def reply_generator():
train_batcher = batcher(args.batch_size, args.train_query_file,
args.train_response_file)
while True:
try:
batch = train_batcher.next()
except StopIteration:
break
(query_src, query_len, reply_src, reply_len, _,
_) = utils.get_variables(batch,
vocab,
args.dec_max_len,
use_cuda=args.use_cuda)
q_avg, r_avg = get_avg_embedding(query_src, query_len, reply_src,
reply_len)
yield r_avg
def eval_model(epoch):
rq_fe = []
qr_fe = []
total_case = 0
cur_time = time.time()
rq_recon = []
qr_recon = []
valid_batcher = batcher(args.batch_size, args.valid_query_file,
args.valid_response_file)
while True:
try:
batch = valid_batcher.next()
except:
break
posts_var, posts_len, response_var, response_len, _, _= utils.get_variables(batch,
vocab, args.dec_max_len, use_cuda=args.use_cuda)
# check batcher, and padding
q_pos, r_pos = norm_data(posts_var, posts_len, response_var, response_len)
rq_recon_error = rq_rbm.valid_op(q_pos, r_pos)
qr_recon_error = qr_rbm.valid_op(r_pos, q_pos)
rq_recon.append(rq_recon_error)
qr_recon.append(qr_recon_error)
rq_fe.append( rq_rbm.hr_free_energy(q_pos, r_pos))
qr_fe.append( qr_rbm.hr_free_energy(r_pos, q_pos))
total_case += len(batch[0])
rq_error = zip(*rq_recon)
qr_error = zip(*qr_recon)
batch_num = len(rq_recon)
logger.info("Valid recon error: %.4f, %.4f, %.4f %.4f" % (sum(rq_error[1])/batch_num,
sum(rq_error[2])/batch_num, sum(qr_error[1])/batch_num, sum(qr_error[2])/batch_num))
logger.info('Valid batch %d: average QR FE %.4f, RQ FE %.4f, QR, FEgap: %.4f, %4.f, RQ FEGap: %.4f, %.4f; (%.1f case/sec)' % (
epoch, sum(qr_fe) / len(qr_fe), sum(rq_fe) / len(rq_fe),
(sum(qr_fe)-sum(qr_error[4])) /len(qr_fe), (sum(qr_fe)-sum(qr_error[5]))/len(qr_fe),
(sum(rq_fe)-sum(rq_error[4])) /len(rq_fe), (sum(rq_fe)-sum(rq_error[5]))/len(rq_fe),
total_case/(time.time()-cur_time)))
logger.info('')
epoch, sum(qr_fe) / len(qr_fe), sum(rq_fe) / len(rq_fe),
(sum(qr_fe)-sum(qr_error[4])) /len(qr_fe), (sum(qr_fe)-sum(qr_error[5]))/len(qr_fe),
(sum(rq_fe)-sum(rq_error[4])) /len(rq_fe), (sum(rq_fe)-sum(rq_error[5]))/len(rq_fe),
total_case/(time.time()-cur_time)))
logger.info('')
eval_model(-1)
for e in range(args.num_epoch):
logger.info('---------------------training--------------------------')
logger.info("Epoch: %d/%d" % (e, args.num_epoch))
print("Epoch: %d/%d" % (e, args.num_epoch))
if e > 5:
rq_rbm.monument = 0.8
qr_rbm.monument = 0.8
train_batcher = batcher(args.batch_size, args.train_query_file,
args.train_response_file)
# train state
rq_fe = []
qr_fe = []
rq_recon = []
qr_recon = []
total_case = 0
cur_time = time.time()
step = 0
while True:
try:
batch = train_batcher.next()
except StopIteration:
break
# train_opt,