关于机器学习:用kerasbert实现多输出参数共享模型

62次阅读

共计 17511 个字符，预计需要花费 44 分钟才能阅读完成。

背景

在 nlp 畛域，预训练模型 bert 堪称是红得发紫。

但当初能搜到的大多数都是 pytorch 写的框架，而且大多都是单输入模型。

所以，本文以 有互相关系的多层标签分类 为背景，用 keras 设计了多输入、参数共享的模型。

keras_bert 根底利用

def batch_iter(data_path, cat_to_id, tokenizer, batch_size=64, shuffle=True):
    """生成批次数据"""
    keras_bert_iter = get_keras_bert_iterator(data_path, cat_to_id, tokenizer)
    while True:
        data_list = []
        for _ in range(batch_size):
            data = next(keras_bert_iter)
            data_list.append(data)
        if shuffle:
            random.shuffle(data_list)
        
        indices_list = []
        segments_list = []
        label_index_list = []
        for data in data_list:
            indices, segments, label_index = data
            indices_list.append(indices)
            segments_list.append(segments)
            label_index_list.append(label_index)
 
        yield [np.array(indices_list), np.array(segments_list)], np.array(label_index_list)
 
def get_model(label_list):
    K.clear_session()
    
    bert_model = load_trained_model_from_checkpoint(bert_paths.config, bert_paths.checkpoint, seq_len=text_max_length)  #加载预训练模型
 
    for l in bert_model.layers:
        l.trainable = True
 
    input_indices = Input(shape=(None,))
    input_segments = Input(shape=(None,))
 
    bert_output = bert_model([input_indices, input_segments])
    bert_cls = Lambda(lambda x: x[:, 0])(bert_output) # 取出 [CLS] 对应的向量用来做分类
    output = Dense(len(label_list), activation='softmax')(bert_cls)
 
    model = Model([input_indices, input_segments], output)
    model.compile(loss='sparse_categorical_crossentropy',
                  optimizer=Adam(1e-5),    #用足够小的学习率
                  metrics=['accuracy'])
    print(model.summary())
    return model
 
early_stopping = EarlyStopping(monitor='val_acc', patience=3)   #早停法，避免过拟合
plateau = ReduceLROnPlateau(monitor="val_acc", verbose=1, mode='max', factor=0.5, patience=2) #当评估指标不在晋升时，缩小学习率
checkpoint = ModelCheckpoint('trained_model/keras_bert_THUCNews.hdf5', monitor='val_acc',verbose=2, save_best_only=True, mode='max', save_weights_only=True) #保留最好的模型
 
def get_step(sample_count, batch_size):
    step = sample_count // batch_size
    if sample_count % batch_size != 0:
        step += 1
    return step
 
batch_size = 4
train_step = get_step(train_sample_count, batch_size)
dev_step = get_step(dev_sample_count, batch_size)
 
train_dataset_iterator = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, batch_size)
dev_dataset_iterator = batch_iter(r"data/keras_bert_dev.txt", cat_to_id, tokenizer, batch_size)
 
model = get_model(categories)
 
#模型训练
model.fit(
    train_dataset_iterator,
    steps_per_epoch=train_step,
    epochs=10,
    validation_data=dev_dataset_iterator,
    validation_steps=dev_step,
    callbacks=[early_stopping, plateau, checkpoint],
    verbose=1
)

多输入、参数共享的模型设计

def batch_iter(data_path, cat_to_id, tokenizer, second_label_list, batch_size=64, shuffle=True):
    """生成批次数据"""
    keras_bert_iter = get_keras_bert_iterator(data_path, cat_to_id, tokenizer, second_label_list)
    while True:
        data_list = []
        for _ in range(batch_size):
            data = next(keras_bert_iter)
            data_list.append(data)
        if shuffle:
            random.shuffle(data_list)
        
        indices_list = []
        segments_list = []
        label_index_list = []
        second_label_list = []
        for data in data_list:
            indices, segments, label_index, second_label = data
            indices_list.append(indices)
            segments_list.append(segments)
            label_index_list.append(label_index)
            second_label_list.append(second_label)
 
        yield [np.array(indices_list), np.array(segments_list)], [np.array(label_index_list), np.array(second_label_list)]
 
def get_model(label_list, second_label_list):
    K.clear_session()
    
    bert_model = load_trained_model_from_checkpoint(bert_paths.config, bert_paths.checkpoint, seq_len=text_max_length)  #加载预训练模型
 
    for l in bert_model.layers:
        l.trainable = True
 
    input_indices = Input(shape=(None,))
    input_segments = Input(shape=(None,))
 
    bert_output = bert_model([input_indices, input_segments])
    bert_cls = Lambda(lambda x: x[:, 0])(bert_output) # 取出 [CLS] 对应的向量用来做分类
    output = Dense(len(label_list), activation='softmax')(bert_cls)
    output_second = Dense(len(second_label_list), activation='softmax')(bert_cls)
 
    model = Model([input_indices, input_segments], [output, output_second])
    model.compile(loss='sparse_categorical_crossentropy',
                  optimizer=Adam(1e-5),    #用足够小的学习率
                  metrics=['accuracy'])
    print(model.summary())
    return model
 
batch_size = 4
train_step = get_step(train_sample_count, batch_size)
dev_step = get_step(dev_sample_count, batch_size)
 
train_dataset_iterator = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, second_label_list, batch_size)
dev_dataset_iterator = batch_iter(r"data/keras_bert_dev.txt", cat_to_id, tokenizer, second_label_list, batch_size)
 
model = get_model(categories, second_label_list)
 
#模型训练
model.fit(
    train_dataset_iterator,
    steps_per_epoch=train_step,
    epochs=10,
    validation_data=dev_dataset_iterator,
    validation_steps=dev_step,
    callbacks=[early_stopping, plateau, checkpoint],
    verbose=1
)

附录

全副源码

import os
import sys
import re
from collections import Counter
import random
 
from tqdm import tqdm
import numpy as np
import tensorflow.keras as keras
from keras_bert import load_vocabulary, load_trained_model_from_checkpoint, Tokenizer, get_checkpoint_paths
from keras_bert.layers import MaskedGlobalMaxPool1D
from keras_bert import load_trained_model_from_checkpoint, Tokenizer
from keras.metrics import top_k_categorical_accuracy
from keras.layers import *
from keras.callbacks import *
from keras.models import Model
import keras.backend as K
from keras.optimizers import Adam
from keras.utils import to_categorical
data_path = "000_text_classifier_tensorflow_textcnn/THUCNews/"
text_max_length = 512
bert_paths = get_checkpoint_paths(r"chinese_L-12_H-768_A-12")

构建原数据文本迭代器

def _read_file(filename):
    """读取一个文件并转换为一行"""
    with open(filename, 'r', encoding='utf-8') as f:
        s = f.read().strip().replace('\n', '。').replace('\t', '').replace('\u3000','')
        return re.sub(r'。+', '。', s)
def get_data_iterator(data_path):
    for category in os.listdir(data_path):
        category_path = os.path.join(data_path, category)
        for file_name in os.listdir(category_path):
            yield _read_file(os.path.join(category_path, file_name)), category
it = get_data_iterator(data_path)
next(it)
('竞彩解析：日本美国争冠死磕 两巴相逢必有生死。周日受注赛事，女足世界杯决赛、美洲杯两场 1 / 4 决赛毫无疑问是全世界球迷和彩民关注的焦点。本届女足世界杯的最大黑马日本队是否一黑到底，发明亚洲奇观？男子足坛霸主美国队是否再次“灭黑”胜利，成就三冠伟业？巴西、巴拉圭狭路相逢，谁又能笑到最初？诸多谜底，在周一凌晨就会揭晓。日本美国争冠死磕。本届女足世界杯，是颠覆与反颠覆之争。夺冠大热门东道主德国队 1 / 4 决赛被日本队加时赛一球而“黑”，另一个夺冠大热门瑞典队则在半决赛被日本队 3:1 彻底打垮。而美国队则保卫着女足豪强的尊严，在 1 / 4 决赛，她们与巴西女足苦战至点球大战，最终以 5:3 淘汰这支迅速崛起的黑马球队，而在半决赛，她们更是 3:1 大胜欧洲黑马法国队。美日两队此次世界杯过程惊人类似，小组赛前两轮全胜，最初一轮输球，1/ 4 决赛同样与对手 90 分钟内战成平局，半决赛竟同样 3:1 大胜对手。此次决战，无论是日本还是美国队夺冠，均将发明女足世界杯新的历史。两巴相逢必有生死。本届美洲杯，让人大跌眼镜的事件太多。巴西、巴拉圭狭路相逢仿佛更具传奇色彩。两队小组赛同分在 B 组，本来两个出线大热门，却双双在前两轮小组赛战平，两队间接交锋就是 2:2 平局，后果双双面临出局危险。最初一轮，巴西队在下半场终于发威，4:2 大胜厄瓜多尔青出于蓝以小组第一出线，而巴拉圭最初一战还是 3:3 战平委内瑞拉取得小组第三，幸运凭借净胜球劣势挤掉 A 组第三名的哥斯达黎加，取得一个八强席位。在小组赛，巴西队是在最初时刻才逼平了巴拉圭，他们的好运气会在淘汰赛再显神威吗？巴拉圭此前 3 轮小组赛仿佛都不足运气，此番又会否被幸运之神弥补一下呢？。另一场美洲杯 1 / 4 决赛，智利队在 C 组小组赛 2 胜 1 平以小组头名升级八强；而委内瑞拉在 B 组是最不被看好的球队，但居然在与巴西、巴拉圭同组的状况下，前两轮就奠定了小组出线权，他们小组 3 战 1 胜 2 平放弃不败战绩，而入球数跟智利一样都是 4 球，只是失球数比智利多了 1 个。但既然他们面对弱小的巴西都能放弃球门不失，此番再创佳绩也不足为怪。',
 '彩票')

构建标签表

def read_category(data_path):
    """读取分类目录，固定"""
    categories = os.listdir(data_path)
 
    cat_to_id = dict(zip(categories, range(len(categories))))
 
    return categories, cat_to_id
categories, cat_to_id = read_category(data_path)
cat_to_id
{'彩票': 0,
 '家居': 1,
 '游戏': 2,
 '股票': 3,
 '科技': 4,
 '社会': 5,
 '财经': 6,
 '时尚': 7,
 '星座': 8,
 '体育': 9,
 '房产': 10,
 '娱乐': 11,
 '时政': 12,
 '教育': 13}
 
 
 
categories
['彩票',
 '家居',
 '游戏',
 '股票',
 '科技',
 '社会',
 '财经',
 '时尚',
 '星座',
 '体育',
 '房产',
 '娱乐',
 '时政',
 '教育']

构建训练、验证、测试集

def build_dataset(data_path, train_path, dev_path, test_path):
    data_iter = get_data_iterator(data_path)
    with open(train_path, 'w', encoding='utf-8') as train_file, \
         open(dev_path, 'w', encoding='utf-8') as dev_file, \
         open(test_path, 'w', encoding='utf-8') as test_file:
        
        for text, label in tqdm(data_iter):
            radio = random.random()
            if radio < 0.8:
                train_file.write(text + "\t" + label + "\n")
            elif radio < 0.9:
                dev_file.write(text + "\t" + label + "\n")
            else:
                test_file.write(text + "\t" + label + "\n")
# build_dataset(data_path, r"data/keras_bert_train.txt", r"data/keras_bert_dev.txt", r"data/keras_bert_test.txt")

获取数据集样本个数

def get_sample_num(data_path):
    count = 0
    with open(data_path, 'r', encoding='utf-8') as data_file:
        for line in tqdm(data_file):
            count += 1
    return count
train_sample_count = get_sample_num(r"data/keras_bert_train.txt")
668858it [00:09, 67648.27it/s]
 
 
dev_sample_count = get_sample_num(r"data/keras_bert_dev.txt")
83721it [00:01, 61733.96it/s]
 
 
test_sample_count = get_sample_num(r"data/keras_bert_test.txt")
83496it [00:01, 72322.53it/s]
 
 
train_sample_count, dev_sample_count, test_sample_count
(668858, 83721, 83496)

构建数据迭代器

def get_text_iterator(data_path):
    with open(data_path, 'r', encoding='utf-8') as data_file:
        for line in data_file:
            data_split = line.strip().split('\t')
            if len(data_split) != 2:
                print(line)
                continue
            yield data_split[0], data_split[1]
it = get_text_iterator(r"data/keras_bert_train.txt")
next(it)
('竞彩解析：日本美国争冠死磕 两巴相逢必有生死。周日受注赛事，女足世界杯决赛、美洲杯两场 1 / 4 决赛毫无疑问是全世界球迷和彩民关注的焦点。本届女足世界杯的最大黑马日本队是否一黑到底，发明亚洲奇观？男子足坛霸主美国队是否再次“灭黑”胜利，成就三冠伟业？巴西、巴拉圭狭路相逢，谁又能笑到最初？诸多谜底，在周一凌晨就会揭晓。日本美国争冠死磕。本届女足世界杯，是颠覆与反颠覆之争。夺冠大热门东道主德国队 1 / 4 决赛被日本队加时赛一球而“黑”，另一个夺冠大热门瑞典队则在半决赛被日本队 3:1 彻底打垮。而美国队则保卫着女足豪强的尊严，在 1 / 4 决赛，她们与巴西女足苦战至点球大战，最终以 5:3 淘汰这支迅速崛起的黑马球队，而在半决赛，她们更是 3:1 大胜欧洲黑马法国队。美日两队此次世界杯过程惊人类似，小组赛前两轮全胜，最初一轮输球，1/ 4 决赛同样与对手 90 分钟内战成平局，半决赛竟同样 3:1 大胜对手。此次决战，无论是日本还是美国队夺冠，均将发明女足世界杯新的历史。两巴相逢必有生死。本届美洲杯，让人大跌眼镜的事件太多。巴西、巴拉圭狭路相逢仿佛更具传奇色彩。两队小组赛同分在 B 组，本来两个出线大热门，却双双在前两轮小组赛战平，两队间接交锋就是 2:2 平局，后果双双面临出局危险。最初一轮，巴西队在下半场终于发威，4:2 大胜厄瓜多尔青出于蓝以小组第一出线，而巴拉圭最初一战还是 3:3 战平委内瑞拉取得小组第三，幸运凭借净胜球劣势挤掉 A 组第三名的哥斯达黎加，取得一个八强席位。在小组赛，巴西队是在最初时刻才逼平了巴拉圭，他们的好运气会在淘汰赛再显神威吗？巴拉圭此前 3 轮小组赛仿佛都不足运气，此番又会否被幸运之神弥补一下呢？。另一场美洲杯 1 / 4 决赛，智利队在 C 组小组赛 2 胜 1 平以小组头名升级八强；而委内瑞拉在 B 组是最不被看好的球队，但居然在与巴西、巴拉圭同组的状况下，前两轮就奠定了小组出线权，他们小组 3 战 1 胜 2 平放弃不败战绩，而入球数跟智利一样都是 4 球，只是失球数比智利多了 1 个。但既然他们面对弱小的巴西都能放弃球门不失，此番再创佳绩也不足为怪。',
 '彩票')
 
 
 
token_dict = load_vocabulary(bert_paths.vocab)
tokenizer = Tokenizer(token_dict)
def get_keras_bert_iterator(data_path, cat_to_id, tokenizer):
    while True:
        data_iter = get_text_iterator(data_path)
        for text, category in data_iter:
            indices, segments = tokenizer.encode(first=text, max_len=text_max_length)
            yield indices, segments, cat_to_id[category]
it = get_keras_bert_iterator(r"data/keras_bert_train.txt", cat_to_id, tokenizer)
# next(it)

构建批次数据迭代器

def batch_iter(data_path, cat_to_id, tokenizer, batch_size=64, shuffle=True):
    """生成批次数据"""
    keras_bert_iter = get_keras_bert_iterator(data_path, cat_to_id, tokenizer)
    while True:
        data_list = []
        for _ in range(batch_size):
            data = next(keras_bert_iter)
            data_list.append(data)
        if shuffle:
            random.shuffle(data_list)
        
        indices_list = []
        segments_list = []
        label_index_list = []
        for data in data_list:
            indices, segments, label_index = data
            indices_list.append(indices)
            segments_list.append(segments)
            label_index_list.append(label_index)
 
        yield [np.array(indices_list), np.array(segments_list)], np.array(label_index_list)
it = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, batch_size=1)
# next(it)
it = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, batch_size=2)
next(it)
([array([[ 101, 4993, 2506, ...,  131,  123,  102],
         [101, 2506, 3696, ..., 1139,  125,  102]]),
  array([[0, 0, 0, ..., 0, 0, 0],
         [0, 0, 0, ..., 0, 0, 0]])],
 array([0, 0]))

定义 base 模型

def get_model(label_list):
    K.clear_session()
    
    bert_model = load_trained_model_from_checkpoint(bert_paths.config, bert_paths.checkpoint, seq_len=text_max_length)  #加载预训练模型
 
    for l in bert_model.layers:
        l.trainable = True
 
    input_indices = Input(shape=(None,))
    input_segments = Input(shape=(None,))
 
    bert_output = bert_model([input_indices, input_segments])
    bert_cls = Lambda(lambda x: x[:, 0])(bert_output) # 取出 [CLS] 对应的向量用来做分类
    output = Dense(len(label_list), activation='softmax')(bert_cls)
 
    model = Model([input_indices, input_segments], output)
    model.compile(loss='sparse_categorical_crossentropy',
                  optimizer=Adam(1e-5),    #用足够小的学习率
                  metrics=['accuracy'])
    print(model.summary())
    return model
early_stopping = EarlyStopping(monitor='val_acc', patience=3)   #早停法，避免过拟合
plateau = ReduceLROnPlateau(monitor="val_acc", verbose=1, mode='max', factor=0.5, patience=2) #当评估指标不在晋升时，缩小学习率
checkpoint = ModelCheckpoint('trained_model/keras_bert_THUCNews.hdf5', monitor='val_acc',verbose=2, save_best_only=True, mode='max', save_weights_only=True) #保留最好的模型

模型训练

def get_step(sample_count, batch_size):
    step = sample_count // batch_size
    if sample_count % batch_size != 0:
        step += 1
    return step
batch_size = 4
train_step = get_step(train_sample_count, batch_size)
dev_step = get_step(dev_sample_count, batch_size)
 
train_dataset_iterator = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, batch_size)
dev_dataset_iterator = batch_iter(r"data/keras_bert_dev.txt", cat_to_id, tokenizer, batch_size)
 
model = get_model(categories)
 
#模型训练
model.fit(
    train_dataset_iterator,
    steps_per_epoch=train_step,
    epochs=10,
    validation_data=dev_dataset_iterator,
    validation_steps=dev_step,
    callbacks=[early_stopping, plateau, checkpoint],
    verbose=1
)
Model: "functional_5"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 512)]        0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            [(None, 512)]        0                                            
__________________________________________________________________________________________________
functional_3 (Functional)       (None, 512, 768)     101677056   input_1[0][0]                    
                                                                 input_2[0][0]                    
__________________________________________________________________________________________________
lambda (Lambda)                 (None, 768)          0           functional_3[0][0]               
__________________________________________________________________________________________________
dense (Dense)                   (None, 14)           10766       lambda[0][0]                     
==================================================================================================
Total params: 101,687,822
Trainable params: 101,687,822
Non-trainable params: 0
__________________________________________________________________________________________________
None
Epoch 1/10
     5/167215 [..............................] - ETA: 775:02:36 - loss: 0.4064 - accuracy: 0.9000
 
 
---------------------------------------------------------------------------

多输入模型

构建数据迭代器

second_label_list = [0, 1, 2]
def get_keras_bert_iterator(data_path, cat_to_id, tokenizer, second_label_list):
    while True:
        data_iter = get_text_iterator(data_path)
        for text, category in data_iter:
            indices, segments = tokenizer.encode(first=text, max_len=text_max_length)
            yield indices, segments, cat_to_id[category], random.choice(second_label_list)
it = get_keras_bert_iterator(r"data/keras_bert_train.txt", cat_to_id, tokenizer, second_label_list)
# next(it)
def batch_iter(data_path, cat_to_id, tokenizer, second_label_list, batch_size=64, shuffle=True):
    """生成批次数据"""
    keras_bert_iter = get_keras_bert_iterator(data_path, cat_to_id, tokenizer, second_label_list)
    while True:
        data_list = []
        for _ in range(batch_size):
            data = next(keras_bert_iter)
            data_list.append(data)
        if shuffle:
            random.shuffle(data_list)
        
        indices_list = []
        segments_list = []
        label_index_list = []
        second_label_list = []
        for data in data_list:
            indices, segments, label_index, second_label = data
            indices_list.append(indices)
            segments_list.append(segments)
            label_index_list.append(label_index)
            second_label_list.append(second_label)
 
        yield [np.array(indices_list), np.array(segments_list)], [np.array(label_index_list), np.array(second_label_list)]
it = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, second_label_list, batch_size=2)
next(it)
([array([[ 101, 4993, 2506, ...,  131,  123,  102],
         [101, 2506, 3696, ..., 1139,  125,  102]]),
  array([[0, 0, 0, ..., 0, 0, 0],
         [0, 0, 0, ..., 0, 0, 0]])],
 [array([0, 0]), array([0, 0])])

定义模型

def get_model(label_list, second_label_list):
    K.clear_session()
    
    bert_model = load_trained_model_from_checkpoint(bert_paths.config, bert_paths.checkpoint, seq_len=text_max_length)  #加载预训练模型
 
    for l in bert_model.layers:
        l.trainable = True
 
    input_indices = Input(shape=(None,))
    input_segments = Input(shape=(None,))
 
    bert_output = bert_model([input_indices, input_segments])
    bert_cls = Lambda(lambda x: x[:, 0])(bert_output) # 取出 [CLS] 对应的向量用来做分类
    output = Dense(len(label_list), activation='softmax')(bert_cls)
    output_second = Dense(len(second_label_list), activation='softmax')(bert_cls)
 
    model = Model([input_indices, input_segments], [output, output_second])
    model.compile(loss='sparse_categorical_crossentropy',
                  optimizer=Adam(1e-5),    #用足够小的学习率
                  metrics=['accuracy'])
    print(model.summary())
    return model
early_stopping = EarlyStopping(monitor='val_acc', patience=3)   #早停法，避免过拟合
plateau = ReduceLROnPlateau(monitor="val_acc", verbose=1, mode='max', factor=0.5, patience=2) #当评估指标不在晋升时，缩小学习率
checkpoint = ModelCheckpoint('trained_model/muilt_keras_bert_THUCNews.hdf5', monitor='val_acc',verbose=2, save_best_only=True, mode='max', save_weights_only=True) #保留最好的模型

模型训练

batch_size = 4
train_step = get_step(train_sample_count, batch_size)
dev_step = get_step(dev_sample_count, batch_size)
 
train_dataset_iterator = batch_iter(r"data/keras_bert_train.txt", cat_to_id, tokenizer, second_label_list, batch_size)
dev_dataset_iterator = batch_iter(r"data/keras_bert_dev.txt", cat_to_id, tokenizer, second_label_list, batch_size)
 
model = get_model(categories, second_label_list)
 
#模型训练
model.fit(
    train_dataset_iterator,
    steps_per_epoch=train_step,
    epochs=10,
    validation_data=dev_dataset_iterator,
    validation_steps=dev_step,
    callbacks=[early_stopping, plateau, checkpoint],
    verbose=1
)
Model: "functional_5"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 512)]        0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            [(None, 512)]        0                                            
__________________________________________________________________________________________________
functional_3 (Functional)       (None, 512, 768)     101677056   input_1[0][0]                    
                                                                 input_2[0][0]                    
__________________________________________________________________________________________________
lambda (Lambda)                 (None, 768)          0           functional_3[0][0]               
__________________________________________________________________________________________________
dense (Dense)                   (None, 14)           10766       lambda[0][0]                     
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 3)            2307        lambda[0][0]                     
==================================================================================================
Total params: 101,690,129
Trainable params: 101,690,129
Non-trainable params: 0
__________________________________________________________________________________________________
None
Epoch 1/10
     7/167215 [..............................] - ETA: 1829:52:33 - loss: 3.1260 - dense_loss: 1.4949 - dense_1_loss: 1.6311 - dense_accuracy: 0.6429 - dense_1_accuracy: 0.3571

正文完
机器学习
发表至： 机器学习
2021-03-06
 0
关于机器学习:直播预告｜FeatureStore-Meetup-V4携手第四范式-腾讯-微软大咖讲师共探特征平台实践
关于机器学习:小样本学习及其在美团场景中的应用
关于机器学习:从钢铁行业数字化管控与超自动化融合看华为WeAutomate能力进化
关于机器学习:ResNetRS谷歌领衔调优ResNet性能全面超越EfficientNet系列-2021-arxiv