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 2020/11/08 23:39:50 UTC
[GitHub] [incubator-tvm] comaniac commented on a change in pull request #6882: [AutoScheduler] Tutorial on auto-scheduling a network for GPU
comaniac commented on a change in pull request #6882:
URL: https://github.com/apache/incubator-tvm/pull/6882#discussion_r519491866
##########
File path: src/auto_scheduler/feature.cc
##########
@@ -1345,11 +1345,6 @@ void GetPerStoreFeaturesFromStates(const Array<State>& states, const SearchTask&
GetPerStoreFeaturesWorkerFunc(task, states[i], max_n_bufs,
&(*features)[i], &error_ct);
});
-
- if (error_ct > 0) {
- std::cerr << "Encountered " << error_ct
- << " errors during feature extraction, which are safely ignored." << std::endl;
- }
Review comment:
Can we keep this message in a lower logging verbose level? If not (which I think might be an issue because DMLC logging system doesn't have DEBUG level), we may need to add a message at the end of the measurement process indicating the case that all states are failed.
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
+in :code:`tvm/python/topi` while does not use existing schedule templates.
+
+Note that this tutorial will not run on Windows or recent versions of macOS. To
+get it to run, you will need to wrap the body of this tutorial in a :code:`if
+__name__ == "__main__":` block.
+"""
+
+import numpy as np
+
+import tvm
+from tvm import relay, auto_scheduler
+import tvm.relay.testing
+from tvm.contrib import graph_runtime
+
+#################################################################
+# Define a Network
+# ----------------
+# First, we need to define the network in relay frontend API.
+# We can load some pre-defined network from :code:`tvm.relay.testing`.
+# We can also load models from MXNet, ONNX, PyTorch, and TensorFlow
+# (see :ref:`front end tutorials<tutorial-frontend>`).
+#
+# Note that although auto-scheduler can work with any layouts,
+# we found that the best performance is typically archived with NHWC layout
+# for convolutional neural networks.
+#
+
+
+def get_network(name, batch_size, layout="NHWC", dtype="float32"):
+ """Get the symbol definition and random weight of a network"""
+
+ # auto-scheduler prefers NHWC layout
+ if layout == "NHWC":
+ image_shape = (224, 224, 3)
+ elif layout == "NCHW":
+ image_shape = (3, 224, 224)
+ else:
+ raise ValueError("Invalid layout: " + layout)
+
+ input_shape = (batch_size,) + image_shape
+ output_shape = (batch_size, 1000)
+
+ if name.startswith("resnet-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name.startswith("resnet3d-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "mobilenet":
+ mod, params = relay.testing.mobilenet.get_workload(
+ batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape
+ )
+ elif name == "squeezenet_v1.1":
+ mod, params = relay.testing.squeezenet.get_workload(
+ version="1.1",
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "inception_v3":
+ input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)
+ mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
+ elif name == "mxnet":
+ # an example for mxnet model
+ from mxnet.gluon.model_zoo.vision import get_model
+
+ assert layout == "NCHW"
+
+ block = get_model("resnet18_v1", pretrained=True)
+ mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
+ net = mod["main"]
+ net = relay.Function(
+ net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
+ )
+ mod = tvm.IRModule.from_expr(net)
+
+ return mod, params, input_shape, output_shape
+
+
+# Define the neural network and compilation target
+network = "resnet-18"
+batch_size = 1
+layout = "NHWC"
+target = tvm.target.Target("cuda")
+dtype = "float32"
+log_file = "%s-%s-B%d.json" % (network, layout, batch_size)
+
+#################################################################
+# Extract Search Tasks
+# --------------------
+# Next, we extract the search tasks and their weights from a network.
+# The weight of a task is the number of appearances of the task's subgraph
+# in the whole network.
+# By using the weight, we can approximate the end-to-end latency of the network
+# as :code:`sum(latency[t] * weight[t])`, where :code:`latency[t]` is the
+# latency of a task and :code:`weight[t]` is the weight of the task.
+
+# Enable auto-scheduler in relay
+auto_scheduler.enable_relay_integration()
+
+# Extract tasks from the network
+print("Extract tasks...")
+mod, params, input_shape, output_shape = get_network(network, batch_size, layout, dtype=dtype)
+tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)
+
+# Define the objective as the end-to-end exeuction time of the network
+objective = lambda costs: sum(c * w for c, w in zip(costs, task_weights))
+
+#################################################################
+# Begin Tuning
+# ------------
+# Now, we set some options of tuning and launch the search tasks
+#
+# * :code:`measure_ctx` launches a different process for measurement to
+# provide isolation. It can protect the master process from GPU crashes
+# happened during measurement and avoid other runtime conflicts.
+# * :code:`min_repeat_ms` defines the minimum duration of one "repeat" in every measurement.
+# This can warmup the GPU, which is necessary to get accurate measurement results.
+# Typically, we recommend a value > 300 ms.
+# * :code:`num_measure_trials` is the number of measurement trials we can use during the tuning.
+# You can set it to a small number (e.g., 200) for a fast demonstrative run.
+# In practice, we recommend setting it round :code:`1000 * len(tasks)`,
+# which is typically enough for the search to converge.
+# For example, there are 21 tasks in resnet-18, so we can set it as 20000 for renset-18.
+# You can adjust this parameter according to your time budget.
+# * In addition, we use :code:`RecordToFile` to dump measurement records into the log file,
+# The measurement records can be used to query the history best, resume the search,
+# and do more analyses later.
+# * see :any:`auto_scheduler.TuningOptions`,
+# :any:`auto_scheduler.LocalRPCMeasureContext` for more parameters.
+#
+
+
+def run_tuning():
+ print("Begin tuning...")
+ measure_ctx = auto_scheduler.LocalRPCMeasureContext(repeat=1, min_repeat_ms=400, timeout=10)
+
+ tuner = auto_scheduler.TaskScheduler(tasks, objective)
+ tune_option = auto_scheduler.TuningOptions(
+ num_measure_trials=200, # change this to 20000 to achieve the best performance
+ runner=measure_ctx.runner,
+ measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
+ )
+
+ tuner.tune(tune_option)
+
+
+# We do not run the tuning in our webpage server since it takes too long.
+# Uncomment the following line to run it by yourself.
+
+# run_tuning()
+
+
+######################################################################
+# .. note:: Explain the printed information during tuning
+#
+# During the tuning, a lot of information will be printed on the screen.
+# They are used for debugging purposes. The most important info is the output
+# of the task scheduler. The following table is a sample output.
+#
+# .. code-block:: c
+#
+# ----------------------------------------------------------------------
+# ------------------------------ [ Task Scheduler ]
+# ----------------------------------------------------------------------
+# | ID | Latency (ms) | Speed (GFLOPS) | Trials |
+# -------------------------------------------------
+# | 0 | 0.014 | 72.07 | 64 |
+# | 1 | 0.185 | 1250.68 | 128 |
+# | 2 | 0.142 | 1626.36 | 192 |
+# | 3 | 0.137 | 1689.42 | 128 |
+# | 4 | 0.097 | 1189.75 | 128 |
+# | 5 | 0.092 | 2505.25 | 128 |
+# | 6 | 0.080 | 2893.08 | 128 |
+# | 7 | 0.119 | 1947.84 | 128 |
+# | 8 | 0.090 | 1292.62 | 64 |
+# | 9 | 0.107 | 2172.30 | 64 |
+# | 10 | 0.095 | 2439.36 | 64 |
+# | 11 | 0.077 | 3003.22 | 64 |
+# | 12 | 0.068 | 1695.13 | 64 |
+# | 13 | 0.058 | 3979.29 | 64 |
+# | 14 | 0.048 | 4859.95 | 128 |
+# | 15 | 0.073 | 3151.76 | 64 |
+# | 16 | 0.056 | 4265.94 | 64 |
+# | 17 | 0.009 | 2754.90 | 64 |
+# | 18 | 0.011 | 1156.08 | 64 |
+# | 19 | 0.013 | 955.80 | 64 |
+# | 20 | 0.029 | 437.71 | 64 |
+# -------------------------------------------------
+# Total latency: 1.649 ms Trials: 1920 Used time : 3598 s Next ID: 9
+#
+# This table lists the latency and speed of all tasks.
+# It also lists the allocation of measurement trials for all tasks.
+# The last line prints the total weighted latency of these tasks,
+# which can be a rough estimation of the end-to-end execution time
+# of the network.
+# The last line also prints the total number of measurement trials,
+# total time spent on auto-tuning and the id of the next task to tune.
+#
+# There will also be some "dmlc::Error"s and CUDA errors. You can safely
+# ignore them if the tuning can continue.
Review comment:
Better to add one more sentence to explain the possible reason. Like this is because auto-scheduler tried an invalid schedule, but this can be safely ignore.
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
Review comment:
```suggestion
schedule templates. In other words, the auto-scheduler only uses the compute declarations
```
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
Review comment:
```suggestion
===========================================
```
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
Review comment:
```suggestion
execution time and prioritizes the one that can reduce the execution time the most.
```
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
+in :code:`tvm/python/topi` while does not use existing schedule templates.
+
+Note that this tutorial will not run on Windows or recent versions of macOS. To
+get it to run, you will need to wrap the body of this tutorial in a :code:`if
+__name__ == "__main__":` block.
+"""
+
+import numpy as np
+
+import tvm
+from tvm import relay, auto_scheduler
+import tvm.relay.testing
+from tvm.contrib import graph_runtime
+
+#################################################################
+# Define a Network
+# ----------------
+# First, we need to define the network in relay frontend API.
+# We can load some pre-defined network from :code:`tvm.relay.testing`.
+# We can also load models from MXNet, ONNX, PyTorch, and TensorFlow
+# (see :ref:`front end tutorials<tutorial-frontend>`).
+#
+# Note that although auto-scheduler can work with any layouts,
+# we found that the best performance is typically archived with NHWC layout
+# for convolutional neural networks.
Review comment:
```suggestion
# and we found that the best performance is typically archived with NHWC layout
# for convolutional neural networks, so we use NHWC layout in this tutorial.
```
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
+in :code:`tvm/python/topi` while does not use existing schedule templates.
+
+Note that this tutorial will not run on Windows or recent versions of macOS. To
+get it to run, you will need to wrap the body of this tutorial in a :code:`if
+__name__ == "__main__":` block.
+"""
+
+import numpy as np
+
+import tvm
+from tvm import relay, auto_scheduler
+import tvm.relay.testing
+from tvm.contrib import graph_runtime
+
+#################################################################
+# Define a Network
+# ----------------
+# First, we need to define the network in relay frontend API.
+# We can load some pre-defined network from :code:`tvm.relay.testing`.
+# We can also load models from MXNet, ONNX, PyTorch, and TensorFlow
+# (see :ref:`front end tutorials<tutorial-frontend>`).
+#
+# Note that although auto-scheduler can work with any layouts,
+# we found that the best performance is typically archived with NHWC layout
+# for convolutional neural networks.
+#
+
+
+def get_network(name, batch_size, layout="NHWC", dtype="float32"):
+ """Get the symbol definition and random weight of a network"""
+
+ # auto-scheduler prefers NHWC layout
+ if layout == "NHWC":
+ image_shape = (224, 224, 3)
+ elif layout == "NCHW":
+ image_shape = (3, 224, 224)
+ else:
+ raise ValueError("Invalid layout: " + layout)
+
+ input_shape = (batch_size,) + image_shape
+ output_shape = (batch_size, 1000)
+
+ if name.startswith("resnet-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name.startswith("resnet3d-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "mobilenet":
+ mod, params = relay.testing.mobilenet.get_workload(
+ batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape
+ )
+ elif name == "squeezenet_v1.1":
+ mod, params = relay.testing.squeezenet.get_workload(
+ version="1.1",
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "inception_v3":
+ input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)
+ mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
+ elif name == "mxnet":
+ # an example for mxnet model
+ from mxnet.gluon.model_zoo.vision import get_model
+
+ assert layout == "NCHW"
+
+ block = get_model("resnet18_v1", pretrained=True)
+ mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
+ net = mod["main"]
+ net = relay.Function(
+ net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
+ )
+ mod = tvm.IRModule.from_expr(net)
+
+ return mod, params, input_shape, output_shape
+
+
+# Define the neural network and compilation target
+network = "resnet-18"
+batch_size = 1
+layout = "NHWC"
+target = tvm.target.Target("cuda")
+dtype = "float32"
+log_file = "%s-%s-B%d.json" % (network, layout, batch_size)
+
+#################################################################
+# Extract Search Tasks
+# --------------------
+# Next, we extract the search tasks and their weights from a network.
+# The weight of a task is the number of appearances of the task's subgraph
+# in the whole network.
+# By using the weight, we can approximate the end-to-end latency of the network
+# as :code:`sum(latency[t] * weight[t])`, where :code:`latency[t]` is the
+# latency of a task and :code:`weight[t]` is the weight of the task.
+
+# Enable auto-scheduler in relay
+auto_scheduler.enable_relay_integration()
+
+# Extract tasks from the network
+print("Extract tasks...")
+mod, params, input_shape, output_shape = get_network(network, batch_size, layout, dtype=dtype)
+tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)
+
+# Define the objective as the end-to-end exeuction time of the network
+objective = lambda costs: sum(c * w for c, w in zip(costs, task_weights))
+
+#################################################################
+# Begin Tuning
+# ------------
+# Now, we set some options of tuning and launch the search tasks
+#
+# * :code:`measure_ctx` launches a different process for measurement to
+# provide isolation. It can protect the master process from GPU crashes
+# happened during measurement and avoid other runtime conflicts.
+# * :code:`min_repeat_ms` defines the minimum duration of one "repeat" in every measurement.
+# This can warmup the GPU, which is necessary to get accurate measurement results.
+# Typically, we recommend a value > 300 ms.
+# * :code:`num_measure_trials` is the number of measurement trials we can use during the tuning.
+# You can set it to a small number (e.g., 200) for a fast demonstrative run.
+# In practice, we recommend setting it round :code:`1000 * len(tasks)`,
+# which is typically enough for the search to converge.
+# For example, there are 21 tasks in resnet-18, so we can set it as 20000 for renset-18.
+# You can adjust this parameter according to your time budget.
+# * In addition, we use :code:`RecordToFile` to dump measurement records into the log file,
+# The measurement records can be used to query the history best, resume the search,
+# and do more analyses later.
+# * see :any:`auto_scheduler.TuningOptions`,
+# :any:`auto_scheduler.LocalRPCMeasureContext` for more parameters.
+#
+
+
+def run_tuning():
+ print("Begin tuning...")
+ measure_ctx = auto_scheduler.LocalRPCMeasureContext(repeat=1, min_repeat_ms=400, timeout=10)
+
+ tuner = auto_scheduler.TaskScheduler(tasks, objective)
+ tune_option = auto_scheduler.TuningOptions(
+ num_measure_trials=200, # change this to 20000 to achieve the best performance
+ runner=measure_ctx.runner,
+ measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
+ )
+
+ tuner.tune(tune_option)
+
+
+# We do not run the tuning in our webpage server since it takes too long.
+# Uncomment the following line to run it by yourself.
+
+# run_tuning()
+
+
+######################################################################
+# .. note:: Explain the printed information during tuning
+#
+# During the tuning, a lot of information will be printed on the screen.
+# They are used for debugging purposes. The most important info is the output
+# of the task scheduler. The following table is a sample output.
+#
+# .. code-block:: c
+#
+# ----------------------------------------------------------------------
+# ------------------------------ [ Task Scheduler ]
+# ----------------------------------------------------------------------
+# | ID | Latency (ms) | Speed (GFLOPS) | Trials |
+# -------------------------------------------------
+# | 0 | 0.014 | 72.07 | 64 |
+# | 1 | 0.185 | 1250.68 | 128 |
+# | 2 | 0.142 | 1626.36 | 192 |
+# | 3 | 0.137 | 1689.42 | 128 |
+# | 4 | 0.097 | 1189.75 | 128 |
+# | 5 | 0.092 | 2505.25 | 128 |
+# | 6 | 0.080 | 2893.08 | 128 |
+# | 7 | 0.119 | 1947.84 | 128 |
+# | 8 | 0.090 | 1292.62 | 64 |
+# | 9 | 0.107 | 2172.30 | 64 |
+# | 10 | 0.095 | 2439.36 | 64 |
+# | 11 | 0.077 | 3003.22 | 64 |
+# | 12 | 0.068 | 1695.13 | 64 |
+# | 13 | 0.058 | 3979.29 | 64 |
+# | 14 | 0.048 | 4859.95 | 128 |
+# | 15 | 0.073 | 3151.76 | 64 |
+# | 16 | 0.056 | 4265.94 | 64 |
+# | 17 | 0.009 | 2754.90 | 64 |
+# | 18 | 0.011 | 1156.08 | 64 |
+# | 19 | 0.013 | 955.80 | 64 |
+# | 20 | 0.029 | 437.71 | 64 |
+# -------------------------------------------------
+# Total latency: 1.649 ms Trials: 1920 Used time : 3598 s Next ID: 9
Review comment:
Maybe "Estimated total latency" could reduce some confusion. Otherwise I guess some people will directly refer to this number and find that it's inconsistent to the final evaluation.
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
+in :code:`tvm/python/topi` while does not use existing schedule templates.
+
+Note that this tutorial will not run on Windows or recent versions of macOS. To
+get it to run, you will need to wrap the body of this tutorial in a :code:`if
+__name__ == "__main__":` block.
+"""
+
+import numpy as np
+
+import tvm
+from tvm import relay, auto_scheduler
+import tvm.relay.testing
+from tvm.contrib import graph_runtime
+
+#################################################################
+# Define a Network
+# ----------------
+# First, we need to define the network in relay frontend API.
+# We can load some pre-defined network from :code:`tvm.relay.testing`.
+# We can also load models from MXNet, ONNX, PyTorch, and TensorFlow
+# (see :ref:`front end tutorials<tutorial-frontend>`).
+#
+# Note that although auto-scheduler can work with any layouts,
+# we found that the best performance is typically archived with NHWC layout
+# for convolutional neural networks.
+#
+
+
+def get_network(name, batch_size, layout="NHWC", dtype="float32"):
+ """Get the symbol definition and random weight of a network"""
+
+ # auto-scheduler prefers NHWC layout
+ if layout == "NHWC":
+ image_shape = (224, 224, 3)
+ elif layout == "NCHW":
+ image_shape = (3, 224, 224)
+ else:
+ raise ValueError("Invalid layout: " + layout)
+
+ input_shape = (batch_size,) + image_shape
+ output_shape = (batch_size, 1000)
+
+ if name.startswith("resnet-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name.startswith("resnet3d-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "mobilenet":
+ mod, params = relay.testing.mobilenet.get_workload(
+ batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape
+ )
+ elif name == "squeezenet_v1.1":
+ mod, params = relay.testing.squeezenet.get_workload(
+ version="1.1",
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "inception_v3":
+ input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)
+ mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
+ elif name == "mxnet":
+ # an example for mxnet model
+ from mxnet.gluon.model_zoo.vision import get_model
+
+ assert layout == "NCHW"
+
+ block = get_model("resnet18_v1", pretrained=True)
+ mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
+ net = mod["main"]
+ net = relay.Function(
+ net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
+ )
+ mod = tvm.IRModule.from_expr(net)
+
+ return mod, params, input_shape, output_shape
+
+
+# Define the neural network and compilation target
+network = "resnet-18"
+batch_size = 1
+layout = "NHWC"
+target = tvm.target.Target("cuda")
+dtype = "float32"
+log_file = "%s-%s-B%d.json" % (network, layout, batch_size)
+
+#################################################################
+# Extract Search Tasks
+# --------------------
+# Next, we extract the search tasks and their weights from a network.
+# The weight of a task is the number of appearances of the task's subgraph
+# in the whole network.
+# By using the weight, we can approximate the end-to-end latency of the network
+# as :code:`sum(latency[t] * weight[t])`, where :code:`latency[t]` is the
+# latency of a task and :code:`weight[t]` is the weight of the task.
+
+# Enable auto-scheduler in relay
+auto_scheduler.enable_relay_integration()
+
+# Extract tasks from the network
+print("Extract tasks...")
+mod, params, input_shape, output_shape = get_network(network, batch_size, layout, dtype=dtype)
+tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)
+
+# Define the objective as the end-to-end exeuction time of the network
+objective = lambda costs: sum(c * w for c, w in zip(costs, task_weights))
+
+#################################################################
+# Begin Tuning
+# ------------
+# Now, we set some options of tuning and launch the search tasks
+#
+# * :code:`measure_ctx` launches a different process for measurement to
+# provide isolation. It can protect the master process from GPU crashes
+# happened during measurement and avoid other runtime conflicts.
+# * :code:`min_repeat_ms` defines the minimum duration of one "repeat" in every measurement.
+# This can warmup the GPU, which is necessary to get accurate measurement results.
+# Typically, we recommend a value > 300 ms.
+# * :code:`num_measure_trials` is the number of measurement trials we can use during the tuning.
+# You can set it to a small number (e.g., 200) for a fast demonstrative run.
+# In practice, we recommend setting it round :code:`1000 * len(tasks)`,
+# which is typically enough for the search to converge.
+# For example, there are 21 tasks in resnet-18, so we can set it as 20000 for renset-18.
+# You can adjust this parameter according to your time budget.
+# * In addition, we use :code:`RecordToFile` to dump measurement records into the log file,
+# The measurement records can be used to query the history best, resume the search,
+# and do more analyses later.
+# * see :any:`auto_scheduler.TuningOptions`,
+# :any:`auto_scheduler.LocalRPCMeasureContext` for more parameters.
+#
+
+
+def run_tuning():
+ print("Begin tuning...")
+ measure_ctx = auto_scheduler.LocalRPCMeasureContext(repeat=1, min_repeat_ms=400, timeout=10)
+
+ tuner = auto_scheduler.TaskScheduler(tasks, objective)
+ tune_option = auto_scheduler.TuningOptions(
+ num_measure_trials=200, # change this to 20000 to achieve the best performance
+ runner=measure_ctx.runner,
+ measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
+ )
+
+ tuner.tune(tune_option)
+
+
+# We do not run the tuning in our webpage server since it takes too long.
+# Uncomment the following line to run it by yourself.
+
+# run_tuning()
+
+
+######################################################################
+# .. note:: Explain the printed information during tuning
+#
+# During the tuning, a lot of information will be printed on the screen.
Review comment:
```suggestion
# During the tuning, a lot of information will be printed on the console.
```
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
+in :code:`tvm/python/topi` while does not use existing schedule templates.
+
+Note that this tutorial will not run on Windows or recent versions of macOS. To
+get it to run, you will need to wrap the body of this tutorial in a :code:`if
+__name__ == "__main__":` block.
+"""
+
+import numpy as np
+
+import tvm
+from tvm import relay, auto_scheduler
+import tvm.relay.testing
+from tvm.contrib import graph_runtime
+
+#################################################################
+# Define a Network
+# ----------------
+# First, we need to define the network in relay frontend API.
+# We can load some pre-defined network from :code:`tvm.relay.testing`.
+# We can also load models from MXNet, ONNX, PyTorch, and TensorFlow
+# (see :ref:`front end tutorials<tutorial-frontend>`).
+#
+# Note that although auto-scheduler can work with any layouts,
+# we found that the best performance is typically archived with NHWC layout
+# for convolutional neural networks.
+#
+
+
+def get_network(name, batch_size, layout="NHWC", dtype="float32"):
+ """Get the symbol definition and random weight of a network"""
+
+ # auto-scheduler prefers NHWC layout
+ if layout == "NHWC":
+ image_shape = (224, 224, 3)
+ elif layout == "NCHW":
+ image_shape = (3, 224, 224)
+ else:
+ raise ValueError("Invalid layout: " + layout)
+
+ input_shape = (batch_size,) + image_shape
+ output_shape = (batch_size, 1000)
+
+ if name.startswith("resnet-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name.startswith("resnet3d-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "mobilenet":
+ mod, params = relay.testing.mobilenet.get_workload(
+ batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape
+ )
+ elif name == "squeezenet_v1.1":
+ mod, params = relay.testing.squeezenet.get_workload(
+ version="1.1",
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "inception_v3":
+ input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)
+ mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
+ elif name == "mxnet":
+ # an example for mxnet model
+ from mxnet.gluon.model_zoo.vision import get_model
+
+ assert layout == "NCHW"
+
+ block = get_model("resnet18_v1", pretrained=True)
+ mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
+ net = mod["main"]
+ net = relay.Function(
+ net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
+ )
+ mod = tvm.IRModule.from_expr(net)
+
+ return mod, params, input_shape, output_shape
+
+
+# Define the neural network and compilation target
+network = "resnet-18"
+batch_size = 1
+layout = "NHWC"
+target = tvm.target.Target("cuda")
+dtype = "float32"
+log_file = "%s-%s-B%d.json" % (network, layout, batch_size)
+
+#################################################################
+# Extract Search Tasks
+# --------------------
+# Next, we extract the search tasks and their weights from a network.
+# The weight of a task is the number of appearances of the task's subgraph
+# in the whole network.
+# By using the weight, we can approximate the end-to-end latency of the network
+# as :code:`sum(latency[t] * weight[t])`, where :code:`latency[t]` is the
+# latency of a task and :code:`weight[t]` is the weight of the task.
+
+# Enable auto-scheduler in relay
+auto_scheduler.enable_relay_integration()
+
+# Extract tasks from the network
+print("Extract tasks...")
+mod, params, input_shape, output_shape = get_network(network, batch_size, layout, dtype=dtype)
+tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)
+
+# Define the objective as the end-to-end exeuction time of the network
+objective = lambda costs: sum(c * w for c, w in zip(costs, task_weights))
+
+#################################################################
+# Begin Tuning
+# ------------
+# Now, we set some options of tuning and launch the search tasks
+#
+# * :code:`measure_ctx` launches a different process for measurement to
+# provide isolation. It can protect the master process from GPU crashes
+# happened during measurement and avoid other runtime conflicts.
+# * :code:`min_repeat_ms` defines the minimum duration of one "repeat" in every measurement.
+# This can warmup the GPU, which is necessary to get accurate measurement results.
+# Typically, we recommend a value > 300 ms.
+# * :code:`num_measure_trials` is the number of measurement trials we can use during the tuning.
+# You can set it to a small number (e.g., 200) for a fast demonstrative run.
+# In practice, we recommend setting it round :code:`1000 * len(tasks)`,
+# which is typically enough for the search to converge.
+# For example, there are 21 tasks in resnet-18, so we can set it as 20000 for renset-18.
+# You can adjust this parameter according to your time budget.
+# * In addition, we use :code:`RecordToFile` to dump measurement records into the log file,
+# The measurement records can be used to query the history best, resume the search,
+# and do more analyses later.
+# * see :any:`auto_scheduler.TuningOptions`,
+# :any:`auto_scheduler.LocalRPCMeasureContext` for more parameters.
+#
+
+
+def run_tuning():
+ print("Begin tuning...")
+ measure_ctx = auto_scheduler.LocalRPCMeasureContext(repeat=1, min_repeat_ms=400, timeout=10)
+
+ tuner = auto_scheduler.TaskScheduler(tasks, objective)
+ tune_option = auto_scheduler.TuningOptions(
+ num_measure_trials=200, # change this to 20000 to achieve the best performance
+ runner=measure_ctx.runner,
+ measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
+ )
+
+ tuner.tune(tune_option)
+
+
+# We do not run the tuning in our webpage server since it takes too long.
+# Uncomment the following line to run it by yourself.
+
+# run_tuning()
+
+
+######################################################################
+# .. note:: Explain the printed information during tuning
+#
+# During the tuning, a lot of information will be printed on the screen.
+# They are used for debugging purposes. The most important info is the output
+# of the task scheduler. The following table is a sample output.
+#
+# .. code-block:: c
+#
+# ----------------------------------------------------------------------
+# ------------------------------ [ Task Scheduler ]
+# ----------------------------------------------------------------------
+# | ID | Latency (ms) | Speed (GFLOPS) | Trials |
+# -------------------------------------------------
+# | 0 | 0.014 | 72.07 | 64 |
+# | 1 | 0.185 | 1250.68 | 128 |
+# | 2 | 0.142 | 1626.36 | 192 |
+# | 3 | 0.137 | 1689.42 | 128 |
+# | 4 | 0.097 | 1189.75 | 128 |
+# | 5 | 0.092 | 2505.25 | 128 |
+# | 6 | 0.080 | 2893.08 | 128 |
+# | 7 | 0.119 | 1947.84 | 128 |
+# | 8 | 0.090 | 1292.62 | 64 |
+# | 9 | 0.107 | 2172.30 | 64 |
+# | 10 | 0.095 | 2439.36 | 64 |
+# | 11 | 0.077 | 3003.22 | 64 |
+# | 12 | 0.068 | 1695.13 | 64 |
+# | 13 | 0.058 | 3979.29 | 64 |
+# | 14 | 0.048 | 4859.95 | 128 |
+# | 15 | 0.073 | 3151.76 | 64 |
+# | 16 | 0.056 | 4265.94 | 64 |
+# | 17 | 0.009 | 2754.90 | 64 |
+# | 18 | 0.011 | 1156.08 | 64 |
+# | 19 | 0.013 | 955.80 | 64 |
+# | 20 | 0.029 | 437.71 | 64 |
+# -------------------------------------------------
+# Total latency: 1.649 ms Trials: 1920 Used time : 3598 s Next ID: 9
+#
+# This table lists the latency and speed of all tasks.
+# It also lists the allocation of measurement trials for all tasks.
+# The last line prints the total weighted latency of these tasks,
+# which can be a rough estimation of the end-to-end execution time
+# of the network.
+# The last line also prints the total number of measurement trials,
+# total time spent on auto-tuning and the id of the next task to tune.
+#
+# There will also be some "dmlc::Error"s and CUDA errors. You can safely
+# ignore them if the tuning can continue.
+
+######################################################################
+# .. note:: Terminate the tuning earlier
+#
+# You can terminate the tuning earlier by forcely killing this process.
+# As long as you get at least one valid schedule for each task in the log file,
+# you should be able to do the compilation (the secion below).
+#
+
+#################################################################
+# Compile and Evaluate
+# --------------------
+# After auto-tuning, we can compile the network with the best schedules we found.
+# All measurement records are dumpled into the log file during auto-tuning,
+# so we can read the log file and load the best schedules.
Review comment:
Better to also mention what happen and what messages you will see if there is no valid schedules in the log file.
##########
File path: tutorials/auto_scheduler/tune_network_cuda.py
##########
@@ -0,0 +1,286 @@
+# 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.
+"""
+Auto-tuning a Neural Network for NVIDIA GPU
+==================================================
+**Author**: `Lianmin Zheng <https://github.com/merrymercy>`_
+
+Auto-tuning for specific devices and workloads is critical for getting the
+best performance. This is a tutorial on how to tune a whole neural
+network for NVIDIA GPU with the auto-scheduler.
+
+To auto-tune a neural network, we partition the network into small subgraphs and
+tune them independently. Each subgraph is treated as one search task.
+A task scheduler slices the time and dynamically allocates time resources to
+these tasks. The task scheduler predicts the impact of each task on the end-to-end
+execution time and prioritizes the one that can reduce the execution time fastest.
+
+For each subgraph, we use the compute declaration in :code:`tvm/python/topi` to
+get the computational DAG in the tensor expression form.
+We then use the auto-scheduler to construct a search space of this DAG and search
+for good schedules (low-level optimizations).
+
+Different from the template-based :ref:`autotvm <tutorials-autotvm-sec>` which relies on
+manual templates to define the search space, the auto-scheduler does not require any
+schedule templates. So the auto-scheduler only uses the compute declarations
+in :code:`tvm/python/topi` while does not use existing schedule templates.
+
+Note that this tutorial will not run on Windows or recent versions of macOS. To
+get it to run, you will need to wrap the body of this tutorial in a :code:`if
+__name__ == "__main__":` block.
+"""
+
+import numpy as np
+
+import tvm
+from tvm import relay, auto_scheduler
+import tvm.relay.testing
+from tvm.contrib import graph_runtime
+
+#################################################################
+# Define a Network
+# ----------------
+# First, we need to define the network in relay frontend API.
+# We can load some pre-defined network from :code:`tvm.relay.testing`.
+# We can also load models from MXNet, ONNX, PyTorch, and TensorFlow
+# (see :ref:`front end tutorials<tutorial-frontend>`).
+#
+# Note that although auto-scheduler can work with any layouts,
+# we found that the best performance is typically archived with NHWC layout
+# for convolutional neural networks.
+#
+
+
+def get_network(name, batch_size, layout="NHWC", dtype="float32"):
+ """Get the symbol definition and random weight of a network"""
+
+ # auto-scheduler prefers NHWC layout
+ if layout == "NHWC":
+ image_shape = (224, 224, 3)
+ elif layout == "NCHW":
+ image_shape = (3, 224, 224)
+ else:
+ raise ValueError("Invalid layout: " + layout)
+
+ input_shape = (batch_size,) + image_shape
+ output_shape = (batch_size, 1000)
+
+ if name.startswith("resnet-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name.startswith("resnet3d-"):
+ n_layer = int(name.split("-")[1])
+ mod, params = relay.testing.resnet.get_workload(
+ num_layers=n_layer,
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "mobilenet":
+ mod, params = relay.testing.mobilenet.get_workload(
+ batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape
+ )
+ elif name == "squeezenet_v1.1":
+ mod, params = relay.testing.squeezenet.get_workload(
+ version="1.1",
+ batch_size=batch_size,
+ layout=layout,
+ dtype=dtype,
+ image_shape=image_shape,
+ )
+ elif name == "inception_v3":
+ input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)
+ mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
+ elif name == "mxnet":
+ # an example for mxnet model
+ from mxnet.gluon.model_zoo.vision import get_model
+
+ assert layout == "NCHW"
+
+ block = get_model("resnet18_v1", pretrained=True)
+ mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
+ net = mod["main"]
+ net = relay.Function(
+ net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
+ )
+ mod = tvm.IRModule.from_expr(net)
+
+ return mod, params, input_shape, output_shape
+
+
+# Define the neural network and compilation target
+network = "resnet-18"
+batch_size = 1
+layout = "NHWC"
+target = tvm.target.Target("cuda")
+dtype = "float32"
+log_file = "%s-%s-B%d.json" % (network, layout, batch_size)
+
+#################################################################
+# Extract Search Tasks
+# --------------------
+# Next, we extract the search tasks and their weights from a network.
+# The weight of a task is the number of appearances of the task's subgraph
+# in the whole network.
+# By using the weight, we can approximate the end-to-end latency of the network
+# as :code:`sum(latency[t] * weight[t])`, where :code:`latency[t]` is the
+# latency of a task and :code:`weight[t]` is the weight of the task.
+
+# Enable auto-scheduler in relay
+auto_scheduler.enable_relay_integration()
+
+# Extract tasks from the network
+print("Extract tasks...")
+mod, params, input_shape, output_shape = get_network(network, batch_size, layout, dtype=dtype)
+tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)
+
+# Define the objective as the end-to-end exeuction time of the network
+objective = lambda costs: sum(c * w for c, w in zip(costs, task_weights))
+
+#################################################################
+# Begin Tuning
+# ------------
+# Now, we set some options of tuning and launch the search tasks
+#
+# * :code:`measure_ctx` launches a different process for measurement to
+# provide isolation. It can protect the master process from GPU crashes
+# happened during measurement and avoid other runtime conflicts.
+# * :code:`min_repeat_ms` defines the minimum duration of one "repeat" in every measurement.
+# This can warmup the GPU, which is necessary to get accurate measurement results.
+# Typically, we recommend a value > 300 ms.
+# * :code:`num_measure_trials` is the number of measurement trials we can use during the tuning.
+# You can set it to a small number (e.g., 200) for a fast demonstrative run.
+# In practice, we recommend setting it round :code:`1000 * len(tasks)`,
+# which is typically enough for the search to converge.
+# For example, there are 21 tasks in resnet-18, so we can set it as 20000 for renset-18.
+# You can adjust this parameter according to your time budget.
+# * In addition, we use :code:`RecordToFile` to dump measurement records into the log file,
+# The measurement records can be used to query the history best, resume the search,
+# and do more analyses later.
+# * see :any:`auto_scheduler.TuningOptions`,
+# :any:`auto_scheduler.LocalRPCMeasureContext` for more parameters.
+#
+
+
+def run_tuning():
+ print("Begin tuning...")
+ measure_ctx = auto_scheduler.LocalRPCMeasureContext(repeat=1, min_repeat_ms=400, timeout=10)
+
+ tuner = auto_scheduler.TaskScheduler(tasks, objective)
+ tune_option = auto_scheduler.TuningOptions(
+ num_measure_trials=200, # change this to 20000 to achieve the best performance
+ runner=measure_ctx.runner,
+ measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
+ )
+
+ tuner.tune(tune_option)
+
+
+# We do not run the tuning in our webpage server since it takes too long.
+# Uncomment the following line to run it by yourself.
+
+# run_tuning()
+
+
+######################################################################
+# .. note:: Explain the printed information during tuning
+#
+# During the tuning, a lot of information will be printed on the screen.
+# They are used for debugging purposes. The most important info is the output
+# of the task scheduler. The following table is a sample output.
+#
+# .. code-block:: c
+#
+# ----------------------------------------------------------------------
+# ------------------------------ [ Task Scheduler ]
+# ----------------------------------------------------------------------
+# | ID | Latency (ms) | Speed (GFLOPS) | Trials |
+# -------------------------------------------------
+# | 0 | 0.014 | 72.07 | 64 |
+# | 1 | 0.185 | 1250.68 | 128 |
+# | 2 | 0.142 | 1626.36 | 192 |
+# | 3 | 0.137 | 1689.42 | 128 |
+# | 4 | 0.097 | 1189.75 | 128 |
+# | 5 | 0.092 | 2505.25 | 128 |
+# | 6 | 0.080 | 2893.08 | 128 |
+# | 7 | 0.119 | 1947.84 | 128 |
+# | 8 | 0.090 | 1292.62 | 64 |
+# | 9 | 0.107 | 2172.30 | 64 |
+# | 10 | 0.095 | 2439.36 | 64 |
+# | 11 | 0.077 | 3003.22 | 64 |
+# | 12 | 0.068 | 1695.13 | 64 |
+# | 13 | 0.058 | 3979.29 | 64 |
+# | 14 | 0.048 | 4859.95 | 128 |
+# | 15 | 0.073 | 3151.76 | 64 |
+# | 16 | 0.056 | 4265.94 | 64 |
+# | 17 | 0.009 | 2754.90 | 64 |
+# | 18 | 0.011 | 1156.08 | 64 |
+# | 19 | 0.013 | 955.80 | 64 |
+# | 20 | 0.029 | 437.71 | 64 |
+# -------------------------------------------------
+# Total latency: 1.649 ms Trials: 1920 Used time : 3598 s Next ID: 9
+#
+# This table lists the latency and speed of all tasks.
+# It also lists the allocation of measurement trials for all tasks.
+# The last line prints the total weighted latency of these tasks,
+# which can be a rough estimation of the end-to-end execution time
+# of the network.
+# The last line also prints the total number of measurement trials,
+# total time spent on auto-tuning and the id of the next task to tune.
+#
+# There will also be some "dmlc::Error"s and CUDA errors. You can safely
+# ignore them if the tuning can continue.
+
+######################################################################
+# .. note:: Terminate the tuning earlier
+#
+# You can terminate the tuning earlier by forcely killing this process.
Review comment:
Since we use fork and RPC mechanism during tuning, simply ctrl+c may cause zombie processes remaining. Maybe we should suggest a way to terminate the process as a good practice.
----------------------------------------------------------------
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