valid_ce.append(test_lio_skel(use, test_model, batch, drop, tr.rng, epoch, tr.batch_size, x_, y_, loader, x_skeleton_))
# save best params
# if valid_ce[-1][1] < 0.25:
res_dir = save_results(train_ce, valid_ce, res_dir, params=params)
if not tr.moved: res_dir = move_results(res_dir)
if valid_ce[-1][1] < best_valid:
save_params(params, res_dir, "best")
save_params(params, res_dir)
if valid_ce[-1][1] < best_valid:
best_valid = valid_ce[-1][1]
# epoch report
epoch_report(epoch, best_valid, time()-time_start, learning_rate.get_value(borrow=True),\
train_ce[-1], valid_ce[-1], res_dir)
# make_plot(train_ce, valid_ce)
if lr.decay_each_epoch:
learning_rate.set_value(float32(learning_rate.get_value(borrow=True)*lr.decay))
# elif lr.decay_if_plateau:
# if epoch - lr_decay_epoch > tr.patience \
# and valid_ce[-1-tr.patience][1] <= valid_ce[-1][1]:
# write("Learning rate decay: validation error stopped improving")
# lr_decay_epoch = epoch
# n_lr_decays +=1
# learning_rate.set_value(float32(learning_rate.get_value(borrow=True)*lr.decay_big))
# if epoch == 0:
# learning_rate.set_value(float32(3e-4))
# else:
tr.batch_size, x_, y_, loader, x_skeleton_))
# save best params
# if valid_ce[-1][1] < 0.25:
res_dir = save_results(train_ce, valid_ce, res_dir, params=params)
if not tr.moved: res_dir = move_results(res_dir)
if valid_ce[-1][1] < best_valid:
save_params(params, res_dir, "best")
save_params(params, res_dir)
if valid_ce[-1][1] < best_valid:
best_valid = valid_ce[-1][1]
# epoch report
epoch_report(epoch, best_valid, time()-time_start, time()-prog_start_time, learning_rate.get_value(borrow=True),\
train_ce[-1], valid_ce[-1], res_dir)
# make_plot(train_ce, valid_ce)
if lr.decay_each_epoch:
learning_rate.set_value(
float32(learning_rate.get_value(borrow=True) * lr.decay))
# elif lr.decay_if_plateau:
# if epoch - lr_decay_epoch > tr.patience \
# and valid_ce[-1-tr.patience][1] <= valid_ce[-1][1]:
# write("Learning rate decay: validation error stopped improving")
# lr_decay_epoch = epoch
# n_lr_decays +=1
# learning_rate.set_value(float32(learning_rate.get_value(borrow=True)*lr.decay_big))
# if epoch == 0:
# learning_rate.set_value(float32(3e-4))