Python caffe2.python.workspace.RunOperatorOnce() Examples
The following are 30
code examples of caffe2.python.workspace.RunOperatorOnce().
You can vote up the ones you like or vote down the ones you don't like,
and go to the original project or source file by following the links above each example.
You may also want to check out all available functions/classes of the module
caffe2.python.workspace
, or try the search function
.
.
Example #1
| Source File: test_spatial_narrow_as_op.py From Detectron with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #2
| Source File: detector.py From Clustered-Object-Detection-in-Aerial-Image with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #3
| Source File: test_spatial_narrow_as_op.py From seg_every_thing with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #4
| Source File: test_spatial_narrow_as_op.py From Clustered-Object-Detection-in-Aerial-Image with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #5
| Source File: test_spatial_narrow_as_op.py From KL-Loss with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #6
| Source File: test_spatial_narrow_as_op.py From masktextspotter.caffe2 with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #7
| Source File: test_batch_permutation_op.py From masktextspotter.caffe2 with Apache License 2.0 | 6 votes |
|
def _run_op_test(self, X, I, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
workspace.FeedBlob('X', X)
workspace.FeedBlob('I', I)
workspace.RunOperatorOnce(op)
Y = workspace.FetchBlob('Y')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.1,
threshold=0.001,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
self.assertTrue(res, 'Grad check failed')
Y_ref = X[I]
np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08)
Example #8
| Source File: loader.py From masktextspotter.caffe2 with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #9
| Source File: mix_loader.py From masktextspotter.caffe2 with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #10
| Source File: detector.py From masktextspotter.caffe2 with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #11
| Source File: loader.py From Detectron-Cascade-RCNN with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #12
| Source File: detector.py From Detectron-Cascade-RCNN with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #13
| Source File: run_add5_op.py From tutorials with Apache License 2.0 | 6 votes |
|
def run_add5_and_add5gradient_op(device):
# clear the workspace before running the operator
workspace.ResetWorkspace()
add5 = core.CreateOperator("Add5",
["X"],
["Y"],
device_option=device)
print("==> Running Add5 op:")
workspace.FeedBlob("X", (np.random.rand(5, 5)), device_option=device)
print("Input of Add5: ", workspace.FetchBlob("X"))
workspace.RunOperatorOnce(add5)
print("Output of Add5: ", workspace.FetchBlob("Y"))
print("\n\n==> Running Add5Gradient op:")
print("Input of Add5Gradient: ", workspace.FetchBlob("Y"))
add5gradient = core.CreateOperator("Add5Gradient",
["Y"],
["Z"],
device_option=device)
workspace.RunOperatorOnce(add5gradient)
print("Output of Add5Gradient: ", workspace.FetchBlob("Z"))
Example #14
| Source File: loader.py From Detectron with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #15
| Source File: detector.py From Detectron with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #16
| Source File: test_spatial_narrow_as_op.py From Detectron-Cascade-RCNN with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #17
| Source File: loader.py From Detectron-DA-Faster-RCNN with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #18
| Source File: detector.py From Detectron-DA-Faster-RCNN with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #19
| Source File: test_spatial_narrow_as_op.py From Detectron-DA-Faster-RCNN with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #20
| Source File: loader.py From CBNet with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #21
| Source File: detector.py From CBNet with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #22
| Source File: test_spatial_narrow_as_op.py From CBNet with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #23
| Source File: test_spatial_narrow_as_op.py From NucleiDetectron with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #24
| Source File: test_batch_permutation_op.py From NucleiDetectron with Apache License 2.0 | 6 votes |
|
def _run_op_test(self, X, I, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
workspace.FeedBlob('X', X)
workspace.FeedBlob('I', I)
workspace.RunOperatorOnce(op)
Y = workspace.FetchBlob('Y')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.1,
threshold=0.001,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
self.assertTrue(res, 'Grad check failed')
Y_ref = X[I]
np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08)
Example #25
| Source File: loader.py From NucleiDetectron with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
dev = c2_utils.CudaDevice(gpu_id)
queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug(
'enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t)
)
t = time.time()
op = core.CreateOperator(
'SafeEnqueueBlobs', [queue_name] + blob_names,
blob_names + [queue_name + '_enqueue_status'],
device_option=dev
)
workspace.RunOperatorOnce(op)
logger.debug(
'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t)
)
Example #26
| Source File: detector.py From NucleiDetectron with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is
the stochastic gradient. Since V is not defined independently of the
learning rate (as it should ideally be), when the learning rate is
changed we should scale the update history V in order to make it
compatible in scale with lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with c2_utils.CudaScope(i):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
Example #27
| Source File: test_spatial_narrow_as_op.py From DetectAndTrack with Apache License 2.0 | 6 votes |
|
def _run_test(self, A, B, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
workspace.FeedBlob('A', A)
workspace.FeedBlob('B', B)
workspace.RunOperatorOnce(op)
C = workspace.FetchBlob('C')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.005,
threshold=0.005,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
self.assertTrue(res, 'Grad check failed')
dims = C.shape
C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08)
Example #28
| Source File: test_batch_permutation_op.py From DetectAndTrack with Apache License 2.0 | 6 votes |
|
def _run_op_test(self, X, I, check_grad=False):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
workspace.FeedBlob('X', X)
workspace.FeedBlob('I', I)
workspace.RunOperatorOnce(op)
Y = workspace.FetchBlob('Y')
if check_grad:
gc = gradient_checker.GradientChecker(
stepsize=0.1,
threshold=0.001,
device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
)
res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
self.assertTrue(res, 'Grad check failed')
Y_ref = X[I]
np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08)
Example #29
| Source File: loader.py From DetectAndTrack with Apache License 2.0 | 6 votes |
|
def enqueue_blobs(self, gpu_id, blob_names, blobs):
"""Put a mini-batch on a BlobsQueue."""
assert len(blob_names) == len(blobs)
t = time.time()
blob_names = [
'gpu_{}/{}'.format(gpu_id, blob_name) for blob_name in blob_names]
dev = core.DeviceOption(caffe2_pb2.CUDA, gpu_id)
for (blob_name, blob) in zip(blob_names, blobs):
workspace.FeedBlob(blob_name, blob, device_option=dev)
logger.debug('enqueue_blobs {}: workspace.FeedBlob: {}'.
format(gpu_id, time.time() - t))
t = time.time()
op = core.CreateOperator(
"EnqueueBlobs",
['gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)] + blob_names,
blob_names, device_option=dev)
workspace.RunOperatorOnce(op)
logger.debug('enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
format(gpu_id, time.time() - t))
Example #30
| Source File: detector.py From DetectAndTrack with Apache License 2.0 | 6 votes |
|
def _CorrectMomentum(self, correction):
"""The MomentumSGDUpdate op implements the update V as
V := mu * V + lr * grad,
where mu is the momentum factor, lr is the learning rate, and grad is the
stochastic gradient. Since V is not defined independently of the learning
rate (as it should ideally be), when the learning rate is changed we should
scale the update history V in order to make it compatible in scale with
lr * grad.
"""
logger.info(
'Scaling update history by {:.6f} (new lr / old lr)'.
format(correction))
for i in range(cfg.NUM_GPUS):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, i)):
for param in self.TrainableParams(gpu_id=i):
op = core.CreateOperator(
'Scale', [param + '_momentum'], [param + '_momentum'],
scale=correction)
workspace.RunOperatorOnce(op)
