def main():
dict_ = pickle.load(open(Params.data_dir + "dictionary.pkl","r"))
vocab_size = dict_.vocab_size
model = Model(is_training = True, vocab_size = vocab_size); print("Built model")
init = False
devdata, dev_ind = get_dev()
if not os.path.isfile(os.path.join(Params.logdir,"checkpoint")):
init = True
glove = np.memmap(Params.data_dir + "glove.np", dtype = np.float32, mode = "r")
glove = np.reshape(glove,(vocab_size,Params.emb_size))
with model.graph.as_default():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sv = tf.train.Supervisor(logdir=Params.logdir,
save_model_secs=0,
global_step = model.global_step,
init_op = model.init_op)
with sv.managed_session(config = config) as sess:
if init: sess.run(model.emb_assign, {model.word_embeddings_placeholder:glove})
pr = Params()
pr.dump_config(Params.__dict__)
for epoch in range(1, Params.num_epochs+1):
train_loss = []
if sv.should_stop(): break
for step in tqdm(range(model.num_batch), total = model.num_batch, ncols=70, leave=False, unit='b'):
_, loss = sess.run([model.train_op, model.mean_loss],
feed_dict={model.dropout: Params.dropout if Params.dropout is not None else 0.0})
train_loss.append(loss)
if step % Params.save_steps == 0:
gs = sess.run(model.global_step)
sv.saver.save(sess, Params.logdir + '/model_epoch_%d_step_%d'%(gs//model.num_batch, gs%model.num_batch))
if step % Params.dev_steps == 0:
EM_ = []
F1_ = []
dev = []
for i in range(Params.dev_batchs):
sample = np.random.choice(dev_ind, Params.batch_size)
feed_dict = {data: devdata[i][sample] for i,data in enumerate(model.data)}
index, dev_loss = sess.run([model.output_index, model.mean_loss], feed_dict = feed_dict)
F1, EM = 0.0, 0.0
for batch in range(Params.batch_size):
f1, em = f1_and_EM(index[batch], devdata[-1][sample][batch], devdata[0][sample][batch], dict_)
F1 += f1
EM += em
F1 /= float(Params.batch_size)
EM /= float(Params.batch_size)
EM_.append(EM)
F1_.append(F1)
dev.append(dev_loss)
EM_ = np.mean(EM_)
F1_ = np.mean(F1_)
dev = np.mean(dev)
sess.run(model.metric_assign,{model.F1_placeholder: F1_, model.EM_placeholder: EM_, model.dev_loss_placeholder: dev})
print("\nTrain_loss: {}\nDev_loss: {}\nDev_Exact_match: {}\nDev_F1_score: {}".format(np.mean(train_loss),dev,EM_,F1_))
train_loss = []
def main():
model = Model(is_training = True); print("Built model")
dict_ = pickle.load(open(Params.data_dir + "dictionary.pkl","r"))
init = False
devdata, dev_ind = get_dev()
if not os.path.isfile(os.path.join(Params.logdir,"checkpoint")):
init = True
glove = np.memmap(Params.data_dir + "glove.np", dtype = np.float32, mode = "r")
glove = np.reshape(glove,(Params.vocab_size,Params.emb_size))
char_glove = np.memmap(Params.data_dir + "glove_char.np",dtype = np.float32, mode = "r")
char_glove = np.reshape(char_glove,(Params.char_vocab_size,Params.emb_size))
with model.graph.as_default():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sv = tf.train.Supervisor(logdir=Params.logdir,
save_model_secs=0,
global_step = model.global_step,
init_op = model.init_op)
with sv.managed_session(config = config) as sess:
if init: sess.run(model.emb_assign, {model.word_embeddings_placeholder:glove, model.char_embeddings_placeholder:char_glove})
for epoch in range(1, Params.num_epochs+1):
if sv.should_stop(): break
for step in tqdm(range(model.num_batch), total = model.num_batch, ncols=70, leave=False, unit='b'):
sess.run(model.train_op)
if step % Params.save_steps == 0:
gs = sess.run(model.global_step)
sv.saver.save(sess, Params.logdir + '/model_epoch_%d_step_%d'%(gs//model.num_batch, gs%model.num_batch))
sample = np.random.choice(dev_ind, Params.batch_size)
feed_dict = {data: devdata[i][sample] for i,data in enumerate(model.data)}
logits, dev_loss = sess.run([model.points_logits, model.mean_loss], feed_dict = feed_dict)
index = np.argmax(logits, axis = 2)
F1, EM = 0.0, 0.0
for batch in range(Params.batch_size):
f1, em = f1_and_EM(index[batch], devdata[8][sample][batch], devdata[0][sample][batch], dict_)
F1 += f1
EM += em
F1 /= float(Params.batch_size)
EM /= float(Params.batch_size)
sess.run(model.metric_assign,{model.F1_placeholder: F1, model.EM_placeholder: EM, model.dev_loss_placeholder: dev_loss})
print("\nDev_loss: {}\nDev_Exact_match: {}\nDev_F1_score: {}".format(dev_loss,EM,F1))