Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge

训练后动态范围量化

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

TensorFlow Lite 现在支持将权重转换为 8 位精度,作为从 TensorFlow GraphDef 到 TensorFlow Lite FlatBuffer 格式的模型转换的一部分。动态范围量化能使模型大小缩减至原来的四分之一。此外,TFLite 支持对激活进行实时量化和反量化以实现以下效果:

  1. 在可用时使用量化内核加快实现速度。
  2. 将计算图不同部分的浮点内核与量化内核混合。

激活始终以浮点进行存储。对于支持量化内核的算子,激活会在处理前动态量化为 8 位精度,并在处理后反量化为浮点精度。根据被转换的模型,这可以提供比纯浮点计算更快的速度。

量化感知训练相比,在此方法中,权重会在训练后量化,激活会在推断时动态量化。因此,不会重新训练模型权重以补偿量化引起的误差。请务必检查量化模型的准确率,以确保下降程度可以接受。

本教程将从头开始训练一个 MNIST 模型,在 TensorFlow 中检查其准确率,然后使用动态范围量化将此模型转换为 Tensorflow Lite FlatBuffer 格式。最后,检查转换后模型的准确率,并将其与原始浮点模型进行比较。

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

训练 TensorFlow 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 21:11:39.064840: 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 21:11:39.072951: 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 21:11:39.073858: 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 21:11:39.075898: 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 21:11:39.076455: 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 21:11:39.077347: 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 21:11:39.078195: 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 21:11:39.672900: 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 21:11:39.674027: 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 21:11:39.674885: 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 21:11:39.675698: 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 21:11:40.547425: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)
2021-08-13 21:11:41.275212: I tensorflow/stream_executor/cuda/cuda_dnn.cc:369] Loaded cuDNN version 8100
2021-08-13 21:11:41.832126: I tensorflow/core/platform/default/subprocess.cc:304] Start cannot spawn child process: No such file or directory
1875/1875 [==============================] - 6s 2ms/step - loss: 0.2566 - accuracy: 0.9299 - val_loss: 0.1167 - val_accuracy: 0.9644
<keras.callbacks.History at 0x7f91c25994d0>

在此示例中,由于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 加载模型:

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
2021-08-13 21:11:46.362395: 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/tmpoxyhfyhm/assets
2021-08-13 21:11:46.730174: 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 21:11:46.730513: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
2021-08-13 21:11:46.730609: I tensorflow/core/grappler/clusters/single_machine.cc:357] Starting new session
2021-08-13 21:11:46.730924: 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 21:11:46.731242: 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 21:11:46.731496: 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 21:11:46.731824: 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 21:11:46.732088: 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 21:11:46.732325: 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 21:11:46.733666: I tensorflow/core/grappler/optimizers/meta_optimizer.cc:1137] Optimization results for grappler item: graph_to_optimize
  function_optimizer: function_optimizer did nothing. time = 0.006ms.
  function_optimizer: function_optimizer did nothing. time = 0.002ms.

2021-08-13 21:11:46.764103: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format.
2021-08-13 21:11:46.764132: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
2021-08-13 21:11:46.767194: 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

要在导出时量化模型,请设置 optimizations 标记以优化大小:

converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_quant_model = converter.convert()
tflite_model_quant_file = tflite_models_dir/"mnist_model_quant.tflite"
tflite_model_quant_file.write_bytes(tflite_quant_model)
INFO:tensorflow:Assets written to: /tmp/tmp8oc3x8_8/assets
INFO:tensorflow:Assets written to: /tmp/tmp8oc3x8_8/assets
2021-08-13 21:11:47.287814: 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 21:11:47.288145: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
2021-08-13 21:11:47.288245: I tensorflow/core/grappler/clusters/single_machine.cc:357] Starting new session
2021-08-13 21:11:47.288615: 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 21:11:47.288947: 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 21:11:47.289201: 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 21:11:47.289534: 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 21:11:47.289829: 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 21:11:47.290070: 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 21:11:47.291429: I tensorflow/core/grappler/optimizers/meta_optimizer.cc:1137] Optimization results for grappler item: graph_to_optimize
  function_optimizer: function_optimizer did nothing. time = 0.006ms.
  function_optimizer: function_optimizer did nothing. time = 0ms.

2021-08-13 21:11:47.324398: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format.
2021-08-13 21:11:47.324429: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
2021-08-13 21:11:47.342477: I tensorflow/lite/tools/optimize/quantize_weights.cc:225] Skipping quantization of tensor sequential/conv2d/Conv2D because it has fewer than 1024 elements (108).
23904

请注意,生成文件的大小约为 1/4

ls -lh {tflite_models_dir}
total 180K
-rw-rw-r-- 1 kbuilder kbuilder 83K Aug 13 21:11 mnist_model.tflite
-rw-rw-r-- 1 kbuilder kbuilder 24K Aug 13 21:11 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 TFLite 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_quant = tf.lite.Interpreter(model_path=str(tflite_model_quant_file))
interpreter_quant.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

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))
0.9644

在动态范围量化模型上重复评估,以获得如下结果:

print(evaluate_model(interpreter_quant))
0.9643

在此示例中,压缩后的模型在准确率方面没有差别。

优化现有模型

带有预激活层的 ResNet (ResNet-v2) 被广泛用于视觉应用。用于 ResNet-v2-101 的预训练冻结计算图可在 Tensorflow Hub 上获得。

您可以通过执行以下代码,使用量化将冻结计算图转换为 TensorFLow Lite FlatBuffer 格式:

import tensorflow_hub as hub

resnet_v2_101 = tf.keras.Sequential([
  keras.layers.InputLayer(input_shape=(224, 224, 3)),
  hub.KerasLayer("https://hub.tensorflow.google.cn/google/imagenet/resnet_v2_101/classification/4")
])

converter = tf.lite.TFLiteConverter.from_keras_model(resnet_v2_101)
# Convert to TF Lite without quantization
resnet_tflite_file = tflite_models_dir/"resnet_v2_101.tflite"
resnet_tflite_file.write_bytes(converter.convert())
WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
INFO:tensorflow:Assets written to: /tmp/tmpzts3lygv/assets
INFO:tensorflow:Assets written to: /tmp/tmpzts3lygv/assets
2021-08-13 21:12:07.060624: 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 21:12:07.060972: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
2021-08-13 21:12:07.061127: I tensorflow/core/grappler/clusters/single_machine.cc:357] Starting new session
2021-08-13 21:12:07.061657: 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 21:12:07.062043: 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 21:12:07.062374: 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 21:12:07.062746: 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 21:12:07.063025: 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 21:12:07.063286: 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 21:12:07.178043: I tensorflow/core/grappler/optimizers/meta_optimizer.cc:1137] Optimization results for grappler item: graph_to_optimize
  function_optimizer: Graph size after: 3495 nodes (2947), 5719 edges (5171), time = 72.187ms.
  function_optimizer: function_optimizer did nothing. time = 1.031ms.

2021-08-13 21:12:13.018055: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format.
2021-08-13 21:12:13.018107: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
178509352
# Convert to TF Lite with quantization
converter.optimizations = [tf.lite.Optimize.DEFAULT]
resnet_quantized_tflite_file = tflite_models_dir/"resnet_v2_101_quantized.tflite"
resnet_quantized_tflite_file.write_bytes(converter.convert())
WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
INFO:tensorflow:Assets written to: /tmp/tmpjjrhzwhh/assets
INFO:tensorflow:Assets written to: /tmp/tmpjjrhzwhh/assets
2021-08-13 21:12:23.433004: 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 21:12:23.433369: I tensorflow/core/grappler/devices.cc:66] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 1
2021-08-13 21:12:23.433475: I tensorflow/core/grappler/clusters/single_machine.cc:357] Starting new session
2021-08-13 21:12:23.433918: 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 21:12:23.434267: 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 21:12:23.434553: 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 21:12:23.434933: 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 21:12:23.435209: 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 21:12:23.435451: 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 21:12:23.550619: I tensorflow/core/grappler/optimizers/meta_optimizer.cc:1137] Optimization results for grappler item: graph_to_optimize
  function_optimizer: Graph size after: 3495 nodes (2947), 5719 edges (5171), time = 70.348ms.
  function_optimizer: function_optimizer did nothing. time = 1.2ms.

2021-08-13 21:12:28.017792: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format.
2021-08-13 21:12:28.017846: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
46256864
ls -lh {tflite_models_dir}/*.tflite
-rw-rw-r-- 1 kbuilder kbuilder  83K Aug 13 21:11 /tmp/mnist_tflite_models/mnist_model.tflite
-rw-rw-r-- 1 kbuilder kbuilder  24K Aug 13 21:11 /tmp/mnist_tflite_models/mnist_model_quant.tflite
-rw-rw-r-- 1 kbuilder kbuilder  25K Aug 13 21:10 /tmp/mnist_tflite_models/mnist_model_quant_16x8.tflite
-rw-rw-r-- 1 kbuilder kbuilder  44K Aug 13 21:08 /tmp/mnist_tflite_models/mnist_model_quant_f16.tflite
-rw-rw-r-- 1 kbuilder kbuilder 171M Aug 13 21:12 /tmp/mnist_tflite_models/resnet_v2_101.tflite
-rw-rw-r-- 1 kbuilder kbuilder  45M Aug 13 21:12 /tmp/mnist_tflite_models/resnet_v2_101_quantized.tflite

模型大小从 171 MB 减小到 43 MB。可以使用为 TFLite 准确率测量提供的脚本来评估此模型在 ImageNet 上的准确率。

优化后模型的 Top-1 准确率为 76.8,与浮点模型相同。