p04896132
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix auto parallel st

This commit is contained in:
lichenever 2020-03-30 21:41:59 +08:00
parent 1a3c06a948
commit 30231eab2f
7 changed files with 22 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tests/st/auto_parallel/onehot_model_parallel.py
View File

@ -130,9 +130,7 @@ class OneHotFactory:
context.reset_auto_parallel_context() context.reset_auto_parallel_context()
assert np.allclose(out_mindspore_single, out_mindspore_parallel, 0.0001, 0.0001) assert np.allclose(out_mindspore_single, out_mindspore_parallel, 0.0001, 0.0001)
@pytest.mark.level0
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_single
def test_reid_onehot_forward_int32_128_depth1024_model_parallel(): def test_reid_onehot_forward_int32_128_depth1024_model_parallel():
fact = OneHotFactory(batch_size=128, fact = OneHotFactory(batch_size=128,
classes=1024, classes=1024,
@ -142,9 +140,7 @@ def test_reid_onehot_forward_int32_128_depth1024_model_parallel():
strategy=((1,device_num),(),())) strategy=((1,device_num),(),()))
fact.forward_cmp() fact.forward_cmp()
@pytest.mark.level0
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_single
def test_reid_onehot_forward_int32_1024_depth128_model_parallel(): def test_reid_onehot_forward_int32_1024_depth128_model_parallel():
fact = OneHotFactory(batch_size=1024, fact = OneHotFactory(batch_size=1024,
classes=128, classes=128,
@ -153,4 +149,3 @@ def test_reid_onehot_forward_int32_1024_depth128_model_parallel():
axis=-1, axis=-1,
strategy=((1,device_num),(),())) strategy=((1,device_num),(),()))
fact.forward_cmp() fact.forward_cmp()

3
tests/st/auto_parallel/run_auto_parallel_loss_expand.sh
View File

@ -18,7 +18,6 @@ BASE_PATH=$(cd "$(dirname $0)"; pwd)
CONFIG_PATH=/home/workspace/mindspore_config CONFIG_PATH=/home/workspace/mindspore_config
export DEVICE_NUM=8 export DEVICE_NUM=8
export RANK_SIZE=$DEVICE_NUM export RANK_SIZE=$DEVICE_NUM
ulimit -n 65535
source ${BASE_PATH}/env.sh source ${BASE_PATH}/env.sh
unset SLOG_PRINT_TO_STDOUT unset SLOG_PRINT_TO_STDOUT
export MINDSPORE_HCCL_CONFIG_PATH=$CONFIG_PATH/hccl/rank_table_${DEVICE_NUM}p.json export MINDSPORE_HCCL_CONFIG_PATH=$CONFIG_PATH/hccl/rank_table_${DEVICE_NUM}p.json
@ -27,7 +26,7 @@ process_pid=()
for((i=0; i<$DEVICE_NUM; i++)); do for((i=0; i<$DEVICE_NUM; i++)); do
rm -rf ${BASE_PATH}/loss_expand${i} rm -rf ${BASE_PATH}/loss_expand${i}
mkdir ${BASE_PATH}/loss_expand${i} mkdir ${BASE_PATH}/loss_expand${i}
cp -r soft_entropy_loss_expand_parallel.py ${BASE_PATH}/loss_expand${i}/ cp -r ${BASE_PATH}/soft_entropy_loss_expand_parallel.py ${BASE_PATH}/loss_expand${i}/
cd ${BASE_PATH}/loss_expand${i} cd ${BASE_PATH}/loss_expand${i}
export RANK_ID=${i} export RANK_ID=${i}
export DEVICE_ID=${i} export DEVICE_ID=${i}

2
tests/st/auto_parallel/run_auto_parallel_resnet50_expand_loss.sh
View File

@ -27,7 +27,7 @@ process_pid=()
for((i=0; i<$DEVICE_NUM; i++)); do for((i=0; i<$DEVICE_NUM; i++)); do
rm -rf ${BASE_PATH}/resnet50_expand_loss${i} rm -rf ${BASE_PATH}/resnet50_expand_loss${i}
mkdir ${BASE_PATH}/resnet50_expand_loss${i} mkdir ${BASE_PATH}/resnet50_expand_loss${i}
cp -r resnet50_expand_loss.py ${BASE_PATH}/resnet50_expand_loss${i}/ cp -r ${BASE_PATH}/resnet50_expand_loss.py ${BASE_PATH}/resnet50_expand_loss${i}/
cd ${BASE_PATH}/resnet50_expand_loss${i} cd ${BASE_PATH}/resnet50_expand_loss${i}
export RANK_ID=${i} export RANK_ID=${i}
export DEVICE_ID=${i} export DEVICE_ID=${i}

2
tests/st/auto_parallel/run_onehot_model_parallel.sh
View File

@ -27,7 +27,7 @@ process_pid=()
for((i=0; i<$DEVICE_NUM; i++)); do for((i=0; i<$DEVICE_NUM; i++)); do
rm -rf ${BASE_PATH}/onehot_model_parallel${i} rm -rf ${BASE_PATH}/onehot_model_parallel${i}
mkdir ${BASE_PATH}/onehot_model_parallel${i} mkdir ${BASE_PATH}/onehot_model_parallel${i}
cp -r onehot_model_parallel.py ${BASE_PATH}/onehot_model_parallel${i}/ cp -r ${BASE_PATH}/onehot_model_parallel.py ${BASE_PATH}/onehot_model_parallel${i}/
cd ${BASE_PATH}/onehot_model_parallel${i} cd ${BASE_PATH}/onehot_model_parallel${i}
export RANK_ID=${i} export RANK_ID=${i}
export DEVICE_ID=${i} export DEVICE_ID=${i}

40
tests/st/auto_parallel/soft_entropy_loss_expand_parallel.py
View File

@ -118,6 +118,9 @@ class Dataset():
def get_dataset_size(self): def get_dataset_size(self):
return self.length return self.length
def get_repeat_count(self):
return self.length
class ModelCallback(Callback): class ModelCallback(Callback):
def __init__(self): def __init__(self):
super(ModelCallback, self).__init__() super(ModelCallback, self).__init__()
@ -177,7 +180,6 @@ class LossFactory():
dataGen = DataGenerator() dataGen = DataGenerator()
self.input_full, self.input_part = dataGen.input_data((batch_size, embed)) self.input_full, self.input_part = dataGen.input_data((batch_size, embed))
self.label_full, self.label_part = dataGen.label_data((batch_size,),embed) self.label_full, self.label_part = dataGen.label_data((batch_size,),embed)
self.expect_out = np.array([0.9205861 , 0.9205861 , 0.9205861 , 0.9201946 , 0.91951686, 0.919343])
def single_matmul_trains(self): def single_matmul_trains(self):
single_callback = ModelCallback() single_callback = ModelCallback()
@ -187,7 +189,8 @@ class LossFactory():
epoch_size = 6 epoch_size = 6
dataset = Dataset(self.input_full, self.label_full) dataset = Dataset(self.input_full, self.label_full)
model.train(epoch_size, dataset, callbacks=single_callback, dataset_sink_mode=False) model.train(epoch_size, dataset, callbacks=single_callback, dataset_sink_mode=False)
print("---loss---",single_callback.loss_list) loss_value = np.array(single_callback.loss_list)
return loss_value
def data_parallel_matmul_trains(self): def data_parallel_matmul_trains(self):
parallel_callback = ModelCallback() parallel_callback = ModelCallback()
@ -199,7 +202,7 @@ class LossFactory():
dataset = Dataset(self.input_part, self.label_part) dataset = Dataset(self.input_part, self.label_part)
model.train(epoch_size, dataset, callbacks=parallel_callback, dataset_sink_mode=False) model.train(epoch_size, dataset, callbacks=parallel_callback, dataset_sink_mode=False)
loss_value = np.array(parallel_callback.loss_list) loss_value = np.array(parallel_callback.loss_list)
assert allclose(loss_value, self.expect_out, 0.00001, 0.00001) return loss_value
def model_parallel_matmul_trains(self): def model_parallel_matmul_trains(self):
parallel_callback = ModelCallback() parallel_callback = ModelCallback()
@ -224,7 +227,7 @@ class LossFactory():
dataset = Dataset(self.input_part, self.label_part) dataset = Dataset(self.input_part, self.label_part)
model.train(epoch_size, dataset, callbacks=parallel_callback, dataset_sink_mode=False) model.train(epoch_size, dataset, callbacks=parallel_callback, dataset_sink_mode=False)
loss_value = np.array(parallel_callback.loss_list) loss_value = np.array(parallel_callback.loss_list)
assert allclose(loss_value, self.expect_out, 0.00001, 0.00001) return loss_value
def mix_parallel_matmul_trains(self): def mix_parallel_matmul_trains(self):
parallel_callback = ModelCallback() parallel_callback = ModelCallback()
@ -249,28 +252,13 @@ class LossFactory():
dataset = Dataset(self.input_part, self.label_part) dataset = Dataset(self.input_part, self.label_part)
model.train(epoch_size, dataset, callbacks=parallel_callback, dataset_sink_mode=False) model.train(epoch_size, dataset, callbacks=parallel_callback, dataset_sink_mode=False)
loss_value = np.array(parallel_callback.loss_list) loss_value = np.array(parallel_callback.loss_list)
assert allclose(loss_value, self.expect_out, 0.00001, 0.00001) return loss_value
@pytest.mark.level0 def test_all_trains():
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_single
def test_matmul_loss_data_parallel_trains():
loss_factory = LossFactory() loss_factory = LossFactory()
context.reset_auto_parallel_context() context.reset_auto_parallel_context()
loss_factory.data_parallel_matmul_trains() single_loss = loss_factory.single_matmul_trains()
model_parallel_loss = loss_factory.model_parallel_matmul_trains()
@pytest.mark.level0 mix_parallel_loss = loss_factory.mix_parallel_matmul_trains()
@pytest.mark.platform_x86_ascend_training assert allclose(single_loss, model_parallel_loss)
@pytest.mark.env_single assert allclose(single_loss, mix_parallel_loss)
def test_matmul_loss_model_parallel_trains():
loss_factory = LossFactory()
context.reset_auto_parallel_context()
loss_factory.model_parallel_matmul_trains()
@pytest.mark.level0
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_single
def test_matmul_loss_mix_parallel_trains():
loss_factory = LossFactory()
context.reset_auto_parallel_context()
loss_factory.mix_parallel_matmul_trains()

4
tests/st/auto_parallel/test_expand_loss.py
View File

@ -18,7 +18,9 @@ import pytest
@pytest.mark.level0 @pytest.mark.level0
@pytest.mark.platform_x86_ascend_training @pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_arm_ascend_training
@pytest.mark.env_single @pytest.mark.env_single
def test_expand_loss(): def test_expand_loss():
ret = os.system("sh run_auto_parallel_loss_expand.sh") sh_path = os.path.split(os.path.realpath(__file__))[0]
ret = os.system(f"sh {sh_path}/run_auto_parallel_loss_expand.sh")
assert(ret==0) assert(ret==0)

3
tests/st/auto_parallel/test_model_parallel_onehot.py
View File

@ -16,9 +16,6 @@
import os import os
import pytest import pytest
@pytest.mark.level0
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_single
def test_expand_loss(): def test_expand_loss():
ret = os.system("sh run_onehot_model_parallel.sh") ret = os.system("sh run_onehot_model_parallel.sh")
assert(ret==0) assert(ret==0)