-
学习流程：
Estimator封装了对机器学习不同阶段的控制，用户无需不断的为新机器学习任务重复编写训练、评估、预测的代码。可以专注于对网络结构的控制。 -
网络结构：
Estimator的网络结构是在model_fn中独立定义的，用户创建的任何网络结构都可以在Estimator的控制下进行机器学习。这可以允许用户很方便的使用别人定义好的model_fn。 -
数据导入：
Estimator的数据导入也是由input_fn独立定义的。例如，用户可以非常方便的只通过改变input_fn的定义，来使用相同的网络结构学习不同的数据。 -
model_fn模型函数必须要有features,mode两个参数，可自己选择加入labels（可以把label也放进features中）。最后要返回特定的tf.estimator.EstimatorSpec()。模型有三个阶段都共用的正向传播部分，和由mode值来控制返回不同tf.estimator.EstimatorSpec的三个分支。 -
预测分支，将想要预测的值写进字典里面
# 创建predictions字典，里面写进所有你想要在预测值输出的数值
# 隐藏层的数值也可以，这里演示了输出所有隐藏层层结果。
# 字典的key是模型，value给的是对应的tensor
predictions = {
"image":features['image'],
"conv1_out":conv1,
"pool1_out":pool1,
"conv2_out":conv2,
"pool2_out":pool2,
"pool2_flat_out":pool2_flat,
"dense1_out":dense1,
"logits":logits,
"classes": tf.argmax(input=logits, axis=1),
"labels": features['label'],
"probabilities": tf.nn.softmax(logits, name="softmax_tensor")
}
# 当mode为tf.estimator.ModeKeys.PREDICT时，我们就让模型返回预测的操作
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
- 训练分支需要loss和train_op操作符。
# 训练和评估时都会用到loss
loss = tf.losses.sparse_softmax_cross_entropy(labels=features['label'], logits=logits)
# 训练分支
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer = tf.train.AdamOptimizer(learning_rate=1e-3)
train_op = optimizer.minimize(
loss=loss,
# global_step用于记录训练了多少步
global_step=tf.train.get_global_step())
# 返回的tf.estimator.EstimatorSpec根据
return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op)
- 评估分支
# 注意评估的时候，模型和训练时一样，是一个循环的loop，不断累积计算评估指标。
# 其中有两个局部变量total和count来控制
# 把网络中的某个tensor结果直接作为字典的value是不好用的
# loss的值是始终做记录的，eval_metric_ops中是额外想要知道的评估指标
eval_metric_ops = {"accuracy": tf.metrics.accuracy(labels=features['label'], predictions=predictions["classes"])}
# 不好用：eval_metric_ops = {"probabilities": predictions["probabilities"]}
return tf.estimator.EstimatorSpec(mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)
创建一个实例
- 训练后的模型参数会保存在
model_dir中，随着训练在目录下生成拥有类似下面内容的checkpoint文件。
# model_dir 表示模型要存到哪里
mnist_classifier = tf.estimator.Estimator(
model_fn=model_fn, model_dir="mnist_model_cnn")
在训练或评估中利用Hook打印中间信息。
-
hooks：如果不送值，则训练过程中不会显示字典中的数值。 -
steps：指定了训练多少次，如果不送值，则训练到dataset API遍历完数据集为止。 -
max_steps：指定了最大训练次数。
# 在训练或评估的循环中，每50次print出一次字典中的数值
tensors_to_log = {"probabilities": "softmax_tensor"}
logging_hook = tf.train.LoggingTensorHook(tensors=tensors_to_log, every_n_iter=50)
mnist_classifier.train(input_fn=train_input_fn, hooks=[logging_hook])
评估
# 训练集
eval_results = mnist_classifier.evaluate(input_fn=train_eval_fn, checkpoint_path=None)
print('train set')
print(eval_results)
# 测试集
# checkpoint_path是可以指定选择那个时刻保存的权重进行评估
eval_results = mnist_classifier.evaluate(input_fn=test_input_fn, checkpoint_path=None)
print('test set')
print(eval_results)
评估
predicts =list(mnist_classifier.predict(input_fn=test_input_fn))
predicts[0].keys()
# 输出为：
dict_keys(['image', 'conv1_out', 'pool1_out', 'conv2_out', 'pool2_out', 'pool2_flat_out', 'dense1_out', 'logits', 'classes', 'labels', 'probabilities'])
如果想要输出第四个卷积
plt.figure(num=4,figsize=(28,28))
for i in range(4):
plt.subplot(1,4,i+1)
plt.imshow(predicts[0]['conv1_out'][:,:,i],cmap = plt.cm.gray)
plt.savefig('conv1_out.png')
一个范例
def model_fn(features, labels, mode, params):
if isinstance(features, dict): # For serving
features = features['feature']
predictions = tf.layers.dense(features, 1)
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode, predictions=predictions)
else:
loss = tf.nn.l2_loss(predictions - labels)
if mode == tf.estimator.ModeKeys.EVAL:
return tf.estimator.EstimatorSpec(
mode, loss=loss)
elif mode == tf.estimator.ModeKeys.TRAIN:
train_op = tf.train.AdamOptimizer(learning_rate=0.5).minimize(
loss, global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(
mode, loss=loss, train_op=train_op)
else:
raise NotImplementedError()
early stopping
函数原型
tf.contrib.estimator.stop_if_no_increase_hook(
estimator,
metric_name,
max_steps_without_increase,
eval_dir=None,
min_steps=0,
run_every_secs=60,
run_every_steps=None
)
-
metric_name: str类型，比如loss或者accuracy. -
max_steps_without_increase: int，如果没有增加的最大长是多少，如果超过了这个最大步长metric还是没有增加那么就会停止。 -
eval_dir：默认是使用estimator.eval_dir目录，用于存放评估的summary file。 -
min_steps：训练的最小步长，如果训练小于这个步长那么永远都不会停止。 -
run_every_secs和run_every_steps：表示多长时间获得步长调用一次should_stop_fn。
import tf_metrics
# Metrics
weights = tf.sequence_mask(nwords)
metrics = {
'acc': tf.metrics.accuracy(tags, pred_ids, weights),
'precision': tf_metrics.precision(tags, pred_ids, num_tags, indices, weights),
'recall': tf_metrics.recall(tags, pred_ids, num_tags, indices, weights),
'f1': tf_metrics.f1(tags, pred_ids, num_tags, indices, weights),
}
# Tensoboard summaries
for metric_name, op in metrics.items():
tf.summary.scalar(metric_name, op[1])
# 1. Define our input_fn
train_inpf = functools.partial(input_fn, 'words.train.txt', 'tags.train.txt',
params, shuffle_and_repeat=True)
eval_inpf = functools.partial(input_fn,'words.testa.txt', 'tags.testa.txt'
params)
# 2. Create a hook
Path(estimator.eval_dir()).mkdir(parents=True, exist_ok=True)
hook = tf.contrib.estimator.stop_if_no_increase_hook(
estimator, 'f1', 500, min_steps=8000, run_every_secs=120)
train_spec = tf.estimator.TrainSpec(input_fn=input_fn, hooks=[hook])
eval_spec = tf.estimator.EvalSpec(input_fn=eval_inpf, throttle_secs=120)
# 3. Train with early stopping
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
第二个范例
-
model_fn的定义有固定的标准，包括其对应的输入参数和返回参数。 -
model_fn定义了模型的网络结果，损失护士，优化函数以及评估参数等一系列的指标，可以根据自身需求进行相应的删减。
def model_fn(features, labels, mode, params):
# Args:
#
# features: This is the x-arg from the input_fn.
# labels: This is the y-arg from the input_fn,
# see e.g. train_input_fn for these two.
# mode: Either TRAIN, EVAL, or PREDICT
# params: User-defined hyper-parameters, e.g. learning-rate.
# Reference to the tensor named "x" in the input-function.
# x = features["images"]
x = tf.feature_column.input_layer(features, params['feature_columns'])
# The convolutional layers expect 4-rank tensors
# but x is a 2-rank tensor, so reshape it.
net = tf.reshape(x, [-1, IMG_HEIGHT, IMG_WIDTH, NUM_CHANNEL])
# First convolutional layer.
net = tf.layers.conv2d(inputs=net, name='layer_conv1',
filters=16, kernel_size=5,
padding='same', activation=tf.nn.relu)
net = tf.layers.max_pooling2d(inputs=net, pool_size=2, strides=2)
# Second convolutional layer.
net = tf.layers.conv2d(inputs=net, name='layer_conv2',
filters=36, kernel_size=5,
padding='same', activation=tf.nn.relu)
net = tf.layers.max_pooling2d(inputs=net, pool_size=2, strides=2)
# Flatten to a 2-rank tensor.
net = tf.contrib.layers.flatten(net)
# Eventually this should be replaced with:
# net = tf.layers.flatten(net)
# First fully-connected / dense layer.
# This uses the ReLU activation function.
net = tf.layers.dense(inputs=net, name='layer_fc1',
units=128, activation=tf.nn.relu)
# Second fully-connected / dense layer.
# This is the last layer so it does not use an activation function.
net = tf.layers.dense(inputs=net, name='layer_fc2',
units=10)
# Logits output of the neural network.
logits = net
# Softmax output of the neural network.
y_pred = tf.nn.softmax(logits=logits)
# Classification output of the neural network.
y_pred_cls = tf.argmax(y_pred, axis=1)
if mode == tf.estimator.ModeKeys.PREDICT:
# If the estimator is supposed to be in prediction-mode
# then use the predicted class-number that is output by
# the neural network. Optimization etc. is not needed.
spec = tf.estimator.EstimatorSpec(mode=mode,
predictions=y_pred_cls)
else:
# Otherwise the estimator is supposed to be in either
# training or evaluation-mode. Note that the loss-function
# is also required in Evaluation mode.
# Define the loss-function to be optimized, by first
# calculating the cross-entropy between the output of
# the neural network and the true labels for the input data.
# This gives the cross-entropy for each image in the batch.
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
# Reduce the cross-entropy batch-tensor to a single number
# which can be used in optimization of the neural network.
loss = tf.reduce_mean(cross_entropy)
# Define the optimizer for improving the neural network.
optimizer = tf.train.AdamOptimizer(learning_rate=params["learning_rate"])
# Get the TensorFlow op for doing a single optimization step.
train_op = optimizer.minimize(
loss=loss, global_step=tf.train.get_global_step())
# Define the evaluation metrics,
# in this case the classification accuracy.
metrics = \
{
"accuracy": tf.metrics.accuracy(labels, y_pred_cls)
}
# Wrap all of this in an EstimatorSpec.
spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=metrics)
return spec
实例化这个Estimaotr
-
params：是传入数据字典。 -
model_dir：是模型和log的保存路径。
params = {"learning_rate": 1e-4,
'feature_columns': feature_columns}
model = tf.estimator.Estimator(model_fn=model_fn,
params=params,
model_dir="./cnn_classifer_dataset/")
接下来我们只需要简单调用train，evaluate，predict方法，就可以实现对模型的训练验证和测试了。
- 训练：
input_fn = lambda: csv_input_fn(\
files_name_pattern= TRAIN_DATA_FILES_PATTERN,mode=tf.estimator.ModeKeys.TRAIN)
# Train the Model
model.train(input_fn, steps=2000)
- 验证
input_fn = lambda: csv_input_fn(files_name_pattern= VAL_DATA_FILES_PATTERN,mode=tf.estimator.ModeKeys.EVAL)
model.evaluate(input_fn,steps=1)
测试
import itertools
input_fn = lambda: csv_input_fn(\
files_name_pattern= TEST_DATA_FILES_PATTERN,mode=tf.estimator.ModeKeys.PREDICT,batch_size=10)
predictions = list(itertools.islice(model.predict(input_fn=input_fn),10))
print('PREDICTIONS',predictions)
范例三
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import argparse
import math
import tensorflow as tf
FLAGS = None
batch_size = 100
def _my_input_fn(filepath, num_epochs):
# image - 784 (=28 x 28) elements of grey-scaled integer value [0, 1]
# label - digit (0, 1, ..., 9)
data_queue = tf.train.string_input_producer(
[filepath],
num_epochs = num_epochs) # data is repeated and it raises OutOfRange when data is over
data_reader = tf.TFRecordReader()
_, serialized_exam = data_reader.read(data_queue)
data_exam = tf.parse_single_example(
serialized_exam,
features={
'image_raw': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64)
})
data_image = tf.decode_raw(data_exam['image_raw'], tf.uint8)
data_image.set_shape([784])
data_image = tf.cast(data_image, tf.float32) * (1. / 255)
data_label = tf.cast(data_exam['label'], tf.int32)
data_batch_image, data_batch_label = tf.train.batch(
[data_image, data_label],
batch_size=batch_size)
return data_batch_image, data_batch_label
def _get_input_fn(filepath, num_epochs):
return lambda: _my_input_fn(filepath, num_epochs)
def _my_model_fn(features, labels, mode):
# with tf.device(...): # You can set device if using GPUs
# define network and inference
# (simple 2 fully connected hidden layer : 784->128->64->10)
with tf.name_scope('hidden1'):
weights = tf.Variable(
tf.truncated_normal(
[784, 128],
stddev=1.0 / math.sqrt(float(784))),
name='weights')
biases = tf.Variable(
tf.zeros([128]),
name='biases')
hidden1 = tf.nn.relu(tf.matmul(features, weights) + biases)
with tf.name_scope('hidden2'):
weights = tf.Variable(
tf.truncated_normal(
[128, 64],
stddev=1.0 / math.sqrt(float(128))),
name='weights')
biases = tf.Variable(
tf.zeros([64]),
name='biases')
hidden2 = tf.nn.relu(tf.matmul(hidden1, weights) + biases)
with tf.name_scope('softmax_linear'):
weights = tf.Variable(
tf.truncated_normal(
[64, 10],
stddev=1.0 / math.sqrt(float(64))),
name='weights')
biases = tf.Variable(
tf.zeros([10]),
name='biases')
logits = tf.matmul(hidden2, weights) + biases
# compute evaluation matrix
predicted_indices = tf.argmax(input=logits, axis=1)
if mode != tf.estimator.ModeKeys.PREDICT:
label_indices = tf.cast(labels, tf.int32)
accuracy = tf.metrics.accuracy(label_indices, predicted_indices)
tf.summary.scalar('accuracy', accuracy[1]) # output to TensorBoard
# compute loss
loss = tf.losses.sparse_softmax_cross_entropy(
labels=labels,
logits=logits)
# define operations
if mode == tf.estimator.ModeKeys.TRAIN:
#global_step = tf.train.create_global_step()
#global_step = tf.contrib.framework.get_or_create_global_step()
global_step = tf.train.get_or_create_global_step()
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=0.07)
train_op = optimizer.minimize(
loss=loss,
global_step=global_step)
return tf.estimator.EstimatorSpec(
mode,
loss=loss,
train_op=train_op)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = {
'accuracy': accuracy
}
return tf.estimator.EstimatorSpec(
mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
if mode == tf.estimator.ModeKeys.PREDICT:
probabilities = tf.nn.softmax(logits, name='softmax_tensor')
predictions = {
'classes': predicted_indices,
'probabilities': probabilities
}
export_outputs = {
'prediction': tf.estimator.export.PredictOutput(predictions)
}
return tf.estimator.EstimatorSpec(
mode,
predictions=predictions,
export_outputs=export_outputs)
def main(_):
# read TF_CONFIG
run_config = tf.contrib.learn.RunConfig()
# define
mnist_fullyconnected_classifier = tf.estimator.Estimator(
model_fn=_my_model_fn,
model_dir=FLAGS.out_dir,
config=run_config)
train_spec = tf.estimator.TrainSpec(
input_fn=_get_input_fn(FLAGS.train_file, 2),
max_steps=60000 * 2 / batch_size)
eval_spec = tf.estimator.EvalSpec(
input_fn=_get_input_fn(FLAGS.test_file, 1),
steps=10000 * 1 / batch_size,
start_delay_secs=0)
# run !
tf.estimator.train_and_evaluate(
mnist_fullyconnected_classifier,
train_spec,
eval_spec
)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument(
'--train_file',
type=str,
default='/home/demouser/train.tfrecords',
help='File path for the training data.')
parser.add_argument(
'--test_file',
type=str,
default='/home/demouser/test.tfrecords',
help='File path for the test data.')
parser.add_argument(
'--out_dir',
type=str,
default='/home/demouser/out',
help='Dir path for the model and checkpoint output.')
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)
- 可以配合不同的hook使用
You cannot see the train loop (while clause for mini-batch) in this code, but you can use custom tf.train.SessionRunHook or tf.train.LoggingTensorHook for your custom code (ex : debug printing, etc) in the train loop.
网友评论