# 张量处理单元 **张量处理单元**（ **TPU** ）是**专用集成电路**（ **ASIC** ），它实现了针对计算要求而优化的硬件电路深度神经网络。 TPU 基于**复杂指令集计算机**（ **CISC** ）指令集，该指令集实现用于训练深度神经网络的复杂任务的高级指令。 TPU 架构的核心在于优化矩阵运算的脉动数组。 The Architecture of TPUImage from: https://cloud.google.com/blog/big-data/2017/05/images/149454602921110/tpu-15.png TensorFlow 提供了一个编译器和软件堆栈，可将 API 调用从 TensorFlow 图转换为 TPU 指令。以下框图描述了在 TPU 堆栈顶部运行的 TensorFlow 模型的体系结构： Image from: https://cloud.google.com/blog/big-data/2017/05/images/149454602921110/tpu-2.pngFor more information on the TPU architecture, read the blog at the following link: [https://cloud.google.com/blog/big-data/2017/05/an-in-depth-look-at-googles-first-tensor-processing-unit-tpu.](https://cloud.google.com/blog/big-data/2017/05/an-in-depth-look-at-googles-first-tensor-processing-unit-tpu) TPU 的 TensorFlow API 位于`tf.contrib.tpu`模块中。为了在 TPU 上构建模型，使用以下三个 TPU 特定的 TensorFlow 模块： * `tpu_config`：`tpu_config`模块允许您创建配置对象，其中包含有关将运行模型的主机的信息。 * `tpu_estimator`：`tpu_estimator`模块将估计器封装在[H​​TG2]类中。要在 TPU 上运行估计器，我们创建此类的对象。 * `tpu_optimizer`：`tpu_optimizer`模块包装优化器。例如，在下面的示例代码中，我们将`tpu_optimizer`类中的 SGD 优化器包装在`tpu_optimizer`类中。 例如，以下代码使用 TF Estimator API 为 TPU 上的 MNIST 数据集构建 CNN 模型： 以下代码改编自 [https://github.com/tensorflow/tpu-demos/blob/master/cloud_tpu/models/mnist/mnist.py.](https://github.com/tensorflow/tpu-demos/blob/master/cloud_tpu/models/mnist/mnist.py) ```py import tensorflow as tf from tensorflow.contrib.tpu.python.tpu import tpu_config from tensorflow.contrib.tpu.python.tpu import tpu_estimator from tensorflow.contrib.tpu.python.tpu import tpu_optimizer learning_rate = 0.01 batch_size = 128 def metric_fn(labels, logits): predictions = tf.argmax(logits, 1) return { "accuracy": tf.metrics.precision( labels=labels, predictions=predictions), } def model_fn(features, labels, mode): if mode == tf.estimator.ModeKeys.PREDICT: raise RuntimeError("mode {} is not supported yet".format(mode)) input_layer = tf.reshape(features, [-1, 28, 28, 1]) conv1 = tf.layers.conv2d( inputs=input_layer, filters=32, kernel_size=[5, 5], padding="same", activation=tf.nn.relu) pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2) conv2 = tf.layers.conv2d( inputs=pool1, filters=64, kernel_size=[5, 5], padding="same", activation=tf.nn.relu) pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2) pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64]) dense = tf.layers.dense(inputs=pool2_flat, units=128, activation=tf.nn.relu) dropout = tf.layers.dropout( inputs=dense, rate=0.4, training=mode == tf.estimator.ModeKeys.TRAIN) logits = tf.layers.dense(inputs=dropout, units=10) onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10) loss = tf.losses.softmax_cross_entropy( onehot_labels=onehot_labels, logits=logits) if mode == tf.estimator.ModeKeys.EVAL: return tpu_estimator.TPUEstimatorSpec( mode=mode, loss=loss, eval_metrics=(metric_fn, [labels, logits])) # Train. decaying_learning_rate = tf.train.exponential_decay(learning_rate, tf.train.get_global_step(), 100000,0.96) optimizer = tpu_optimizer.CrossShardOptimizer( tf.train.GradientDescentOptimizer( learning_rate=decaying_learning_rate)) train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step()) return tpu_estimator.TPUEstimatorSpec(mode=mode, loss=loss, train_op=train_op) def get_input_fn(filename): def input_fn(params): batch_size = params["batch_size"] def parser(serialized_example): features = tf.parse_single_example( serialized_example, features={ "image_raw": tf.FixedLenFeature([], tf.string), "label": tf.FixedLenFeature([], tf.int64), }) image = tf.decode_raw(features["image_raw"], tf.uint8) image.set_shape([28 * 28]) image = tf.cast(image, tf.float32) * (1\. / 255) - 0.5 label = tf.cast(features["label"], tf.int32) return image, label dataset = tf.data.TFRecordDataset( filename, buffer_size=FLAGS.dataset_reader_buffer_size) dataset = dataset.map(parser).cache().repeat() dataset = dataset.apply( tf.contrib.data.batch_and_drop_remainder(batch_size)) images, labels = dataset.make_one_shot_iterator().get_next() return images, labels return input_fn # TPU config master = 'local' #URL of the TPU instance model_dir = '/home/armando/models/mnist' n_iterations = 50 # number of iterations per TPU training loop n_shards = 8 # number of TPU chips run_config = tpu_config.RunConfig( master=master, evaluation_master=master, model_dir=model_dir, session_config=tf.ConfigProto( allow_soft_placement=True, log_device_placement=True ), tpu_config=tpu_config.TPUConfig(n_iterations, n_shards ) ) estimator = tpu_estimator.TPUEstimator( model_fn=model_fn, use_tpu=True, train_batch_size=batch_size, eval_batch_size=batch_size, config=run_config) train_file = '/home/armando/datasets/mnist/train' # input data file train_steps = 1000 # number of steps to train for estimator.train(input_fn=get_input_fn(train_file), max_steps=train_steps ) eval_file = '/home/armando/datasets/mnist/test' # test data file eval_steps = 10 estimator.evaluate(input_fn=get_input_fn(eval_file), steps=eval_steps ) ``` 有关在 TPU 上构建模型的更多示例，请访问以下链接：[https://github.com/tensorflow/tpu-demos.](https://github.com/tensorflow/tpu-demos)