def dump_xmodel(self):
if self.quant_mode < 2:
return
compiler = CompilerFactory.get_compiler("xmodel")
xmodel_dir = os.path.join(self._output_dir, "xmodel")
generic_utils.mkdir_if_not_exist(xmodel_dir)
compile_args = []
if self.lstm:
for direction, graph in self._cell_graphs:
compile_args.append((graph, {'direction': direction}))
else:
compile_args.append((graph, {}))
for graph, attr_kwargs in compile_args:
for node in graph.nodes:
# TODO(yuwang): Set out tensor shape in parser.
# Maybe specify shape for all tensors?
# Input shape must be specified for xgraph shape inference.
if node.op.type == OpTypes.INPUT:
node.out_tensors[0].shape = node.op.attr['shape']
try:
compiler.do_compile(
graph,
os.path.join(xmodel_dir, graph.name),
quant_config_info=self.quant_config,
graph_attr_kwargs=attr_kwargs)
except Exception as e:
print('[ERROR] Failed to dump xmodel: {}'.format(e))
return
print('[INFO] Successfully convert nndct graph to xmodel!')
def dump_xmodel(self, deploy_check=False):
"""
`dump xmodel for LSTM cell`
"""
quantizer = GLOBAL_MAP.get_ele(NNDCT_KEYS.QUANTIZER)
if quantizer and quantizer.quant_mode > 1:
compiler = CompilerFactory.get_compiler("xmodel")
xmodel_dir = os.path.join(self._export_folder, "xmodel")
create_work_dir(xmodel_dir)
for info in self._modules_info.values():
for l_num, layer_graph in enumerate(info["layers_graph"]):
for lstm_direction, graph in layer_graph.items():
try:
compiler.do_compile(
nndct_graph=graph,
quant_config_info=quantizer.quant_config,
output_file_name=os.path.join(xmodel_dir, graph.name),
graph_attr_kwargs={"direction": lstm_direction})
except Exception as e:
print(
f"[NNDCT_ERROR]:failed convert nndct graph to xmodel({str(e)})."
)
else:
print("[NNDCT_NOTE]:Successfully convert nndct graph to xmodel!")
if deploy_check:
print("[NNDCT_NOTE]: Dumping checking data...")
checker = DeployChecker(
output_dir_name=self._export_folder, data_format="txt")
# get timestep output
for name, info in self._layers_info.items():
cell = info["cell_module"]
layer = info["layer_module"]
graph = info["graph"]
if layer.input is None:
warnings.warn(
f"[NNDCT_WARNING]: Provide inputs for '{name}' when do deploy checking",
RuntimeWarning)
continue
set_outputs_recorder_status(cell, True)
layer(layer.input, layer.initial_state, layer.batch_lengths)
for timestep in range(layer.input.size()[1]):
enable_dump_weight = True if timestep == 0 else False
update_nndct_blob_data(cell, graph, timestep)
checker.update_dump_folder(f"{graph.name}/frame_{timestep}")
checker.dump_nodes_output(
graph,
quantizer.quant_config,
round_method=quantizer.quant_opt['round_method'],
enable_dump_weight=enable_dump_weight)
set_outputs_recorder_status(cell, False)
print("[NNDCT_NOTE]: Finsh dumping data.")
def dump_xmodel(output_dir="quantize_result", deploy_check=False):
r"""converts module to xmodel for deployment
compilation only works when quantm model = 2.
The xmodel and some checking data will be generated under work dir.
Args:
deploy_check(bool): if true, can dump blobs and parameters of model for deployment verification
Returns:
None
"""
quantizer = GLOBAL_MAP.get_ele(NNDCT_KEYS.QUANTIZER)
if quantizer and quantizer.quant_mode > 1:
nndct_utils.create_work_dir(output_dir)
# compile to xmodel
compiler = CompilerFactory.get_compiler("xmodel")
NndctScreenLogger().info("=>Converting to xmodel ...")
deploy_graphs = get_deploy_graph_list(quantizer.quant_model,
quantizer.Nndctgraph)
depoly_infos = compiler.get_deloy_graph_infos(quantizer, deploy_graphs)
for depoly_info in depoly_infos:
try:
compiler.do_compile(depoly_info.dev_graph,
quant_config_info=depoly_info.quant_info,
output_file_name=os.path.join(
output_dir,
depoly_info.dev_graph.name))
except AddXopError as e:
NndctScreenLogger().error(
f"Failed convert graph '{depoly_info.dev_graph.name}' to xmodel({str(e)})."
)
# dump data for accuracy check
if deploy_check:
NndctScreenLogger().info(
f"=>Dumping '{depoly_info.dev_graph.name}'' checking data..."
)
checker = DeployChecker(output_dir_name=output_dir)
checker.update_dump_folder(f"{depoly_info.dev_graph.name}")
checker.dump_nodes_output(
depoly_info.dev_graph,
depoly_info.quant_info,
round_method=quantizer.quant_opt['round_method'],
select_batch=False)
NndctScreenLogger().info(
f"=>Finsh dumping data.({checker.dump_folder})")
set_outputs_recorder_status(quantizer.quant_model, False)
def dump_xmodel(output_dir="quantize_result", deploy_check=False):
r"""converts module to xmodel for deployment
compilation only works when quantm model = 2.
The xmodel and some checking data will be generated under work dir.
Args:
deploy_check(bool): if true, can dump blobs and parameters of model for deployment verification
Returns:
None
"""
quantizer = GLOBAL_MAP.get_ele(NNDCT_KEYS.QUANTIZER)
if quantizer and quantizer.quant_mode > 1:
nndct_utils.create_work_dir(output_dir)
# compile to xmodel
try:
compiler = CompilerFactory.get_compiler("xmodel")
NndctScreenLogger().info("=>Converting to xmodel ...")
compiler.do_compile(nndct_graph=quantizer.Nndctgraph,
quant_config_info=quantizer.quant_config,
output_file_name=os.path.join(
output_dir, quantizer.Nndctgraph.name))
except AddXopError as e:
NndctScreenLogger().error(
f"Failed convert nndct graph to xmodel({str(e)}).")
else:
NndctScreenLogger().info(
f"=>Successfully convert to xmodel.({compiler.xmodel_file})")
# dump data for accuracy checkvim
if deploy_check:
NndctScreenLogger().info("=>Dumping checking data...")
update_nndct_blob_data(quantizer.quant_model, quantizer.Nndctgraph)
checker = DeployChecker(output_dir_name=output_dir)
checker.dump_nodes_output(
quantizer.Nndctgraph,
quantizer.quant_config,
round_method=quantizer.quant_opt['round_method'])
set_outputs_recorder_status(quantizer.quant_model, False)
NndctScreenLogger().info(
f"=>Finsh dumping data.({checker.dump_folder})")
def dump_xmodel(output_dir="quantize_result", deploy_check=False, lstm_app=False):
r"""converts module to xmodel for deployment
compilation only works when quantm model = 2.
The xmodel and some checking data will be generated under work dir.
Args:
deploy_check(bool): if true, can dump blobs and parameters of model for deployment verification
Returns:
None
"""
quantizer = GLOBAL_MAP.get_ele(NNDCT_KEYS.QUANTIZER)
if quantizer and quantizer.quant_mode > 1:
nndct_utils.create_work_dir(output_dir)
# compile to xmodel
compiler = CompilerFactory.get_compiler("xmodel")
NndctScreenLogger().info("=>Converting to xmodel ...")
deploy_graphs = get_deploy_graph_list(quantizer.quant_model, quantizer.Nndctgraph)
#depoly_infos = compiler.get_deloy_graph_infos(quantizer, deploy_graphs)
xmodel_depoly_infos, dump_deploy_infos = compiler.get_xmodel_and_dump_infos(quantizer, deploy_graphs)
if not lstm_app:
for node in xmodel_depoly_infos[0].dev_graph.nodes:
error_out = False
if node.op.type not in [NNDCT_OP.INPUT, NNDCT_OP.QUANT_STUB]:
continue
for i, tensor in enumerate(node.out_tensors):
if tensor.shape and tensor.shape[0] != 1:
NndctScreenLogger().error(f"Batch size must be 1 when exporting xmodel.")
error_out = True
break
if error_out:
break
for depoly_info in dump_deploy_infos:
# dump data for accuracy check
if deploy_check:
NndctScreenLogger().info(f"=>Dumping '{depoly_info.dev_graph.name}'' checking data...")
if lstm_app:
checker = DeployChecker(output_dir_name=output_dir, data_format='txt')
checker.update_dump_folder(f"{depoly_info.dev_graph.name}/frame_0")
select_batch = True
else:
checker = DeployChecker(output_dir_name=output_dir)
checker.update_dump_folder(f"{depoly_info.dev_graph.name}")
select_batch = False
checker.dump_nodes_output(
depoly_info.dev_graph,
depoly_info.quant_info,
round_method=quantizer.quant_opt['round_method'], select_batch=select_batch)
NndctScreenLogger().info(f"=>Finsh dumping data.({checker.dump_folder})")
for depoly_info in xmodel_depoly_infos:
try:
xgraph = compiler.do_compile(
depoly_info.dev_graph,
quant_config_info=depoly_info.quant_info,
output_file_name=os.path.join(output_dir, depoly_info.dev_graph.name))
except AddXopError as e:
NndctScreenLogger().error(f"Failed convert graph '{depoly_info.dev_graph.name}' to xmodel.")
raise e
compiler.verify_xmodel(depoly_info.dev_graph, xgraph)
set_outputs_recorder_status(quantizer.quant_model, False)