def get_optimized_symbol(executor):
"""
Take an executor's underlying symbol graph and return its generated optimized version.
Parameters
----------
executor :
An executor for which you want to see an optimized symbol. Getting an optimized symbol
is useful to compare and verify the work TensorRT has done against a legacy behaviour.
Returns
-------
symbol : nnvm::Symbol
The nnvm symbol optimized.
"""
handle = SymbolHandle()
try:
check_call(
_LIB.MXExecutorGetOptimizedSymbol(executor.handle,
ctypes.byref(handle)))
result = Symbol(handle=handle)
return result
except MXNetError:
logging.error(
'Error while trying to fetch TRT optimized symbol for graph. Please ensure '
'build was compiled with MXNET_USE_TENSORRT enabled.')
raise
def _check_subgraph_exe3(sym, subgraph_backend, op_names):
"""Use the partitioned sym to bind an executor and compare the outputs
with those of the original executor"""
out = SymbolHandle()
check_call(_LIB.MXBuildSubgraphByOpNames(sym.handle, c_str(subgraph_backend), mx_uint(len(op_names)),
c_str_array(op_names), ctypes.byref(out)))
partitioned_sym = Symbol(out)
input_names = sym.list_inputs()
arg_names = sym.list_arguments()
aux_names = sym.list_auxiliary_states()
assert partitioned_sym.list_inputs() == input_names
assert partitioned_sym.list_arguments() == arg_names
assert partitioned_sym.list_auxiliary_states() == aux_names
arg_shapes, _, aux_shapes = sym.infer_shape()
arg_array = [mx.nd.random.uniform(shape=shape) for shape in arg_shapes]
aux_array = [mx.nd.random.uniform(shape=shape) for shape in aux_shapes]
exe = sym.bind(ctx=mx.current_context(), args=arg_array, aux_states=aux_array, grad_req='null')
partitioned_exe = partitioned_sym.bind(ctx=mx.current_context(), args=arg_array,
aux_states=aux_array, grad_req='null')
exe.forward()
partitioned_exe.forward()
assert len(exe.outputs) == len(partitioned_exe.outputs)
for i in range(len(exe.outputs)):
assert_almost_equal((exe.outputs[i] - partitioned_exe.outputs[i]).abs().sum().asnumpy(),
np.zeros(shape=(1,)))
def _check_subgraph_exe1(sym, subgraph_backend, op_names):
"""Use the partitioned sym to simple_bind an executor and compare the outputs
with those of the original executor"""
out = SymbolHandle()
check_call(_LIB.MXBuildSubgraphByOpNames(sym.handle, c_str(subgraph_backend), mx_uint(len(op_names)),
c_str_array(op_names), ctypes.byref(out)))
partitioned_sym = Symbol(out)
assert partitioned_sym.list_inputs() == sym.list_inputs()
assert partitioned_sym.list_arguments() == sym.list_arguments()
assert partitioned_sym.list_auxiliary_states() == sym.list_auxiliary_states()
exe = sym.simple_bind(ctx=mx.current_context(), grad_req='null')
partitioned_exe = partitioned_sym.simple_bind(ctx=mx.current_context(), grad_req='null')
input_names = sym.list_inputs()
for name in input_names:
if name in exe.arg_dict:
exe.arg_dict[name][:] = mx.nd.random.uniform(shape=exe.arg_dict[name].shape)
partitioned_exe.arg_dict[name][:] = exe.arg_dict[name]
else:
assert name in exe.aux_dict
exe.aux_dict[name][:] = mx.nd.random.uniform(shape=exe.aux_dict[name].shape)
partitioned_exe.aux_dict[name][:] = exe.aux_dict[name]
exe.forward()
partitioned_exe.forward()
assert len(exe.outputs) == len(partitioned_exe.outputs)
for i in range(len(exe.outputs)):
assert_almost_equal((exe.outputs[i] - partitioned_exe.outputs[i]).abs().sum().asnumpy(),
np.zeros(shape=(1,)))
def _quantize_symbol(sym,
excluded_symbols=None,
offline_params=None,
quantized_dtype='int8'):
"""Given a symbol object representing a neural network of data type FP32,
quantize it into a INT8 network.
Parameters
----------
sym : Symbol
FP32 neural network symbol.
excluded_sym_names : list of strings
A list of strings representing the names of the symbols that users want to excluding
from being quantized.
offline_params : list of strs
Names of the parameters that users want to quantize offline. It's always recommended to
quantize parameters offline so that quantizing parameters during the inference can be
avoided.
quantized_dtype: str
The quantized destination type for input data.
"""
num_excluded_symbols = 0
if excluded_symbols is not None:
assert isinstance(excluded_symbols, list)
num_excluded_symbols = len(excluded_symbols)
else:
excluded_symbols = []
num_offline = 0
offline = []
if offline_params is not None:
num_offline = len(offline_params)
for k in offline_params:
offline.append(c_str(k))
out = SymbolHandle()
check_call(
_LIB.MXQuantizeSymbol(sym.handle, ctypes.byref(out),
mx_uint(num_excluded_symbols),
c_str_array(excluded_symbols),
mx_uint(num_offline),
c_array(ctypes.c_char_p, offline),
c_str(quantized_dtype), ctypes.c_bool(True)))
return Symbol(out)
def calibrate_quantized_sym(qsym, th_dict):
if th_dict is None or len(th_dict) == 0:
return qsym
num_layer_outputs = len(th_dict)
layer_output_names = []
min_vals = []
max_vals = []
for k, v in th_dict.items():
layer_output_names.append(k)
min_vals.append(v[0])
max_vals.append(v[1])
calibrated_sym = SymbolHandle()
check_call(
_LIB.MXSetCalibTableToQuantizedSymbol(
qsym.handle, mx_uint(num_layer_outputs),
c_str_array(layer_output_names), c_array(ctypes.c_float, min_vals),
c_array(ctypes.c_float, max_vals), ctypes.byref(calibrated_sym)))
return Symbol(calibrated_sym)
def _calibrate_quantized_sym(qsym, th_dict):
"""Given a dictionary containing the thresholds for quantizing the layers,
set the thresholds into the quantized symbol as the params of requantize operators.
"""
if th_dict is None or len(th_dict) == 0:
return qsym
num_layer_outputs = len(th_dict)
layer_output_names = []
min_vals = []
max_vals = []
for k, v in th_dict.items():
layer_output_names.append(k)
min_vals.append(v[0])
max_vals.append(v[1])
calibrated_sym = SymbolHandle()
check_call(
_LIB.MXSetCalibTableToQuantizedSymbol(
qsym.handle, mx_uint(num_layer_outputs),
c_str_array(layer_output_names), c_array(ctypes.c_float, min_vals),
c_array(ctypes.c_float, max_vals), ctypes.byref(calibrated_sym)))
return Symbol(calibrated_sym)
def quantize_symbol(sym,
excluded_symbols=[],
offline_params=[],
quantized_dtype='uint8',
calib_quantize_op=False):
"""
Quantize symbol.
:param sym: mxnet.symbol.Symbol
The symbol to quantize.
:param excluded_symbols: list of str
The names of symbols to exclude.
:param offline_params: list of str
The names of parameters to quantize offline.
:param quantized_dtype: {"int8", "uint8"}
The data type that you will quantize to.
:param calib_quantize_op: bool
Calibrate or not.(Only for quantization online.
:return: mxnet.symbol.Symbol
The symbol that has been quantized.
"""
assert isinstance(excluded_symbols, list)
num_excluded_symbols = len(excluded_symbols)
# exclude = [s.handle for s in excluded_symbols]
assert isinstance(offline_params, list)
offline = [c_str(k) for k in offline_params]
num_offline = len(offline)
out = SymbolHandle()
check_call(
_LIB.MXQuantizeSymbol(sym.handle, ctypes.byref(out),
mx_uint(num_excluded_symbols),
c_str_array(excluded_symbols),
mx_uint(num_offline),
c_array(ctypes.c_char_p, offline),
c_str(quantized_dtype),
ctypes.c_bool(calib_quantize_op)))
return Symbol(out)