Quantization Aware Training with NNCF, using TensorFlow Framework¶
The goal of this notebook to demonstrate how to use the Neural Network Compression Framework NNCF 8-bit quantization to optimize a TensorFlow model for inference with OpenVINO Toolkit. The optimization process contains the following steps: * Transform the original FP32 model to INT8 * Use fine-tuning to restore the accuracy * Export optimized and original models to Frozen Graph and then to OpenVINO * Measure and compare the performance of models
For more advanced usage, please refer to these examples.
We selected the ResNet-18 model with Imagenette dataset. Imagenette is a subset of 10 easily classified classes from the Imagenet dataset. Using the smaller model and dataset will speed up training and download time.
Imports and Settings¶
Import NNCF and all auxiliary packages from your Python code. Set a name for the model, input image size, used batch size, and the learning rate. Also define paths where Frozen Graph and OpenVINO IR versions of the models will be stored.
NOTE: All NNCF logging messages below ERROR level (INFO and WARNING) are disabled to simplify the tutorial. For production use, it is recommended to enable logging, by removing
set_log_level(logging.ERROR).
from pathlib import Path
import logging
import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.python.keras import layers
from tensorflow.python.keras import models
from nncf import NNCFConfig
from nncf.tensorflow.helpers.model_creation import create_compressed_model
from nncf.tensorflow.initialization import register_default_init_args
from nncf.common.utils.logger import set_log_level
set_log_level(logging.ERROR)
MODEL_DIR = Path("model")
OUTPUT_DIR = Path("output")
MODEL_DIR.mkdir(exist_ok=True)
OUTPUT_DIR.mkdir(exist_ok=True)
BASE_MODEL_NAME = "ResNet-18"
fp32_h5_path = Path(MODEL_DIR / (BASE_MODEL_NAME + "_fp32")).with_suffix(".h5")
fp32_sm_path = Path(OUTPUT_DIR / (BASE_MODEL_NAME + "_fp32"))
fp32_ir_path = Path(OUTPUT_DIR / "saved_model").with_suffix(".xml")
int8_pb_path = Path(OUTPUT_DIR / (BASE_MODEL_NAME + "_int8")).with_suffix(".pb")
int8_pb_name = Path(BASE_MODEL_NAME + "_int8").with_suffix(".pb")
int8_ir_path = int8_pb_path.with_suffix(".xml")
BATCH_SIZE = 128
IMG_SIZE = (64, 64) # Default Imagenet image size
NUM_CLASSES = 10 # For Imagenette dataset
LR = 1e-5
MEAN_RGB = (0.485 * 255, 0.456 * 255, 0.406 * 255) # From Imagenet dataset
STDDEV_RGB = (0.229 * 255, 0.224 * 255, 0.225 * 255) # From Imagenet dataset
fp32_pth_url = "https://storage.openvinotoolkit.org/repositories/nncf/openvino_notebook_ckpts/305_resnet18_imagenette_fp32_v1.h5"
_ = tf.keras.utils.get_file(fp32_h5_path.resolve(), fp32_pth_url)
print(f'Absolute path where the model weights are saved:\n {fp32_h5_path.resolve()}')
Downloading data from https://storage.openvinotoolkit.org/repositories/nncf/openvino_notebook_ckpts/305_resnet18_imagenette_fp32_v1.h5
134610944/134604992 [==============================] - 3s 0us/step
Absolute path where the model weights are saved:
/home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/model/ResNet-18_fp32.h5
Dataset Preprocessing¶
Download and prepare Imagenette 160px dataset. - Number of classes: 10 - Download size: 94.18 MiB | Split | Examples | |————–|———-| | ‘train’ | 12,894 | | ‘validation’ | 500 |
datasets, datasets_info = tfds.load('imagenette/160px', shuffle_files=True, as_supervised=True, with_info=True,
read_config=tfds.ReadConfig(shuffle_seed=0))
train_dataset, validation_dataset = datasets['train'], datasets['validation']
fig = tfds.show_examples(train_dataset, datasets_info)
[1mDownloading and preparing dataset 94.18 MiB (download: 94.18 MiB, generated: Unknown size, total: 94.18 MiB) to /home/runner/tensorflow_datasets/imagenette/160px/0.1.0...[0m
Dl Completed...: 0 url [00:00, ? url/s]
Dl Size...: 0 MiB [00:00, ? MiB/s]
Extraction completed...: 0 file [00:00, ? file/s]
Generating splits...: 0%| | 0/2 [00:00<?, ? splits/s]
Generating train examples...: 0%| | 0/12894 [00:00<?, ? examples/s]
Shuffling imagenette-train.tfrecord...: 0%| | 0/12894 [00:00<?, ? examples/s]
Generating validation examples...: 0%| | 0/500 [00:00<?, ? examples/s]
Shuffling imagenette-validation.tfrecord...: 0%| | 0/500 [00:00<?, ? examples/s]
[1mDataset imagenette downloaded and prepared to /home/runner/tensorflow_datasets/imagenette/160px/0.1.0. Subsequent calls will reuse this data.[0m
def preprocessing(image, label):
image = tf.image.resize(image, IMG_SIZE)
image = image - MEAN_RGB
image = image / STDDEV_RGB
label = tf.one_hot(label, NUM_CLASSES)
return image, label
train_dataset = (train_dataset.map(preprocessing, num_parallel_calls=tf.data.experimental.AUTOTUNE)
.batch(BATCH_SIZE)
.prefetch(tf.data.experimental.AUTOTUNE))
validation_dataset = (validation_dataset.map(preprocessing, num_parallel_calls=tf.data.experimental.AUTOTUNE)
.batch(BATCH_SIZE)
.prefetch(tf.data.experimental.AUTOTUNE))
Define a Floating-Point Model¶
def residual_conv_block(filters, stage, block, strides=(1, 1), cut='pre'):
def layer(input_tensor):
x = layers.BatchNormalization(epsilon=2e-5)(input_tensor)
x = layers.Activation('relu')(x)
# defining shortcut connection
if cut == 'pre':
shortcut = input_tensor
elif cut == 'post':
shortcut = layers.Conv2D(filters, (1, 1), strides=strides, kernel_initializer='he_uniform',
use_bias=False)(x)
# continue with convolution layers
x = layers.ZeroPadding2D(padding=(1, 1))(x)
x = layers.Conv2D(filters, (3, 3), strides=strides, kernel_initializer='he_uniform', use_bias=False)(x)
x = layers.BatchNormalization(epsilon=2e-5)(x)
x = layers.Activation('relu')(x)
x = layers.ZeroPadding2D(padding=(1, 1))(x)
x = layers.Conv2D(filters, (3, 3), kernel_initializer='he_uniform', use_bias=False)(x)
# add residual connection
x = layers.Add()([x, shortcut])
return x
return layer
def ResNet18(input_shape=None):
"""Instantiates the ResNet18 architecture."""
img_input = layers.Input(shape=input_shape, name='data')
# ResNet18 bottom
x = layers.BatchNormalization(epsilon=2e-5, scale=False)(img_input)
x = layers.ZeroPadding2D(padding=(3, 3))(x)
x = layers.Conv2D(64, (7, 7), strides=(2, 2), kernel_initializer='he_uniform', use_bias=False)(x)
x = layers.BatchNormalization(epsilon=2e-5)(x)
x = layers.Activation('relu')(x)
x = layers.ZeroPadding2D(padding=(1, 1))(x)
x = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(x)
# ResNet18 body
repetitions = (2, 2, 2, 2)
for stage, rep in enumerate(repetitions):
for block in range(rep):
filters = 64 * (2 ** stage)
if block == 0 and stage == 0:
x = residual_conv_block(filters, stage, block, strides=(1, 1), cut='post')(x)
elif block == 0:
x = residual_conv_block(filters, stage, block, strides=(2, 2), cut='post')(x)
else:
x = residual_conv_block(filters, stage, block, strides=(1, 1), cut='pre')(x)
x = layers.BatchNormalization(epsilon=2e-5)(x)
x = layers.Activation('relu')(x)
# ResNet18 top
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dense(NUM_CLASSES)(x)
x = layers.Activation('softmax')(x)
# Create model
model = models.Model(img_input, x)
return model
IMG_SHAPE = IMG_SIZE + (3,)
model = ResNet18(input_shape=IMG_SHAPE)
Pre-train Floating-Point Model¶
Using NNCF for model compression assumes that the user has a pre-trained model and a training pipeline.
NOTE For the sake of simplicity of the tutorial, we propose to skip FP32 model training and load the weights that are provided.
# Load the floating-point weights
model.load_weights(fp32_h5_path)
# Compile the floating-point model
model.compile(loss=tf.keras.losses.CategoricalCrossentropy(label_smoothing=0.1),
metrics=[tf.keras.metrics.CategoricalAccuracy(name='acc@1')])
# Validate the floating-point model
test_loss, acc_fp32 = model.evaluate(validation_dataset,
callbacks=tf.keras.callbacks.ProgbarLogger(stateful_metrics=['acc@1']))
print(f"\nAccuracy of FP32 model: {acc_fp32:.3f}")
4/4 [==============================] - 1s 362ms/sample - loss: 0.9807 - acc@1: 0.8220
Accuracy of FP32 model: 0.822
Save the floating-point model to the saved model, which will be later used for conversion to OpenVINO IR and further performance measurement.
model.save(fp32_sm_path)
print(f'Absolute path where the model is saved:\n {fp32_sm_path.resolve()}')
/opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/tensorflow/python/keras/utils/generic_utils.py:494: CustomMaskWarning: Custom mask layers require a config and must override get_config. When loading, the custom mask layer must be passed to the custom_objects argument.
warnings.warn('Custom mask layers require a config and must override '
INFO:tensorflow:Assets written to: output/ResNet-18_fp32/assets
INFO:tensorflow:Assets written to: output/ResNet-18_fp32/assets
Absolute path where the model is saved:
/home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output/ResNet-18_fp32
Create and Initialize Quantization¶
NNCF enables compression-aware training by integrating into regular training pipelines. The framework is designed so that modifications to your original training code are minor. Quantization is the simplest scenario and requires only 3 modifications.
Configure NNCF parameters to specify compression
nncf_config_dict = {
"input_info": {"sample_size": [1, 3] + list(IMG_SIZE)},
"log_dir": str(OUTPUT_DIR), # log directory for NNCF-specific logging outputs
"compression": {
"algorithm": "quantization", # specify the algorithm here
},
}
nncf_config = NNCFConfig.from_dict(nncf_config_dict)
Provide data loader to initialize the values of quantization ranges and determine which activation should be signed or unsigned from the collected statistics using a given number of samples.
nncf_config = register_default_init_args(nncf_config=nncf_config,
data_loader=train_dataset,
batch_size=BATCH_SIZE)
Create a wrapped model ready for compression fine-tuning from a pre-trained FP32 model and configuration object.
compression_ctrl, model = create_compressed_model(model, nncf_config)
Evaluate the new model on the validation set after initialization of quantization. The accuracy should be not far from the accuracy of the floating-point FP32 model for a simple case like the one we are demonstrating now.
# Compile the int8 model
model.compile(optimizer=tf.keras.optimizers.Adam(lr=LR),
loss=tf.keras.losses.CategoricalCrossentropy(label_smoothing=0.1),
metrics=[tf.keras.metrics.CategoricalAccuracy(name='acc@1')])
# Validate the int8 model
test_loss, test_acc = model.evaluate(validation_dataset,
callbacks=tf.keras.callbacks.ProgbarLogger(stateful_metrics=['acc@1']))
print(f"\nAccuracy of INT8 model after initialization: {test_acc:.3f}")
/opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:374: UserWarning: The lr argument is deprecated, use learning_rate instead. warnings.warn(
4/4 [==============================] - 2s 546ms/sample - loss: 0.9786 - acc@1: 0.8160
Accuracy of INT8 model after initialization: 0.816
Fine-tune the Compressed Model¶
At this step, a regular fine-tuning process is applied to further improve quantized model accuracy. Normally, several epochs of tuning are required with a small learning rate, the same that is usually used at the end of the training of the original model. No other changes in the training pipeline are required. Here is a simple example.
# Train the int8 model
model.fit(train_dataset,
epochs=2)
# Validate the int8 model
test_loss, acc_int8 = model.evaluate(validation_dataset,
callbacks=tf.keras.callbacks.ProgbarLogger(stateful_metrics=['acc@1']))
print(f"\nAccuracy of INT8 model after fine-tuning: {acc_int8:.3f}")
print(f"\nAccuracy drop of tuned INT8 model over pre-trained FP32 model: {acc_fp32 - acc_int8:.3f}")
Epoch 1/2
101/101 [==============================] - 465s 5s/step - loss: 0.7135 - acc@1: 0.9304
Epoch 2/2
101/101 [==============================] - 419s 4s/step - loss: 0.6804 - acc@1: 0.9493
4/4 [==============================] - 1s 312ms/sample - loss: 0.9788 - acc@1: 0.8220
Accuracy of INT8 model after fine-tuning: 0.822
Accuracy drop of tuned INT8 model over pre-trained FP32 model: 0.000
Save the INT8 model to the frozen graph (saved model does not work with quantized model for now). Frozen graph will be later used for conversion to OpenVINO IR and further performance measurement.
compression_ctrl.export_model(int8_pb_path, 'frozen_graph')
print(f'Absolute path where the int8 model is saved:\n {int8_pb_path.resolve()}')
Absolute path where the int8 model is saved:
/home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output/ResNet-18_int8.pb
Export Frozen Graph Models to OpenVINO Intermediate Representation (IR)¶
Call the OpenVINO Model Optimizer tool to convert the Saved Model and Frozen Graph models to OpenVINO IR. The models are saved to the current directory.
See the Model Optimizer Developer Guide for more information about Model Optimizer.
Executing this command may take a while. There may be some errors or
warnings in the output. Model Optimization successfully export to IR if
the last lines of the output include:
[ SUCCESS ] Generated IR version 10 model
!mo --framework=tf --input_shape=[1,64,64,3] --input=data --saved_model_dir=$fp32_sm_path --output_dir=$OUTPUT_DIR
Model Optimizer arguments:
Common parameters:
- Path to the Input Model: None
- Path for generated IR: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output
- IR output name: saved_model
- Log level: ERROR
- Batch: Not specified, inherited from the model
- Input layers: data
- Output layers: Not specified, inherited from the model
- Input shapes: [1,64,64,3]
- Source layout: Not specified
- Target layout: Not specified
- Layout: Not specified
- Mean values: Not specified
- Scale values: Not specified
- Scale factor: Not specified
- Precision of IR: FP32
- Enable fusing: True
- User transformations: Not specified
- Reverse input channels: False
- Enable IR generation for fixed input shape: False
- Use the transformations config file: None
Advanced parameters:
- Force the usage of legacy Frontend of Model Optimizer for model conversion into IR: False
- Force the usage of new Frontend of Model Optimizer for model conversion into IR: False
TensorFlow specific parameters:
- Input model in text protobuf format: False
- Path to model dump for TensorBoard: None
- List of shared libraries with TensorFlow custom layers implementation: None
- Update the configuration file with input/output node names: None
- Use configuration file used to generate the model with Object Detection API: None
- Use the config file: None
OpenVINO runtime found in: /opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/openvino
OpenVINO runtime version: 2022.1.0-7019-cdb9bec7210-releases/2022/1
Model Optimizer version: 2022.1.0-7019-cdb9bec7210-releases/2022/1
[ SUCCESS ] Generated IR version 11 model.
[ SUCCESS ] XML file: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output/saved_model.xml
[ SUCCESS ] BIN file: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output/saved_model.bin
[ SUCCESS ] Total execution time: 17.04 seconds.
[ SUCCESS ] Memory consumed: 892 MB.
It's been a while, check for a new version of Intel(R) Distribution of OpenVINO(TM) toolkit here https://software.intel.com/content/www/us/en/develop/tools/openvino-toolkit/download.html?cid=other&source=prod&campid=ww_2022_bu_IOTG_OpenVINO-2022-1&content=upg_all&medium=organic or on the GitHub*
[ INFO ] The model was converted to IR v11, the latest model format that corresponds to the source DL framework input/output format. While IR v11 is backwards compatible with OpenVINO Inference Engine API v1.0, please use API v2.0 (as of 2022.1) to take advantage of the latest improvements in IR v11.
Find more information about API v2.0 and IR v11 at https://docs.openvino.ai
!mo --framework=tf --input_shape=[1,64,64,3] --input=Placeholder --input_model=$int8_pb_path --output_dir=$OUTPUT_DIR
Model Optimizer arguments:
Common parameters:
- Path to the Input Model: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output/ResNet-18_int8.pb
- Path for generated IR: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output
- IR output name: ResNet-18_int8
- Log level: ERROR
- Batch: Not specified, inherited from the model
- Input layers: Placeholder
- Output layers: Not specified, inherited from the model
- Input shapes: [1,64,64,3]
- Source layout: Not specified
- Target layout: Not specified
- Layout: Not specified
- Mean values: Not specified
- Scale values: Not specified
- Scale factor: Not specified
- Precision of IR: FP32
- Enable fusing: True
- User transformations: Not specified
- Reverse input channels: False
- Enable IR generation for fixed input shape: False
- Use the transformations config file: None
Advanced parameters:
- Force the usage of legacy Frontend of Model Optimizer for model conversion into IR: False
- Force the usage of new Frontend of Model Optimizer for model conversion into IR: False
TensorFlow specific parameters:
- Input model in text protobuf format: False
- Path to model dump for TensorBoard: None
- List of shared libraries with TensorFlow custom layers implementation: None
- Update the configuration file with input/output node names: None
- Use configuration file used to generate the model with Object Detection API: None
- Use the config file: None
OpenVINO runtime found in: /opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/openvino
OpenVINO runtime version: 2022.1.0-7019-cdb9bec7210-releases/2022/1
Model Optimizer version: 2022.1.0-7019-cdb9bec7210-releases/2022/1
[ ERROR ] Exception occurred during running replacer "REPLACEMENT_ID" (<class 'openvino.tools.mo.front.user_data_repack.UserDataRepack'>): No node with name Placeholder
Benchmark Model Performance by Computing Inference Time¶
Finally, we will measure the inference performance of the FP32 and INT8 models. To do this, we use Benchmark Tool - OpenVINO’s inference performance measurement tool. By default, Benchmark Tool runs inference for 60 seconds in asynchronous mode on CPU. It returns inference speed as latency (milliseconds per image) and throughput (frames per second) values.
NOTE: In this notebook we run benchmark_app for 15 seconds to give a quick indication of performance. For more accurate performance, we recommended running benchmark_app in a terminal/command prompt after closing other applications. Run
benchmark_app -m model.xml -d CPUto benchmark async inference on CPU for one minute. Change CPU to GPU to benchmark on GPU. Runbenchmark_app --helpto see an overview of all command line options.
def parse_benchmark_output(benchmark_output):
parsed_output = [line for line in benchmark_output if not (line.startswith(r"[") or line.startswith(" ") or line == "")]
print(*parsed_output, sep='\n')
print('Benchmark FP32 model (IR)')
benchmark_output = ! benchmark_app -m $fp32_ir_path -d CPU -api async -t 15
parse_benchmark_output(benchmark_output)
print('\nBenchmark INT8 model (IR)')
benchmark_output = ! benchmark_app -m $int8_ir_path -d CPU -api async -t 15
parse_benchmark_output(benchmark_output)
Benchmark FP32 model (IR)
Count: 4556 iterations
Duration: 15004.59 ms
Latency:
Throughput: 303.64 FPS
Benchmark INT8 model (IR)
Traceback (most recent call last):
RuntimeError: Model file /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/305-tensorflow-quantization-aware-training/output/ResNet-18_int8.xml cannot be opened!
Show CPU Information for reference
from openvino.runtime import Core
ie = Core()
ie.get_property(device_name='CPU', name="FULL_DEVICE_NAME")
'Intel(R) Xeon(R) Platinum 8272CL CPU @ 2.60GHz'