add lookahead and radam, add use api
This commit is contained in:
myzrshjy
parent
4b95b2f77f
commit
90d204dba3
14
README.md
14
README.md
@ -110,6 +110,20 @@ step3: goto # Train&Usage(调用) and Predict&Usage(调用)
|
||||
* transformer模型: [https://github.com/CyberZHG/keras-transformer](https://github.com/CyberZHG/keras-transformer)
|
||||
* keras_albert_model: [https://github.com/TinkerMob/keras_albert_model](https://github.com/TinkerMob/keras_albert_model)
|
||||
|
||||
|
||||
# 训练简单调用:
|
||||
```python
|
||||
from keras_textclassification import train
|
||||
train(graph='TextCNN', # 必填, 算法名, 可选"ALBERT","BERT","XLNET","FASTTEXT","TEXTCNN","CHARCNN",
|
||||
# "TEXTRNN","RCNN","DCNN","DPCNN","VDCNN","CRNN","DEEPMOJI",
|
||||
# "SELFATTENTION", "HAN","CAPSULE","TRANSFORMER"
|
||||
label=17, # 必填, 类别数, 训练集和测试集合必须一样
|
||||
path_train_data=None, # 必填, 训练数据文件, csv格式, 必须含'label,ques'头文件, 详见keras_textclassification/data
|
||||
path_dev_data=None, # 必填, 测试数据文件, csv格式, 必须含'label,ques'头文件, 详见keras_textclassification/data
|
||||
rate=1, # 可填, 训练数据选取比例
|
||||
hyper_parameters=None) # 可填, json格式, 超参数, 默认embedding为'char','random'
|
||||
```
|
||||
|
||||
# Train&Usage(调用)
|
||||
|
||||
```python
|
||||
|
||||
@ -7,6 +7,8 @@
|
||||
|
||||
from keras_textclassification.data_preprocess.generator_preprocess import PreprocessGenerator
|
||||
from keras_textclassification.data_preprocess.text_preprocess import save_json
|
||||
from keras_textclassification.keras_layers.keras_lookahead import Lookahead
|
||||
from keras_textclassification.keras_layers.keras_radam import RAdam
|
||||
from keras.callbacks import ModelCheckpoint, EarlyStopping
|
||||
from keras.optimizers import Adam
|
||||
from keras import backend as K
|
||||
@ -48,6 +50,7 @@ class graph:
|
||||
self.path_hyper_parameters = hyper_parameters_model.get('path_hyper_parameters', "path_hyper_parameters") # 超参数保存地址
|
||||
self.path_fineture = hyper_parameters_model.get('path_fineture', "path_fineture") # embedding层保存地址, 例如静态词向量、动态词向量、微调bert层等
|
||||
self.patience = hyper_parameters_model.get('patience', 3) # 早停, 2-3就可以了
|
||||
self.optimizer_name = hyper_parameters_model.get('optimizer_name', 'RAdam,Lookahead') # 早停, 2-3就可以了
|
||||
if self.gpu_memory_fraction:
|
||||
# keras, tensorflow控制GPU使用率等
|
||||
import tensorflow as tf
|
||||
@ -103,9 +106,20 @@ class graph:
|
||||
构建优化器、损失函数和评价函数
|
||||
:return:
|
||||
"""
|
||||
self.model.compile(optimizer=Adam(lr=self.lr, beta_1=0.9, beta_2=0.999, decay=0.0),
|
||||
loss=self.loss,
|
||||
metrics=[self.metrics])
|
||||
if self.optimizer_name.upper() == "ADAM":
|
||||
self.model.compile(optimizer=Adam(lr=self.lr, beta_1=0.9, beta_2=0.999, decay=0.0),
|
||||
loss=self.loss,
|
||||
metrics=[self.metrics]) # Any optimize
|
||||
elif self.optimizer_name.upper() == "RADAM":
|
||||
self.model.compile(optimizer=RAdam(lr=self.lr, beta_1=0.9, beta_2=0.999, decay=0.0),
|
||||
loss=self.loss,
|
||||
metrics=[self.metrics]) # Any optimize
|
||||
else:
|
||||
self.model.compile(optimizer=RAdam(lr=self.lr, beta_1=0.9, beta_2=0.999, decay=0.0),
|
||||
loss=self.loss,
|
||||
metrics=[self.metrics]) # Any optimize
|
||||
lookahead = Lookahead(k=5, alpha=0.5) # Initialize Lookahead
|
||||
lookahead.inject(self.model) # add into model
|
||||
|
||||
def fit(self, x_train, y_train, x_dev, y_dev):
|
||||
"""
|
||||
|
||||
77
keras_textclassification/keras_layers/keras_lookahead.py
Normal file
77
keras_textclassification/keras_layers/keras_lookahead.py
Normal file
@ -0,0 +1,77 @@
|
||||
# !/usr/bin/python
|
||||
# -*- coding: utf-8 -*-
|
||||
# @time : 2019/11/12 16:14
|
||||
# @author : Mo
|
||||
# @function: lookahead of keras
|
||||
# @codefrom: https://github.com/bojone/keras_lookahead
|
||||
|
||||
|
||||
from keras import backend as K
|
||||
|
||||
|
||||
class Lookahead(object):
|
||||
"""Add the [Lookahead Optimizer](https://arxiv.org/abs/1907.08610) functionality for [keras](https://keras.io/).
|
||||
"""
|
||||
|
||||
def __init__(self, k=5, alpha=0.5):
|
||||
self.k = k
|
||||
self.alpha = alpha
|
||||
self.count = 0
|
||||
|
||||
def inject(self, model):
|
||||
"""Inject the Lookahead algorithm for the given model.
|
||||
The following code is modified from keras's _make_train_function method.
|
||||
See: https://github.com/keras-team/keras/blob/master/keras/engine/training.py#L497
|
||||
"""
|
||||
if not hasattr(model, 'train_function'):
|
||||
raise RuntimeError('You must compile your model before using it.')
|
||||
|
||||
model._check_trainable_weights_consistency()
|
||||
|
||||
if model.train_function is None:
|
||||
inputs = (model._feed_inputs +
|
||||
model._feed_targets +
|
||||
model._feed_sample_weights)
|
||||
if model._uses_dynamic_learning_phase():
|
||||
inputs += [K.learning_phase()]
|
||||
fast_params = model._collected_trainable_weights
|
||||
|
||||
with K.name_scope('training'):
|
||||
with K.name_scope(model.optimizer.__class__.__name__):
|
||||
training_updates = model.optimizer.get_updates(
|
||||
params=fast_params,
|
||||
loss=model.total_loss)
|
||||
slow_params = [K.variable(p) for p in fast_params]
|
||||
fast_updates = (model.updates +
|
||||
training_updates +
|
||||
model.metrics_updates)
|
||||
|
||||
slow_updates, copy_updates = [], []
|
||||
for p, q in zip(fast_params, slow_params):
|
||||
slow_updates.append(K.update(q, q + self.alpha * (p - q)))
|
||||
copy_updates.append(K.update(p, q))
|
||||
|
||||
# Gets loss and metrics. Updates weights at each call.
|
||||
fast_train_function = K.function(
|
||||
inputs,
|
||||
[model.total_loss] + model.metrics_tensors,
|
||||
updates=fast_updates,
|
||||
name='fast_train_function',
|
||||
**model._function_kwargs)
|
||||
|
||||
def F(inputs):
|
||||
self.count += 1
|
||||
R = fast_train_function(inputs)
|
||||
if self.count % self.k == 0:
|
||||
K.batch_get_value(slow_updates)
|
||||
K.batch_get_value(copy_updates)
|
||||
return R
|
||||
|
||||
model.train_function = F
|
||||
|
||||
if __name__ == '__main__':
|
||||
gg = 0
|
||||
# useage
|
||||
# model.compile(optimizer=Adam(1e-3), loss='mse') # Any optimizer
|
||||
# lookahead = Lookahead(k=5, alpha=0.5) # Initialize Lookahead
|
||||
# lookahead.inject(model) # add into model
|
||||
96
keras_textclassification/keras_layers/keras_radam.py
Normal file
96
keras_textclassification/keras_layers/keras_radam.py
Normal file
@ -0,0 +1,96 @@
|
||||
# !/usr/bin/python
|
||||
# -*- coding: utf-8 -*-
|
||||
# @time : 2019/11/12 16:12
|
||||
# @author : Mo
|
||||
# @function: radam of keras
|
||||
# @codefrom: https://github.com/bojone/keras_radam
|
||||
|
||||
|
||||
from keras.legacy import interfaces
|
||||
from keras.optimizers import Optimizer
|
||||
import keras.backend as K
|
||||
|
||||
|
||||
class RAdam(Optimizer):
|
||||
"""RAdam optimizer.
|
||||
Default parameters follow those provided in the original Adam paper.
|
||||
# Arguments
|
||||
lr: float >= 0. Learning rate.
|
||||
beta_1: float, 0 < beta < 1. Generally close to 1.
|
||||
beta_2: float, 0 < beta < 1. Generally close to 1.
|
||||
epsilon: float >= 0. Fuzz factor. If `None`, defaults to `K.epsilon()`.
|
||||
decay: float >= 0. Learning rate decay over each update.
|
||||
amsgrad: boolean. Whether to apply the AMSGrad variant of this
|
||||
algorithm from the paper "On the Convergence of Adam and
|
||||
Beyond".
|
||||
# References
|
||||
- [RAdam - A Method for Stochastic Optimization]
|
||||
(https://arxiv.org/abs/1908.03265)
|
||||
- [On The Variance Of The Adaptive Learning Rate And Beyond]
|
||||
(https://arxiv.org/abs/1908.03265)
|
||||
"""
|
||||
|
||||
def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
|
||||
epsilon=None, decay=0., **kwargs):
|
||||
super(RAdam, self).__init__(**kwargs)
|
||||
with K.name_scope(self.__class__.__name__):
|
||||
self.iterations = K.variable(0, dtype='int64', name='iterations')
|
||||
self.lr = K.variable(lr, name='lr')
|
||||
self.beta_1 = K.variable(beta_1, name='beta_1')
|
||||
self.beta_2 = K.variable(beta_2, name='beta_2')
|
||||
self.decay = K.variable(decay, name='decay')
|
||||
if epsilon is None:
|
||||
epsilon = K.epsilon()
|
||||
self.epsilon = epsilon
|
||||
self.initial_decay = decay
|
||||
|
||||
@interfaces.legacy_get_updates_support
|
||||
def get_updates(self, loss, params):
|
||||
grads = self.get_gradients(loss, params)
|
||||
self.updates = [K.update_add(self.iterations, 1)]
|
||||
|
||||
lr = self.lr
|
||||
if self.initial_decay > 0:
|
||||
lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
|
||||
K.dtype(self.decay))))
|
||||
|
||||
t = K.cast(self.iterations, K.floatx()) + 1
|
||||
beta_1_t = K.pow(self.beta_1, t)
|
||||
beta_2_t = K.pow(self.beta_2, t)
|
||||
rho = 2 / (1 - self.beta_2) - 1
|
||||
rho_t = rho - 2 * t * beta_2_t / (1 - beta_2_t)
|
||||
r_t = K.sqrt(
|
||||
K.relu(rho_t - 4) * K.relu(rho_t - 2) * rho / ((rho - 4) * (rho - 2) * rho_t)
|
||||
)
|
||||
flag = K.cast(rho_t > 4, K.floatx())
|
||||
|
||||
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
|
||||
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
|
||||
self.weights = [self.iterations] + ms + vs
|
||||
|
||||
for p, g, m, v in zip(params, grads, ms, vs):
|
||||
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
|
||||
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
|
||||
mhat_t = m_t / (1 - beta_1_t)
|
||||
vhat_t = K.sqrt(v_t / (1 - beta_2_t))
|
||||
p_t = p - lr * mhat_t * (flag * r_t / (vhat_t + self.epsilon) + (1 - flag))
|
||||
|
||||
self.updates.append(K.update(m, m_t))
|
||||
self.updates.append(K.update(v, v_t))
|
||||
new_p = p_t
|
||||
|
||||
# Apply constraints.
|
||||
if getattr(p, 'constraint', None) is not None:
|
||||
new_p = p.constraint(new_p)
|
||||
|
||||
self.updates.append(K.update(p, new_p))
|
||||
return self.updates
|
||||
|
||||
def get_config(self):
|
||||
config = {'lr': float(K.get_value(self.lr)),
|
||||
'beta_1': float(K.get_value(self.beta_1)),
|
||||
'beta_2': float(K.get_value(self.beta_2)),
|
||||
'decay': float(K.get_value(self.decay)),
|
||||
'epsilon': self.epsilon}
|
||||
base_config = super(RAdam, self).get_config()
|
||||
return dict(list(base_config.items()) + list(config.items()))
|
||||
@ -26,12 +26,12 @@ import time
|
||||
def train(hyper_parameters=None, rate=1.0):
|
||||
if not hyper_parameters:
|
||||
hyper_parameters = {
|
||||
'len_max': 2, # 句子最大长度, 固定推荐20-50, bert越长会越慢, 占用空间也会变大, 小心OOM
|
||||
'len_max': 32, # 句子最大长度, 固定推荐20-50, bert越长会越慢, 占用空间也会变大, 小心OOM
|
||||
'embed_size': 768, # 字/词向量维度, bert取768, word取300, char可以更小些
|
||||
'vocab_size': 20000, # 这里随便填的，会根据代码里修改
|
||||
'trainable': False, # embedding是静态的还是动态的, 即控制可不可以微调
|
||||
'level_type': 'char', # 级别, 最小单元, 字/词, 填 'char' or 'word', 注意:word2vec模式下训练语料要首先切好
|
||||
'embedding_type': 'bert', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
'embedding_type': 'random', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
'gpu_memory_fraction': 0.76, #gpu使用率
|
||||
'model': {'label': 17, # 类别数
|
||||
'batch_size': 2, # 批处理尺寸, 感觉原则上越大越好,尤其是样本不均衡的时候, batch_size设置影响比较大
|
||||
@ -45,6 +45,7 @@ def train(hyper_parameters=None, rate=1.0):
|
||||
'activate_classify': 'softmax', # 最后一个layer, 即分类激活函数
|
||||
'loss': 'categorical_crossentropy', # 损失函数
|
||||
'metrics': 'accuracy', # 保存更好模型的评价标准
|
||||
'optimizer_name': 'RAdam', # 优化器, 可选['Adam', 'Radam', 'RAdam,Lookahead']
|
||||
'is_training': True, # 训练后者是测试模型
|
||||
'model_path': path_model,
|
||||
# 模型地址, loss降低则保存的依据, save_best_only=True, save_weights_only=True
|
||||
|
||||
@ -51,10 +51,16 @@ def pred_tet(path_hyper_parameter=path_hyper_parameters, path_test=None, rate=1.
|
||||
for x_one in x:
|
||||
count += 1
|
||||
ques_embed = ra_ed.sentence2idx(x_one)
|
||||
if hyper_parameters['embedding_type'] == 'bert': # bert数据处理, token
|
||||
|
||||
if hyper_parameters['embedding_type'] in ['bert', 'albert']:
|
||||
x_val_1 = np.array([ques_embed[0]])
|
||||
x_val_2 = np.array([ques_embed[1]])
|
||||
x_val = [x_val_1, x_val_2]
|
||||
elif hyper_parameters['embedding_type'] == 'xlnet':
|
||||
x_val_1 = np.array([ques_embed[0]])
|
||||
x_val_2 = np.array([ques_embed[1]])
|
||||
x_val_3 = np.array([ques_embed[2]])
|
||||
x_val = [x_val_1, x_val_2, x_val_3]
|
||||
else:
|
||||
x_val = ques_embed
|
||||
# 预测
|
||||
|
||||
@ -26,16 +26,16 @@ import time
|
||||
def train(hyper_parameters=None, rate=1.0):
|
||||
if not hyper_parameters:
|
||||
hyper_parameters = {
|
||||
'len_max': 50, # 句子最大长度, 固定推荐20-50, bert越长会越慢, 占用空间也会变大, 本地win10-4G设为20就好, 过大小心OOM
|
||||
'embed_size': 768, # 字/词向量维度, bert取768, word取300, char可以更小些
|
||||
'len_max': 55, # 句子最大长度, 固定推荐20-50, bert越长会越慢, 占用空间也会变大, 本地win10-4G设为20就好, 过大小心OOM
|
||||
'embed_size': 100, # 字/词向量维度, bert取768, word取300, char可以更小些
|
||||
'vocab_size': 20000, # 这里随便填的，会根据代码里修改
|
||||
'trainable': True, # embedding是静态的还是动态的, 即控制可不可以微调
|
||||
'level_type': 'char', # 级别, 最小单元, 字/词, 填 'char' or 'word', 注意:word2vec模式下训练语料要首先切好
|
||||
'embedding_type': 'bert', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
'trainable': False, # embedding是静态的还是动态的, 即控制可不可以微调
|
||||
'level_type': 'word', # 级别, 最小单元, 字/词, 填 'char' or 'word', 注意:word2vec模式下训练语料要首先切好
|
||||
'embedding_type': 'word2vec', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
'gpu_memory_fraction': 0.66, #gpu使用率
|
||||
'model': {'label': 17, # 类别数
|
||||
'batch_size': 2, # 批处理尺寸, 感觉原则上越大越好,尤其是样本不均衡的时候, batch_size设置影响比较大
|
||||
'dropout': 0.32, # 随机失活, 概率
|
||||
'dropout': 0.5, # 随机失活, 概率
|
||||
'decay_step': 100, # 学习率衰减step, 每N个step衰减一次
|
||||
'decay_rate': 0.9, # 学习率衰减系数, 乘法
|
||||
'epochs': 20, # 训练最大轮次
|
||||
|
||||
@ -17,7 +17,7 @@ from keras import regularizers
|
||||
from keras_textclassification.base.graph import graph
|
||||
|
||||
|
||||
class DeepMoji(graph):
|
||||
class DeepMojiGraph(graph):
|
||||
def __init__(self, hyper_parameters):
|
||||
"""
|
||||
初始化
|
||||
|
||||
@ -20,7 +20,7 @@ from keras_textclassification.conf.path_config import path_baidu_qa_2019_train,
|
||||
# 数据预处理, 删除文件目录下文件
|
||||
from keras_textclassification.data_preprocess.text_preprocess import PreprocessText, delete_file
|
||||
# 模型图
|
||||
from keras_textclassification.m10_DeepMoji.graph import DeepMoji as Graph
|
||||
from keras_textclassification.m10_DeepMoji.graph import DeepMojiGraph as Graph
|
||||
# 计算时间
|
||||
import time
|
||||
|
||||
|
||||
@ -73,7 +73,14 @@ def pred_tet(path_hyper_parameter=path_hyper_parameters, path_test=None, rate=1.
|
||||
# 评估
|
||||
report_predict = classification_report(index_y, index_pred,
|
||||
target_names=target_names, digits=9)
|
||||
|
||||
count = 0
|
||||
for i in range(len(index_pred)):
|
||||
if index_y[i]==index_pred[i]:
|
||||
count += 1
|
||||
print(report_predict)
|
||||
print('accuracy:{}'.format(count/len(index_y)))
|
||||
|
||||
print("耗时:" + str(time.time() - time_start))
|
||||
|
||||
|
||||
@ -118,29 +125,15 @@ def pred_input(path_hyper_parameter=path_hyper_parameters):
|
||||
|
||||
if __name__=="__main__":
|
||||
# 测试集预测
|
||||
pred_tet(path_test=path_baidu_qa_2019_valid, rate=1) # sample条件下设为1,否则训练语料可能会很少
|
||||
pred_tet(rate=1) # sample条件下设为1,否则训练语料可能会很少
|
||||
|
||||
# 可输入 input 预测
|
||||
pred_input()
|
||||
|
||||
# 92218/92218 [==============================] - 138s 2ms/step - loss: 1.0619 - acc: 0.7534 - val_loss: 3.3621 - val_acc: 0.3412
|
||||
# Epoch 00020: val_loss improved from 3.36474 to 3.36213, saving model to D:/workspace/pythonMyCode/django_project/Keras-TextClassification/keras_textclassification/data/model/fast_text/model_fast_text.h5
|
||||
# 耗时:2830.3001523017883
|
||||
|
||||
# precision recall f1-score support
|
||||
#
|
||||
# 电子 0.428571429 0.375000000 0.400000000 8
|
||||
# 社会 0.250000000 0.166666667 0.200000000 12
|
||||
# 烦恼 0.500000000 0.800000000 0.615384615 20
|
||||
# 电脑 0.700000000 0.549019608 0.615384615 51
|
||||
# 汽车 0.333333333 0.600000000 0.428571429 5
|
||||
# 商业 0.781250000 0.714285714 0.746268657 35
|
||||
# 文化 0.000000000 0.000000000 0.000000000 7
|
||||
# 健康 0.727272727 0.655737705 0.689655172 61
|
||||
# 育儿 0.142857143 0.200000000 0.166666667 5
|
||||
# 教育 0.638297872 0.517241379 0.571428571 58
|
||||
# 娱乐 0.390243902 0.400000000 0.395061728 40
|
||||
# 生活 0.487179487 0.387755102 0.431818182 49
|
||||
# 体育 0.666666667 0.400000000 0.500000000 5
|
||||
# 游戏 0.658333333 0.858695652 0.745283019 92
|
||||
#
|
||||
# accuracy 0.589285714 448
|
||||
# macro avg 0.478857564 0.473171559 0.464680190 448
|
||||
# weighted avg 0.595133343 0.589285714 0.583989557 448
|
||||
# 92218/92218 [==============================] - 161s 2ms/step - loss: 0.0449 - acc: 0.9887 - val_loss: 0.7595 - val_acc: 0.8114
|
||||
# Epoch 00018: val_loss improved from 0.76525 to 0.75950, saving model to D:/workspace/pythonMyCode/django_project/Keras-TextClassification/keras_textclassification/data/model/fast_text/model_fast_text.h5
|
||||
# Epoch 19/20
|
||||
|
||||
@ -31,19 +31,19 @@ def train(hyper_parameters=None, rate=1.0):
|
||||
"""
|
||||
if not hyper_parameters:
|
||||
hyper_parameters = {
|
||||
'len_max': 50, # 句子最大长度, 固定 推荐20-50
|
||||
'len_max': 55, # 句子最大长度, 固定 推荐20-50
|
||||
'embed_size': 300, # 字/词向量维度
|
||||
'vocab_size': 20000, # 这里随便填的，会根据代码里修改
|
||||
'trainable': True, # embedding是静态的还是动态的
|
||||
'level_type': 'char', # 级别, 最小单元, 字/词, 填 'char' or 'word'
|
||||
'embedding_type': 'random', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
'gpu_memory_fraction': 0.66, # gpu使用率
|
||||
'model': {'label': 17, # 类别数
|
||||
'batch_size': 64, # 批处理尺寸, 感觉原则上越大越好,尤其是样本不均衡的时候, batch_size设置影响比较大
|
||||
'trainable': False, # embedding是静态的还是动态的
|
||||
'level_type': 'word', # 级别, 最小单元, 字/词, 填 'char' or 'word'
|
||||
'embedding_type': 'word2vec', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
'gpu_memory_fraction': 0.76, # gpu使用率
|
||||
'model': {'label': 90, # 类别数
|
||||
'batch_size': 256, # 批处理尺寸, 感觉原则上越大越好,尤其是样本不均衡的时候, batch_size设置影响比较大
|
||||
'dropout': 0.5, # 随机失活, 概率
|
||||
'decay_step': 100, # 学习率衰减step, 每N个step衰减一次
|
||||
'decay_rate': 0.9, # 学习率衰减系数, 乘法
|
||||
'epochs': 50, # 训练最大轮次
|
||||
'epochs': 20, # 训练最大轮次
|
||||
'patience': 5, # 早停,2-3就好
|
||||
'lr': 5e-4, # 学习率, 对训练会有比较大的影响, 如果准确率一直上不去,可以考虑调这个参数
|
||||
'l2': 1e-9, # l2正则化
|
||||
@ -90,10 +90,4 @@ def train(hyper_parameters=None, rate=1.0):
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
train(rate=1) # sample条件下设为1,否则训练语料可能会很少
|
||||
# 注意: 4G的1050Ti的GPU、win10下batch_size=32,len_max=20, gpu<=0.87, 应该就可以bert-fineture了。
|
||||
# 全量数据训练一轮(batch_size=32),就能达到80%准确率(验证集), 效果还是不错的
|
||||
# win10下出现过错误,gpu、len_max、batch_size配小一点就好:ailed to allocate 3.56G (3822520832 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory
|
||||
# 参数较多,不适合用bert,会比较慢和OOM
|
||||
|
||||
# 速度很慢呐，字/词维度(embed_size)一设大就OOM了，不喜欢用
|
||||
train(rate=1)
|
||||
|
||||
@ -76,7 +76,7 @@ def train(graph='TextCNN', label=17, rate=1.0, hyper_parameters=None, path_train
|
||||
hyper_parameters_real = {
|
||||
'len_max': 50, # 句子最大长度, 固定 推荐20-50
|
||||
'trainable': True, # embedding是静态的还是动态的
|
||||
'embed_size': 768, # 字/词向量维度
|
||||
'embed_size': 64, # 字/词向量维度
|
||||
'vocab_size': 20000, # 这里随便填的，会根据代码里修改
|
||||
'level_type': 'char', # 级别, 最小单元, 字/词, 填 'char' or 'word'
|
||||
'embedding_type': 'random', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
||||
@ -160,4 +160,4 @@ def train(graph='TextCNN', label=17, rate=1.0, hyper_parameters=None, path_train
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
train(graph='TextCNN', label=17, rate=1, hyper_parameters=None, path_train_data=None, path_dev_data=None)
|
||||
train(graph='TextCNN', label=17, rate=1, path_train_data=None, path_dev_data=None,hyper_parameters=None)
|
||||
|
||||
2
setup.py
2
setup.py
@ -25,7 +25,7 @@ with codecs.open('requirements.txt', 'r', 'utf8') as reader:
|
||||
install_requires = list(map(lambda x: x.strip(), reader.readlines()))
|
||||
|
||||
setup(name=NAME,
|
||||
version='0.1.4',
|
||||
version='0.1.5',
|
||||
description=DESCRIPTION,
|
||||
long_description=long_description,
|
||||
long_description_content_type="text/markdown",
|
||||
|
||||
@ -73,6 +73,11 @@ def pred_tet(path_hyper_parameter=path_hyper_parameters, path_test=None, rate=1.
|
||||
# 评估
|
||||
report_predict = classification_report(index_y, index_pred,
|
||||
target_names=target_names, digits=9)
|
||||
count = 0
|
||||
for i in range(len(index_pred)):
|
||||
if index_y[i] == index_pred[i]:
|
||||
count += 1
|
||||
print(count)
|
||||
print(report_predict)
|
||||
print("耗时:" + str(time.time() - time_start))
|
||||
|
||||
@ -123,4 +128,7 @@ if __name__=="__main__":
|
||||
# 可输入 input 预测
|
||||
pred_input()
|
||||
|
||||
|
||||
# 180/180 [==============================] - 56s 313ms/step - loss: 0.0696 - acc: 0.9784 - val_loss: 0.8469 - val_acc: 0.8155
|
||||
#
|
||||
# Epoch 00016: val_loss improved from 0.85291 to 0.84695, saving model to D:/workspace/pythonMyCode/django_project/Keras-TextClassification/keras_textclassification/data/model/fast_text/model_fast_text.h5
|
||||
# Epoch 17/20
|
||||
|
||||
16
test/tet_train.py
Normal file
16
test/tet_train.py
Normal file
@ -0,0 +1,16 @@
|
||||
# !/usr/bin/python
|
||||
# -*- coding: utf-8 -*-
|
||||
# @time : 2019/11/12 16:45
|
||||
# @author : Mo
|
||||
# @function:
|
||||
|
||||
|
||||
from keras_textclassification import train
|
||||
train(graph='TextCNN', # 必填, 算法名, 可选"ALBERT","BERT","XLNET","FASTTEXT","TEXTCNN","CHARCNN",
|
||||
# "TEXTRNN","RCNN","DCNN","DPCNN","VDCNN","CRNN","DEEPMOJI",
|
||||
# "SELFATTENTION", "HAN","CAPSULE","TRANSFORMER"
|
||||
label=17, # 必填, 类别数, 训练集和测试集合必须一样
|
||||
path_train_data=None, # 必填, 训练数据文件, csv格式, 必须含'label,ques'头文件, 详见keras_textclassification/data
|
||||
path_dev_data=None, # 必填, 测试数据文件, csv格式, 必须含'label,ques'头文件, 详见keras_textclassification/data
|
||||
rate=1, # 可填, 训练数据选取比例
|
||||
hyper_parameters=None) # 可填, json格式, 超参数, 默认embedding为'char','random'
|
||||
Loading…
Reference in New Issue
Block a user