使用 int16 激活值进行训练后整数量化

在 TensorFlow.org上查看 在 Google Colab 中运行 在 GitHub 上查看源代码 下载笔记本

概述

现在,将模型从 TensorFlow 转换为 TensorFlow Lite 的 FlatBuffer 格式时,TensorFlow Lite 支持将激活转换为 16 位整数值,同时将权重转换为 8 位整数值。我们将此模式称为“16x8 量化模式”。当激活对量化敏感时,此模式可以大幅提高量化模型的准确率,同时还可以将模型大小缩减四分之一至四分之三。此外,这种完全量化的模型可供仅支持整数的硬件加速器使用。

一些可以从这种训练后量化模式受益的示例模型包括:

  • 超分辨率,
  • 音频信号处理,如噪声消除和波束成形,
  • 图像降噪,
  • 基于单张图像的 HDR 重建

在本教程中,您将从头开始训练一个 MNIST 模型,并在 TensorFlow 中检查其准确率,然后使用此模式将该模型转换为 Tensorflow Lite FlatBuffer。最后,您将检查转换的模型的准确率,并将其与原始 float32 模型进行对比。请注意,本示例旨在演示此模式的用法,并不会展现与 TensorFlow Lite 中提供的其他量化技术相比的优势。

构建 MNIST 模型

设置

import logging
logging.getLogger("tensorflow").setLevel(logging.DEBUG)

import tensorflow as tf
from tensorflow import keras
import numpy as np
import pathlib

检查 16x8 量化模式是否可用

tf.lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8
<OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8: 'EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8'>

训练并导出模型

# Load MNIST dataset
mnist = keras.datasets.mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# Normalize the input image so that each pixel value is between 0 to 1.
train_images = train_images / 255.0
test_images = test_images / 255.0

# Define the model architecture
model = keras.Sequential([
  keras.layers.InputLayer(input_shape=(28, 28)),
  keras.layers.Reshape(target_shape=(28, 28, 1)),
  keras.layers.Conv2D(filters=12, kernel_size=(3, 3), activation=tf.nn.relu),
  keras.layers.MaxPooling2D(pool_size=(2, 2)),
  keras.layers.Flatten(),
  keras.layers.Dense(10)
])

# Train the digit classification model
model.compile(optimizer='adam',
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
model.fit(
  train_images,
  train_labels,
  epochs=1,
  validation_data=(test_images, test_labels)
)
2021-08-13 21:10:34.217149: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.226026: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.226944: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.228589: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2021-08-13 21:10:34.229187: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.230116: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.231004: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.820411: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.821367: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.822256: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:34.823079: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 14648 MB memory:  -> device: 0, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:05.0, compute capability: 7.0
2021-08-13 21:10:35.698373: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)
2021-08-13 21:10:36.427390: I tensorflow/stream_executor/cuda/cuda_dnn.cc:369] Loaded cuDNN version 8100
2021-08-13 21:10:36.948287: I tensorflow/core/platform/default/subprocess.cc:304] Start cannot spawn child process: No such file or directory
1875/1875 [==============================] - 6s 2ms/step - loss: 0.2773 - accuracy: 0.9217 - val_loss: 0.1299 - val_accuracy: 0.9622
<keras.callbacks.History at 0x7fbf585e3b10>

在此示例中,您只对模型进行了一个周期的训练,因此只训练到约 96% 的准确率。

转换为 TensorFlow Lite 模型

现在,您可以使用 Python TFLiteConverter 将训练的模型转换为 TensorFlow Lite 模型。

现在,使用 TFliteConverter 将模型转换为默认的 float32 格式:

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
2021-08-13 21:10:41.486118: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
INFO:tensorflow:Assets written to: /tmp/tmpyetg5jcr/assets
2021-08-13 21:10:41.855311: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:41.855664: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
2021-08-13 21:10:41.855778: I tensorflow/core/grappler/clusters/single_machine.cc:357] Starting new session
2021-08-13 21:10:41.856128: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:41.856572: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:41.856861: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:41.857220: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:41.857524: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:41.857790: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 14648 MB memory:  -> device: 0, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:05.0, compute capability: 7.0
2021-08-13 21:10:41.859122: I tensorflow/core/grappler/optimizers/meta_optimizer.cc:1137] Optimization results for grappler item: graph_to_optimize
  function_optimizer: function_optimizer did nothing. time = 0.007ms.
  function_optimizer: function_optimizer did nothing. time = 0.001ms.

2021-08-13 21:10:41.889267: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format.
2021-08-13 21:10:41.889304: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
2021-08-13 21:10:41.892454: I tensorflow/compiler/mlir/tensorflow/utils/dump_mlir_util.cc:210] disabling MLIR crash reproducer, set env var `MLIR_CRASH_REPRODUCER_DIRECTORY` to enable.

将其写入 .tflite 文件:

tflite_models_dir = pathlib.Path("/tmp/mnist_tflite_models/")
tflite_models_dir.mkdir(exist_ok=True, parents=True)
tflite_model_file = tflite_models_dir/"mnist_model.tflite"
tflite_model_file.write_bytes(tflite_model)
84500

要改为将模型量化为 16x8 量化模式,首先将 optimizations 标记设置为使用默认优化。然后将 16x8 量化模式指定为目标规范中要求的受支持运算:

converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_ops = [tf.lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8]

对于 int8 训练后量化,通过将转换器选项 inference_input(output)_type 设置为 tf.int16,可以产生全整数量化模型。

设置校准数据:

mnist_train, _ = tf.keras.datasets.mnist.load_data()
images = tf.cast(mnist_train[0], tf.float32) / 255.0
mnist_ds = tf.data.Dataset.from_tensor_slices((images)).batch(1)
def representative_data_gen():
  for input_value in mnist_ds.take(100):
    # Model has only one input so each data point has one element.
    yield [input_value]
converter.representative_dataset = representative_data_gen

最后,像往常一样转换模型。请注意,为了方便调用,转换后的模型默认仍将使用浮点输入和输出。

tflite_16x8_model = converter.convert()
tflite_model_16x8_file = tflite_models_dir/"mnist_model_quant_16x8.tflite"
tflite_model_16x8_file.write_bytes(tflite_16x8_model)
INFO:tensorflow:Assets written to: /tmp/tmpfhuog02w/assets
INFO:tensorflow:Assets written to: /tmp/tmpfhuog02w/assets
2021-08-13 21:10:42.892609: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:42.892955: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
2021-08-13 21:10:42.893072: I tensorflow/core/grappler/clusters/single_machine.cc:357] Starting new session
2021-08-13 21:10:42.893454: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:42.893801: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:42.894061: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:42.894381: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:42.894644: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-13 21:10:42.894888: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 14648 MB memory:  -> device: 0, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:05.0, compute capability: 7.0
2021-08-13 21:10:42.896278: I tensorflow/core/grappler/optimizers/meta_optimizer.cc:1137] Optimization results for grappler item: graph_to_optimize
  function_optimizer: function_optimizer did nothing. time = 0.008ms.
  function_optimizer: function_optimizer did nothing. time = 0.001ms.

2021-08-13 21:10:42.927173: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format.
2021-08-13 21:10:42.927204: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
24768

请注意,生成文件的大小约为原来的 1/3

ls -lh {tflite_models_dir}
total 180K
-rw-rw-r-- 1 kbuilder kbuilder 83K Aug 13 21:10 mnist_model.tflite
-rw-rw-r-- 1 kbuilder kbuilder 24K Aug 13 21:09 mnist_model_quant.tflite
-rw-rw-r-- 1 kbuilder kbuilder 25K Aug 13 21:10 mnist_model_quant_16x8.tflite
-rw-rw-r-- 1 kbuilder kbuilder 44K Aug 13 21:08 mnist_model_quant_f16.tflite

运行 TensorFlow Lite 模型

使用 Python TensorFlow Lite 解释器运行 TensorFlow Lite 模型。

将模型加载到解释器中

interpreter = tf.lite.Interpreter(model_path=str(tflite_model_file))
interpreter.allocate_tensors()
interpreter_16x8 = tf.lite.Interpreter(model_path=str(tflite_model_16x8_file))
interpreter_16x8.allocate_tensors()

在单个图像上测试模型

test_image = np.expand_dims(test_images[0], axis=0).astype(np.float32)

input_index = interpreter.get_input_details()[0]["index"]
output_index = interpreter.get_output_details()[0]["index"]

interpreter.set_tensor(input_index, test_image)
interpreter.invoke()
predictions = interpreter.get_tensor(output_index)
import matplotlib.pylab as plt

plt.imshow(test_images[0])
template = "True:{true}, predicted:{predict}"
_ = plt.title(template.format(true= str(test_labels[0]),
                              predict=str(np.argmax(predictions[0]))))
plt.grid(False)

png

test_image = np.expand_dims(test_images[0], axis=0).astype(np.float32)

input_index = interpreter_16x8.get_input_details()[0]["index"]
output_index = interpreter_16x8.get_output_details()[0]["index"]

interpreter_16x8.set_tensor(input_index, test_image)
interpreter_16x8.invoke()
predictions = interpreter_16x8.get_tensor(output_index)
plt.imshow(test_images[0])
template = "True:{true}, predicted:{predict}"
_ = plt.title(template.format(true= str(test_labels[0]),
                              predict=str(np.argmax(predictions[0]))))
plt.grid(False)

png

评估模型

# A helper function to evaluate the TF Lite model using "test" dataset.
def evaluate_model(interpreter):
  input_index = interpreter.get_input_details()[0]["index"]
  output_index = interpreter.get_output_details()[0]["index"]

  # Run predictions on every image in the "test" dataset.
  prediction_digits = []
  for test_image in test_images:
    # Pre-processing: add batch dimension and convert to float32 to match with
    # the model's input data format.
    test_image = np.expand_dims(test_image, axis=0).astype(np.float32)
    interpreter.set_tensor(input_index, test_image)

    # Run inference.
    interpreter.invoke()

    # Post-processing: remove batch dimension and find the digit with highest
    # probability.
    output = interpreter.tensor(output_index)
    digit = np.argmax(output()[0])
    prediction_digits.append(digit)

  # Compare prediction results with ground truth labels to calculate accuracy.
  accurate_count = 0
  for index in range(len(prediction_digits)):
    if prediction_digits[index] == test_labels[index]:
      accurate_count += 1
  accuracy = accurate_count * 1.0 / len(prediction_digits)

  return accuracy
print(evaluate_model(interpreter))
0.9622

在 16x8 量化模型上重复评估:

# NOTE: This quantization mode is an experimental post-training mode,
# it does not have any optimized kernels implementations or
# specialized machine learning hardware accelerators. Therefore,
# it could be slower than the float interpreter.
print(evaluate_model(interpreter_16x8))
0.9625

在此示例中,您已将模型量化为 16x8 模型,准确率没有任何差异,但文件大小只有原来的 1/3。