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 2020/01/09 15:31:49 UTC
[GitHub] [incubator-tvm] vinx13 opened a new pull request #4667: [Tutorial]
Deploy Quantized Model on CUDA
vinx13 opened a new pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667
This tutorial demonstrates how to import a model using Relay frontend, run quantization and calibration passes, and perform quantized inference.
ref https://github.com/apache/incubator-tvm/issues/4435
cc @tqchen @masahi @anijain2305 @ZihengJiang @tmoreau89
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] masahi commented on a change in pull request #4667:
[Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
masahi commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r364966229
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorials, we
+# only use a few samples for calibration.
+
+calibration_samples = 10
+
+def calibrate_dataset():
+ val_data, batch_fn = get_val_data()
+ val_data.reset()
+ for i, batch in enumerate(val_data):
+ if i * batch_size >= calibration_samples:
+ break
+ data, _ = batch_fn(batch)
+ yield {'data': data}
+
+
+###############################################################################
+# Import the model
+# ----------------
+# We use the Relay MxNet frontent to import a model from the Gluon model zoo.
Review comment:
frontend
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] tmoreau89 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
tmoreau89 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365398358
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,156 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+Automatic quantization is one of the quantization mode in TVM. More details of the quantization story in TVM can be found `here <https://discuss.tvm.ai/t/quantization-story/3920>`_.
+In this tutorial, we will import a GluonCV pre-trained model on ImageNet to
+Relay, quantize the Relay model and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorial, we
+# only use a few samples for calibration.
+
+calibration_samples = 10
+
+def calibrate_dataset():
+ val_data, batch_fn = get_val_data()
+ val_data.reset()
+ for i, batch in enumerate(val_data):
+ if i * batch_size >= calibration_samples:
+ break
+ data, _ = batch_fn(batch)
+ yield {'data': data}
+
+
+###############################################################################
+# Import the model
+# ----------------
+# We use the Relay MxNet frontend to import a model from the Gluon model zoo.
+def get_model():
+ gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ data_shape = (batch_size, 3, img_size, img_size)
+ mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
+ return mod, params
+
+
+###############################################################################
+# Quantize the Model
+# ------------------
+# In quantization, we need to find the scale for each weight and intermediate
+# feature map tensor of each layer.
+#
+# For weights, the scales are directly calculated based on the value of the
+# weights. Two modes are supported: `power2` and `max`. Both modes find the
+# maximum value within the weight tensor first. In `power2` mode, the maximum
+# is rounded down to power of two. If the scales of both weights and
+# intermediate feature maps are power of two, we can leverage bit shifting for
+# multiplications. This make it computationally more efficient. In `max` mode,
+# the maximum is used as the scale. Without rounding, `max` mode might have
+# better accuracy in some cases. When the scales are not power of two, fixed
Review comment:
powers
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] tmoreau89 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
tmoreau89 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365397553
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,156 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+Automatic quantization is one of the quantization mode in TVM. More details of the quantization story in TVM can be found `here <https://discuss.tvm.ai/t/quantization-story/3920>`_.
Review comment:
mode -> modes
details of -> details on
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r364867350
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorials, we
Review comment:
tutorials -> tutorial
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] tmoreau89 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
tmoreau89 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365398006
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,156 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+Automatic quantization is one of the quantization mode in TVM. More details of the quantization story in TVM can be found `here <https://discuss.tvm.ai/t/quantization-story/3920>`_.
+In this tutorial, we will import a GluonCV pre-trained model on ImageNet to
+Relay, quantize the Relay model and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
Review comment:
"should be an"
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] masahi commented on a change in pull request #4667:
[Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
masahi commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r364973536
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorials, we
+# only use a few samples for calibration.
+
+calibration_samples = 10
+
+def calibrate_dataset():
+ val_data, batch_fn = get_val_data()
+ val_data.reset()
+ for i, batch in enumerate(val_data):
+ if i * batch_size >= calibration_samples:
+ break
+ data, _ = batch_fn(batch)
+ yield {'data': data}
+
+
+###############################################################################
+# Import the model
+# ----------------
+# We use the Relay MxNet frontent to import a model from the Gluon model zoo.
+def get_model():
+ gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ data_shape = (batch_size, 3, img_size, img_size)
+ mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
+ return mod, params
+
+
+###############################################################################
+# Quantize the Model
+# ------------------
+# In quantization, we need to find the scale for each weight and output tensor.
+# For weights, the scales are directly calculated based on the value of the
+# weights. Two modes are supported: `power2` and `max`. Both modes find the
+# maximum value within the weight tensor first. In `power2` mode, the maximum
+# is rounded down to power of two. If the scales of both weights and outputs
+# are power of two, we can leverage bit shifting for multiplications. This make
+# it computationally more efficient. In `max` mode, the maximum is used as the
+# scale. Without rounding, `max` mode might have better accuracy in some cases.
+# When the scales are not power of two, fixed point multiplications will
+# be used.
+#
+# For outputs, we can find the scales with data-aware quantization.
+# Data-aware quantization takes a calibration dataset as the input argument.
+# Scales are calculated by minimizing the KL divergence of between the data
Review comment:
of between -> between
I'd prefer "KL divergence between distribution of activation before and after quantization"
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] vinx13 commented on a change in pull request #4667:
[Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
vinx13 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365245824
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorials, we
+# only use a few samples for calibration.
+
+calibration_samples = 10
+
+def calibrate_dataset():
+ val_data, batch_fn = get_val_data()
+ val_data.reset()
+ for i, batch in enumerate(val_data):
+ if i * batch_size >= calibration_samples:
+ break
+ data, _ = batch_fn(batch)
+ yield {'data': data}
+
+
+###############################################################################
+# Import the model
+# ----------------
+# We use the Relay MxNet frontent to import a model from the Gluon model zoo.
+def get_model():
+ gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ data_shape = (batch_size, 3, img_size, img_size)
+ mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
+ return mod, params
+
+
+###############################################################################
+# Quantize the Model
+# ------------------
+# In quantization, we need to find the scale for each weight and output tensor.
+# For weights, the scales are directly calculated based on the value of the
+# weights. Two modes are supported: `power2` and `max`. Both modes find the
+# maximum value within the weight tensor first. In `power2` mode, the maximum
+# is rounded down to power of two. If the scales of both weights and outputs
+# are power of two, we can leverage bit shifting for multiplications. This make
+# it computationally more efficient. In `max` mode, the maximum is used as the
+# scale. Without rounding, `max` mode might have better accuracy in some cases.
+# When the scales are not power of two, fixed point multiplications will
+# be used.
+#
+# For outputs, we can find the scales with data-aware quantization.
+# Data-aware quantization takes a calibration dataset as the input argument.
+# Scales are calculated by minimizing the KL divergence of between the data
Review comment:
it is not necessarily 'activation', so i updated to 'output of each layer'
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r364868831
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
Review comment:
Not used
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r364867851
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorials, we
+# only use a few samples for calibration.
+
+calibration_samples = 10
+
+def calibrate_dataset():
+ val_data, batch_fn = get_val_data()
+ val_data.reset()
+ for i, batch in enumerate(val_data):
+ if i * batch_size >= calibration_samples:
+ break
+ data, _ = batch_fn(batch)
+ yield {'data': data}
+
+
+###############################################################################
+# Import the model
+# ----------------
+# We use the Relay MxNet frontent to import a model from the Gluon model zoo.
+def get_model():
+ gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ data_shape = (batch_size, 3, img_size, img_size)
+ mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
+ return mod, params
+
+
+###############################################################################
+# Quantize the Model
+# ------------------
+# In quantization, we need to find the scale for each weight and output tensor.
+# For weights, the scales are directly calculated based on the value of the
+# weights. Two modes are supported: `power2` and `max`. Both modes find the
+# maximum value within the weight tensor first. In `power2` mode, the maximum
+# is rounded down to power of two. If the scales of both weights and outputs
+# are power of two, we can leverage bit shifting for multiplications. This make
+# it computationally more efficient. In `max` mode, the maximum is used as the
+# scale. Without rounding, `max` mode might have better accuracy in some cases.
+# When the scales are not power of two, fixed point multiplications will
+# be used.
+#
+# For outputs, we can find the scales with data-aware quantization.
Review comment:
outputs --> intermediate feature maps
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] masahi merged pull request #4667: [Tutorial] Deploy
Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
masahi merged pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] tmoreau89 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
tmoreau89 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365397868
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,156 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+Automatic quantization is one of the quantization mode in TVM. More details of the quantization story in TVM can be found `here <https://discuss.tvm.ai/t/quantization-story/3920>`_.
+In this tutorial, we will import a GluonCV pre-trained model on ImageNet to
+Relay, quantize the Relay model and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
Review comment:
validation set
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] tmoreau89 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
tmoreau89 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365397726
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,156 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+Automatic quantization is one of the quantization mode in TVM. More details of the quantization story in TVM can be found `here <https://discuss.tvm.ai/t/quantization-story/3920>`_.
Review comment:
also long line
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] masahi commented on issue #4667: [Tutorial] Deploy
Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
masahi commented on issue #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#issuecomment-573352336
Thanks @vinx13 @anijain2305 @tmoreau89
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] masahi commented on a change in pull request #4667:
[Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
masahi commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r365293197
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
+We will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model
+and then perform the inference.
+"""
+
+import tvm
+from tvm import relay
+from tvm.relay import quantize as qtz
+import mxnet as mx
+from tvm.contrib.download import download_testdata
+from mxnet import gluon
+import logging
+import os
+
+batch_size = 1
+model_name = "resnet18_v1"
+target = 'cuda'
+ctx = tvm.context(target)
+
+###############################################################################
+# Prepare the Dataset
+# -------------------
+# We will demonstrate how to prepare the calibration dataset for quantization.
+# We first download the validate set of ImageNet and pre-process the dataset.
+calibration_rec = download_testdata(
+ 'http://data.mxnet.io.s3-website-us-west-1.amazonaws.com/data/val_256_q90.rec',
+ 'val_256_q90.rec')
+
+def get_val_data(num_workers=4):
+ mean_rgb = [123.68, 116.779, 103.939]
+ std_rgb = [58.393, 57.12, 57.375]
+
+ def batch_fn(batch):
+ return batch.data[0].asnumpy(), batch.label[0].asnumpy()
+
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ val_data = mx.io.ImageRecordIter(
+ path_imgrec=calibration_rec,
+ preprocess_threads=num_workers,
+ shuffle=False,
+ batch_size=batch_size,
+ resize=256,
+ data_shape=(3, img_size, img_size),
+ mean_r=mean_rgb[0],
+ mean_g=mean_rgb[1],
+ mean_b=mean_rgb[2],
+ std_r=std_rgb[0],
+ std_g=std_rgb[1],
+ std_b=std_rgb[2],
+ )
+ return val_data, batch_fn
+
+
+###############################################################################
+# The calibration dataset should be a iterable object. We define the
+# calibration dataset as a generator object in Python. In this tutorials, we
+# only use a few samples for calibration.
+
+calibration_samples = 10
+
+def calibrate_dataset():
+ val_data, batch_fn = get_val_data()
+ val_data.reset()
+ for i, batch in enumerate(val_data):
+ if i * batch_size >= calibration_samples:
+ break
+ data, _ = batch_fn(batch)
+ yield {'data': data}
+
+
+###############################################################################
+# Import the model
+# ----------------
+# We use the Relay MxNet frontent to import a model from the Gluon model zoo.
+def get_model():
+ gluon_model = gluon.model_zoo.vision.get_model(model_name, pretrained=True)
+ img_size = 299 if model_name == 'inceptionv3' else 224
+ data_shape = (batch_size, 3, img_size, img_size)
+ mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
+ return mod, params
+
+
+###############################################################################
+# Quantize the Model
+# ------------------
+# In quantization, we need to find the scale for each weight and output tensor.
+# For weights, the scales are directly calculated based on the value of the
+# weights. Two modes are supported: `power2` and `max`. Both modes find the
+# maximum value within the weight tensor first. In `power2` mode, the maximum
+# is rounded down to power of two. If the scales of both weights and outputs
+# are power of two, we can leverage bit shifting for multiplications. This make
+# it computationally more efficient. In `max` mode, the maximum is used as the
+# scale. Without rounding, `max` mode might have better accuracy in some cases.
+# When the scales are not power of two, fixed point multiplications will
+# be used.
+#
+# For outputs, we can find the scales with data-aware quantization.
+# Data-aware quantization takes a calibration dataset as the input argument.
+# Scales are calculated by minimizing the KL divergence of between the data
Review comment:
By activation I do mean 'output of each layer' or intermediate feature map, not necessary the output of activation function like relu. I think this is a standard terminology in the literature (they talk about "weight 8bit, acitvation 32 bit" etc).
----------------------------------------------------------------
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
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4667: [Tutorial] Deploy Quantized Model on CUDA
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4667: [Tutorial] Deploy Quantized Model on CUDA
URL: https://github.com/apache/incubator-tvm/pull/4667#discussion_r364866984
##########
File path: tutorials/frontend/deploy_quantized.py
##########
@@ -0,0 +1,154 @@
+# 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.
+"""
+Deploy a Quantized Model on Cuda
+================================
+**Author**: `Wuwei Lin <https://github.com/vinx13>`_
+
+This article is an introductory tutorial of automatic quantization with TVM.
Review comment:
Maybe add a link to this discuss forum - https://discuss.tvm.ai/t/quantization-story/3920 to give a high-level idea of whats automatic quantization.
----------------------------------------------------------------
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
With regards,
Apache Git Services