def _lazy_init():
global _initialized, _cudart, _original_pid, _queued_calls
if _initialized:
return
if _in_bad_fork:
from sys import version_info
if version_info < (3, 4):
msg = ("To use CUDA with multiprocessing, you must use Python "
"3.4+ and the 'spawn' start method")
else:
msg = ("To use CUDA with multiprocessing, you must use the "
"'spawn' start method")
raise RuntimeError("Cannot re-initialize CUDA in forked subprocess. " +
msg)
_check_driver()
torch._C._cuda_init()
_cudart = _load_cudart()
_cudart.cudaGetErrorName.restype = ctypes.c_char_p
_cudart.cudaGetErrorString.restype = ctypes.c_char_p
_original_pid = os.getpid()
_initialized = True
# Important to do this after _initialized, since some queued calls
# may themselves call _lazy_init()
for queued_call, orig_traceback in _queued_calls:
try:
queued_call()
except Exception as e:
msg = (
"CUDA call failed lazily at initialization with error: {}\n\n"
"CUDA call was originally invoked at:\n\n{}").format(
str(e), orig_traceback)
raise_from(DeferredCudaCallError(msg), e)
def __call__(cls, *args, **kwargs):
# NOTE: this is called whenever an instance of this class is created
# The super call below will call __new__ and __init__, and we will
# patch things later.
try:
obj = super(CompiledModuleMeta,
cls).__call__(*args, **kwargs)
except TypeError as e:
# If this fails here, the user probably didn't use this as a class decorator
if "super" in str(e):
raise_from(
TypeError(
"torch.jit.compile must be used as a class decorator; "
"using it on an already defined class is not valid."
"\n\nOriginal error: {}".format(str(e))),
e)
else:
raise
compiled_fn = torch._C.CompiledFunction(
nderivs, optimize, obj.forward, arg.__name__)
compiled_fn.enabled = enabled
obj.compiled_fn = compiled_fn
obj.forward = lambda *args: compiled_fn(
args, list(obj.parameters()))
obj.has_trace_for = lambda *args: compiled_fn.has_trace_for(
args, list(obj.parameters()))
return obj
def _lazy_init():
global _initialized, _cudart, _original_pid, _queued_calls
if _initialized:
return
if _in_bad_fork:
from sys import version_info
if version_info < (3, 4):
msg = ("To use CUDA with multiprocessing, you must use Python "
"3.4+ and the 'spawn' start method")
else:
msg = ("To use CUDA with multiprocessing, you must use the "
"'spawn' start method")
raise RuntimeError(
"Cannot re-initialize CUDA in forked subprocess. " + msg)
_check_driver()
torch._C._cuda_init()
_cudart = _load_cudart()
_cudart.cudaGetErrorName.restype = ctypes.c_char_p
_cudart.cudaGetErrorString.restype = ctypes.c_char_p
_original_pid = os.getpid()
_initialized = True
# Important to do this after _initialized, since some queued calls
# may themselves call _lazy_init()
for queued_call, orig_traceback in _queued_calls:
try:
queued_call()
except Exception as e:
msg = ("CUDA call failed lazily at initialization with error: {}\n\n"
"CUDA call was originally invoked at:\n\n{}").format(str(e), orig_traceback)
raise_from(DeferredCudaCallError(msg), e)
def exception_catcher(self, batch):
try:
yield
except Exception as e:
wrapper = EngineException(
epoch=self._epochs,
iteration=self._iterations,
batch=self.batch_id_fn(batch) if self.batch_id_fn else batch,
cause=e,
)
raise_from(wrapper, e)
def _parse(self, value, function, fmt):
"""
Parse a string into a value, and format a nice ValueError if it fails.
Returns `function(value)`.
Any `ValueError` raised is catched and a new `ValueError` is raised
with message `fmt.format(e)`, where `e` is the caught `ValueError`.
"""
try:
return function(value)
except ValueError as e:
raise_from(ValueError(fmt.format(e)), None)
def _read_annotations(self, csv_reader, classes):
result = {}
for line, row in enumerate(csv_reader):
line += 1
try:
img_file, x1, y1, x2, y2, class_name = row[:6]
except ValueError:
raise_from(
ValueError(
'line {}: format should be \'img_file,x1,y1,x2,y2,class_name\' or \'img_file,,,,,\''
.format(line)), None)
if img_file not in result:
result[img_file] = []
# If a row contains only an image path, it's an image without annotations.
if (x1, y1, x2, y2, class_name) == ('', '', '', '', ''):
continue
x1 = self._parse(x1, int,
'line {}: malformed x1: {{}}'.format(line))
y1 = self._parse(y1, int,
'line {}: malformed y1: {{}}'.format(line))
x2 = self._parse(x2, int,
'line {}: malformed x2: {{}}'.format(line))
y2 = self._parse(y2, int,
'line {}: malformed y2: {{}}'.format(line))
# Check that the bounding box is valid.
if x2 <= x1:
raise ValueError(
'line {}: x2 ({}) must be higher than x1 ({})'.format(
line, x2, x1))
if y2 <= y1:
raise ValueError(
'line {}: y2 ({}) must be higher than y1 ({})'.format(
line, y2, y1))
# check if the current class name is correctly present
if class_name not in classes:
raise ValueError(
'line {}: unknown class name: \'{}\' (classes: {})'.format(
line, class_name, classes))
result[img_file].append({
'x1': x1,
'x2': x2,
'y1': y1,
'y2': y2,
'class': class_name
})
return result
def _lazy_init():
global _initialized, _queued_calls
if is_initialized() or hasattr(_tls, 'is_initializing'):
return
with _initialization_lock:
# We be double-checked locking, boys! This is OK because
# the above test was GIL protected anyway. The inner test
# is for when a thread blocked on some other thread which was
# doing the initialization; when they get the lock, they will
# find there is nothing left to do.
if is_initialized():
return
# It is important to prevent other threads from entering _lazy_init
# immediately, while we are still guaranteed to have the GIL, because some
# of the C calls we make below will release the GIL
if _is_in_bad_fork():
from sys import version_info
if version_info < (3, 4):
msg = ("To use CUDA with multiprocessing, you must use Python "
"3.4+ and the 'spawn' start method")
else:
msg = ("To use CUDA with multiprocessing, you must use the "
"'spawn' start method")
raise RuntimeError(
"Cannot re-initialize CUDA in forked subprocess. " + msg)
if not hasattr(torch._C, '_cuda_getDeviceCount'):
raise AssertionError("Torch not compiled with CUDA enabled")
if _cudart is None:
raise AssertionError(
"libcudart functions unavailable. It looks like you have a broken build?"
)
# This function throws if there's a driver initialization error, no GPUs
# are found or any other error occurs
torch._C._cuda_init()
# Some of the queued calls may reentrantly call _lazy_init();
# we need to just return without initializing in that case.
# However, we must not let any *other* threads in!
_tls.is_initializing = True
try:
for queued_call, orig_traceback in _queued_calls:
try:
queued_call()
except Exception as e:
msg = (
"CUDA call failed lazily at initialization with error: {}\n\n"
"CUDA call was originally invoked at:\n\n{}").format(
str(e), orig_traceback)
raise_from(DeferredCudaCallError(msg), e)
finally:
delattr(_tls, 'is_initializing')
_initialized = True
def _lazy_init():
global _initialized, _cudart, _original_pid, _queued_calls
if _initialized or hasattr(_tls, 'is_initializing'):
return
with _initialization_lock:
# We be double-checked locking, boys! This is OK because
# the above test was GIL protected anyway. The inner test
# is for when a thread blocked on some other thread which was
# doing the initialization; when they get the lock, they will
# find there is nothing left to do.
if _initialized:
return
# It is important to prevent other threads from entering _lazy_init
# immediately, while we are still guaranteed to have the GIL, because some
# of the C calls we make below will release the GIL
if _in_bad_fork:
from sys import version_info
if version_info < (3, 4):
msg = ("To use CUDA with multiprocessing, you must use Python "
"3.4+ and the 'spawn' start method")
else:
msg = ("To use CUDA with multiprocessing, you must use the "
"'spawn' start method")
raise RuntimeError(
"Cannot re-initialize CUDA in forked subprocess. " + msg)
_check_driver()
torch._C._cuda_init()
_cudart = _load_cudart()
_cudart.cudaGetErrorName.restype = ctypes.c_char_p
_cudart.cudaGetErrorString.restype = ctypes.c_char_p
_original_pid = os.getpid()
# Some of the queued calls may reentrantly call _lazy_init();
# we need to just return without initializing in that case.
# However, we must not let any *other* threads in!
_tls.is_initializing = True
try:
for queued_call, orig_traceback in _queued_calls:
try:
queued_call()
except Exception as e:
msg = (
"CUDA call failed lazily at initialization with error: {}\n\n"
"CUDA call was originally invoked at:\n\n{}").format(
str(e), orig_traceback)
raise_from(DeferredCudaCallError(msg), e)
finally:
delattr(_tls, 'is_initializing')
_initialized = True
def __new_init(self, *args, **kwargs):
try:
# __old_init is assumed to handle super call
self.__old_init(*args, **kwargs)
except TypeError as e:
# If this fails here, the user probably didn't use this as a class
# decorator
if "super" in str(e):
raise_from(TypeError("torch.jit.compile must be used as a class decorator; "
"using it on an already defined class is not valid."
"\n\nOriginal error: {}".format(str(e))), e)
else:
raise
model_name = self.__model_name if self.__model_name else type(self).__name__
self.__name = "jit_{}_{}".format(model_name, _CompiledMixin.__next_id)
_CompiledMixin.__next_id += 1
self.__ktrace_cache = {}
self.__next_ktrace_id = 0
def __init__(self, *args, **kwargs):
torch._C.CompiledFunction.__init__(self,
nderivs, optimize, enabled,
self.forward,
arg.__name__)
try:
old_init(self, *args, **kwargs)
except TypeError as e:
# If this fails here, the user probably didn't use this as a class decorator
if "super" in str(e):
raise_from(TypeError("torch.jit.compile must be used as a class decorator; "
"using it on an already defined class is not valid."
"\n\nOriginal error: {}".format(str(e))), e)
else:
raise
# NOTE: This can't be done in CompiledFunction constructor,
# because self.parameters() isn't well defined by then
# (Module constructor hasn't run yet).
self.set_captured_vars(list(self.parameters()))
def __init__(self, *args, **kwargs):
torch._C.CompiledFunction.__init__(self,
nderivs, optimize, enabled,
self.forward,
arg.__name__)
try:
old_init(self, *args, **kwargs)
except TypeError as e:
# If this fails here, the user probably didn't use this as a class decorator
if "super" in str(e):
raise_from(TypeError("torch.jit.compile must be used as a class decorator; "
"using it on an already defined class is not valid."
"\n\nOriginal error: {}".format(str(e))), e)
else:
raise
# NOTE: This can't be done in CompiledFunction constructor,
# because self.parameters() isn't well defined by then
# (Module constructor hasn't run yet).
self.set_captured_vars(list(self.parameters()))
def __call__(cls, *args, **kwargs):
# NOTE: this is called whenever an instance of this class is created
# The super call below will call __new__ and __init__, and we will
# patch things later.
try:
obj = super(CompiledModuleMeta, cls).__call__(*args, **kwargs)
except TypeError as e:
# If this fails here, the user probably didn't use this as a class decorator
if "super" in str(e):
raise_from(TypeError("torch.jit.compile must be used as a class decorator; "
"using it on an already defined class is not valid."
"\n\nOriginal error: {}".format(str(e))), e)
else:
raise
compiled_fn = torch._C.CompiledFunction(nderivs, optimize,
obj.forward,
arg.__name__)
compiled_fn.enabled = enabled
obj.compiled_fn = compiled_fn
obj.forward = lambda *args: compiled_fn(args, list(obj.parameters()))
obj.has_trace_for = lambda *args: compiled_fn.has_trace_for(args, list(obj.parameters()))
return obj
def _lazy_init():
global _initialized, _original_pid, _queued_calls
if _initialized or hasattr(_tls, 'is_initializing'):
return
with _initialization_lock:
# We be double-checked locking, boys! This is OK because
# the above test was GIL protected anyway. The inner test
# is for when a thread blocked on some other thread which was
# doing the initialization; when they get the lock, they will
# find there is nothing left to do.
if _initialized:
return
# It is important to prevent other threads from entering _lazy_init
# immediately, while we are still guaranteed to have the GIL, because some
# of the C calls we make below will release the GIL
if _in_bad_fork:
raise RuntimeError(
"Cannot re-initialize HammerBlade in forked subprocess.")
# TODO: enable this
# _check_driver()
torch._C._hammerblade_init()
_original_pid = os.getpid()
# Some of the queued calls may reentrantly call _lazy_init();
# we need to just return without initializing in that case.
# However, we must not let any *other* threads in!
_tls.is_initializing = True
try:
for queued_call, orig_traceback in _queued_calls:
try:
queued_call()
except Exception as e:
msg = ("HammerBlade call failed lazily at initialization with error: {}\n\n"
"HammerBlade call was originally invoked at:\n\n{}").format(str(e), orig_traceback)
raise_from(DeferredHammerBladeCallError(msg), e)
finally:
delattr(_tls, 'is_initializing')
_initialized = True