def report(y_true, y_pred, class_names, threshold=0.5):
tp = np.logical_and(y_true, y_pred)
fp = np.logical_and(np.logical_not(y_true), y_pred)
fn = np.logical_and(y_true, np.logical_not(y_pred))
pre = np.sum(tp.astype(float), axis=1)/np.sum(np.logical_or(tp, fp).astype(float), axis=1)
rec = np.sum(tp.astype(float), axis=1)/np.sum(np.logical_or(tp, fn).astype(float), axis=1)
f1 = (2.0 * pre * rec)/(pre + rec)
count = np.sum(y_true, axis=0)
auc_array = []
ret = '%45s\t P\t R\t F1\t AUC\t C\n' % ' '
ret += '\t' + '-' * 80 + '\n'
for i, (c, p, r, f, s) in enumerate(zip(class_names, pre[5], rec[5], f1[5], count)):
auc = util.calc_auc_pr(pre[:, i], rec[:, i])
auc_array.append(auc)
ret += '%45s\t%.4f\t%.4f\t%.4f\t%.4f\t%4d\n' % (c, p, r, f, auc, s)
ret += '\t' + '-' * 80 + '\n'
auc = np.array(auc_array)
try:
idx = list(np.linspace(0, 1, 11)).index(threshold)
except ValueError:
idx = 5
p = pre[idx, np.isfinite(pre[idx])]*count[np.isfinite(pre[idx])]
r = rec[idx, np.isfinite(rec[idx])]*count[np.isfinite(rec[idx])]
f = f1[idx, np.isfinite(f1[idx])]*count[np.isfinite(f1[idx])]
a = auc[np.isfinite(auc)]*count[np.isfinite(auc)]
ret += '%45s\t%.4f\t%.4f\t%.4f\t%.4f\t%4d\n' % ('avg.',
np.sum(p)/count.sum(dtype=float),
np.sum(r)/count.sum(dtype=float),
np.sum(f)/count.sum(dtype=float),
np.sum(a)/count.sum(dtype=float),
count.sum())
return ret
def dev_step(mtest, sess):
dev_loss = []
dev_auc = []
dev_f1_score = []
# create batch
test_batches = util.batch_iter(test_data,
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
for batch in test_batches:
left_batch, right_batch, y_batch, n_batch = zip(*batch)
feed = {
mtest.left: np.array(left_batch),
mtest.right: np.array(right_batch),
mtest.labels: np.array(y_batch)
}
if FLAGS.negative:
feed[mtest.negative] = np.array(n_batch)
loss_value, eval_value = sess.run(
[mtest.total_loss, mtest.eval_op], feed_dict=feed)
dev_loss.append(loss_value)
pre, rec = zip(*eval_value)
dev_auc.append(util.calc_auc_pr(pre, rec))
dev_f1_score.append((2.0 * pre[5] * rec[5]) /
(pre[5] + rec[5])) # threshold = 0.5
return (np.mean(dev_loss), np.mean(dev_auc),
np.mean(dev_f1_score))
def dev_step(mtest, sess):
dev_loss = []
dev_auc = []
dev_f1_score = []
# create batch
test_batches = util.batch_iter(test_data,
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
for batch in test_batches:
x_batch, y_batch, _ = zip(*batch)
loss_value, eval_value = sess.run(
[mtest.total_loss, mtest.eval_op],
feed_dict={
mtest.inputs: np.array(x_batch),
mtest.labels: np.array(y_batch)
})
dev_loss.append(loss_value)
pre, rec = zip(*eval_value)
# look at the 5th index, which corresponds to a threshold = 0.5
threshold = 5
dev_auc.append(util.calc_auc_pr(pre, rec, threshold))
dev_f1_score.append(
(2.0 * pre[threshold] * rec[threshold]) /
(pre[threshold] + rec[threshold]))
return np.mean(dev_loss), np.mean(dev_auc), np.mean(
dev_f1_score)
def dev_step(mtest, sess):
dev_loss = []
dev_auc = []
dev_f1_score = []
# create batch
test_batches = util.batch_iter(test_data,
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
for batch in test_batches:
x_batch, y_batch, _ = zip(*batch)
#a_batch = np.ones((len(batch), 1), dtype=np.float32) / len(batch) # average
loss_value, eval_value = sess.run(
[mtest.total_loss, mtest.eval_op],
feed_dict={
mtest.inputs: np.array(x_batch),
mtest.labels: np.array(y_batch)
})
dev_loss.append(loss_value)
pre, rec = zip(*eval_value)
dev_auc.append(util.calc_auc_pr(pre, rec))
dev_f1_score.append((2.0 * pre[5] * rec[5]) /
(pre[5] + rec[5])) # threshold = 0.5
return (np.mean(dev_loss), np.mean(dev_auc),
np.mean(dev_f1_score))
def evaluate(eval_data, config):
""" Build evaluation graph and run. """
with tf.Graph().as_default():
with tf.variable_scope('cnn'):
if config.has_key('contextwise') and config['contextwise']:
import cnn_context
m = cnn_context.Model(config, is_train=False)
else:
import cnn
m = cnn.Model(config, is_train=False)
saver = tf.train.Saver(tf.all_variables())
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(config['train_dir'])
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise IOError("Loading checkpoint file failed!")
#embeddings = sess.run(tf.all_variables())[0]
print "\nStart evaluation\n"
#losses = []
#precision = []
#recall = []
#batches = util.batch_iter(eval_data, batch_size=config['batch_size'], num_epochs=1, shuffle=False)
#for batch in batches:
if config.has_key('contextwise') and config['contextwise']:
left_batch, middle_batch, right_batch, y_batch, _ = zip(
*eval_data)
feed = {
m.left: np.array(left_batch),
m.middle: np.array(middle_batch),
m.right: np.array(right_batch),
m.labels: np.array(y_batch)
}
else:
x_batch, y_batch, _ = zip(*eval_data)
feed = {
m.inputs: np.array(x_batch),
m.labels: np.array(y_batch)
}
loss, eval = sess.run([m.total_loss, m.eval_op], feed_dict=feed)
#losses.append(loss)
pre, rec = zip(*eval)
#precision.append(pre)
#recall.append(rec)
avg_precision = np.mean(np.array(pre))
avg_recall = np.mean(np.array(rec))
auc = util.calc_auc_pr(pre, rec)
f1 = (2.0 * pre[5] * rec[5]) / (pre[5] + rec[5])
print '%s: loss = %.6f, f1 = %.4f, auc = %.4f' % (datetime.now(),
loss, f1, auc)
return pre, rec
def evaluate(eval_data, config):
""" Build evaluation graph and run. """
with tf.Graph().as_default():
with tf.variable_scope('cnn'):
if config.has_key('contextwise') and config['contextwise']:
import cnn_context
m = cnn_context.Model(config, is_train=False)
else:
import cnn
m = cnn.Model(config, is_train=False)
saver = tf.train.Saver(tf.global_variables())
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(config['train_dir'])
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise IOError("Loading checkpoint file failed!")
print "\nStart evaluation on test set ...\n"
if config.has_key('contextwise') and config['contextwise']:
left_batch, middle_batch, right_batch, y_batch, _ = zip(
*eval_data)
feed = {
m.left: np.array(left_batch),
m.middle: np.array(middle_batch),
m.right: np.array(right_batch),
m.labels: np.array(y_batch)
}
else:
x_batch, y_batch, _ = zip(*eval_data)
feed = {
m.inputs: np.array(x_batch),
m.labels: np.array(y_batch)
}
loss, eval = sess.run([m.total_loss, m.eval_op], feed_dict=feed)
pre, rec = zip(*eval)
auc = util.calc_auc_pr(pre, rec)
f1 = (2.0 * pre[5] * rec[5]) / (pre[5] + rec[5])
print '%s: loss = %.6f, p = %.4f, r = %4.4f, f1 = %.4f, auc = %.4f' % (
datetime.now(), loss, pre[5], rec[5], f1, auc)
return pre, rec
def train(train_data, test_data):
# train_dir
timestamp = str(int(time.time()))
out_dir = os.path.abspath(os.path.join(FLAGS.train_dir, timestamp))
# save flags
if not os.path.exists(out_dir):
os.mkdir(out_dir)
FLAGS._parse_flags()
config = dict(FLAGS.__flags.items())
# Window_size must not be larger than the sent_len
if config['sent_len'] < config['max_window']:
config['max_window'] = config['sent_len']
# flag to restore the contextwise model
config['split'] = True
# save flags
config['train_dir'] = out_dir
util.dump_to_file(os.path.join(out_dir, 'flags.cPickle'), config)
# display parameter settings
print 'Parameters:'
for k, v in config.iteritems():
print '\t' + k + '=' + str(v)
# max number of steps
num_batches_per_epoch = int(
np.ceil(float(len(train_data)) / FLAGS.batch_size))
max_steps = num_batches_per_epoch * FLAGS.num_epochs
with tf.Graph().as_default():
with tf.variable_scope('cnn', reuse=None):
m = cnn_split.Model(config, is_train=True)
with tf.variable_scope('cnn', reuse=True):
mtest = cnn_split.Model(config, is_train=False)
# checkpoint
saver = tf.train.Saver(tf.all_variables())
save_path = os.path.join(out_dir, 'model.ckpt')
summary_op = tf.merge_all_summaries()
# session
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
with sess.as_default():
train_summary_writer = tf.train.SummaryWriter(os.path.join(
out_dir, "train"),
graph=sess.graph)
dev_summary_writer = tf.train.SummaryWriter(os.path.join(
out_dir, "dev"),
graph=sess.graph)
sess.run(tf.initialize_all_variables())
# assign pretrained embeddings
if FLAGS.use_pretrain:
print "Initializing model with pretrained embeddings ..."
pretrained_embedding = np.load(
os.path.join(FLAGS.data_dir, 'emb.npy'))
m.assign_embedding(sess, pretrained_embedding)
# initialize parameters
current_lr = FLAGS.init_lr
lowest_loss_value = float("inf")
decay_step_counter = 0
global_step = 0
# evaluate on dev set
def dev_step(mtest, sess):
dev_loss = []
dev_auc = []
dev_f1_score = []
# create batch
test_batches = util.batch_iter(test_data,
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
for batch in test_batches:
left_batch, right_batch, y_batch, n_batch = zip(*batch)
feed = {
mtest.left: np.array(left_batch),
mtest.right: np.array(right_batch),
mtest.labels: np.array(y_batch)
}
if FLAGS.negative:
feed[mtest.negative] = np.array(n_batch)
loss_value, eval_value = sess.run(
[mtest.total_loss, mtest.eval_op], feed_dict=feed)
dev_loss.append(loss_value)
pre, rec = zip(*eval_value)
dev_auc.append(util.calc_auc_pr(pre, rec))
dev_f1_score.append((2.0 * pre[5] * rec[5]) /
(pre[5] + rec[5])) # threshold = 0.5
return (np.mean(dev_loss), np.mean(dev_auc),
np.mean(dev_f1_score))
# train loop
print "\nStart training (save checkpoints in %s)\n" % out_dir
train_loss = []
train_auc = []
train_f1_score = []
train_batches = util.batch_iter(train_data,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs)
for batch in train_batches:
batch_size = len(batch)
m.assign_lr(sess, current_lr)
global_step += 1
left_batch, right_batch, y_batch, n_batch = zip(*batch)
feed = {
m.left: np.array(left_batch),
m.right: np.array(right_batch),
m.labels: np.array(y_batch)
}
if FLAGS.negative:
feed[m.negative] = np.array(n_batch)
start_time = time.time()
_, loss_value, eval_value = sess.run(
[m.train_op, m.total_loss, m.eval_op], feed_dict=feed)
proc_duration = time.time() - start_time
train_loss.append(loss_value)
pre, rec = zip(*eval_value)
auc = util.calc_auc_pr(pre, rec)
f1 = (2.0 * pre[5] * rec[5]) / (pre[5] + rec[5]
) # threshold = 0.5
train_auc.append(auc)
train_f1_score.append(f1)
assert not np.isnan(loss_value), "Model loss is NaN."
# print log
if global_step % FLAGS.log_step == 0:
examples_per_sec = batch_size / proc_duration
format_str = '%s: step %d/%d, f1 = %.4f, auc = %.4f, loss = %.4f ' + \
'(%.1f examples/sec; %.3f sec/batch), lr: %.6f'
print format_str % (datetime.now(), global_step, max_steps,
f1, auc, loss_value, examples_per_sec,
proc_duration, current_lr)
# write summary
if global_step % FLAGS.summary_step == 0:
summary_str = sess.run(summary_op)
train_summary_writer.add_summary(summary_str, global_step)
dev_summary_writer.add_summary(summary_str, global_step)
# summary loss, f1
train_summary_writer.add_summary(_summary_for_scalar(
'loss', np.mean(train_loss)),
global_step=global_step)
train_summary_writer.add_summary(_summary_for_scalar(
'auc', np.mean(train_auc)),
global_step=global_step)
train_summary_writer.add_summary(_summary_for_scalar(
'f1', np.mean(train_f1_score)),
global_step=global_step)
dev_loss, dev_auc, dev_f1 = dev_step(mtest, sess)
dev_summary_writer.add_summary(_summary_for_scalar(
'loss', dev_loss),
global_step=global_step)
dev_summary_writer.add_summary(_summary_for_scalar(
'auc', dev_auc),
global_step=global_step)
dev_summary_writer.add_summary(_summary_for_scalar(
'f1', dev_f1),
global_step=global_step)
print "\n===== write summary ====="
print "%s: step %d/%d: train_loss = %.6f, train_auc = %.4f train_f1 = %.4f" \
% (datetime.now(), global_step, max_steps,
np.mean(train_loss), np.mean(train_auc), np.mean(train_f1_score))
print "%s: step %d/%d: dev_loss = %.6f, dev_auc = %.4f dev_f1 = %.4f\n" \
% (datetime.now(), global_step, max_steps, dev_loss, dev_auc, dev_f1)
# reset container
train_loss = []
train_auc = []
train_f1_score = []
# decay learning rate if necessary
if loss_value < lowest_loss_value:
lowest_loss_value = loss_value
decay_step_counter = 0
else:
decay_step_counter += 1
if decay_step_counter >= FLAGS.tolerance_step:
current_lr *= FLAGS.lr_decay
print '%s: step %d/%d, Learning rate decays to %.5f' % \
(datetime.now(), global_step, max_steps, current_lr)
decay_step_counter = 0
# stop learning if learning rate is too low
if current_lr < 1e-5:
break
# save checkpoint
if global_step % FLAGS.checkpoint_step == 0:
saver.save(sess, save_path, global_step=global_step)
saver.save(sess, save_path, global_step=global_step)
def train(train_data, test_data):
# train_dir
timestamp = str(int(time.time()))
out_dir = os.path.abspath(os.path.join(FLAGS.train_dir, timestamp))
# save flags
if not os.path.exists(out_dir):
os.mkdir(out_dir)
FLAGS._parse_flags()
config = dict(FLAGS.__flags.items())
# Window_size must not be larger than the sent_len
if config['sent_len'] < config['max_window']:
config['max_window'] = config['sent_len']
util.dump_to_file(os.path.join(out_dir, 'flags.cPickle'), config)
print("Parameters:")
for k, v in config.items():
print('%20s %r' % (k, v))
num_batches_per_epoch = int(
np.ceil(float(len(train_data)) / FLAGS.batch_size))
max_steps = num_batches_per_epoch * FLAGS.num_epochs
with tf.Graph().as_default():
with tf.variable_scope('cnn', reuse=None):
m = cnn.Model(config, is_train=True)
with tf.variable_scope('cnn', reuse=True):
mtest = cnn.Model(config, is_train=False)
# checkpoint
saver = tf.train.Saver(tf.global_variables())
save_path = os.path.join(out_dir, 'model.ckpt')
summary_op = tf.summary.merge_all()
# session
with tf.Session().as_default() as sess:
proj_config = tf.contrib.tensorboard.plugins.projector.ProjectorConfig(
)
embedding = proj_config.embeddings.add()
embedding.tensor_name = m.W_emb.name
embedding.metadata_path = os.path.join(FLAGS.data_dir, 'vocab.txt')
train_summary_writer = tf.summary.FileWriter(os.path.join(
out_dir, "train"),
graph=sess.graph)
dev_summary_writer = tf.summary.FileWriter(os.path.join(
out_dir, "dev"),
graph=sess.graph)
tf.contrib.tensorboard.plugins.projector.visualize_embeddings(
train_summary_writer, proj_config)
tf.contrib.tensorboard.plugins.projector.visualize_embeddings(
dev_summary_writer, proj_config)
sess.run(tf.global_variables_initializer())
# assign pretrained embeddings
if FLAGS.use_pretrain:
print("Initialize model with pretrained embeddings...")
pretrained_embedding = np.load(
os.path.join(FLAGS.data_dir, 'emb.npy'))
m.assign_embedding(sess, pretrained_embedding)
# initialize parameters
current_lr = FLAGS.init_lr
lowest_loss_value = float("inf")
decay_step_counter = 0
global_step = 0
# evaluate on dev set
def dev_step(mtest, sess):
dev_loss = []
dev_auc = []
dev_f1_score = []
# create batch
test_batches = util.batch_iter(test_data,
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
for batch in test_batches:
x_batch, y_batch, _ = zip(*batch)
loss_value, eval_value = sess.run(
[mtest.total_loss, mtest.eval_op],
feed_dict={
mtest.inputs: np.array(x_batch),
mtest.labels: np.array(y_batch)
})
dev_loss.append(loss_value)
pre, rec = zip(*eval_value)
# look at the 5th index, which corresponds to a threshold = 0.5
threshold = 5
dev_auc.append(util.calc_auc_pr(pre, rec, threshold))
dev_f1_score.append(
(2.0 * pre[threshold] * rec[threshold]) /
(pre[threshold] + rec[threshold]))
return np.mean(dev_loss), np.mean(dev_auc), np.mean(
dev_f1_score)
# train loop
print("\nStart training (save checkpoints in %s)\n" % out_dir)
train_loss = []
train_auc = []
train_f1_score = []
train_batches = util.batch_iter(train_data,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs)
for batch in train_batches:
batch_size = len(batch)
m.assign_lr(sess, current_lr)
global_step += 1
x_batch, y_batch, a_batch = zip(*batch)
feed = {
m.inputs: np.array(x_batch),
m.labels: np.array(y_batch)
}
if FLAGS.attention:
feed[m.attention] = np.array(a_batch)
start_time = time.time()
_, loss_value, eval_value = sess.run(
[m.train_op, m.total_loss, m.eval_op], feed_dict=feed)
proc_duration = time.time() - start_time
train_loss.append(loss_value)
pre, rec = zip(*eval_value)
# look at the 5th index, which corresponds to a threshold = 0.5
threshold = 5
auc = util.calc_auc_pr(pre, rec, threshold)
f1 = (2.0 * pre[threshold] *
rec[threshold]) / (pre[threshold] + rec[threshold])
train_auc.append(auc)
train_f1_score.append(f1)
assert not np.isnan(loss_value), "Model loss is NaN."
# print log
if global_step % FLAGS.log_step == 0:
examples_per_sec = batch_size / proc_duration
format_str = '%s: step %d/%d, f1 = %.4f, auc = %.4f, loss = %.4f ' + \
'(%.1f examples/sec; %.3f sec/batch), lr: %.6f'
print(format_str % (datetime.now(), global_step, max_steps,
f1, auc, loss_value, examples_per_sec,
proc_duration, current_lr))
# write summary
if global_step % FLAGS.summary_step == 0:
summary_str = sess.run(summary_op)
train_summary_writer.add_summary(summary_str, global_step)
dev_summary_writer.add_summary(summary_str, global_step)
# summary loss, f1
train_summary_writer.add_summary(_summary_for_scalar(
'loss', np.mean(train_loss)),
global_step=global_step)
train_summary_writer.add_summary(_summary_for_scalar(
'auc', np.mean(train_auc)),
global_step=global_step)
train_summary_writer.add_summary(_summary_for_scalar(
'f1', np.mean(train_f1_score)),
global_step=global_step)
dev_loss, dev_auc, dev_f1 = dev_step(mtest, sess)
dev_summary_writer.add_summary(_summary_for_scalar(
'loss', dev_loss),
global_step=global_step)
dev_summary_writer.add_summary(_summary_for_scalar(
'auc', dev_auc),
global_step=global_step)
dev_summary_writer.add_summary(_summary_for_scalar(
'f1', dev_f1),
global_step=global_step)
print("\n===== write summary =====")
print("%s: step %d/%d: train_loss = %.6f, train_auc = %.4f, train_f1 = %.4f" \
% (datetime.now(), global_step, max_steps,
np.mean(train_loss), np.mean(train_auc), np.mean(train_f1_score)))
print("%s: step %d/%d: dev_loss = %.6f, dev_auc = %.4f, dev_f1 = %.4f\n" \
% (datetime.now(), global_step, max_steps, dev_loss, dev_auc, dev_f1))
# reset container
train_loss = []
train_auc = []
train_f1_score = []
# decay learning rate if necessary
if loss_value < lowest_loss_value:
lowest_loss_value = loss_value
decay_step_counter = 0
else:
decay_step_counter += 1
if decay_step_counter >= FLAGS.tolerance_step:
current_lr *= FLAGS.lr_decay
print('%s: step %d/%d, Learning rate decays to %.5f' % \
(datetime.now(), global_step, max_steps, current_lr))
decay_step_counter = 0
# stop learning if learning rate is too low
if current_lr < 1e-5:
break
# save checkpoint
if global_step % FLAGS.checkpoint_step == 0:
saver.save(sess, save_path, global_step=global_step)
saver.save(sess, save_path, global_step=global_step)
def evaluate(eval_data, config):
""" Build evaluation graph and run. """
with tf.Graph().as_default():
with tf.variable_scope('cnn'):
if config.has_key('contextwise') and config['contextwise']:
import cnn_context
m = cnn_context.Model(config, is_train=False)
else:
import cnn
m = cnn.Model(config, is_train=False)
saver = tf.train.Saver(tf.global_variables())
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(config['train_dir'])
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise IOError("Loading checkpoint file failed!")
#embeddings = sess.run(tf.global_variables())[0]
print "\nStart evaluation\n"
#losses = []
#precision = []
#recall = []
#batches = util.batch_iter(eval_data, batch_size=config['batch_size'], num_epochs=1, shuffle=False)
#for batch in batches:
if config.has_key('contextwise') and config['contextwise']:
left_batch, middle_batch, right_batch, y_batch, _ = zip(*eval_data)
feed = {m.left: np.array(left_batch),
m.middle: np.array(middle_batch),
m.right: np.array(right_batch),
m.labels: np.array(y_batch)}
else:
x_batch, y_batch, _ = zip(*eval_data)
feed = {m.inputs: np.array(x_batch), m.labels: np.array(y_batch)}
loss, eval, actual_output, eval_per_class = sess.run([m.total_loss, m.eval_op, m.scores, m.eval_class_op], feed_dict=feed)
#losses.append(loss)
pre, rec = zip(*eval)
#precision.append(pre)
#recall.append(rec)
avg_precision = np.mean(np.array(pre))
avg_recall = np.mean(np.array(rec))
auc = util.calc_auc_pr(pre, rec)
f1 = (2.0 * pre[5] * rec[5]) / (pre[5] + rec[5])
print '%s: Overall\nloss = %.6f, f1 = %.4f, auc = %.4f' % (datetime.now(), loss, f1, auc)
pre_per_class, rec_per_class = zip(*eval_per_class)
num_class = len(pre_per_class)
for class_i in range(num_class):
current_pre = pre_per_class[class_i]
current_rec = rec_per_class[class_i]
current_auc = util.calc_auc_pr(current_pre, current_rec)
current_f1 = (2.0 * current_pre[5] * current_rec[5]) / (current_pre[5] + current_rec[5])
print 'Class "%s": precision = %.4f, recall = %.4f, f1 = %.4f, auc = %.4f' % (CLASS_NAMES[class_i], current_pre[5], current_rec[5], current_f1, current_auc)
x_batch = np.array(x_batch)
y_batch = np.array(y_batch)
# Now calculate the true probability distribution using softmax.
actual_output_exp = np.exp(actual_output)
actual_output_softmax = actual_output_exp / np.sum(actual_output_exp, axis=1, keepdims=True)
plot_precision_recall(y_batch,actual_output_softmax)
return pre, rec, x_batch, y_batch, actual_output
def train(train_data, test_data, FLAGS = tf.app.flags.FLAGS):
# # train_dir
# timestamp = str(int(time.time()))
# out_dir = os.path.abspath(os.path.join(FLAGS.train_dir, timestamp))
#
# # save flags
# if not os.path.exists(out_dir):
# os.mkdir(out_dir)
# FLAGS._parse_flags()
# config = dict(FLAGS.__flags.items())
#
# # Window_size must not be larger than the sent_len
# if config['sent_len'] < config['max_window']:
# config['max_window'] = config['sent_len']
#
# util.dump_to_file(os.path.join(out_dir, 'flags.cPickle'), config)
train_x = get_key_phrases(train_data)
_, train_y, _ = zip(*train_data)
test_x = get_key_phrases(test_data)
_, test_y, _ = zip(*test_data)
# # assign pretrained embeddings
# if FLAGS.use_pretrain:
print "Initialize model with pretrained embeddings..."
print("Please don't forget to change the vocab size to the corresponding on in the embedding.")
pretrained_embedding = np.load(os.path.join(FLAGS.data_dir, 'emb.npy'))
train_x = key_phrase_indices_to_embedding(train_x, pretrained_embedding)
test_x = key_phrase_indices_to_embedding(test_x, pretrained_embedding)
# Use SVM. But SVM does not output a probability
# train_y = np.argmax(train_y, axis=1)
# test_y = np.argmax(test_y, axis=1)
# clf = svm.SVC(class_weight='balanced')
# clf.fit(train_x, train_y)
# predicted_test_y = clf.predict(test_x)
# Use fully connected multilayer nn.
config = dict(FLAGS.__flags.items())
# train_dir
timestamp = str(int(time.time()))
out_dir = os.path.abspath(os.path.join(FLAGS.train_dir, timestamp))
# save flags
if not os.path.exists(out_dir):
os.mkdir(out_dir)
FLAGS._parse_flags()
config = dict(FLAGS.__flags.items())
util.dump_to_file(os.path.join(out_dir, 'flags.cPickle'), config)
num_batches_per_epoch = int(np.ceil(float(len(train_data))/FLAGS.batch_size))
max_steps = num_batches_per_epoch * FLAGS.num_epochs
with tf.Graph().as_default():
with tf.variable_scope('multilayer', reuse=None):
m = multilayer.Model(config, is_train=True)
with tf.variable_scope('multilayer', reuse=True):
mtest = multilayer.Model(config, is_train=False)
# checkpoint
saver = tf.train.Saver(tf.global_variables(), max_to_keep=1)
save_path = os.path.join(out_dir, 'model.ckpt')
try:
summary_op = tf.summary.merge_all()
except:
summary_op = tf.merge_all_summaries()
# session
config = tf.ConfigProto(log_device_placement=FLAGS.log_device_placement)
if FLAGS.gpu_percentage > 0:
config.gpu_options.per_process_gpu_memory_fraction = FLAGS.gpu_percentage
else:
config = tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement,
device_count={'GPU': 0}
)
sess = tf.Session(config=config)
with sess.as_default():
train_summary_writer = tf.train.SummaryWriter(os.path.join(out_dir, "train"), graph=sess.graph)
dev_summary_writer = tf.train.SummaryWriter(os.path.join(out_dir, "dev"), graph=sess.graph)
try:
sess.run(tf.global_variables_initializer())
except:
sess.run(tf.initialize_all_variables())
# # assign pretrained embeddings
# if FLAGS.use_pretrain:
# print "Initialize model with pretrained embeddings..."
# print("Please don't forget to change the vocab size to the corresponding on in the embedding.")
# pretrained_embedding = np.load(os.path.join(FLAGS.data_dir, 'emb.npy'))
# m.assign_embedding(sess, pretrained_embedding)
# initialize parameters
current_lr = FLAGS.init_lr
lowest_loss_value = float("inf")
decay_step_counter = 0
global_step = 0
# evaluate on dev set
def dev_step(mtest, sess):
dev_loss = []
dev_auc = []
dev_f1_score = []
# create batch
test_batches = util.batch_iter(zip(test_x, test_y), batch_size=FLAGS.batch_size, num_epochs=1, shuffle=False)
for batch in test_batches:
x_batch, y_batch, = zip(*batch)
# a_batch = np.ones((len(batch), 1), dtype=np.float32) / len(batch) # average
loss_value, eval_value = sess.run([mtest.total_loss, mtest.eval_op],
feed_dict={mtest.inputs: np.array(x_batch),
mtest.labels: np.array(y_batch)})
dev_loss.append(loss_value)
pre, rec = zip(*eval_value)
dev_auc.append(util.calc_auc_pr(pre, rec))
dev_f1_score.append((2.0 * pre[5] * rec[5]) / (pre[5] + rec[5])) # threshold = 0.5
return (np.mean(dev_loss), np.mean(dev_auc), np.mean(dev_f1_score))
# train loop
print "\nStart training (save checkpoints in %s)\n" % out_dir
train_loss = []
train_auc = []
train_f1_score = []
train_batches = util.batch_iter(zip(train_x, train_y), batch_size=FLAGS.batch_size, num_epochs=FLAGS.num_epochs)
for batch in train_batches:
batch_size = len(batch)
m.assign_lr(sess, current_lr)
global_step += 1
x_batch, y_batch, = zip(*batch)
feed = {m.inputs: np.array(x_batch), m.labels: np.array(y_batch)}
start_time = time.time()
_, loss_value, eval_value = sess.run([m.train_op, m.total_loss, m.eval_op], feed_dict=feed)
proc_duration = time.time() - start_time
train_loss.append(loss_value)
pre, rec = zip(*eval_value)
auc = util.calc_auc_pr(pre, rec)
f1 = (2.0 * pre[5] * rec[5]) / (pre[5] + rec[5]) # threshold = 0.5
train_auc.append(auc)
train_f1_score.append(f1)
assert not np.isnan(loss_value), "Model loss is NaN."
# print log
if global_step % FLAGS.log_step == 0:
examples_per_sec = batch_size / proc_duration
format_str = '%s: step %d/%d, f1 = %.4f, auc = %.4f, loss = %.4f ' + \
'(%.1f examples/sec; %.3f sec/batch), lr: %.6f'
print format_str % (datetime.now(), global_step, max_steps, f1, auc, loss_value,
examples_per_sec, proc_duration, current_lr)
# write summary
if global_step % FLAGS.summary_step == 0:
summary_str = sess.run(summary_op)
train_summary_writer.add_summary(summary_str, global_step)
dev_summary_writer.add_summary(summary_str, global_step)
# summary loss, f1
train_summary_writer.add_summary(
_summary_for_scalar('loss', np.mean(train_loss)), global_step=global_step)
train_summary_writer.add_summary(
_summary_for_scalar('auc', np.mean(train_auc)), global_step=global_step)
train_summary_writer.add_summary(
_summary_for_scalar('f1', np.mean(train_f1_score)), global_step=global_step)
dev_loss, dev_auc, dev_f1 = dev_step(mtest, sess)
dev_summary_writer.add_summary(
_summary_for_scalar('loss', dev_loss), global_step=global_step)
dev_summary_writer.add_summary(
_summary_for_scalar('auc', dev_auc), global_step=global_step)
dev_summary_writer.add_summary(
_summary_for_scalar('f1', dev_f1), global_step=global_step)
print "\n===== write summary ====="
print "%s: step %d/%d: train_loss = %.6f, train_auc = %.4f, train_f1 = %.4f" \
% (datetime.now(), global_step, max_steps,
np.mean(train_loss), np.mean(train_auc), np.mean(train_f1_score))
print "%s: step %d/%d: dev_loss = %.6f, dev_auc = %.4f, dev_f1 = %.4f\n" \
% (datetime.now(), global_step, max_steps, dev_loss, dev_auc, dev_f1)
# reset container
train_loss = []
train_auc = []
train_f1_score = []
# decay learning rate if necessary
if loss_value < lowest_loss_value:
lowest_loss_value = loss_value
decay_step_counter = 0
else:
decay_step_counter += 1
if decay_step_counter >= FLAGS.tolerance_step:
current_lr *= FLAGS.lr_decay
print '%s: step %d/%d, Learning rate decays to %.5f' % \
(datetime.now(), global_step, max_steps, current_lr)
decay_step_counter = 0
# stop learning if learning rate is too low
if current_lr < 1e-5:
break
# save checkpoint
if global_step % FLAGS.checkpoint_step == 0:
saver.save(sess, save_path, global_step=global_step)
saver.save(sess, save_path, global_step=global_step)
# Lastly evaluate the test set
loss_value, predicted_test_y_logits = sess.run([mtest.total_loss, mtest.scores],
feed_dict={mtest.inputs: np.array(test_x),
mtest.labels: np.array(test_y)})
predicted_test_y = np.argmax(predicted_test_y_logits, axis=1)
test_y = np.argmax(test_y, axis=1)
result = (predicted_test_y == test_y)
accuracy = np.sum(result.astype(np.int32)) / float(result.shape[0])
print("Overall %f%% answers were correct. " %(float(accuracy * 100)))
epsilon = 0.00000001
num_categories = 3
true_positive_per_category = [np.bitwise_and(test_y==category_i, predicted_test_y==category_i) for category_i in range(num_categories)]
false_positive_per_category = [np.bitwise_and(test_y!=category_i, predicted_test_y==category_i) for category_i in range(num_categories)]
true_negative_per_category = [np.bitwise_and(test_y!=category_i, predicted_test_y!=category_i) for category_i in range(num_categories)]
false_negative_per_category = [np.bitwise_and(test_y==category_i, predicted_test_y!=category_i) for category_i in range(num_categories)]
precision_per_category = [np.sum(true_positive_per_category[category_i].astype(np.int32)) /
float(np.sum(true_positive_per_category[category_i].astype(np.int32)) +
np.sum(false_positive_per_category[category_i].astype(np.int32)) + epsilon)
for category_i in range(num_categories)]
recall_per_category = [np.sum(true_positive_per_category[category_i].astype(np.int32)) /
float(np.sum(true_positive_per_category[category_i].astype(np.int32)) +
np.sum(false_negative_per_category[category_i].astype(np.int32)) + epsilon)
for category_i in range(num_categories)]
f1_per_category = [2 / (1 / (precision_per_category[category_i] + epsilon) +
1 / (recall_per_category[category_i] + epsilon)) for category_i in range(num_categories)]
for category_i in range(num_categories):
print("Category %d has f1 score: %f, precision: %f, and recall %f"
%(category_i, f1_per_category[category_i], precision_per_category[category_i], recall_per_category[category_i]))
return test_x, predicted_test_y_logits
