## @package workspace
# Module caffe2.python.workspace
import collections
import contextlib
from google.protobuf.message import Message
from multiprocessing import Process
import os
from collections import defaultdict
import logging
import numpy as np
from past.builtins import basestring
import shutil
import socket
import tempfile
from caffe2.proto import caffe2_pb2
from caffe2.python import scope, utils
from caffe2.python.lazy import TriggerLazyImport
import caffe2.python._import_c_extension as C
logger = logging.getLogger(__name__)
Blobs = C.blobs
ResetBlob = C.reset_blob
CreateBlob = C.create_blob
CurrentWorkspace = C.current_workspace
DeserializeBlob = C.deserialize_blob
GlobalInit = C.global_init
HasBlob = C.has_blob
RegisteredOperators = C.registered_operators
SerializeBlob = C.serialize_blob
SwitchWorkspace = C.switch_workspace
RootFolder = C.root_folder
Workspaces = C.workspaces
BenchmarkNet = C.benchmark_net
BenchmarkNetOnce = C.benchmark_net_once
GetStats = C.get_stats
CreateOfflineTensor = C.create_offline_tensor
operator_tracebacks = defaultdict(dict)
is_asan = C.is_asan
has_cuda_support = C.has_cuda_support
has_hip_support = C.has_hip_support
has_gpu_support = C.has_gpu_support
if has_cuda_support:
GpuDeviceType = caffe2_pb2.CUDA
NumCudaDevices = C.num_cuda_devices
# This is a duplicate of NumCudaDevices. Remove
# NumCudaDevices once replaced everywhere in the code
NumGpuDevices = C.num_cuda_devices
GetCUDAVersion = C.get_cuda_version
GetCuDNNVersion = C.get_cudnn_version
def GetGpuPeerAccessPattern():
return np.asarray(C.get_cuda_peer_access_pattern())
GetDeviceProperties = C.get_device_properties
GetGPUMemoryInfo = C.get_gpu_memory_info
else:
NumCudaDevices = lambda: 0 # noqa
GetCUDAVersion = lambda: 0 # noqa
GetCuDNNVersion = lambda: 0 # noqa
if has_hip_support:
GpuDeviceType = caffe2_pb2.HIP
NumGpuDevices = C.num_hip_devices
GetHIPVersion = C.get_hip_version
def GetGpuPeerAccessPattern():
return np.asarray(C.get_hip_peer_access_pattern())
GetDeviceProperties = C.get_device_properties
GetGPUMemoryInfo = C.get_gpu_memory_info
if not has_gpu_support:
# setting cuda as the default GpuDeviceType as some tests
# like core, scope tests use GpuDeviceType even without gpu support
GpuDeviceType = caffe2_pb2.CUDA
NumGpuDevices = lambda: 0 # noqa
GetDeviceProperties = lambda x: None # noqa
GetGpuPeerAccessPattern = lambda: np.array([]) # noqa
GetGPUMemoryInfo = lambda: None # noqa
IsNUMAEnabled = C.is_numa_enabled
GetNumNUMANodes = C.get_num_numa_nodes
GetBlobNUMANode = C.get_blob_numa_node
GetBlobSizeBytes = C.get_blob_size_bytes
def FillRandomNetworkInputs(net, input_dims, input_types):
C.fill_random_network_inputs(net.Proto().SerializeToString(), input_dims, input_types)
def _GetFreeFlaskPort():
"""Get a free flask port."""
# We will prefer to use 5000. If not, we will then pick a random port.
sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
result = sock.connect_ex(('127.0.0.1', 5000))
if result == 0:
return 5000
else:
s = socket.socket()
s.bind(('', 0))
port = s.getsockname()[1]
s.close()
# Race condition: between the interval we close the socket and actually
# start a mint process, another process might have occupied the port. We
# don't do much here as this is mostly for convenience in research
# rather than 24x7 service.
return port
def StartMint(root_folder=None, port=None):
"""Start a mint instance.
TODO(Yangqing): this does not work well under ipython yet. According to
https://github.com/ipython/ipython/issues/5862
writing up some fix is a todo item.
"""
from caffe2.python.mint import app
if root_folder is None:
# Get the root folder from the current workspace
root_folder = C.root_folder()
if port is None:
port = _GetFreeFlaskPort()
process = Process(
target=app.main,
args=(
['-p', str(port), '-r', root_folder],
)
)
process.start()
print('Mint running at http://{}:{}'.format(socket.getfqdn(), port))
return process
def StringifyProto(obj):
"""Stringify a protocol buffer object.
Inputs:
obj: a protocol buffer object, or a Pycaffe2 object that has a Proto()
function.
Outputs:
string: the output protobuf string.
Raises:
AttributeError: if the passed in object does not have the right attribute.
"""
if isinstance(obj, basestring):
return obj
else:
if isinstance(obj, Message):
# First, see if this object is a protocol buffer, which we can
# simply serialize with the SerializeToString() call.
return obj.SerializeToString()
elif hasattr(obj, 'Proto'):
return obj.Proto().SerializeToString()
else:
raise ValueError("Unexpected argument to StringifyProto of type " +
type(obj).__name__)
def ResetWorkspace(root_folder=None):
if root_folder is None:
# Reset the workspace, but keep the current root folder setting.
return C.reset_workspace(C.root_folder())
else:
if not os.path.exists(root_folder):
os.makedirs(root_folder)
return C.reset_workspace(root_folder)
def CreateNet(net, overwrite=False, input_blobs=None):
TriggerLazyImport()
if input_blobs is None:
input_blobs = []
for input_blob in input_blobs:
C.create_blob(input_blob)
return CallWithExceptionIntercept(
C.create_net,
C.Workspace.current._last_failed_op_net_position,
GetNetName(net),
StringifyProto(net), overwrite,
)
def Predictor(init_net, predict_net):
return C.Predictor(StringifyProto(init_net), StringifyProto(predict_net))
def GetOperatorCost(operator, blobs):
return C.get_operator_cost(StringifyProto(operator), blobs)
def RunOperatorOnce(operator):
return C.run_operator_once(StringifyProto(operator))
def RunOperatorMultiple(operator, num_runs):
return C.run_operator_multiple(StringifyProto(operator), num_runs)
def RunOperatorsOnce(operators):
for op in operators:
success = RunOperatorOnce(op)
if not success:
return False
return True
def ClearGlobalNetObserver():
return C.clear_global_net_observer()
def CallWithExceptionIntercept(func, op_id_fetcher, net_name, *args, **kwargs):
try:
return func(*args, **kwargs)
except Exception:
op_id = op_id_fetcher()
net_tracebacks = operator_tracebacks.get(net_name, None)
logger.warning(
'Original python traceback for operator `{}` in network '
'`{}` in exception above (most recent call last):'.format(
op_id, net_name))
if net_tracebacks and op_id in net_tracebacks:
tb = net_tracebacks[op_id]
for line in reversed(tb):
logger.warning(' File "{}", line {}, in {}'.format(
line[0], line[1], line[2]))
raise
def RunNetOnce(net):
return CallWithExceptionIntercept(
C.run_net_once,
C.Workspace.current._last_failed_op_net_position,
GetNetName(net),
StringifyProto(net),
)
def RunNet(name, num_iter=1, allow_fail=False):
"""Runs a given net.
Inputs:
name: the name of the net, or a reference to the net.
num_iter: number of iterations to run
allow_fail: if True, does not assert on net exec failure but returns False
Returns:
True or an exception.
"""
return CallWithExceptionIntercept(
C.run_net,
C.Workspace.current._last_failed_op_net_position,
GetNetName(name),
StringifyNetName(name), num_iter, allow_fail,
)
def RunPlan(plan_or_step):
# TODO(jiayq): refactor core.py/workspace.py to avoid circular deps
import caffe2.python.core as core
if isinstance(plan_or_step, core.ExecutionStep):
plan_or_step = core.Plan(plan_or_step)
return C.run_plan(StringifyProto(plan_or_step))
def RunPlanInBackground(plan_or_step):
# TODO(jiayq): refactor core.py/workspace.py to avoid circular deps
import caffe2.python.core as core
if isinstance(plan_or_step, core.ExecutionStep):
plan_or_step = core.Plan(plan_or_step)
return C.run_plan_in_background(StringifyProto(plan_or_step))
def InferShapesAndTypes(nets, blob_dimensions=None, nets_proto=False,
blob_types=None):
"""Infers the shapes and types for the specified nets.
Inputs:
nets: the list of nets
blob_dimensions (optional): a dictionary of blobs and their dimensions.
If not specified, the workspace blobs are used.
nets_proto (optional): a boolean flag indicating whether the protobuffer
representation is passed to the routine.
Returns:
A tuple of (shapes, types) dictionaries keyed by blob name.
"""
if nets_proto:
net_protos = [StringifyProto(n) for n in nets]
else:
net_protos = [StringifyProto(n.Proto()) for n in nets]
if blob_dimensions is None:
assert blob_types is None
blobdesc_prototxt = C.infer_shapes_and_types_from_workspace(net_protos)
elif blob_types is None:
blobdesc_prototxt = C.infer_shapes_and_types_from_map(
net_protos, blob_dimensions
)
else:
blobdesc_prototxt = C.infer_shapes_and_types_from_map(
net_protos, blob_dimensions, blob_types
)
blobdesc_proto = caffe2_pb2.TensorShapes()
blobdesc_proto.ParseFromString(blobdesc_prototxt)
shapes = {}
types = {}
for ts in blobdesc_proto.shapes:
if not ts.unknown_shape:
shapes[ts.name] = list(ts.dims)
types[ts.name] = ts.data_type
return (shapes, types)
def _StringifyName(name, expected_type):
if isinstance(name, basestring):
return name
assert type(name).__name__ == expected_type, \
"Expected a string or %s" % expected_type
return str(name)
def StringifyBlobName(name):
return _StringifyName(name, "BlobReference")
def StringifyNetName(name):
return _StringifyName(name, "Net")
def GetNetName(net):
if isinstance(net, basestring):
return net
if type(net).__name__ == "Net" or type(net).__name__ == "NetWithShapeInference":
return net.Name()
if isinstance(net, caffe2_pb2.NetDef):
return net.name
raise Exception("Not a Net object: {}".format(str(net)))
def FeedBlob(name, arr, device_option=None):
Loading ...