val)
pos_llh_array.append(pos_llh_val)
mul_llh_array.append(mul_llh_val)
llh_array = np.concatenate(llh_array, axis=0)
pos_llh_array = np.concatenate(pos_llh_array, axis=0)
mul_llh_array = np.array(mul_llh_array)
print("Loss and pos_loss mean: ", np.mean(llh_array),
np.mean(pos_llh_array))
return llh_array, pos_llh_array, mul_llh_array
def evaluate_emb(reviews, batch_feeder, model, config):
graph = tf.Graph()
with graph.as_default():
tf.set_random_seed(27)
# construct model graph
print('Building evaluation graph...')
builder = GraphBuilder()
inputs, outputs, model_param = builder.construct_model_graph(
None, config, model,
training=False) # necessary sizes are in `model'
init = tf.global_variables_initializer()
with tf.Session(graph=graph) as session:
# We must initialize all variables before we use them.
print('Initializing evaluation graph...')
init.run()
pos_llh_list = []
neg_llh_list = []
review_size = reviews.shape[0]
print('Calculating llh of instances...')
for step in xrange(review_size):
index, label = batch_feeder(reviews[step])
if index.size <= 1: # neglect views with only one entry
continue
feed_dict = {
inputs['input_ind']: index,
inputs['input_label']: label
}
llh_item, nums = session.run(
(outputs['llh_item'], outputs['num_items']),
feed_dict=feed_dict)
num_pos = nums[0]
num_neg = nums[1]
pos_llh_list.append(llh_item[0:num_pos])
neg_llh_list.append(llh_item[num_pos:])
pos_llh = np.concatenate(pos_llh_list)
neg_llh = np.concatenate(neg_llh_list)
mean_llh = (np.sum(pos_llh) + np.sum(neg_llh)) / (pos_llh.shape[0] +
neg_llh.shape[0])
print("ELBO mean on the test set: ", mean_llh)
return dict(pos_llh=pos_llh, neg_llh=neg_llh)
def fit_emb(reviews, config, init_model):
np.random.seed(27)
use_valid_set = True
if use_valid_set:
reviews, valid_reviews = separate_valid(reviews, 0.1)
graph = tf.Graph()
with graph.as_default():
tf.set_random_seed(27)
builder = GraphBuilder()
inputs, outputs, model_param = builder.construct_model_graph(
reviews, config, init_model, training=True)
optimizer = tf.train.AdagradOptimizer(0.05).minimize(
outputs['objective'])
init = tf.global_variables_initializer()
with tf.Session(graph=graph) as session:
# We must initialize all variables before we use them.
init.run()
nprint = 5000
val_accum = np.array([0.0, 0.0])
train_logg = np.zeros([int(config['max_iter'] / nprint) + 1, 3])
review_size = reviews['scores'].shape[0]
for step in xrange(1, config['max_iter'] + 1):
rind = np.random.choice(review_size)
atts, indices, labels = generate_batch(reviews, rind)
if indices.size <= 1: # neglect views with only one entry
raise Exception(
'Row %d of the data has only one non-zero entry.' % rind)
feed_dict = {
inputs['input_att']: atts,
inputs['input_ind']: indices,
inputs['input_label']: labels
}
_, llh_val, obj_val, debug_val = session.run(
(optimizer, outputs['llh'], outputs['objective'],
outputs['debugv']),
feed_dict=feed_dict)
val_accum = val_accum + np.array([llh_val, obj_val])
# print loss every nprint iterations
if step % nprint == 0 or np.isnan(llh_val) or np.isinf(llh_val):
valid_llh = 0.0
break_flag = False
if use_valid_set:
valid_llh = validate(valid_reviews, session, inputs,
outputs)
#if ivalid > 0 and valid_llh[ivalid] < valid_llh[ivalid - 1]: # performance becomes worse
# print('validation llh: ', valid_llh[ivalid - 1], ' vs ', valid_llh[ivalid])
# break_flag = True
# record the three values
ibatch = int(step / nprint)
train_logg[ibatch, :] = np.append(val_accum / nprint,
valid_llh)
val_accum[:] = 0.0 # reset the accumulater
print("iteration[", step,
"]: average llh, obj, and valid_llh are ",
train_logg[ibatch, :])
if np.isnan(llh_val) or np.isinf(llh_val):
print('Loss value is ', llh_val,
', and the debug value is ', debug_val)
raise Exception('Bad values')
if break_flag:
break
# save model parameters to dict
model = dict(alpha=model_param['alpha'].eval(),
rho=model_param['rho'].eval(),
invmu=model_param['invmu'].eval(),
weight=model_param['weight'].eval(),
nbr=model_param['nbr'].eval())
return model, train_logg
def fit_emb(reviews, batch_feeder, config):
do_log_save = True
do_profiling = True
log_save_path = 'log'
# separate a validation set
use_valid_set = True
if use_valid_set:
reviews, valid_reviews = separate_valid(reviews, 0.1)
# build model graph
with tf.device('/gpu:0'):
graph = tf.Graph()
with graph.as_default():
tf.set_random_seed(27)
builder = GraphBuilder()
problem_size = {
'num_reviews': reviews.shape[0],
'num_items': reviews.shape[1]
}
inputs, outputs, model_param = builder.construct_model_graph(
problem_size, config, init_model=None, training=True)
model_vars = [
model_param['alpha'], model_param['rho'],
model_param['intercept'], model_param['prior_logit']
]
optimizer = tf.train.AdagradOptimizer(0.1).minimize(
outputs['objective'], var_list=model_vars)
if config['model'] in ['context_select']:
net_vars = builder.infer_net.param_list()
net_optimizer = tf.train.AdagradOptimizer(0.1).minimize(
outputs['objective'], var_list=net_vars)
init = tf.global_variables_initializer()
# for visualization
vis_conf = projector.ProjectorConfig()
embedding = vis_conf.embeddings.add()
embedding.tensor_name = model_param['alpha'].name
# optimize the model
with tf.Session(graph=graph) as session:
# initialize all variables
init.run()
# Merge all the summaries and write them out to /tmp/mnist_logs (by defaul)
if do_log_save:
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(log_save_path, session.graph)
projector.visualize_embeddings(train_writer, vis_conf)
else:
merged = []
if do_profiling:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
nprint = config['nprint']
val_accum = np.array([0.0, 0.0])
count_accum = 0
train_logg = np.zeros([int(config['max_iter'] / nprint) + 1, 3])
review_size = reviews.shape[0]
for step in xrange(1, config['max_iter'] + 1):
rind = np.random.choice(review_size)
indices, labels = batch_feeder(reviews[rind])
if indices.shape[0] <= 1: # neglect views with only one entry
raise Exception(
'Row %d of the data has only one non-zero entry.' % rind)
feed_dict = {
inputs['input_ind']: indices,
inputs['input_label']: labels
}
if config['model'] in ['context_select']:
_, net_debugv, summary = session.run(
(net_optimizer, outputs['debugv'], merged),
feed_dict=feed_dict)
else:
net_debugv = ''
_, llh_val, nums, obj_val, debug_val, summary = session.run((optimizer, outputs['llh_sum'], outputs['num_items'], \
outputs['objective'], outputs['debugv'], merged), feed_dict=feed_dict)
if do_log_save:
train_writer.add_summary(summary, step)
# record llh, and objective
val_accum = val_accum + np.array([llh_val, obj_val])
count_accum = count_accum + (nums[0] + nums[1])
# print loss every nprint iterations
if step % nprint == 0 or np.isnan(llh_val) or np.isinf(llh_val):
# do validation
valid_llh = 0.0
if use_valid_set:
valid_llh = validate(valid_reviews, batch_feeder, session,
inputs, outputs)
# record the three values
ibatch = int(step / nprint)
train_logg[ibatch, :] = np.array([
val_accum[0] / count_accum, val_accum[1] / nprint,
valid_llh
])
val_accum[:] = 0.0 # reset the accumulater
count_accum = 0
print("iteration[", step,
"]: average llh, obj, valid_llh, and debug_val are ",
train_logg[ibatch, :], debug_val, net_debugv)
#check nan value
if np.isnan(llh_val) or np.isinf(llh_val):
print('Loss value is ', llh_val,
', and the debug value is ', debug_val)
raise Exception('Bad values')
model = get_model(model_param, session, config)
if do_log_save:
tf.train.Saver().save(session, log_save_path, step)
# Create the Timeline object, and write it to a json
if do_profiling:
tl = timeline.Timeline(run_metadata.step_stats)
ctf = tl.generate_chrome_trace_format()
with open(
log_save_path + '/timeline_step%d.json' %
(step / nprint), 'w') as f:
f.write(ctf)
pickle.dump(model['alpha'], open('model_%d.json' & step, 'wb'),
-1)
model = get_model(model_param, session, config)
return model, train_logg