Quantization Aware Training with NNCF, using TensorFlow Framework¶

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS. To run without installing anything, click the “Open in Colab” button.

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:

Transforming the original FP32 model to INT8
Using fine-tuning to restore the accuracy.
Exporting optimized and original models to Frozen Graph and then to OpenVINO.
Measuring and comparing the performance of models.

For more advanced usage, refer to these examples.

This tutorial uses 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.

Table of contents:

Imports and Settings
Dataset Preprocessing
Define a Floating-Point Model
Pre-train a Floating-Point Model
Create and Initialize Quantization
Fine-tune the Compressed Model
Export Models to OpenVINO Intermediate Representation (IR)
Benchmark Model Performance by Computing Inference 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).

!pip install -q "openvino-dev>=2023.0.0" "nncf>=2.5.0"
!pip install -q "tensorflow-datasets>=4.8.0"

DEPRECATION: pytorch-lightning 1.6.5 has a non-standard dependency specifier torch>=1.8.*. pip 23.3 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pytorch-lightning or contact the author to suggest that they release a version with a conforming dependency specifiers. Discussion can be found at https://github.com/pypa/pip/issues/12063
DEPRECATION: pytorch-lightning 1.6.5 has a non-standard dependency specifier torch>=1.8.*. pip 23.3 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pytorch-lightning or contact the author to suggest that they release a version with a conforming dependency specifiers. Discussion can be found at https://github.com/pypa/pip/issues/12063
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
pytorch-lightning 1.6.5 requires protobuf<=3.20.1, but you have protobuf 3.20.3 which is incompatible.

from pathlib import Path
import logging

import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras import layers
from tensorflow.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.logging.logger import set_log_level
from openvino.runtime import serialize
from openvino.tools import mo

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_ir_path = Path(OUTPUT_DIR / "saved_model").with_suffix(".xml")
int8_pb_path = Path(OUTPUT_DIR / (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()}')

2023-08-16 01:17:34.103410: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2023-08-16 01:17:34.137361: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-08-16 01:17:34.726614: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino
WARNING:nncf:NNCF provides best results with tensorflow==2.11.*, while current tensorflow version is 2.12.0. If you encounter issues, consider switching to tensorflow==2.11.*
Downloading data from https://storage.openvinotoolkit.org/repositories/nncf/openvino_notebook_ckpts/305_resnet18_imagenette_fp32_v1.h5
134604992/134604992 [==============================] - 30s 0us/step
Absolute path where the model weights are saved:
 /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/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)

2023-08-16 01:18:08.016585: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1956] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...
2023-08-16 01:18:08.132762: I tensorflow/core/common_runtime/executor.cc:1197] [/device:CPU:0] (DEBUG INFO) Executor start aborting (this does not indicate an error and you can ignore this message): INVALID_ARGUMENT: You must feed a value for placeholder tensor 'Placeholder/_1' with dtype string and shape [1]
     [[{{node Placeholder/_1}}]]
2023-08-16 01:18:08.133087: I tensorflow/core/common_runtime/executor.cc:1197] [/device:CPU:0] (DEBUG INFO) Executor start aborting (this does not indicate an error and you can ignore this message): INVALID_ARGUMENT: You must feed a value for placeholder tensor 'Placeholder/_0' with dtype string and shape [1]
     [[{{node Placeholder/_0}}]]
2023-08-16 01:18:08.170026: W tensorflow/core/kernels/data/cache_dataset_ops.cc:856] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to dataset.cache().take(k).repeat(). You should use dataset.take(k).cache().repeat() instead.

../_images/305-tensorflow-quantization-aware-training-with-output_6_1.png

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 the model.
    model = models.Model(img_input, x)

    return model

IMG_SHAPE = IMG_SIZE + (3,)
fp32_model = ResNet18(input_shape=IMG_SHAPE)

Pre-train a 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, it is recommended to skip FP32 model training and load the weights that are provided.

# Load the floating-point weights.
fp32_model.load_weights(fp32_h5_path)

# Compile the floating-point model.
fp32_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 = fp32_model.evaluate(
    validation_dataset,
    callbacks=tf.keras.callbacks.ProgbarLogger(stateful_metrics=['acc@1'])
)
print(f"\nAccuracy of FP32 model: {acc_fp32:.3f}")

2023-08-16 01:18:09.025847: I tensorflow/core/common_runtime/executor.cc:1197] [/device:CPU:0] (DEBUG INFO) Executor start aborting (this does not indicate an error and you can ignore this message): INVALID_ARGUMENT: You must feed a value for placeholder tensor 'Placeholder/_1' with dtype string and shape [1]
     [[{{node Placeholder/_1}}]]
2023-08-16 01:18:09.026203: I tensorflow/core/common_runtime/executor.cc:1197] [/device:CPU:0] (DEBUG INFO) Executor start aborting (this does not indicate an error and you can ignore this message): INVALID_ARGUMENT: You must feed a value for placeholder tensor 'Placeholder/_0' with dtype string and shape [1]
     [[{{node Placeholder/_0}}]]

4/4 [==============================] - 1s 229ms/sample - loss: 0.9807 - acc@1: 0.8220

Accuracy of FP32 model: 0.822

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),  # The log directory for NNCF-specific logging outputs.
    "compression": {
        "algorithm": "quantization",  # Specify the algorithm here.
    },
}
nncf_config = NNCFConfig.from_dict(nncf_config_dict)

Provide a 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 a configuration object.

compression_ctrl, int8_model = create_compressed_model(fp32_model, nncf_config)

2023-08-16 01:18:11.729441: I tensorflow/core/common_runtime/executor.cc:1197] [/device:CPU:0] (DEBUG INFO) Executor start aborting (this does not indicate an error and you can ignore this message): INVALID_ARGUMENT: You must feed a value for placeholder tensor 'Placeholder/_4' with dtype int64 and shape [1]
     [[{{node Placeholder/_4}}]]
2023-08-16 01:18:11.729828: I tensorflow/core/common_runtime/executor.cc:1197] [/device:CPU:0] (DEBUG INFO) Executor start aborting (this does not indicate an error and you can ignore this message): INVALID_ARGUMENT: You must feed a value for placeholder tensor 'Placeholder/_3' with dtype int64 and shape [1]
     [[{{node Placeholder/_3}}]]
2023-08-16 01:18:12.738622: W tensorflow/core/kernels/data/cache_dataset_ops.cc:856] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to dataset.cache().take(k).repeat(). You should use dataset.take(k).cache().repeat() instead.
2023-08-16 01:18:13.389616: W tensorflow/core/kernels/data/cache_dataset_ops.cc:856] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to dataset.cache().take(k).repeat(). You should use dataset.take(k).cache().repeat() instead.
2023-08-16 01:18:21.360841: W tensorflow/core/kernels/data/cache_dataset_ops.cc:856] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to dataset.cache().take(k).repeat(). You should use dataset.take(k).cache().repeat() instead.

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 being demonstrated here.

# Compile the INT8 model.
int8_model.compile(
    optimizer=tf.keras.optimizers.Adam(learning_rate=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 = int8_model.evaluate(
    validation_dataset,
    callbacks=tf.keras.callbacks.ProgbarLogger(stateful_metrics=['acc@1'])
)

4/4 [==============================] - 1s 301ms/sample - loss: 0.9766 - acc@1: 0.8120

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.

print(f"\nAccuracy of INT8 model after initialization: {test_acc:.3f}")

# Train the INT8 model.
int8_model.fit(train_dataset, epochs=2)

# Validate the INT8 model.
test_loss, acc_int8 = 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}")

Accuracy of INT8 model after initialization: 0.812
Epoch 1/2
101/101 [==============================] - 49s 417ms/step - loss: 0.7134 - acc@1: 0.9299
Epoch 2/2
101/101 [==============================] - 42s 414ms/step - loss: 0.6807 - acc@1: 0.9489
4/4 [==============================] - 1s 144ms/sample - loss: 0.9760 - acc@1: 0.8160

Accuracy of INT8 model after fine-tuning: 0.816

Accuracy drop of tuned INT8 model over pre-trained FP32 model: 0.006

Export Models to OpenVINO Intermediate Representation (IR) ⇑¶

Use model conversion Python API to convert the models to OpenVINO IR.

For more information about model conversion, see this page.

Executing this command may take a while.

model_ir_fp32 = mo.convert_model(
    fp32_model,
    input_shape=[1, 64, 64, 3],
)

2023-08-16 01:19:54.530759: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 2
2023-08-16 01:19:54.530838: I tensorflow/core/grappler/clusters/single_machine.cc:358] Starting new session
2023-08-16 01:19:54.651453: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1956] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...

model_ir_int8 = mo.convert_model(
    int8_model,
    input_shape=[1, 64, 64, 3],
)

2023-08-16 01:19:56.200644: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 2
2023-08-16 01:19:56.200714: I tensorflow/core/grappler/clusters/single_machine.cc:358] Starting new session
2023-08-16 01:19:56.202200: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1956] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...

Benchmark Model Performance by Computing Inference Time ⇑¶

Finally, measure the inference performance of the FP32 and INT8 models, using Benchmark Tool - an inference performance measurement tool in OpenVINO. 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

This notebook runs benchmark_app for 15 seconds to give a quick indication of performance. For more accurate performance, it is recommended to run benchmark_app in a terminal/command prompt after closing other applications. Run benchmark_app -m model.xml -d CPU to benchmark async inference on CPU for one minute. Change CPU to GPU to benchmark on GPU. Run benchmark_app --help to see an overview of all command-line options.

serialize(model_ir_fp32, str(fp32_ir_path))
serialize(model_ir_int8, str(int8_ir_path))


def parse_benchmark_output(benchmark_output):
    parsed_output = [line for line in benchmark_output if 'FPS' in 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)
[ INFO ] Throughput:   2831.57 FPS

Benchmark INT8 model (IR)
[ INFO ] Throughput:   11769.65 FPS

Show CPU Information for reference.

from openvino.runtime import Core

ie = Core()
ie.get_property('CPU', "FULL_DEVICE_NAME")

'Intel(R) Core(TM) i9-10920X CPU @ 3.50GHz'