def _compare_result_with_origin_model(self,
check_func,
use_cuda,
delta2=1e-5,
compare_seperately=True):
if use_cuda and not core.is_compiled_with_cuda():
return
func_1_first_loss, func_1_last_loss = self.check_network_convergence(
seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(),
use_cuda=use_cuda,
use_reduce=False,
optimizer=seresnext_net.optimizer)
func_2_first_loss, func_2_last_loss = check_func(
seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(),
use_cuda=use_cuda)
if compare_seperately:
for loss in zip(func_1_first_loss, func_2_first_loss):
self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
for loss in zip(func_1_last_loss, func_2_last_loss):
self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
else:
self.assertAlmostEquals(np.mean(func_1_first_loss),
func_2_first_loss[0],
delta=1e-5)
self.assertAlmostEquals(np.mean(func_1_last_loss),
func_2_last_loss[0],
delta=delta2)
def test_seresnet(self):
with self.program_scope_guard():
self.check_prune_correctness(
method=seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda=False),
optimizer=seresnext_net.optimizer)
def _compare_reduce_and_allreduce(self, use_cuda, delta2=1e-5):
if use_cuda and not core.is_compiled_with_cuda():
return
all_reduce_first_loss, all_reduce_last_loss = self.check_network_convergence(
seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(),
use_cuda=use_cuda,
use_reduce=False,
optimizer=seresnext_net.optimizer)
reduce_first_loss, reduce_last_loss = self.check_network_convergence(
seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(),
use_cuda=use_cuda,
use_reduce=True,
optimizer=seresnext_net.optimizer)
for loss in zip(all_reduce_first_loss, reduce_first_loss):
self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
for loss in zip(all_reduce_last_loss, reduce_last_loss):
self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
if not use_cuda:
return
all_reduce_first_loss_seq, all_reduce_last_loss_seq = self.check_network_convergence(
seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(),
use_cuda=use_cuda,
use_reduce=False,
optimizer=seresnext_net.optimizer,
enable_sequential_execution=True)
reduce_first_loss_seq, reduce_last_loss_seq = self.check_network_convergence(
seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(),
use_cuda=use_cuda,
use_reduce=True,
optimizer=seresnext_net.optimizer,
enable_sequential_execution=True)
for loss in zip(all_reduce_first_loss, all_reduce_first_loss_seq):
self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
for loss in zip(all_reduce_last_loss, all_reduce_last_loss_seq):
self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
for loss in zip(reduce_first_loss, reduce_first_loss_seq):
self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
for loss in zip(reduce_last_loss, reduce_last_loss_seq):
self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
for loss in zip(all_reduce_first_loss_seq, reduce_first_loss_seq):
self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
for loss in zip(all_reduce_last_loss_seq, reduce_last_loss_seq):
self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
