You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/09/29 14:12:40 UTC
[GitHub] [incubator-tvm] FrozenGene commented on a change in pull request #6578: [tvmc] Introduce 'run' subcommand (part 4/4)
FrozenGene commented on a change in pull request #6578:
URL: https://github.com/apache/incubator-tvm/pull/6578#discussion_r496740342
##########
File path: python/tvm/driver/tvmc/runner.py
##########
@@ -0,0 +1,450 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""
+Provides support to run compiled networks both locally and remotely.
+"""
+import json
+import logging
+import os
+import tarfile
+import tempfile
+
+import numpy as np
+import tvm
+from tvm import rpc
+from tvm.autotvm.measure import request_remote
+from tvm.contrib import graph_runtime as runtime
+from tvm.contrib.debugger import debug_runtime
+
+from . import common
+from .common import TVMCException
+from .main import register_parser
+
+
+# pylint: disable=invalid-name
+logger = logging.getLogger("TVMC")
+
+
+@register_parser
+def add_run_parser(subparsers):
+ """ Include parser for 'run' subcommand """
+
+ parser = subparsers.add_parser("run", help="run a compiled module")
+ parser.set_defaults(func=drive_run)
+
+ # TODO --device needs to be extended and tested to support other targets,
+ # like 'cl', 'webgpu', etc (@leandron)
+ parser.add_argument(
+ "--device",
+ choices=["cpu", "gpu"],
+ default="cpu",
+ help="target device to run the compiled module",
+ )
+ parser.add_argument(
+ "--fill-mode",
+ choices=["zeros", "ones", "random"],
+ default="zeros",
Review comment:
fill zeros is not a good default value. let us fill random values by default.
##########
File path: python/tvm/driver/tvmc/runner.py
##########
@@ -0,0 +1,450 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""
+Provides support to run compiled networks both locally and remotely.
+"""
+import json
+import logging
+import os
+import tarfile
+import tempfile
+
+import numpy as np
+import tvm
+from tvm import rpc
+from tvm.autotvm.measure import request_remote
+from tvm.contrib import graph_runtime as runtime
+from tvm.contrib.debugger import debug_runtime
+
+from . import common
+from .common import TVMCException
+from .main import register_parser
+
+
+# pylint: disable=invalid-name
+logger = logging.getLogger("TVMC")
+
+
+@register_parser
+def add_run_parser(subparsers):
+ """ Include parser for 'run' subcommand """
+
+ parser = subparsers.add_parser("run", help="run a compiled module")
+ parser.set_defaults(func=drive_run)
+
+ # TODO --device needs to be extended and tested to support other targets,
+ # like 'cl', 'webgpu', etc (@leandron)
+ parser.add_argument(
+ "--device",
+ choices=["cpu", "gpu"],
+ default="cpu",
+ help="target device to run the compiled module",
+ )
+ parser.add_argument(
+ "--fill-mode",
+ choices=["zeros", "ones", "random"],
+ default="zeros",
+ help="fill all input tensors with values",
+ )
+ parser.add_argument("-i", "--inputs", help="path to the .npz input file")
+ parser.add_argument("-o", "--outputs", help="path to the .npz output file")
+ parser.add_argument(
+ "--print-time", action="store_true", help="record and print the execution time(s)"
+ )
+ parser.add_argument(
+ "--print-top",
+ metavar="N",
+ type=int,
+ help="print the top n values and indices of the output tensor",
+ )
+ parser.add_argument(
+ "--profile", action="store_true", help="generate profiling data from the runtime execution"
+ )
+ parser.add_argument("--repeat", metavar="N", type=int, default=1, help="repeat the run n times")
+ parser.add_argument(
+ "--rpc-key",
+ nargs=1,
+ help="the RPC tracker key of the target device",
+ )
+ parser.add_argument(
+ "--rpc-tracker",
+ nargs=1,
+ help="hostname (required) and port (optional, defaults to 9090) of the RPC tracker, "
+ "e.g. '192.168.0.100:9999'",
+ )
+ parser.add_argument("FILE", help="path to the compiled module file")
+
+
+def drive_run(args):
+ """Invoke runner module with command line arguments
+
+ Parameters
+ ----------
+ args: argparse.Namespace
+ Arguments from command line parser.
+ """
+ inputs = {}
+ if args.inputs:
+ inputs = np.load(args.inputs)
+
+ rpc_hostname, rpc_port = common.tracker_host_port_from_cli(args.rpc_tracker)
+
+ outputs, times = run_module(
+ args.FILE,
+ rpc_hostname,
+ rpc_port,
+ args.rpc_key,
+ inputs=inputs,
+ device=args.device,
+ fill_mode=args.fill_mode,
+ repeat=args.repeat,
+ profile=args.profile,
+ )
+
+ if args.print_time:
+ stat_table = format_times(times)
+ # print here is intentional
+ print(stat_table)
+
+ if args.print_top:
+ top_results = get_top_results(outputs, args.print_top)
+ # print here is intentional
+ print(top_results)
+
+ if args.outputs:
+ # Save the outputs
+ np.savez(args.outputs, **outputs)
+
+
+def get_input_info(graph_str, params):
+ """Return the 'shape' and 'dtype' dictionaries for the input
+ tensors of a compiled module.
+
+ Parameters
+ ----------
+ graph_str : str
+ JSON graph of the module serialized as a string.
+ params : bytearray
+ Params serialized as a bytearray.
+
+ Returns
+ -------
+ shape_dict : dict
+ Shape dictionary - {input_name: tuple}.
+ dtype_dict : dict
+ dtype dictionary - {input_name: dtype}.
+ """
+ # NOTE - We can't simply get the input tensors from a TVM graph
+ # because weight tensors are treated equivalently. Therefore, to
+ # find the input tensors we look at the 'arg_nodes' in the graph
+ # (which are either weights or inputs) and check which ones don't
+ # appear in the params (where the weights are stored). These nodes
+ # are therefore inferred to be input tensors.
+
+ shape_dict = {}
+ dtype_dict = {}
+ # Use a special function to load the binary params back into a dict
+ load_arr = tvm.get_global_func("tvm.relay._load_param_dict")(params)
+ param_names = [v.name for v in load_arr]
+ graph = json.loads(graph_str)
+ for node_id in graph["arg_nodes"]:
+ node = graph["nodes"][node_id]
+ # If a node is not in the params, infer it to be an input node
+ name = node["name"]
+ if name not in param_names:
+ shape_dict[name] = graph["attrs"]["shape"][1][node_id]
+ dtype_dict[name] = graph["attrs"]["dltype"][1][node_id]
+
+ logger.debug("collecting graph input shape and type:")
+ logger.debug("graph input shape: %s", shape_dict)
+ logger.debug("graph input type: %s", dtype_dict)
+
+ return shape_dict, dtype_dict
+
+
+def generate_tensor_data(shape, dtype, fill_mode):
+ """Generate data to produce a tensor of given shape and dtype.
+
+ Random data generation depends on the dtype. For int8 types,
+ random integers in the range 0->255 are generated. For all other
+ types, random floats are generated in the range -1->1 and then
+ cast to the appropriate dtype.
+
+ This is used to quickly generate some data to input the models, as
+ a way to check that compiled module is sane for running.
+
+ Parameters
+ ----------
+ shape : tuple
+ The shape of the tensor.
+ dtype : str
+ The dtype of the tensor.
+ fill_mode : str
+ The fill-mode to use, either "zeros", "ones" or "random".
+
+ Returns
+ -------
+ tensor : np.array
+ The generated tensor as a np.array.
+ """
+ if fill_mode == "zeros":
+ tensor = np.zeros(shape=shape, dtype=dtype)
+ elif fill_mode == "ones":
+ tensor = np.ones(shape=shape, dtype=dtype)
+ elif fill_mode == "random":
+ if "int8" in dtype:
+ tensor = np.random.randint(256, size=shape, dtype=dtype)
Review comment:
let us restrict to 128. If the model is int8 (not uint8). 256 will make us overflow.
##########
File path: python/tvm/driver/tvmc/runner.py
##########
@@ -0,0 +1,450 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""
+Provides support to run compiled networks both locally and remotely.
+"""
+import json
+import logging
+import os
+import tarfile
+import tempfile
+
+import numpy as np
+import tvm
+from tvm import rpc
+from tvm.autotvm.measure import request_remote
+from tvm.contrib import graph_runtime as runtime
+from tvm.contrib.debugger import debug_runtime
+
+from . import common
+from .common import TVMCException
+from .main import register_parser
+
+
+# pylint: disable=invalid-name
+logger = logging.getLogger("TVMC")
+
+
+@register_parser
+def add_run_parser(subparsers):
+ """ Include parser for 'run' subcommand """
+
+ parser = subparsers.add_parser("run", help="run a compiled module")
+ parser.set_defaults(func=drive_run)
+
+ # TODO --device needs to be extended and tested to support other targets,
+ # like 'cl', 'webgpu', etc (@leandron)
+ parser.add_argument(
+ "--device",
+ choices=["cpu", "gpu"],
+ default="cpu",
+ help="target device to run the compiled module",
+ )
+ parser.add_argument(
+ "--fill-mode",
+ choices=["zeros", "ones", "random"],
+ default="zeros",
+ help="fill all input tensors with values",
+ )
+ parser.add_argument("-i", "--inputs", help="path to the .npz input file")
+ parser.add_argument("-o", "--outputs", help="path to the .npz output file")
+ parser.add_argument(
+ "--print-time", action="store_true", help="record and print the execution time(s)"
+ )
+ parser.add_argument(
+ "--print-top",
+ metavar="N",
+ type=int,
+ help="print the top n values and indices of the output tensor",
+ )
+ parser.add_argument(
+ "--profile", action="store_true", help="generate profiling data from the runtime execution"
+ )
+ parser.add_argument("--repeat", metavar="N", type=int, default=1, help="repeat the run n times")
+ parser.add_argument(
+ "--rpc-key",
+ nargs=1,
+ help="the RPC tracker key of the target device",
+ )
+ parser.add_argument(
+ "--rpc-tracker",
+ nargs=1,
+ help="hostname (required) and port (optional, defaults to 9090) of the RPC tracker, "
+ "e.g. '192.168.0.100:9999'",
+ )
+ parser.add_argument("FILE", help="path to the compiled module file")
+
+
+def drive_run(args):
+ """Invoke runner module with command line arguments
+
+ Parameters
+ ----------
+ args: argparse.Namespace
+ Arguments from command line parser.
+ """
+ inputs = {}
+ if args.inputs:
+ inputs = np.load(args.inputs)
+
+ rpc_hostname, rpc_port = common.tracker_host_port_from_cli(args.rpc_tracker)
+
+ outputs, times = run_module(
+ args.FILE,
+ rpc_hostname,
+ rpc_port,
+ args.rpc_key,
+ inputs=inputs,
+ device=args.device,
+ fill_mode=args.fill_mode,
+ repeat=args.repeat,
+ profile=args.profile,
+ )
+
+ if args.print_time:
+ stat_table = format_times(times)
+ # print here is intentional
+ print(stat_table)
+
+ if args.print_top:
+ top_results = get_top_results(outputs, args.print_top)
+ # print here is intentional
+ print(top_results)
+
+ if args.outputs:
+ # Save the outputs
+ np.savez(args.outputs, **outputs)
+
+
+def get_input_info(graph_str, params):
+ """Return the 'shape' and 'dtype' dictionaries for the input
+ tensors of a compiled module.
+
+ Parameters
+ ----------
+ graph_str : str
+ JSON graph of the module serialized as a string.
+ params : bytearray
+ Params serialized as a bytearray.
+
+ Returns
+ -------
+ shape_dict : dict
+ Shape dictionary - {input_name: tuple}.
+ dtype_dict : dict
+ dtype dictionary - {input_name: dtype}.
+ """
+ # NOTE - We can't simply get the input tensors from a TVM graph
+ # because weight tensors are treated equivalently. Therefore, to
+ # find the input tensors we look at the 'arg_nodes' in the graph
+ # (which are either weights or inputs) and check which ones don't
+ # appear in the params (where the weights are stored). These nodes
+ # are therefore inferred to be input tensors.
+
+ shape_dict = {}
+ dtype_dict = {}
+ # Use a special function to load the binary params back into a dict
+ load_arr = tvm.get_global_func("tvm.relay._load_param_dict")(params)
+ param_names = [v.name for v in load_arr]
+ graph = json.loads(graph_str)
+ for node_id in graph["arg_nodes"]:
+ node = graph["nodes"][node_id]
+ # If a node is not in the params, infer it to be an input node
+ name = node["name"]
+ if name not in param_names:
+ shape_dict[name] = graph["attrs"]["shape"][1][node_id]
+ dtype_dict[name] = graph["attrs"]["dltype"][1][node_id]
+
+ logger.debug("collecting graph input shape and type:")
+ logger.debug("graph input shape: %s", shape_dict)
+ logger.debug("graph input type: %s", dtype_dict)
+
+ return shape_dict, dtype_dict
+
+
+def generate_tensor_data(shape, dtype, fill_mode):
+ """Generate data to produce a tensor of given shape and dtype.
+
+ Random data generation depends on the dtype. For int8 types,
+ random integers in the range 0->255 are generated. For all other
+ types, random floats are generated in the range -1->1 and then
+ cast to the appropriate dtype.
+
+ This is used to quickly generate some data to input the models, as
+ a way to check that compiled module is sane for running.
+
+ Parameters
+ ----------
+ shape : tuple
+ The shape of the tensor.
+ dtype : str
+ The dtype of the tensor.
+ fill_mode : str
+ The fill-mode to use, either "zeros", "ones" or "random".
+
+ Returns
+ -------
+ tensor : np.array
+ The generated tensor as a np.array.
+ """
+ if fill_mode == "zeros":
+ tensor = np.zeros(shape=shape, dtype=dtype)
+ elif fill_mode == "ones":
+ tensor = np.ones(shape=shape, dtype=dtype)
+ elif fill_mode == "random":
+ if "int8" in dtype:
+ tensor = np.random.randint(256, size=shape, dtype=dtype)
+ else:
+ tensor = np.random.uniform(-1, 1, size=shape).astype(dtype)
+ else:
+ raise TVMCException("unknown fill-mode: {}".format(fill_mode))
+
+ return tensor
+
+
+def make_inputs_dict(inputs, shape_dict, dtype_dict, fill_mode):
+ """Make the inputs dictionary for a graph.
+
+ Use data from 'inputs' where specified. For input tensors
+ where no data has been given, generate data according to the
+ chosen fill-mode.
+
+ Parameters
+ ----------
+ inputs : dict
+ Input data dictionary - {input_name: np.array}.
+ shape_dict : dict
+ Shape dictionary - {input_name: tuple}.
+ dtype_dict : dict
+ dtype dictionary - {input_name: dtype}.
+ fill_mode : str
+ The fill-mode to use when generating tensor data.
+ Can be either "zeros", "ones" or "random".
+
+ Returns
+ -------
+ inputs_dict : dict
+ Complete inputs dictionary - {input_name: np.array}.
+ """
+ logger.debug("creating inputs dict")
+
+ # First check all the keys in inputs exist in the graph
+ for input_name in inputs:
+ if input_name not in shape_dict.keys():
+ raise TVMCException("the input tensor '{}' is not in the graph".format(input_name))
+
+ # Now construct the input dict, generating tensors where no
+ # data already exists in 'inputs'
+ inputs_dict = {}
+ for input_name in shape_dict:
+ if input_name in inputs.keys():
+ logger.debug("setting input '%s' with user input data", input_name)
+ inputs_dict[input_name] = inputs[input_name]
+ else:
+ shape = shape_dict[input_name]
+ dtype = dtype_dict[input_name]
+
+ logger.debug(
+ "generating data for input '%s' (shape: %s, dtype: %s), using fill-mode '%s'",
+ input_name,
+ shape,
+ dtype,
+ fill_mode,
+ )
+ data = generate_tensor_data(shape, dtype, fill_mode)
+ inputs_dict[input_name] = data
+
+ return inputs_dict
+
+
+def run_module(
+ module_file,
+ hostname,
+ port=9090,
+ rpc_key=None,
+ device=None,
+ inputs=None,
+ fill_mode="zeros",
+ repeat=1,
+ profile=False,
+):
+ """Run a compiled graph runtime module locally or remotely with
+ optional input values.
+
+ If input tensors are not specified explicitly, they can be filled
+ with zeroes, ones or random data.
+
+ Parameters
+ ----------
+ module_file : str
+ The path to the module file (a .tar file).
+ hostname : str
+ The hostname of the target device on which to run.
+ port : int, optional
+ The port of the target device on which to run.
+ rpc_key : str, optional
+ The tracker key of the target device. If this is set, it
+ will be assumed that remote points to a tracker.
+ device: str, optional
+ the device (e.g. "cpu" or "gpu") to be targeted by the RPC
+ session, local or remote).
+ inputs : dict, optional
+ A dictionary of {input_name: np.array} storing the input
+ tensors to the network. If this is not specified, data will
+ be automatically generated for the input tensors.
+ fill_mode : str, optional
+ The fill-mode to use when generating data for input tensors.
+ Valid options are "zeros", "ones" and "random".
+ Defaults to "zeros".
+ repeat : int, optional
+ How many times to repeat the run.
+ profile : bool
+ Whether to profile the run with the debug runtime.
+
+ Returns
+ -------
+ outputs : dict
+ a dictionary with output tensors, generated by the module
+ times : list of str
+ execution times generated by the time evaluator
+ """
+ if not inputs:
+ inputs = {}
+
+ with tempfile.TemporaryDirectory() as tmp_dir:
+ logger.debug("extracting module file %s", module_file)
+ t = tarfile.open(module_file)
Review comment:
let us support both mode (tar and .so). Because not all remote devices has compiler.
##########
File path: python/tvm/driver/tvmc/runner.py
##########
@@ -0,0 +1,450 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""
+Provides support to run compiled networks both locally and remotely.
+"""
+import json
+import logging
+import os
+import tarfile
+import tempfile
+
+import numpy as np
+import tvm
+from tvm import rpc
+from tvm.autotvm.measure import request_remote
+from tvm.contrib import graph_runtime as runtime
+from tvm.contrib.debugger import debug_runtime
+
+from . import common
+from .common import TVMCException
+from .main import register_parser
+
+
+# pylint: disable=invalid-name
+logger = logging.getLogger("TVMC")
+
+
+@register_parser
+def add_run_parser(subparsers):
+ """ Include parser for 'run' subcommand """
+
+ parser = subparsers.add_parser("run", help="run a compiled module")
+ parser.set_defaults(func=drive_run)
+
+ # TODO --device needs to be extended and tested to support other targets,
+ # like 'cl', 'webgpu', etc (@leandron)
+ parser.add_argument(
+ "--device",
+ choices=["cpu", "gpu"],
+ default="cpu",
+ help="target device to run the compiled module",
+ )
+ parser.add_argument(
+ "--fill-mode",
+ choices=["zeros", "ones", "random"],
+ default="zeros",
+ help="fill all input tensors with values",
+ )
+ parser.add_argument("-i", "--inputs", help="path to the .npz input file")
+ parser.add_argument("-o", "--outputs", help="path to the .npz output file")
+ parser.add_argument(
+ "--print-time", action="store_true", help="record and print the execution time(s)"
+ )
+ parser.add_argument(
+ "--print-top",
+ metavar="N",
+ type=int,
+ help="print the top n values and indices of the output tensor",
+ )
+ parser.add_argument(
+ "--profile", action="store_true", help="generate profiling data from the runtime execution"
Review comment:
I think we should mention users need to turn on debug graph runtime in the config.cmake.
##########
File path: python/tvm/driver/tvmc/runner.py
##########
@@ -0,0 +1,450 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""
+Provides support to run compiled networks both locally and remotely.
+"""
+import json
+import logging
+import os
+import tarfile
+import tempfile
+
+import numpy as np
+import tvm
+from tvm import rpc
+from tvm.autotvm.measure import request_remote
+from tvm.contrib import graph_runtime as runtime
+from tvm.contrib.debugger import debug_runtime
+
+from . import common
+from .common import TVMCException
+from .main import register_parser
+
+
+# pylint: disable=invalid-name
+logger = logging.getLogger("TVMC")
+
+
+@register_parser
+def add_run_parser(subparsers):
+ """ Include parser for 'run' subcommand """
+
+ parser = subparsers.add_parser("run", help="run a compiled module")
+ parser.set_defaults(func=drive_run)
+
+ # TODO --device needs to be extended and tested to support other targets,
+ # like 'cl', 'webgpu', etc (@leandron)
+ parser.add_argument(
+ "--device",
+ choices=["cpu", "gpu"],
+ default="cpu",
+ help="target device to run the compiled module",
+ )
+ parser.add_argument(
+ "--fill-mode",
+ choices=["zeros", "ones", "random"],
+ default="zeros",
+ help="fill all input tensors with values",
+ )
+ parser.add_argument("-i", "--inputs", help="path to the .npz input file")
+ parser.add_argument("-o", "--outputs", help="path to the .npz output file")
+ parser.add_argument(
+ "--print-time", action="store_true", help="record and print the execution time(s)"
+ )
+ parser.add_argument(
+ "--print-top",
+ metavar="N",
+ type=int,
+ help="print the top n values and indices of the output tensor",
+ )
+ parser.add_argument(
+ "--profile", action="store_true", help="generate profiling data from the runtime execution"
+ )
+ parser.add_argument("--repeat", metavar="N", type=int, default=1, help="repeat the run n times")
+ parser.add_argument(
+ "--rpc-key",
+ nargs=1,
+ help="the RPC tracker key of the target device",
+ )
+ parser.add_argument(
+ "--rpc-tracker",
+ nargs=1,
+ help="hostname (required) and port (optional, defaults to 9090) of the RPC tracker, "
+ "e.g. '192.168.0.100:9999'",
+ )
+ parser.add_argument("FILE", help="path to the compiled module file")
+
+
+def drive_run(args):
+ """Invoke runner module with command line arguments
+
+ Parameters
+ ----------
+ args: argparse.Namespace
+ Arguments from command line parser.
+ """
+ inputs = {}
+ if args.inputs:
+ inputs = np.load(args.inputs)
+
+ rpc_hostname, rpc_port = common.tracker_host_port_from_cli(args.rpc_tracker)
+
+ outputs, times = run_module(
+ args.FILE,
+ rpc_hostname,
+ rpc_port,
+ args.rpc_key,
+ inputs=inputs,
+ device=args.device,
+ fill_mode=args.fill_mode,
+ repeat=args.repeat,
+ profile=args.profile,
+ )
+
+ if args.print_time:
+ stat_table = format_times(times)
+ # print here is intentional
+ print(stat_table)
+
+ if args.print_top:
+ top_results = get_top_results(outputs, args.print_top)
+ # print here is intentional
+ print(top_results)
+
+ if args.outputs:
+ # Save the outputs
+ np.savez(args.outputs, **outputs)
+
+
+def get_input_info(graph_str, params):
+ """Return the 'shape' and 'dtype' dictionaries for the input
+ tensors of a compiled module.
+
+ Parameters
+ ----------
+ graph_str : str
+ JSON graph of the module serialized as a string.
+ params : bytearray
+ Params serialized as a bytearray.
+
+ Returns
+ -------
+ shape_dict : dict
+ Shape dictionary - {input_name: tuple}.
+ dtype_dict : dict
+ dtype dictionary - {input_name: dtype}.
+ """
+ # NOTE - We can't simply get the input tensors from a TVM graph
+ # because weight tensors are treated equivalently. Therefore, to
+ # find the input tensors we look at the 'arg_nodes' in the graph
+ # (which are either weights or inputs) and check which ones don't
+ # appear in the params (where the weights are stored). These nodes
+ # are therefore inferred to be input tensors.
+
+ shape_dict = {}
+ dtype_dict = {}
+ # Use a special function to load the binary params back into a dict
+ load_arr = tvm.get_global_func("tvm.relay._load_param_dict")(params)
+ param_names = [v.name for v in load_arr]
+ graph = json.loads(graph_str)
+ for node_id in graph["arg_nodes"]:
+ node = graph["nodes"][node_id]
+ # If a node is not in the params, infer it to be an input node
+ name = node["name"]
+ if name not in param_names:
+ shape_dict[name] = graph["attrs"]["shape"][1][node_id]
+ dtype_dict[name] = graph["attrs"]["dltype"][1][node_id]
+
+ logger.debug("collecting graph input shape and type:")
+ logger.debug("graph input shape: %s", shape_dict)
+ logger.debug("graph input type: %s", dtype_dict)
+
+ return shape_dict, dtype_dict
+
+
+def generate_tensor_data(shape, dtype, fill_mode):
+ """Generate data to produce a tensor of given shape and dtype.
+
+ Random data generation depends on the dtype. For int8 types,
+ random integers in the range 0->255 are generated. For all other
+ types, random floats are generated in the range -1->1 and then
+ cast to the appropriate dtype.
+
+ This is used to quickly generate some data to input the models, as
+ a way to check that compiled module is sane for running.
+
+ Parameters
+ ----------
+ shape : tuple
+ The shape of the tensor.
+ dtype : str
+ The dtype of the tensor.
+ fill_mode : str
+ The fill-mode to use, either "zeros", "ones" or "random".
+
+ Returns
+ -------
+ tensor : np.array
+ The generated tensor as a np.array.
+ """
+ if fill_mode == "zeros":
+ tensor = np.zeros(shape=shape, dtype=dtype)
+ elif fill_mode == "ones":
+ tensor = np.ones(shape=shape, dtype=dtype)
+ elif fill_mode == "random":
+ if "int8" in dtype:
+ tensor = np.random.randint(256, size=shape, dtype=dtype)
+ else:
+ tensor = np.random.uniform(-1, 1, size=shape).astype(dtype)
+ else:
+ raise TVMCException("unknown fill-mode: {}".format(fill_mode))
+
+ return tensor
+
+
+def make_inputs_dict(inputs, shape_dict, dtype_dict, fill_mode):
+ """Make the inputs dictionary for a graph.
+
+ Use data from 'inputs' where specified. For input tensors
+ where no data has been given, generate data according to the
+ chosen fill-mode.
+
+ Parameters
+ ----------
+ inputs : dict
+ Input data dictionary - {input_name: np.array}.
+ shape_dict : dict
+ Shape dictionary - {input_name: tuple}.
+ dtype_dict : dict
+ dtype dictionary - {input_name: dtype}.
+ fill_mode : str
+ The fill-mode to use when generating tensor data.
+ Can be either "zeros", "ones" or "random".
+
+ Returns
+ -------
+ inputs_dict : dict
+ Complete inputs dictionary - {input_name: np.array}.
+ """
+ logger.debug("creating inputs dict")
+
+ # First check all the keys in inputs exist in the graph
+ for input_name in inputs:
+ if input_name not in shape_dict.keys():
+ raise TVMCException("the input tensor '{}' is not in the graph".format(input_name))
+
+ # Now construct the input dict, generating tensors where no
+ # data already exists in 'inputs'
+ inputs_dict = {}
+ for input_name in shape_dict:
+ if input_name in inputs.keys():
+ logger.debug("setting input '%s' with user input data", input_name)
+ inputs_dict[input_name] = inputs[input_name]
+ else:
+ shape = shape_dict[input_name]
+ dtype = dtype_dict[input_name]
+
+ logger.debug(
+ "generating data for input '%s' (shape: %s, dtype: %s), using fill-mode '%s'",
+ input_name,
+ shape,
+ dtype,
+ fill_mode,
+ )
+ data = generate_tensor_data(shape, dtype, fill_mode)
+ inputs_dict[input_name] = data
+
+ return inputs_dict
+
+
+def run_module(
+ module_file,
+ hostname,
+ port=9090,
+ rpc_key=None,
+ device=None,
+ inputs=None,
+ fill_mode="zeros",
Review comment:
Same comment as before.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org