如何在MNIST上使用tf.contrib.model_pruning？

Question

我很难使用Tensorflow的剪枝库，也没有找到很多有用的例子，所以我正在寻找帮助来修剪一个在MNIST数据集上训练过的简单模型。如果有人可以帮助修复我的尝试，或者提供一个如何在MNIST上使用这个库的例子，我将非常感激。

我的代码的前半部分是相当标准的，除了我的模型有两个隐藏层，300个单元宽，使用layers.masked_fully_connected进行剪枝。

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，我尝试定义必要的剪枝操作，但是我得到了一个错误。

############ 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)

这一行出错：

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

_device="/job:localhost/replica:0/task:0/device:GPU:0"]] (回溯参见上文)：您必须为占位符张量' Placeholder_1‘提供一个值，其中包含dtype InvalidArgumentError InvalidArgumentError ?,10[Node: Placeholder_1= Placeholderdtype=DT_FLOAT，shape=?,10，Placeholder_1 [Node: global_step/_57 = _Recv_start_time=0，client_terminated=false，recv_device="/job:localhost/replica:0/task:0/device:CPU:0"，]send_device="/job:localhost/replica:0/task:0/device:GPU:0"，send_device_incarnation=1，tensor_name="edge_71_global_step"，tensor_type=DT_INT64，_device="/job:localhost/replica:0/task:0/device:CPU:0"]

我猜想它需要一种不同类型的操作来代替train_op，但是我还没有发现任何有效的调整。

同样，如果你有一个不同的工作例子，修剪一个在MNIST上训练的模型，我会考虑这个答案。

Answer 1

最简单的剪枝库示例，我可以得到工作，我想我应该张贴在这里，以防它帮助其他noobie谁有困难的文档。

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()))

Answer 2

罗曼尼基辛要求的代码，可以保存模型，这是一个轻微的扩展，我原来的答案。

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
model_path_unpruned = "Model_Saves/Unpruned.ckpt"
model_path_pruned = "Model_Saves/Pruned.ckpt"

# 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()

# Create a saver for writing training checkpoints.
saver = tf.train.Saver()

with tf.Session() as sess:

    # Uncomment the following if you don't have a trained model yet
    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()))

    # Saves the model before pruning
    saver.save(sess, model_path_unpruned)

    # Resets the session and restores the saved model
    sess.run(tf.initialize_all_variables())
    saver.restore(sess, model_path_unpruned)

    # 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()))

    # Saves the model after pruning
    saver.save(sess, model_path_pruned)

    # 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()))