Merge pull request #1068 from IRDonch/quantizer

Roman Donchenko · web-flow · commit 45dc60a09eb8 · 2020-04-27T17:21:07.000+03:00
Add a model quantizer script
diff --git a/models/public/alexnet/model.yml b/models/public/alexnet/model.yml
@@ -46,4 +46,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/alexnet.caffemodel
   - --input_proto=$dl_dir/alexnet.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/BVLC/caffe/master/models/bvlc_alexnet/readme.md
diff --git a/models/public/googlenet-v1/model.yml b/models/public/googlenet-v1/model.yml
@@ -47,4 +47,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/googlenet-v1.caffemodel
   - --input_proto=$dl_dir/googlenet-v1.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/BVLC/caffe/master/models/bvlc_googlenet/readme.md
diff --git a/models/public/googlenet-v3-pytorch/model.yml b/models/public/googlenet-v3-pytorch/model.yml
@@ -45,4 +45,5 @@ model_optimizer_args:
   - --reverse_input_channels
   - --output=prob
   - --input_model=$conv_dir/googlenet-v3.onnx
+quantizable: yes
 license: https://raw.githubusercontent.com/pytorch/vision/master/LICENSE
diff --git a/models/public/mobilenet-v1-1.0-224/model.yml b/models/public/mobilenet-v1-1.0-224/model.yml
@@ -37,4 +37,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/mobilenet-v1-1.0-224.caffemodel
   - --input_proto=$dl_dir/mobilenet-v1-1.0-224.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/shicai/MobileNet-Caffe/26a8b8c0afb6114a07c1c9e4f550e4e0dd8cced1/LICENSE
diff --git a/models/public/mobilenet-v2-pytorch/model.yml b/models/public/mobilenet-v2-pytorch/model.yml
@@ -52,4 +52,5 @@ model_optimizer_args:
   - --reverse_input_channels
   - --output=prob
   - --input_model=$conv_dir/mobilenet-v2.onnx
+quantizable: yes
 license: https://raw.githubusercontent.com/tonylins/pytorch-mobilenet-v2/master/LICENSE
diff --git a/models/public/mobilenet-v2/model.yml b/models/public/mobilenet-v2/model.yml
@@ -33,4 +33,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/mobilenet-v2.caffemodel
   - --input_proto=$dl_dir/mobilenet-v2.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/shicai/MobileNet-Caffe/26a8b8c0afb6114a07c1c9e4f550e4e0dd8cced1/LICENSE
diff --git a/models/public/resnet-101/model.yml b/models/public/resnet-101/model.yml
@@ -32,4 +32,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/resnet-101.caffemodel
   - --input_proto=$dl_dir/resnet-101.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/KaimingHe/deep-residual-networks/master/LICENSE
diff --git a/models/public/resnet-152/model.yml b/models/public/resnet-152/model.yml
@@ -32,4 +32,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/resnet-152.caffemodel
   - --input_proto=$dl_dir/resnet-152.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/KaimingHe/deep-residual-networks/master/LICENSE
diff --git a/models/public/resnet-18-pytorch/model.yml b/models/public/resnet-18-pytorch/model.yml
@@ -45,4 +45,5 @@ model_optimizer_args:
   - --reverse_input_channels
   - --output=prob
   - --input_model=$conv_dir/resnet-18-pytorch.onnx
+quantizable: yes
 license: https://raw.githubusercontent.com/pytorch/vision/master/LICENSE
diff --git a/models/public/resnet-50-pytorch/model.yml b/models/public/resnet-50-pytorch/model.yml
@@ -45,4 +45,5 @@ model_optimizer_args:
   - --reverse_input_channels
   - --output=prob
   - --input_model=$conv_dir/resnet-v1-50.onnx
+quantizable: yes
 license: https://raw.githubusercontent.com/pytorch/vision/master/LICENSE
diff --git a/models/public/resnet-50/model.yml b/models/public/resnet-50/model.yml
@@ -32,4 +32,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/resnet-50.caffemodel
   - --input_proto=$dl_dir/resnet-50.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/KaimingHe/deep-residual-networks/master/LICENSE
diff --git a/models/public/squeezenet1.1-caffe2/model.yml b/models/public/squeezenet1.1-caffe2/model.yml
@@ -39,4 +39,5 @@ model_optimizer_args:
   - --input=data
   - --mean_values=data[103.96,116.78,123.68]
   - --input_model=$conv_dir/squeezenet1.1-caffe2.onnx
+quantizable: yes
 license: https://raw.githubusercontent.com/caffe2/models/master/LICENSE
diff --git a/models/public/squeezenet1.1/model.yml b/models/public/squeezenet1.1/model.yml
@@ -48,4 +48,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/squeezenet1.1.caffemodel
   - --input_proto=$dl_dir/squeezenet1.1.prototxt
 framework: caffe
+quantizable: yes
 license: https://raw.githubusercontent.com/forresti/SqueezeNet/master/LICENSE
diff --git a/models/public/vgg16/model.yml b/models/public/vgg16/model.yml
@@ -45,4 +45,5 @@ model_optimizer_args:
   - --input_model=$dl_dir/vgg16.caffemodel
   - --input_proto=$dl_dir/vgg16.prototxt
 framework: caffe
+quantizable: yes
 license: https://creativecommons.org/licenses/by/4.0/legalcode.txt
diff --git a/tools/downloader/README.md b/tools/downloader/README.md
@@ -11,6 +11,9 @@ based on configuration files in the models' directories.
 * `converter.py` (model converter) converts the models that are not in the
   Inference Engine IR format into that format using Model Optimizer.
 
+* `quantizer.py` (model quantizer) quantizes full-precision models in the IR
+  format into low-precision versions using Post-Training Optimization Toolkit.
+
 * `info_dumper.py` (model information dumper) prints information about the models
   in a stable machine-readable format.
 
@@ -67,17 +70,18 @@ The basic usage is to run the script like this:
 ./downloader.py --all
 ```
 
-This will download all models into a directory tree rooted in the current
-directory. To download into a different directory, use the `-o`/`--output_dir`
-option:
+This will download all models. The `--all` option can be replaced with
+other filter options to download only a subset of models. See the "Shared options"
+section.
+
+By default, the script will download models into a directory tree rooted
+in the current directory. To download into a different directory, use
+the `-o`/`--output_dir` option:
 
 ```sh
 ./downloader.py --all --output_dir my/download/directory
 ```
 
-The `--all` option can be replaced with other filter options to download only
-a subset of models. See the "Shared options" section.
-
 You may use `--precisions` flag to specify comma separated precisions of weights
 to be downloaded.
 
@@ -221,6 +225,9 @@ This will convert all models into the Inference Engine IR format. Models that
 were originally in that format are ignored. Models in PyTorch and Caffe2 formats will be
 converted in ONNX format first.
 
+The `--all` option can be replaced with other filter options to convert only
+a subset of models. See the "Shared options" section.
+
 The current directory must be the root of a download tree created by the model
 downloader. To specify a different download tree path, use the `-d`/`--download_dir`
 option:
@@ -237,9 +244,6 @@ into a different directory tree, use the `-o`/`--output_dir` option:
 ```
 >Note: models in intermediate format are placed to this directory too.
 
-The `--all` option can be replaced with other filter options to convert only
-a subset of models. See the "Shared options" section.
-
 By default, the script will produce models in every precision that is supported
 for conversion. To only produce models in a specific precision, use the `--precisions`
 option:
@@ -295,6 +299,100 @@ To do this, use the `--dry_run` option:
 See the "Shared options" section for information on other options accepted by
 the script.
 
+Model quantizer usage
+---------------------
+
+Before you run the model quantizer, you must prepare a directory with
+the datasets required for the quantization process. This directory will be
+referred to as `<DATASET_DIR>` below. See the "Dataset directory layout"
+section for information on the expected contents of that directory.
+
+The basic usage is to run the script like this:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR>
+```
+
+This will quantize all models for which quantization is supported. Other models
+are ignored.
+
+The `--all` option can be replaced with other filter options to quantize only
+a subset of models. See the "Shared options" section.
+
+The current directory must be the root of a tree of model files create by the model
+converter. To specify a different model tree path, use the `--model_dir` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --model_dir my/model/directory
+```
+
+By default, the quantized models are placed into the same model tree. To place them
+into a different directory tree, use the `-o`/`--output_dir` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --output_dir my/output/directory
+```
+
+By default, the script will produce models in every precision that is supported
+as a quantization output. To only produce models in a specific precision, use
+the `--precisions` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --precisions=FP16-INT8
+```
+
+The script will attempt to locate Post-Training Optimization Toolkit using
+the environment variables set by the OpenVINO&trade; toolkit's `setupvars.sh`/`setupvars.bat`
+script. You can override this heuristic with the `--pot` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --pot my/openvino/path/post_training_optimization_toolkit/main.py
+```
+
+By default, the script will run Post-Training Optimization Toolkit using the same
+Python executable that was used to run the script itself. To use a different
+Python executable, use the `-p`/`--python` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --python my/python
+```
+
+It's possible to specify a target device for Post-Training Optimization Toolkit
+to optimize for, by using the `--target_device` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --target_device VPU
+```
+
+The supported values are those accepted by the "target_device" option in
+Post-Training Optimization Toolkit's config files. If this option is unspecified,
+Post-Training Optimization Toolkit's default is used.
+
+The script can print the quantization commands without actually running them.
+To do this, use the `--dry_run` option:
+
+```sh
+./quantizer.py --all --dataset_dir <DATASET_DIR> --dry_run
+```
+
+With this option specified, the configuration file for Post-Training Optimization
+Toolkit will still be created, so that you can inspect it.
+
+See the "Shared options" section for information on other options accepted by
+the script.
+
+### Dataset directory layout
+
+Currently, all models for which quantization is supported require the
+[ILSVRC 2012](http://image-net.org/challenges/LSVRC/2012/index) validation
+dataset. This means that `<DATASET_DIR>` must contain the following entries:
+
+* A subdirectory named `ILSVRC2012_img_val` containing the ILSVRC 2012
+  validation images. To obtain these images, follow the
+  [instructions at the ILSVRC 2012 website](http://image-net.org/challenges/LSVRC/2012/signup).
+
+* `val.txt` from <http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz>.
+
 Model information dumper usage
 ------------------------------
 
@@ -363,7 +461,7 @@ describing a single model. Each such object has the following keys:
 Shared options
 --------------
 
-The are certain options that both tools accept.
+The are certain options that all tools accept.
 
 `-h`/`--help` can be used to print a help message:
 
diff --git a/tools/downloader/common.py b/tools/downloader/common.py
@@ -16,6 +16,7 @@
 import contextlib
 import fnmatch
 import json
+import platform
 import re
 import shlex
 import shutil
@@ -61,6 +62,9 @@
     'semantic_segmentation',
 }
 
+KNOWN_QUANTIZED_PRECISIONS = {p + '-INT8': p for p in ['FP16', 'FP32']}
+assert KNOWN_QUANTIZED_PRECISIONS.keys() <= KNOWN_PRECISIONS
+
 RE_MODEL_NAME = re.compile(r'[0-9a-zA-Z._-]+')
 RE_SHA256SUM = re.compile(r'[0-9a-fA-F]{64}')
 
@@ -307,13 +311,14 @@ def apply(self, reporter, output_dir):
 Postproc.types['unpack_archive'] = PostprocUnpackArchive
 
 class Model:
-    def __init__(self, name, subdirectory, files, postprocessing, mo_args, framework,
+    def __init__(self, name, subdirectory, files, postprocessing, mo_args, quantizable, framework,
                  description, license_url, precisions, task_type, conversion_to_onnx_args):
         self.name = name
         self.subdirectory = subdirectory
         self.files = files
         self.postprocessing = postprocessing
         self.mo_args = mo_args
+        self.quantizable = quantizable
         self.framework = framework
         self.description = description
         self.license_url = license_url
@@ -389,13 +394,17 @@ def deserialize(cls, model, name, subdirectory):
 
                 precisions = set(files_per_precision.keys())
 
+            quantizable = model.get('quantizable', False)
+            if not isinstance(quantizable, bool):
+                raise DeserializationError('"quantizable": expected a boolean, got {!r}'.format(quantizable))
+
             description = validate_string('"description"', model['description'])
 
             license_url = validate_string('"license"', model['license'])
 
             task_type = validate_string_enum('"task_type"', model['task_type'], KNOWN_TASK_TYPES)
 
-            return cls(name, subdirectory, files, postprocessing, mo_args, framework,
+            return cls(name, subdirectory, files, postprocessing, mo_args, quantizable, framework,
                 description, license_url, precisions, task_type, conversion_to_onnx_args)
 
 def load_models(args):
@@ -482,3 +491,17 @@ def load_models_from_args(parser, args):
                 models[model.name] = model
 
         return list(models.values())
+
+def quote_arg_windows(arg):
+    if not arg: return '""'
+    if not re.search(r'\s|"', arg): return arg
+    # On Windows, only backslashes that precede a quote or the end of the argument must be escaped.
+    return '"' + re.sub(r'(\\+)$', r'\1\1', re.sub(r'(\\*)"', r'\1\1\\"', arg)) + '"'
+
+if platform.system() == 'Windows':
+    quote_arg = quote_arg_windows
+else:
+    quote_arg = shlex.quote
+
+def command_string(args):
+    return ' '.join(map(quote_arg, args))
diff --git a/tools/downloader/converter.py b/tools/downloader/converter.py
@@ -17,10 +17,8 @@
 import argparse
 import concurrent.futures
 import os
-import platform
 import queue
 import re
-import shlex
 import string
 import subprocess
 import sys
@@ -77,17 +75,6 @@ def cancel(self):
         self._future.cancel()
 
 
-def quote_windows(arg):
-    if not arg: return '""'
-    if not re.search(r'\s|"', arg): return arg
-    # On Windows, only backslashes that precede a quote or the end of the argument must be escaped.
-    return '"' + re.sub(r'(\\+)$', r'\1\1', re.sub(r'(\\*)"', r'\1\1\\"', arg)) + '"'
-
-if platform.system() == 'Windows':
-    quote_arg = quote_windows
-else:
-    quote_arg = shlex.quote
-
 def convert_to_onnx(context, model, output_dir, args):
     context.printf('========= {}Converting {} to ONNX',
                    '(DRY RUN) ' if args.dry_run else '', model.name)
@@ -97,7 +84,7 @@ def convert_to_onnx(context, model, output_dir, args):
                                for arg in model.conversion_to_onnx_args]
     cmd = [str(args.python), str(Path(__file__).absolute().parent / model.converter_to_onnx), *conversion_to_onnx_args]
 
-    context.printf('Conversion to ONNX command: {}', ' '.join(map(quote_arg, cmd)))
+    context.printf('Conversion to ONNX command: {}', common.command_string(cmd))
     context.printf('')
 
     success = True if args.dry_run else context.subprocess(cmd)
@@ -206,7 +193,7 @@ def convert(context, model):
             context.printf('========= {}Converting {} to IR ({})',
                 '(DRY RUN) ' if args.dry_run else '', model.name, model_precision)
 
-            context.printf('Conversion command: {}', ' '.join(map(quote_arg, mo_cmd)))
+            context.printf('Conversion command: {}', common.command_string(mo_cmd))
 
             if not args.dry_run:
                 context.printf('', flush=True)
diff --git a/tools/downloader/quantizer.py b/tools/downloader/quantizer.py
