拿来主义:看我的代码,我是在模型acc和验证数据集val_acc都达到99.8%时候才终止训练。

import numpy as np
import tflearn
from tflearn.layers.core import dropout
from tflearn.layers.normalization import batch_normalization
from tflearn.data_utils import to_categorical
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, confusion_matrix
import sys class EarlyStoppingCallback(tflearn.callbacks.Callback):
def __init__(self, val_acc_thresh):
""" Note: We are free to define our init function however we please. """
# Store a validation accuracy threshold, which we can compare against
# the current validation accuracy at, say, each epoch, each batch step, etc.
self.val_acc_thresh = val_acc_thresh def on_epoch_end(self, training_state):
"""
This is the final method called in trainer.py in the epoch loop.
We can stop training and leave without losing any information with a simple exception.
"""
#print dir(training_state)
print("Terminating training at the end of epoch", training_state.epoch)
if training_state.val_acc >= self.val_acc_thresh and training_state.acc_value >= self.val_acc_thresh:
raise StopIteration
def on_train_end(self, training_state):
"""
Furthermore, tflearn will then immediately call this method after we terminate training,
(or when training ends regardless). This would be a good time to store any additional
information that tflearn doesn't store already.
"""
print("Successfully left training! Final model accuracy:", training_state.acc_value) if __name__ == "__main__":
training_data = []
with open("feature_with_dnn_todo.dat") as f:
training_data = [parse_line(line) for line in f] X = training_data
org_labels = [1 if int(x[0])==2.0 else 0 for x in X]
labels = to_categorical(org_labels, nb_classes=2)
data = [x[1:] for x in X]
input_dim = len(data[0]) X = data
Y = labels print "X len:", len(X), "Y len:", len(Y)
trainX, testX, trainY, testY = train_test_split(X, Y, test_size=0.2, random_state=42)
print trainX[0]
print trainY[0]
print testX[-1]
print testY[-1] # Build neural network
net = tflearn.input_data(shape=[None, input_dim])
# RMSProp | epoch: 100 | loss: 0.25209 - acc: 0.9109 | val_loss: 0.19742 - val_acc: 0.9392 -- iter: 14084/14084 remove unwanted_cols 2
# | RMSProp | epoch: 100 | loss: 0.29420 - acc: 0.9075 | val_loss: 0.14464 - val_acc: 0.9551 -- iter: 14084/14084
net = batch_normalization(net)
dense1 = tflearn.fully_connected(net, 64, activation='tanh',
regularizer='L2', weight_decay=0.001)
dropout1 = tflearn.dropout(dense1, 0.8)
dense2 = tflearn.fully_connected(dropout1, 64, activation='tanh',
regularizer='L2', weight_decay=0.001)
dropout2 = tflearn.dropout(dense2, 0.8)
softmax = tflearn.fully_connected(dropout2, 2, activation='softmax') # Regression using SGD with learning rate decay and Top-3 accuracy
net = tflearn.regression(softmax, optimizer="rmsprop", learning_rate=0.001, loss='categorical_crossentropy') """
#| Adam | epoch: 100 | loss: 0.15578 - acc: 0.9419 | val_loss: 0.16620 - val_acc: 0.9392 -- iter: 14084/14084
net = batch_normalization(net)
net = tflearn.fully_connected(net, input_dim)
net = tflearn.fully_connected(net, 128, activation='tanh')
net = dropout(net, 0.5)
net = tflearn.fully_connected(net, 2, activation='softmax')
net = tflearn.regression(net, optimizer='adam', learning_rate=0.001,
loss='categorical_crossentropy', name='target')
"""
# Define model
model = tflearn.DNN(net)
# Start training (apply gradient descent algorithm)
# Initialize our callback with desired accuracy threshold.
early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.998)
try:
model.fit(trainX, trainY, validation_set=(testX, testY), n_epoch=500, batch_size=8, show_metric=True, callbacks=early_stopping_cb)
except StopIteration as e:
print "pass"
filename = 'dns_tunnel998.tflearn'
model.save(filename)
model.load(filename)
#model.fit(trainX, trainY, validation_set=(testX, testY), show_metric=True, batch_size=1024, n_epoch=5)
#model.fit(trainX, trainY, validation_set=(testX, testY), show_metric=True, batch_size=1024, n_epoch=5)
y_predict_list = model.predict(X)
y_predict = []
for i in y_predict_list:
#print i[0]
if i[0] >= 0.5:
y_predict.append(0)
else:
y_predict.append(1) print(classification_report(org_labels, y_predict))
print confusion_matrix(org_labels, y_predict)

The EarlyStoppingCallback Class

I show a proof-of-concept version of early stopping below. This is the simplest possible case: just stop training after the first epoch no matter what. It is up to the user to decide the conditions they want to trigger the stopping on.

class EarlyStoppingCallback(tflearn.callbacks.Callback):
def __init__(self, val_acc_thresh):
""" Note: We are free to define our init function however we please. """
# Store a validation accuracy threshold, which we can compare against
# the current validation accuracy at, say, each epoch, each batch step, etc.
self.val_acc_thresh = val_acc_thresh def on_epoch_end(self, training_state):
"""
This is the final method called in trainer.py in the epoch loop.
We can stop training and leave without losing any information with a simple exception.
"""
print("Terminating training at the end of epoch", training_state.epoch)
raise StopIteration def on_train_end(self, training_state):
"""
Furthermore, tflearn will then immediately call this method after we terminate training,
(or when training ends regardless). This would be a good time to store any additional
information that tflearn doesn't store already.
"""
print("Successfully left training! Final model accuracy:", training_state.acc_value) # Initialize our callback with desired accuracy threshold.
early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.5)

Result: Train the Model and Stop Early

try:
# Give it to our trainer and let it fit the data.
trainer.fit(feed_dicts={X: trainX, Y: trainY},
val_feed_dicts={X: testX, Y: testY},
n_epoch=1,
show_metric=True, # Calculate accuracy and display at every step.
callbacks=early_stopping_cb)
except StopIteration:
print("Caught callback exception. Returning control to user program.")
Training Step: 860  | total loss: [1m[32m1.73372[0m[0m
| Optimizer | epoch: 002 | loss: 1.73372 - acc: 0.8196 | val_loss: 1.87058 - val_acc: 0.8011 -- iter: 55000/55000
Training Step: 860 | total loss: [1m[32m1.73372[0m[0m
| Optimizer | epoch: 002 | loss: 1.73372 - acc: 0.8196 | val_loss: 1.87058 - val_acc: 0.8011 -- iter: 55000/55000
--
Terminating training at the end of epoch 2
Successfully left training! Final model accuracy: 0.8196054697036743
Caught callback exception. Returning control to user program.

Appendix

For my own reference, this is the code I started with before tinkering with the early stopping solution above.

from __future__ import division, print_function, absolute_import

import os
import sys
import tempfile
import urllib
import collections
import math import numpy as np
import tensorflow as tf
from scipy.io import arff import tflearn
from sklearn.utils import shuffle
from sklearn.metrics import roc_auc_score
from tflearn.data_utils import shuffle, to_categorical
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.normalization import local_response_normalization, batch_normalization
from tflearn.layers.estimator import regression
import tflearn.datasets.mnist as mnist # Load the data and handle any preprocessing here.
X, Y, testX, testY = mnist.load_data(one_hot=True)
X, Y = shuffle(X, Y)
X = X.reshape([-1, 28, 28, 1])
testX = testX.reshape([-1, 28, 28, 1]) # Define our network architecture: a simple 2-layer network of the form
# InputImages -> Fully Connected -> Softmax
out_readin1 = input_data(shape=[None,28,28,1])
out_fully_connected2 = fully_connected(out_readin1, 10)
out_softmax3 = fully_connected(out_fully_connected2, 10, activation='softmax') hash='f0c188c3777519fb93f1a825ca758a0c'
scriptid='MNIST-f0c188c3777519fb93f1a825ca758a0c' # Define our training metrics.
network = regression(out_softmax3,
optimizer='adam',
learning_rate=0.01,
loss='categorical_crossentropy',
name='target')
model = tflearn.DNN(network, tensorboard_verbose=3)
try:
model.fit(X, Y, n_epoch=1, validation_set=(testX, testY),
snapshot_epoch=False,
show_metric=True,
run_id=scriptid,callbacks=early_stopping_cb)
except StopIteration:
print("Caught callback exception. Returning control to user program.") prediction = model.predict(testX)
auc=roc_auc_score(testY, prediction, average='macro', sample_weight=None)
accuracy=model.evaluate(testX,testY) print("Accuracy:", accuracy)
print("ROC AUC Score:", auc)
Training Step: 860  | total loss: [1m[32m0.30941[0m[0m
| Adam | epoch: 001 | loss: 0.30941 - acc: 0.9125 -- iter: 55000/55000
Terminating training at the end of epoch 1
Successfully left training! Final model accuracy: 0.9125033020973206
Caught callback exception. Returning control to user program.
Accuracy: [0.90410000000000001]
ROC AUC Score: 0.992379719297

参考:http://mckinziebrandon.me/TensorflowNotebooks/2016/11/19/tflearn-only.html

TFLearn

19 Nov 2016

Examples::Extending Tensorflow::Trainer

import tensorflow as tf
import tflearn
import tflearn.datasets.mnist as mnist trainX, trainY, testX, testY = mnist.load_data(one_hot=True)
hdf5 not supported (please install/reinstall h5py)
Extracting mnist/train-images-idx3-ubyte.gz
Extracting mnist/train-labels-idx1-ubyte.gz
Extracting mnist/t10k-images-idx3-ubyte.gz
Extracting mnist/t10k-labels-idx1-ubyte.gz

Define the Architecture (Basic Tensorflow)

# Because I don't feel like retyping stuff.
def tfp(shape):
return tf.placeholder("float", shape)
def tfrn(shape, name):
return tf.Variable(tf.random_normal(shape), name=name) # Define the inputs/outputs/weights as usual.
X, Y = tfp([None, 784]), tfp([None, 10])
W1, W2, W3 = tfrn([784, 256], 'W1'), tfrn([256, 256], 'W2'), tfrn([256, 10], 'W3')
b1, b2, b3 = tfrn([256], 'b1'), tfrn([256], 'b2'), tfrn([10], 'b3') # Multilayer perceptron.
def dnn(x):
x = tf.tanh(tf.add(tf.matmul(x, W1), b1))
x = tf.tanh(tf.add(tf.matmul(x, W2), b2))
x = tf.add(tf.matmul(x, W3), b3)
return x
net = dnn(X) loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net, Y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
accuracy = tf.reduce_mean(tf.cast(
tf.equal( tf.argmax(net, 1), tf.argmax(Y, 1) ), tf.float32),
name='acc')

Using a TFLearn Trainer

trainop = tflearn.TrainOp(loss=loss, optimizer=optimizer, metric=accuracy, batch_size=128)
trainer = tflearn.Trainer(train_ops=trainop, tensorboard_verbose=1)
trainer.fit({X: trainX, Y: trainY}, val_feed_dicts={X: testX, Y: testY},
n_epoch=2, show_metric=True)
Training Step: 860  | total loss: [1m[32m1.73376[0m[0m
| Optimizer | epoch: 002 | loss: 1.73376 - acc: 0.8053 | val_loss: 1.78279 - val_acc: 0.8015 -- iter: 55000/55000
Training Step: 860 | total loss: [1m[32m1.73376[0m[0m
| Optimizer | epoch: 002 | loss: 1.73376 - acc: 0.8053 | val_loss: 1.78279 - val_acc: 0.8015 -- iter: 55000/55000
--

Training Callbacks

One suggestion for early stopping with tflearn (made by owner of tflearn repository) is to define a custom callback that raises an exception when we want to stop training. I’ve written a small snippet below as an example.

class EarlyStoppingCallback(tflearn.callbacks.Callback):
def __init__(self, acc_thresh):
"""
Args:
acc_thresh - if our accuracy > acc_thresh, terminate training.
"""
self.acc_thresh = acc_thresh
self.accs = [] def on_epoch_end(self, training_state):
""" """
self.accs.append(training_state.global_acc)
if training_state.val_acc is not None and training_state.val_acc < self.acc_thresh:
raise StopIteration
cb = EarlyStoppingCallback(acc_thresh=0.5)
trainer.fit({X: trainX, Y: trainY}, val_feed_dicts={X: testX, Y: testY},
n_epoch=3, show_metric=True, snapshot_epoch=False,
callbacks=cb)
Training Step: 3965  | total loss: [1m[32m0.33810[0m[0m
| Optimizer | epoch: 010 | loss: 0.33810 - acc: 0.9455 -- iter: 55000/55000
GOODBYE --------------------------------------------------------------------------- StopIteration Traceback (most recent call last) <ipython-input-24-9c383c6f5a8b> in <module>()
2 trainer.fit({X: trainX, Y: trainY}, val_feed_dicts={X: testX, Y: testY},
3 n_epoch=3, show_metric=True, snapshot_epoch=False,
----> 4 callbacks=cb) /usr/local/lib/python3.5/dist-packages/tflearn/helpers/trainer.py in fit(self, feed_dicts, n_epoch, val_feed_dicts, show_metric, snapshot_step, snapshot_epoch, shuffle_all, dprep_dict, daug_dict, excl_trainops, run_id, callbacks)
315
316 # Epoch end
--> 317 caller.on_epoch_end(self.training_state)
318
319 finally: /usr/local/lib/python3.5/dist-packages/tflearn/callbacks.py in on_epoch_end(self, training_state)
67 def on_epoch_end(self, training_state):
68 for callback in self.callbacks:
---> 69 callback.on_epoch_end(training_state)
70
71 def on_train_end(self, training_state): <ipython-input-23-d44cbdbc0814> in on_epoch_end(self, training_state)
13 if True:
14 print("GOODBYE")
---> 15 raise StopIteration StopIteration:
cb.accs
[None]

参考:

Early Stopping with TensorFlow and TFLearn

20 Nov 2016

import tensorflow as tf
import tflearn
import tflearn.datasets.mnist as mnist trainX, trainY, testX, testY = mnist.load_data(one_hot=True)
hdf5 not supported (please install/reinstall h5py)
Extracting mnist/train-images-idx3-ubyte.gz
Extracting mnist/train-labels-idx1-ubyte.gz
Extracting mnist/t10k-images-idx3-ubyte.gz
Extracting mnist/t10k-labels-idx1-ubyte.gz
n_features = 784
n_hidden = 256
n_classes = 10 # Define the inputs/outputs/weights as usual.
X = tf.placeholder("float", [None, n_features])
Y = tf.placeholder("float", [None, n_classes]) # Define the connections/weights and biases between layers.
W1 = tf.Variable(tf.random_normal([n_features, n_hidden]), name='W1')
W2 = tf.Variable(tf.random_normal([n_hidden, n_hidden]), name='W2')
W3 = tf.Variable(tf.random_normal([n_hidden, n_classes]), name='W3') b1 = tf.Variable(tf.random_normal([n_hidden]), name='b1')
b2 = tf.Variable(tf.random_normal([n_hidden]), name='b2')
b3 = tf.Variable(tf.random_normal([n_classes]), name='b3') # Define the operations throughout the network.
net = tf.tanh(tf.add(tf.matmul(X, W1), b1))
net = tf.tanh(tf.add(tf.matmul(net, W2), b2))
net = tf.add(tf.matmul(net, W3), b3) # Define the optimization problem.
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net, Y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
accuracy = tf.reduce_mean(tf.cast(
tf.equal(tf.argmax(net, 1), tf.argmax(Y, 1) ), tf.float32), name='acc')

Early Stopping

Training Setup

In tflearn, we can train our model with a tflearn.Trainer object: “Generic class to handle any TensorFlow graph training. It requires the use of TrainOp to specify all optimization parameters.”

  • TrainOp represents a set of operation used for optimizing a network.

  • Example: Time to initialize our trainer to work with our MNIST network. Below we create a TrainOp object that is then used for the purpose of telling our trainer

    1. Our loss function. (softmax cross entropy with logits)
    2. Our optimizer. (GradientDescentOptimizer)
    3. Our evaluation [tensor] metric. (classification accuracy)
trainop = tflearn.TrainOp(loss=loss, optimizer=optimizer, metric=accuracy, batch_size=128)
trainer = tflearn.Trainer(train_ops=trainop, tensorboard_verbose=1)

Callbacks

The Callbacks interface describes a set of methods that we can implement ourselves that will be called during runtime. Below are our options, where here we will be primarily concerned with the on_epoch_end() method. * __ Methods __ :

    def on_train_begin(self, training_state):
def on_epoch_begin(self, training_state):
def on_batch_begin(self, training_state):
def on_sub_batch_begin(self, training_state):
def on_sub_batch_end(self, training_state, train_index=0):
def on_batch_end(self, training_state, snapshot=False):
def on_epoch_end(self, training_state):
def on_train_end(self, training_state):
  • TrainingState: Notice that each method requires us to pass a training_state object as an argument. These useful helpers will be able to provide us with the information we need to determine when to stop training. Below is a list of the instance variables we can access with a training_state object:

    • self.epoch
    • self.step
    • self.current_iter
    • self.acc_value
    • self.loss_value
    • self.val_acc
    • self.val_loss
    • self.best_accuracy
    • self.global_acc
    • self.global_loss
  • Implementing our Callback: Let’s say we want to stop training when the validation accuracy reaches a certain threshold. Below, we implement the code required to define such a callback and fit the MNIST data.
class EarlyStoppingCallback(tflearn.callbacks.Callback):
def __init__(self, val_acc_thresh):
""" Note: We are free to define our init function however we please. """
self.val_acc_thresh = val_acc_thresh def on_epoch_end(self, training_state):
""" """
# Apparently this can happen.
if training_state.val_acc is None: return
if training_state.val_acc > self.val_acc_thresh:
raise StopIteration
# Initializae our callback.
early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.5)
# Give it to our trainer and let it fit the data.
trainer.fit(feed_dicts={X: trainX, Y: trainY},
val_feed_dicts={X: testX, Y: testY},
n_epoch=2,
show_metric=True, # Calculate accuracy and display at every step.
snapshot_epoch=False,
callbacks=early_stopping_cb)
Training Step: 1720  | total loss: [1m[32m0.81290[0m[0m
| Optimizer | epoch: 004 | loss: 0.81290 - acc_2: 0.8854 -- iter: 55000/55000

Using tf.contrib.learn instead

Iris data loading/tutorial prep

Note: can also load via: ```python import csv import random import numpy as np from sklearn import datasets from sklearn.cross_validation import train_test_split

X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.33, random_state=42) iris = datasets.load_iris() print(iris.data.shape) print(“Xt”, X_train.shape, “Yt”, y_train.shape) ```

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function # Suppress the massive amount of warnings.
tf.logging.set_verbosity(tf.logging.ERROR) # Data sets
IRIS_TRAINING = "iris_training.csv"
IRIS_TEST = "iris_test.csv" # Load datasets.
training_set = tf.contrib.learn.datasets.base.load_csv_with_header(filename=IRIS_TRAINING,
target_dtype=np.int,
features_dtype=np.float32)
test_set = tf.contrib.learn.datasets.base.load_csv_with_header(filename=IRIS_TEST,
target_dtype=np.int,
features_dtype=np.float32) # Specify that all features have real-value data
feature_columns = [tf.contrib.layers.real_valued_column("", dimension=4)] # Build 3 layer DNN with 10, 20, 10 units respectively.
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[10, 20, 10],
n_classes=3,
model_dir="/tmp/iris_model") # Fit model.
classifier.fit(x=X_train,
y=y_train,
steps=2000) # Evaluate accuracy.
accuracy_score = classifier.evaluate(x=X_test, y=y_test)["accuracy"]
print('Accuracy: {0:f}'.format(accuracy_score)) # Classify two new flower samples.
new_samples = np.array([[6.4, 3.2, 4.5, 1.5], [5.8, 3.1, 5.0, 1.7]], dtype=np.float32) y = classifier.predict(new_samples)
print('Predictions: {}'.format(str(y)))
Accuracy: 0.980000
Predictions: [1 1]

Validation Monitors

# Vanilla version
validation_monitor = tf.contrib.learn.monitors.ValidationMonitor(test_set.data,
test_set.target,
every_n_steps=50) classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[10, 20, 10],
n_classes=3,
model_dir="/tmp/iris_model",
config=tf.contrib.learn.RunConfig(
save_checkpoints_secs=1)) classifier.fit(x=training_set.data,
y=training_set.target,
steps=2000,
monitors=[validation_monitor])
Estimator(params={'dropout': None, 'hidden_units': [10, 20, 10], 'weight_column_name': None, 'feature_columns': [_RealValuedColumn(column_name='', dimension=4, default_value=None, dtype=tf.float32, normalizer=None)], 'optimizer': 'Adagrad', 'n_classes': 3, 'activation_fn': <function relu at 0x7f8568caa598>, 'num_ps_replicas': 0, 'gradient_clip_norm': None, 'enable_centered_bias': True})

Customizing the Evaluation Metrics and Stopping Early

If we run the code below, it stops early! Warning: You’re going to see a lot of WARNING print outputs from tf. I guess this tutorial is a bit out of date. But that’s not what we care abot here, we just want that early stopping! The important output to notice is

INFO:tensorflow:Validation (step 22556): accuracy = 0.966667, global_step = 22535, loss = 0.2767
INFO:tensorflow:Stopping. Best step: 22356 with loss = 0.2758353650569916.
validation_metrics = {"accuracy": tf.contrib.metrics.streaming_accuracy,
"precision": tf.contrib.metrics.streaming_precision,
"recall": tf.contrib.metrics.streaming_recall} validation_monitor = tf.contrib.learn.monitors.ValidationMonitor(
test_set.data,
test_set.target,
every_n_steps=50,
#metrics=validation_metrics,
early_stopping_metric='loss',
early_stopping_metric_minimize=True,
early_stopping_rounds=200) tf.logging.set_verbosity(tf.logging.ERROR)
classifier.fit(x=training_set.data,
y=training_set.target,
steps=2000,
monitors=[validation_monitor])
Estimator(params={'dropout': None, 'hidden_units': [10, 20, 10], 'weight_column_name': None, 'feature_columns': [_RealValuedColumn(column_name='', dimension=4, default_value=None, dtype=tf.float32, normalizer=None)], 'optimizer': 'Adagrad', 'n_classes': 3, 'activation_fn': <function relu at 0x7f8568caa598>, 'num_ps_replicas': 0, 'gradient_clip_norm': None, 'enable_centered_bias': True})

TFLearn 在给定模型精度时候提前终止训练的更多相关文章

  1. MindSpore模型精度调优实战:如何更快定位精度问题

    摘要:为大家梳理了针对常见精度问题的调试调优指南,将以"MindSpore模型精度调优实战"系列文章的形式分享出来,帮助大家轻松定位精度问题,快速优化模型精度. 本文分享自华为云社 ...

  2. MindSpore模型精度调优实战:常用的定位精度调试调优思路

    摘要:在模型的开发过程中,精度达不到预期常常让人头疼.为了帮助用户解决模型调试调优的问题,我们为MindSpore量身定做了可视化调试调优组件:MindInsight. 本文分享自华为云社区<技 ...

  3. ES5 forEach()用法和提前终止遍历

    forEach()方法从头到尾的遍历数组,为每个元素调用指定的函数,第一个参数接收的是一个函数,第二个参数是可选的,如果有第二个参数,则调用的函数被看作是第二个参数的方法(第二个参数可以作为第一个调用 ...

  4. forEach 如何提前终止 跳出运行

    forEach 如何提前终止 跳出运行 try{ arr.forEach(function(item,index){ if (...) { foreach.break=new Error(" ...

  5. MindSpore模型精度调优实践

    MindSpore模型精度调优实践 引论:在模型的开发过程中,精度达不到预期常常让人头疼.为了帮助用户解决模型调试调优的问题,为MindSpore量身定做了可视化调试调优组件:MindInsight. ...

  6. 从Word Embedding到Bert模型—自然语言处理中的预训练技术发展史(转载)

    转载 https://zhuanlan.zhihu.com/p/49271699 首发于深度学习前沿笔记 写文章   从Word Embedding到Bert模型—自然语言处理中的预训练技术发展史 张 ...

  7. zz从Word Embedding到Bert模型—自然语言处理中的预训练技术发展史

    从Word Embedding到Bert模型—自然语言处理中的预训练技术发展史 Bert最近很火,应该是最近最火爆的AI进展,网上的评价很高,那么Bert值得这么高的评价吗?我个人判断是值得.那为什么 ...

  8. TensorFlow从1到2(七)线性回归模型预测汽车油耗以及训练过程优化

    线性回归模型 "回归"这个词,既是Regression算法的名称,也代表了不同的计算结果.当然结果也是由算法决定的. 不同于前面讲过的多个分类算法或者逻辑回归,线性回归模型的结果是 ...

  9. pytorch入门2.2构建回归模型初体验(开始训练)

    pytorch入门2.x构建回归模型系列: pytorch入门2.0构建回归模型初体验(数据生成) pytorch入门2.1构建回归模型初体验(模型构建) pytorch入门2.2构建回归模型初体验( ...

随机推荐

  1. Python学习-range的用法

    range() 函数的用法 range(start,end,step):可以参见已连串的数字,常与for循环配合使用 参数详解如下 start:开始创建的起始位置,默认为0 end:开始创建的结束位置 ...

  2. Python 软件开发目录规范

    目录规范: ATM  #工程文件夹 ------| bin  #用来存放可执行文件的 |----  start.py conf  #用来存放配置信息的 |----  settings.py lib  ...

  3. tcpdump用于抓取tcp数据包

    一.简单使用:-c监听次数.-v打印详情.host后接监听地址 1.1.监听 tcpdump -c -v host www.baidu.com 1.2.访问被监听的网址: 1.3.查看监听的数据:

  4. PHP 反射API

    出处:http://blog.csdn.net/hguisu/article/details/7357421 PHP5添加了一项新的功能:Reflection.这个功能使得phper可以reverse ...

  5. ACdream 1063 字典树

    ACdream 1063 字典树 平衡树 神奇的cxlove有一颗平衡树,其树之神奇无法用语言来描述 OrzOrz. 这棵树支持3种操作: 1.加入一个数到树中,维护平衡树的合法性: 2.给一个数X, ...

  6. Mybatis_HelloWorld

    [梗概] 1.根据XML配置文件(全局配置文件,有数据源的一些运行信息)创建一个SqlSessionFactory对象. 2.sql映射文件:配置了每一个sql以及sql的封装规则. 3.将sql映射 ...

  7. 【spring boot 系列】spring data jpa 全面解析(实践 + 源码分析)

    前言 本文将从示例.原理.应用3个方面介绍spring data jpa. 以下分析基于spring boot 2.0 + spring 5.0.4版本源码 概述 JPA是什么? JPA (Java ...

  8. Linux下汇编语言学习笔记30 ---

    这是17年暑假学习Linux汇编语言的笔记记录,参考书目为清华大学出版社 Jeff Duntemann著 梁晓辉译<汇编语言基于Linux环境>的书,喜欢看原版书的同学可以看<Ass ...

  9. memcache适用和不适用场景[转载]

    适用memcached的业务场景:1)如果网站包含了访问量很大的动态网页,因而数据库的负载将会很高.由于大部分数据库请求都是读操作,那么memcached可以显著地减小数据库负载.2)如果数据库服务器 ...

  10. NIST的安全内容自动化协议(SCAP)以及SCAP中文社区简介

    https://blog.csdn.net/langkew/article/details/8795530?utm_source=tuicool&utm_medium=referral