如何在MNIST上使用tf.contrib.model_pruning？

Question

我正在努力使用Tensorflow的修剪库，但没有找到许多有用的示例，因此我正在寻找帮助以修剪在MNIST数据集上训练的简单模型。如果有人可以帮助解决我的问题或提供如何在MNIST上使用该库的示例，我将不胜感激。

我的代码的前半部分是非常标准的，除了我的模型有2个隐藏的图层，其宽度layers.masked_fully_connected为300单位，用于修剪。

import tensorflow as tf
from tensorflow.contrib.model_pruning.python import pruning
from tensorflow.contrib.model_pruning.python.layers import layers
from tensorflow.examples.tutorials.mnist import input_data

# Import dataset
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

# Define Placeholders
image = tf.placeholder(tf.float32, [None, 784])
label = tf.placeholder(tf.float32, [None, 10])

# Define the model
layer1 = layers.masked_fully_connected(image, 300)
layer2 = layers.masked_fully_connected(layer1, 300)
logits = tf.contrib.layers.fully_connected(layer2, 10, tf.nn.relu)

# Loss function
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=label))

# Training op
train_op = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss)

# Accuracy ops
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(label, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

然后，我尝试定义必要的修剪操作，但出现错误。

############ Pruning Operations ##############
# Create global step variable
global_step = tf.contrib.framework.get_or_create_global_step()

# Create a pruning object using the pruning specification
pruning_hparams = pruning.get_pruning_hparams()
p = pruning.Pruning(pruning_hparams, global_step=global_step)

# Mask Update op
mask_update_op = p.conditional_mask_update_op()

# Set up the specification for model pruning
prune_train = tf.contrib.model_pruning.train(train_op=train_op, logdir=None, mask_update_op=mask_update_op)

此行错误：

prune_train = tf.contrib.model_pruning.train(train_op=train_op, logdir=None, mask_update_op=mask_update_op)

InvalidArgumentError（请参见上面的回溯）：您必须使用dtype float和shape [？，10]来输入占位符张量'Placeholder_1'的值[[Node：Placeholder_1 = Placeholderdtype = DT_FLOAT，shape = [?, 10]，_device =“ / job：本地主机/副本：0 /任务：0 /设备：GPU：0“]] [[节点：global_step / _57 = _Recv_start_time = 0，client_terminated = false，recv_device =” / job：localhost /副本：0 /任务：0 /设备：CPU：0“，send_device =” /作业：本地主机/副本：0 /任务：0 /设备：GPU：0“，send_device_incarnation = 1，tensor_name =” edge_71_global_step“，tensor_type = DT_INT64，_device =” / job：本地主机/副本：0 /任务：0 /设备：CPU：0“]]

我假设它想要代替train_op的其他类型的操作，但是我没有找到任何有效的调整方法。

同样，如果您有一个其他的工作示例修剪了在MNIST上训练的模型，我会认为这是一个答案。

Answer 1

我可以使用的最简单的修剪库示例，我想我会将其发布在这里，以防它帮助其他一些在文档方面遇到困难的新手。

import tensorflow as tf
from tensorflow.contrib.model_pruning.python import pruning
from tensorflow.contrib.model_pruning.python.layers import layers
from tensorflow.examples.tutorials.mnist import input_data

epochs = 250
batch_size = 55000 # Entire training set

# Import dataset
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
batches = int(len(mnist.train.images) / batch_size)

# Define Placeholders
image = tf.placeholder(tf.float32, [None, 784])
label = tf.placeholder(tf.float32, [None, 10])

# Define the model
layer1 = layers.masked_fully_connected(image, 300)
layer2 = layers.masked_fully_connected(layer1, 300)
logits = layers.masked_fully_connected(layer2, 10)

# Create global step variable (needed for pruning)
global_step = tf.train.get_or_create_global_step()
reset_global_step_op = tf.assign(global_step, 0)

# Loss function
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=label))

# Training op, the global step is critical here, make sure it matches the one used in pruning later
# running this operation increments the global_step
train_op = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss, global_step=global_step)

# Accuracy ops
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(label, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

# Get, Print, and Edit Pruning Hyperparameters
pruning_hparams = pruning.get_pruning_hparams()
print("Pruning Hyperparameters:", pruning_hparams)

# Change hyperparameters to meet our needs
pruning_hparams.begin_pruning_step = 0
pruning_hparams.end_pruning_step = 250
pruning_hparams.pruning_frequency = 1
pruning_hparams.sparsity_function_end_step = 250
pruning_hparams.target_sparsity = .9

# Create a pruning object using the pruning specification, sparsity seems to have priority over the hparam
p = pruning.Pruning(pruning_hparams, global_step=global_step, sparsity=.9)
prune_op = p.conditional_mask_update_op()

with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())

    # Train the model before pruning (optional)
    for epoch in range(epochs):
        for batch in range(batches):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            sess.run(train_op, feed_dict={image: batch_xs, label: batch_ys})

        # Calculate Test Accuracy every 10 epochs
        if epoch % 10 == 0:
            acc_print = sess.run(accuracy, feed_dict={image: mnist.test.images, label: mnist.test.labels})
            print("Un-pruned model step %d test accuracy %g" % (epoch, acc_print))

    acc_print = sess.run(accuracy, feed_dict={image: mnist.test.images, label: mnist.test.labels})
    print("Pre-Pruning accuracy:", acc_print)
    print("Sparsity of layers (should be 0)", sess.run(tf.contrib.model_pruning.get_weight_sparsity()))

    # Reset the global step counter and begin pruning
    sess.run(reset_global_step_op)
    for epoch in range(epochs):
        for batch in range(batches):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            # Prune and retrain
            sess.run(prune_op)
            sess.run(train_op, feed_dict={image: batch_xs, label: batch_ys})

        # Calculate Test Accuracy every 10 epochs
        if epoch % 10 == 0:
            acc_print = sess.run(accuracy, feed_dict={image: mnist.test.images, label: mnist.test.labels})
            print("Pruned model step %d test accuracy %g" % (epoch, acc_print))
            print("Weight sparsities:", sess.run(tf.contrib.model_pruning.get_weight_sparsity()))

    # Print final accuracy
    acc_print = sess.run(accuracy, feed_dict={image: mnist.test.images, label: mnist.test.labels})
    print("Final accuracy:", acc_print)
    print("Final sparsity by layer (should be 0)", sess.run(tf.contrib.model_pruning.get_weight_sparsity()))