未验证 提交 9d66385f 编写于 作者: H Huihuang Zheng 提交者: GitHub

[FixUnitTest] Fix a Batch Size Random Unit Test Failure (#24718)

The PR: https://github.com/PaddlePaddle/Paddle/pull/24651 seems causes new random failure of unit test test_parallel_executor_seresnext_base_cpu. The reason is that smaller batch size causes random optimization of neural network. I distinguished cpu/gpu batch size to fix the unittest.
上级 f9c7f58e
...@@ -172,8 +172,11 @@ def optimizer(learning_rate=0.01): ...@@ -172,8 +172,11 @@ def optimizer(learning_rate=0.01):
model = SE_ResNeXt50Small model = SE_ResNeXt50Small
def batch_size(): def batch_size(use_cuda):
return 8 if use_cuda:
# Paddle uses 8GB P4 GPU for unittest so we decreased the batch size.
return 8
return 12
def iter(use_cuda): def iter(use_cuda):
...@@ -183,9 +186,9 @@ def iter(use_cuda): ...@@ -183,9 +186,9 @@ def iter(use_cuda):
gpu_img, gpu_label = init_data( gpu_img, gpu_label = init_data(
batch_size=batch_size(), img_shape=img_shape, label_range=999) batch_size=batch_size(use_cuda=True), img_shape=img_shape, label_range=999)
cpu_img, cpu_label = init_data( cpu_img, cpu_label = init_data(
batch_size=batch_size(), img_shape=img_shape, label_range=999) batch_size=batch_size(use_cuda=False), img_shape=img_shape, label_range=999)
feed_dict_gpu = {"image": gpu_img, "label": gpu_label} feed_dict_gpu = {"image": gpu_img, "label": gpu_label}
feed_dict_cpu = {"image": cpu_img, "label": cpu_label} feed_dict_cpu = {"image": cpu_img, "label": cpu_label}
......
...@@ -32,7 +32,7 @@ class TestResnetBase(TestParallelExecutorBase): ...@@ -32,7 +32,7 @@ class TestResnetBase(TestParallelExecutorBase):
seresnext_net.model, seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda), feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda), iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(), batch_size=seresnext_net.batch_size(use_cuda),
use_cuda=use_cuda, use_cuda=use_cuda,
use_reduce=False, use_reduce=False,
optimizer=seresnext_net.optimizer) optimizer=seresnext_net.optimizer)
...@@ -41,7 +41,7 @@ class TestResnetBase(TestParallelExecutorBase): ...@@ -41,7 +41,7 @@ class TestResnetBase(TestParallelExecutorBase):
seresnext_net.model, seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda), feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda), iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(), batch_size=seresnext_net.batch_size(use_cuda),
use_cuda=use_cuda) use_cuda=use_cuda)
if compare_seperately: if compare_seperately:
......
...@@ -28,7 +28,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase): ...@@ -28,7 +28,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
seresnext_net.model, seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda), feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda), iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(), batch_size=seresnext_net.batch_size(use_cuda),
use_cuda=use_cuda, use_cuda=use_cuda,
use_reduce=False, use_reduce=False,
optimizer=seresnext_net.optimizer) optimizer=seresnext_net.optimizer)
...@@ -36,7 +36,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase): ...@@ -36,7 +36,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
seresnext_net.model, seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda), feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda), iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(), batch_size=seresnext_net.batch_size(use_cuda),
use_cuda=use_cuda, use_cuda=use_cuda,
use_reduce=True, use_reduce=True,
optimizer=seresnext_net.optimizer) optimizer=seresnext_net.optimizer)
...@@ -53,7 +53,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase): ...@@ -53,7 +53,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
seresnext_net.model, seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda), feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda), iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(), batch_size=seresnext_net.batch_size(use_cuda),
use_cuda=use_cuda, use_cuda=use_cuda,
use_reduce=False, use_reduce=False,
optimizer=seresnext_net.optimizer, optimizer=seresnext_net.optimizer,
...@@ -63,7 +63,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase): ...@@ -63,7 +63,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
seresnext_net.model, seresnext_net.model,
feed_dict=seresnext_net.feed_dict(use_cuda), feed_dict=seresnext_net.feed_dict(use_cuda),
iter=seresnext_net.iter(use_cuda), iter=seresnext_net.iter(use_cuda),
batch_size=seresnext_net.batch_size(), batch_size=seresnext_net.batch_size(use_cuda),
use_cuda=use_cuda, use_cuda=use_cuda,
use_reduce=True, use_reduce=True,
optimizer=seresnext_net.optimizer, optimizer=seresnext_net.optimizer,
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册