OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-07 12:21:27 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

quick fix inplace blob bug

Browse Source 
Summary: fixing the case where the init net will initialize same blob twice. I made an exception by allowing inplace blob among ops if the blob keeps on the same device. This should fix this problem in a generalized way as most of our training is only on CPU now.

Reviewed By: dzhulgakov

Differential Revision: D5450564

fbshipit-source-id: 525c4c9a2e5216a70dbd1229da2d9f8a58b89e47
This commit is contained in:
Yiming Wu 2017-07-23 02:14:03 -07:00 committed by Facebook Github Bot
parent 920c553ac0
commit b51e0ec0c2
2 changed files with 19 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
caffe2/python/core.py
View File

@ -2090,6 +2090,11 @@ def copy_func_between_devices(src, dst):
def device_equal(src, dst):
'''
We are using this fucntion instead of == operator because optional-value
comparison between empty device_options and {device_type:0, cuda_gpu_id:0}
returns not equal in some cases.
'''
return src.device_type == dst.device_type and src.cuda_gpu_id == dst.cuda_gpu_id
@ -2182,12 +2187,17 @@ def InjectCrossDeviceCopies(net, blob_to_device=None):
# Enforcing no reuse blob between operators. In-place blob usage in an
# op is allowed. This is based on the assumption that in-place op has
# same device info
for out_blob in op.output:
if out_blob in blob_to_device and out_blob not in op.input:
for out_blob, device in zip(op.output, output_dev):
if out_blob in blob_to_device and (
out_blob not in op.input and
not device_equal(blob_to_device[out_blob], device)
):
raise RuntimeError(
"In-place blob: {} is not supported between operators. "
"Failed op:\n {}".
format(out_blob, op)
"In-place blob: {} is not supported between operators "
"with different device option previous:{} now: {}. "
"Failed op:\n {}".format(
out_blob, blob_to_device[out_blob], device, op
)
)
blob_to_device.update({o: d for d, o in zip(output_dev, op.output)})
new_op = caffe2_pb2.OperatorDef()

4
caffe2/python/core_test.py
View File

@ -787,6 +787,10 @@ external_input: "data"
self.assertEqual(op.input[0], 'param_cuda_1')
net.Relu('nonsense_input', 'moment')
# should not raise inplace error
core.InjectCrossDeviceCopies(net)
with core.DeviceScope(device_option):
net.Relu('nonsense_input_gpu', 'moment')
with self.assertRaises(RuntimeError):
core.InjectCrossDeviceCopies(net)