365天深度学习训练营] 第7周：咖啡豆识别

前期工作

本文为365天深度学习训练营内部限免文章
参考本文所写记录性文章，请在文章开头保留以下内容

• 本文为365天深度学习训练营 中的学习记录博客
• 参考文章：365天深度学习训练营-第7周：咖啡豆识别（训练营内部成员可读）
• 原作者：K同学啊|接辅导、项目定制

我的环境：

• 语言环境：Python3.6.5
• 编译器：jupyter lab
• 深度学习环境：TensorFlow2.4.1

⏲往期文章：

• 5天学习计划-第6周：好莱坞明星识别

• 5天学习计划-第5周：运动鞋品牌识别

• 难度：夯实基础

• 语言：Python3、TensorFlow2

• 时间：9月5-9月9日

要求：

1. 自己搭建VGG-16网络框架
2. 调用官方的VGG-16网络框架

拔高（可选）：

1. 验证集准确率达到100%
2. 使用PPT画出VGG-16算法框架图（发论文需要这项技能）

探索（难度有点大）

1. 在不影响准确率的前提下轻量化模型

设置GPU

如果使用的是CPU可以忽略这步

import tensorflow as tf

gpus = tf.config.list_physical_devices("GPU")

if gpus:
    tf.config.experimental.set_memory_growth(gpus[0], True)  #设置GPU显存用量按需使用
    tf.config.set_visible_devices([gpus[0]],"GPU")

导入数据

from tensorflow       import keras
from tensorflow.keras import layers,models
import numpy             as np
import matplotlib.pyplot as plt
import os,PIL,pathlib

# 这里需要更换成相应的地址
data_dir = "./49-data/"
data_dir = pathlib.Path(data_dir)

image_count = len(list(data_dir.glob('*/*.png')))

print("图片总数为：",image_count)

图片总数为： 1200

import torch
import torch.nn as nn
import os,PIL,pathlib
from PIL import Image
from torchvision import transforms, datasets

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

数据预处理

加载数据

使用image_dataset_from_directory方法将磁盘中的数据加载到tf.data.Dataset中

batch_size = 32
img_height = 224
img_width = 224

"""
关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789
"""
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.2,
    subset="training",
    seed=123,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 1200 files belonging to 4 classes.
Using 960 files for training.

"""
关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789
"""
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.2,
    subset="validation",
    seed=123,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 1200 files belonging to 4 classes.
Using 240 files for validation.

我们可以通过class_names输出数据集的标签。标签将按字母顺序对应于目录名称。

class_names = train_ds.class_names
print(class_names)

['Dark', 'Green', 'Light', 'Medium']

可视化数据

plt.figure(figsize=(10, 4))  # 图形的宽为10高为5

for images, labels in train_ds.take(1):
    for i in range(10):
        
        ax = plt.subplot(2, 5, i + 1)  

        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(class_names[labels[i]])
        
        plt.axis("off")

for image_batch, labels_batch in train_ds:
    print(image_batch.shape)
    print(labels_batch.shape)
    break

(32, 224, 224, 3)
(32,)

配置数据集

• shuffle() ：打乱数据，关于此函数的详细介绍可以参考：https://zhuanlan.zhihu.com/p/42417456
• prefetch() ：预取数据，加速运行，其详细介绍可以参考我前两篇文章，里面都有讲解。
• cache() ：将数据集缓存到内存当中，加速运行

AUTOTUNE = tf.data.AUTOTUNE

train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds   = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

normalization_layer = layers.experimental.preprocessing.Rescaling(1./255)

train_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))
val_ds   = val_ds.map(lambda x, y: (normalization_layer(x), y))

image_batch, labels_batch = next(iter(val_ds))
first_image = image_batch[0]

# 查看归一化后的数据
print(np.min(first_image), np.max(first_image))

0.0 1.0

构建VGG-16网络

在官方模型与自建模型之间进行二选一就可以了，选着一个注释掉另外一个。

VGG优缺点分析：

• VGG优点

VGG的结构非常简洁，整个网络都使用了同样大小的卷积核尺寸（3x3）和最大池化尺寸（2x2）。

• VGG缺点

1)训练时间过长，调参难度大。2)需要的存储容量大，不利于部署。例如存储VGG-16权重值文件的大小为500多MB，不利于安装到嵌入式系统中。

官方模型

官网模型调用这块我放到后面几篇文章中，下面主要讲一下VGG-16

# model = tf.keras.applications.VGG16(weights='imagenet')
# model.summary()

自建模型

from tensorflow.keras import layers, models, Input
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout

def VGG16(nb_classes, input_shape):
    input_tensor = Input(shape=input_shape)
    # 1st block
    x = Conv2D(64, (3,3), activation='relu', padding='same',name='block1_conv1')(input_tensor)
    x = Conv2D(64, (3,3), activation='relu', padding='same',name='block1_conv2')(x)
    x = MaxPooling2D((2,2), strides=(2,2), name = 'block1_pool')(x)
    # 2nd block
    x = Conv2D(128, (3,3), activation='relu', padding='same',name='block2_conv1')(x)
    x = Conv2D(128, (3,3), activation='relu', padding='same',name='block2_conv2')(x)
    x = MaxPooling2D((2,2), strides=(2,2), name = 'block2_pool')(x)
    # 3rd block
    x = Conv2D(256, (3,3), activation='relu', padding='same',name='block3_conv1')(x)
    x = Conv2D(256, (3,3), activation='relu', padding='same',name='block3_conv2')(x)
    x = Conv2D(256, (3,3), activation='relu', padding='same',name='block3_conv3')(x)
    x = MaxPooling2D((2,2), strides=(2,2), name = 'block3_pool')(x)
    # 4th block
    x = Conv2D(512, (3,3), activation='relu', padding='same',name='block4_conv1')(x)
    x = Conv2D(512, (3,3), activation='relu', padding='same',name='block4_conv2')(x)
    x = Conv2D(512, (3,3), activation='relu', padding='same',name='block4_conv3')(x)
    x = MaxPooling2D((2,2), strides=(2,2), name = 'block4_pool')(x)
    # 5th block
    x = Conv2D(512, (3,3), activation='relu', padding='same',name='block5_conv1')(x)
    x = Conv2D(512, (3,3), activation='relu', padding='same',name='block5_conv2')(x)
    x = Conv2D(512, (3,3), activation='relu', padding='same',name='block5_conv3')(x)
    x = MaxPooling2D((2,2), strides=(2,2), name = 'block5_pool')(x)
    # full connection
    x = Flatten()(x)
    x = Dense(4096, activation='relu',  name='fc1')(x)
    x = Dense(4096, activation='relu', name='fc2')(x)
    output_tensor = Dense(nb_classes, activation='softmax', name='predictions')(x)

    model = Model(input_tensor, output_tensor)
    return model

model=VGG16(len(class_names), (img_width, img_height, 3))
model.summary()

2022-09-09 18:18:38.287389: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 411041792 exceeds 10% of free system memory.


Model: "model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(None, 224, 224, 3)]     0         
                                                                 
 block1_conv1 (Conv2D)       (None, 224, 224, 64)      1792      
                                                                 
 block1_conv2 (Conv2D)       (None, 224, 224, 64)      36928     
                                                                 
 block1_pool (MaxPooling2D)  (None, 112, 112, 64)      0         
                                                                 
 block2_conv1 (Conv2D)       (None, 112, 112, 128)     73856     
                                                                 
 block2_conv2 (Conv2D)       (None, 112, 112, 128)     147584    
                                                                 
 block2_pool (MaxPooling2D)  (None, 56, 56, 128)       0         
                                                                 
 block3_conv1 (Conv2D)       (None, 56, 56, 256)       295168    
                                                                 
 block3_conv2 (Conv2D)       (None, 56, 56, 256)       590080    
                                                                 
 block3_conv3 (Conv2D)       (None, 56, 56, 256)       590080    
                                                                 
 block3_pool (MaxPooling2D)  (None, 28, 28, 256)       0         
                                                                 
 block4_conv1 (Conv2D)       (None, 28, 28, 512)       1180160   
                                                                 
 block4_conv2 (Conv2D)       (None, 28, 28, 512)       2359808   
                                                                 
 block4_conv3 (Conv2D)       (None, 28, 28, 512)       2359808   
                                                                 
 block4_pool (MaxPooling2D)  (None, 14, 14, 512)       0         
                                                                 
 block5_conv1 (Conv2D)       (None, 14, 14, 512)       2359808   
                                                                 
 block5_conv2 (Conv2D)       (None, 14, 14, 512)       2359808   
                                                                 
 block5_conv3 (Conv2D)       (None, 14, 14, 512)       2359808   
                                                                 
 block5_pool (MaxPooling2D)  (None, 7, 7, 512)         0         
                                                                 
 flatten (Flatten)           (None, 25088)             0         
                                                                 
 fc1 (Dense)                 (None, 4096)              102764544 
                                                                 
 fc2 (Dense)                 (None, 4096)              16781312  
                                                                 
 predictions (Dense)         (None, 4)                 16388     
                                                                 
=================================================================
Total params: 134,276,932
Trainable params: 134,276,932
Non-trainable params: 0
_________________________________________________________________


2022-09-09 18:18:38.623610: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 411041792 exceeds 10% of free system memory.
2022-09-09 18:18:38.650108: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 411041792 exceeds 10% of free system memory.

网络结构图

参加了365天深度学习训练营的同学可以在语雀中查看网络结构图

VGG-16的结构说明：

● 13个卷积层（Convolutional Layer），分别用blockX_convX表示

● 3个全连接层（Fully connected Layer），分别用fcX与predictions表示

● 5个池化层（Pool layer），分别用blockX_pool表示

VGG-16包含了16个隐藏层（13个卷积层和3个全连接层），故称为VGG-16

自己画的结构图

编译

设置初始学习率

initial_learning_rate = 1e-4

lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate,
    decay_steps=30,
    decay_rate=0.92,
    staircase=True
)

# 设置优化器
opt = tf.keras.optimizers.Adam(learning_rate=initial_learning_rate)

model.compile(optimizer=opt,
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

设置早停

model.compile(optimizer=opt,
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

from tensorflow.keras.callbacks import ModelCheckpoint,EarlyStopping

checkpointer = ModelCheckpoint('best_model2.h5',
                               monitor='val_accuracy',
                               verbose=1,
                               save_best_only=True,
                               save_weights_only=True)

earlystopper = EarlyStopping(monitor='val_accuracy',
                             min_delta=0.001,
                             patience=10,
                             verbose=1     )

训练模型

epochs = 100

history = model.fit(train_ds,validation_data=val_ds,
          epochs=epochs,callbacks=[checkpointer,earlystopper])

Epoch 1/100


/home/liangjie/anaconda3/lib/python3.9/site-packages/keras/backend.py:5581: UserWarning: "`sparse_categorical_crossentropy` received `from_logits=True`, but the `output` argument was produced by a Softmax activation and thus does not represent logits. Was this intended?
  output, from_logits = _get_logits(
2022-09-09 18:18:52.862014: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 411041792 exceeds 10% of free system memory.
2022-09-09 18:18:52.911589: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 411041792 exceeds 10% of free system memory.


30/30 [==============================] - ETA: 0s - loss: 1.3866 - accuracy: 0.2583
Epoch 1: val_accuracy improved from -inf to 0.22083, saving model to best_model2.h5
30/30 [==============================] - 283s 9s/step - loss: 1.3866 - accuracy: 0.2583 - val_loss: 1.3780 - val_accuracy: 0.2208
Epoch 2/100
30/30 [==============================] - ETA: 0s - loss: 1.1149 - accuracy: 0.4042
Epoch 2: val_accuracy improved from 0.22083 to 0.56250, saving model to best_model2.h5
30/30 [==============================] - 280s 9s/step - loss: 1.1149 - accuracy: 0.4042 - val_loss: 0.8386 - val_accuracy: 0.5625
Epoch 3/100
30/30 [==============================] - ETA: 0s - loss: 0.7181 - accuracy: 0.6333
Epoch 3: val_accuracy improved from 0.56250 to 0.77500, saving model to best_model2.h5
30/30 [==============================] - 284s 9s/step - loss: 0.7181 - accuracy: 0.6333 - val_loss: 0.5654 - val_accuracy: 0.7750
Epoch 4/100
30/30 [==============================] - ETA: 0s - loss: 0.5410 - accuracy: 0.7490
Epoch 4: val_accuracy did not improve from 0.77500
30/30 [==============================] - 290s 10s/step - loss: 0.5410 - accuracy: 0.7490 - val_loss: 0.5881 - val_accuracy: 0.7042
Epoch 5/100
30/30 [==============================] - ETA: 0s - loss: 0.4812 - accuracy: 0.7500
Epoch 5: val_accuracy improved from 0.77500 to 0.84583, saving model to best_model2.h5
30/30 [==============================] - 288s 10s/step - loss: 0.4812 - accuracy: 0.7500 - val_loss: 0.4021 - val_accuracy: 0.8458
Epoch 6/100
30/30 [==============================] - ETA: 0s - loss: 0.3147 - accuracy: 0.8760
Epoch 6: val_accuracy improved from 0.84583 to 0.91667, saving model to best_model2.h5
30/30 [==============================] - 281s 9s/step - loss: 0.3147 - accuracy: 0.8760 - val_loss: 0.2322 - val_accuracy: 0.9167
Epoch 7/100
30/30 [==============================] - ETA: 0s - loss: 0.1761 - accuracy: 0.9417
Epoch 7: val_accuracy improved from 0.91667 to 0.95833, saving model to best_model2.h5
30/30 [==============================] - 284s 9s/step - loss: 0.1761 - accuracy: 0.9417 - val_loss: 0.0905 - val_accuracy: 0.9583
Epoch 8/100
30/30 [==============================] - ETA: 0s - loss: 0.1180 - accuracy: 0.9604
Epoch 8: val_accuracy improved from 0.95833 to 0.97083, saving model to best_model2.h5
30/30 [==============================] - 283s 9s/step - loss: 0.1180 - accuracy: 0.9604 - val_loss: 0.0926 - val_accuracy: 0.9708
Epoch 9/100
30/30 [==============================] - ETA: 0s - loss: 0.1047 - accuracy: 0.9615
Epoch 9: val_accuracy improved from 0.97083 to 0.97917, saving model to best_model2.h5
30/30 [==============================] - 280s 9s/step - loss: 0.1047 - accuracy: 0.9615 - val_loss: 0.0477 - val_accuracy: 0.9792
Epoch 10/100
30/30 [==============================] - ETA: 0s - loss: 0.0889 - accuracy: 0.9792
Epoch 10: val_accuracy did not improve from 0.97917
30/30 [==============================] - 276s 9s/step - loss: 0.0889 - accuracy: 0.9792 - val_loss: 0.0630 - val_accuracy: 0.9792
Epoch 11/100
30/30 [==============================] - ETA: 0s - loss: 0.0399 - accuracy: 0.9885
Epoch 11: val_accuracy did not improve from 0.97917
30/30 [==============================] - 273s 9s/step - loss: 0.0399 - accuracy: 0.9885 - val_loss: 0.0775 - val_accuracy: 0.9792
Epoch 12/100
30/30 [==============================] - ETA: 0s - loss: 0.0562 - accuracy: 0.9781
Epoch 12: val_accuracy did not improve from 0.97917
30/30 [==============================] - 275s 9s/step - loss: 0.0562 - accuracy: 0.9781 - val_loss: 0.1950 - val_accuracy: 0.9333
Epoch 13/100
30/30 [==============================] - ETA: 0s - loss: 0.1094 - accuracy: 0.9604
Epoch 13: val_accuracy did not improve from 0.97917
30/30 [==============================] - 282s 9s/step - loss: 0.1094 - accuracy: 0.9604 - val_loss: 0.6036 - val_accuracy: 0.8000
Epoch 14/100
30/30 [==============================] - ETA: 0s - loss: 0.1230 - accuracy: 0.9688
Epoch 14: val_accuracy did not improve from 0.97917
30/30 [==============================] - 285s 9s/step - loss: 0.1230 - accuracy: 0.9688 - val_loss: 0.3477 - val_accuracy: 0.9000
Epoch 15/100
30/30 [==============================] - ETA: 0s - loss: 0.0766 - accuracy: 0.9719
Epoch 15: val_accuracy did not improve from 0.97917
30/30 [==============================] - 284s 9s/step - loss: 0.0766 - accuracy: 0.9719 - val_loss: 0.1054 - val_accuracy: 0.9667
Epoch 16/100
30/30 [==============================] - ETA: 0s - loss: 0.0280 - accuracy: 0.9875
Epoch 16: val_accuracy did not improve from 0.97917
30/30 [==============================] - 286s 10s/step - loss: 0.0280 - accuracy: 0.9875 - val_loss: 0.0859 - val_accuracy: 0.9750
Epoch 17/100
30/30 [==============================] - ETA: 0s - loss: 0.0543 - accuracy: 0.9833
Epoch 17: val_accuracy did not improve from 0.97917
30/30 [==============================] - 285s 9s/step - loss: 0.0543 - accuracy: 0.9833 - val_loss: 0.1131 - val_accuracy: 0.9667
Epoch 18/100
30/30 [==============================] - ETA: 0s - loss: 0.0530 - accuracy: 0.9833
Epoch 18: val_accuracy did not improve from 0.97917
30/30 [==============================] - 289s 10s/step - loss: 0.0530 - accuracy: 0.9833 - val_loss: 0.0772 - val_accuracy: 0.9792
Epoch 19/100
30/30 [==============================] - ETA: 0s - loss: 0.0113 - accuracy: 0.9969
Epoch 19: val_accuracy improved from 0.97917 to 0.98750, saving model to best_model2.h5
30/30 [==============================] - 284s 9s/step - loss: 0.0113 - accuracy: 0.9969 - val_loss: 0.0545 - val_accuracy: 0.9875
Epoch 20/100
30/30 [==============================] - ETA: 0s - loss: 0.0115 - accuracy: 0.9958
Epoch 20: val_accuracy did not improve from 0.98750
30/30 [==============================] - 282s 9s/step - loss: 0.0115 - accuracy: 0.9958 - val_loss: 0.0745 - val_accuracy: 0.9833
Epoch 21/100
30/30 [==============================] - ETA: 0s - loss: 0.0140 - accuracy: 0.9948
Epoch 21: val_accuracy did not improve from 0.98750
30/30 [==============================] - 281s 9s/step - loss: 0.0140 - accuracy: 0.9948 - val_loss: 0.2310 - val_accuracy: 0.9583
Epoch 22/100
30/30 [==============================] - ETA: 0s - loss: 0.1326 - accuracy: 0.9510
Epoch 22: val_accuracy did not improve from 0.98750
30/30 [==============================] - 286s 10s/step - loss: 0.1326 - accuracy: 0.9510 - val_loss: 0.1169 - val_accuracy: 0.9542
Epoch 23/100
30/30 [==============================] - ETA: 0s - loss: 0.0357 - accuracy: 0.9896
Epoch 23: val_accuracy did not improve from 0.98750
30/30 [==============================] - 282s 9s/step - loss: 0.0357 - accuracy: 0.9896 - val_loss: 0.0613 - val_accuracy: 0.9750
Epoch 24/100
30/30 [==============================] - ETA: 0s - loss: 0.0148 - accuracy: 0.9948
Epoch 24: val_accuracy did not improve from 0.98750
30/30 [==============================] - 273s 9s/step - loss: 0.0148 - accuracy: 0.9948 - val_loss: 0.0496 - val_accuracy: 0.9875
Epoch 25/100
30/30 [==============================] - ETA: 0s - loss: 0.2279 - accuracy: 0.9240
Epoch 25: val_accuracy did not improve from 0.98750
30/30 [==============================] - 282s 9s/step - loss: 0.2279 - accuracy: 0.9240 - val_loss: 0.3698 - val_accuracy: 0.8625
Epoch 26/100
30/30 [==============================] - ETA: 0s - loss: 0.1049 - accuracy: 0.9688
Epoch 26: val_accuracy did not improve from 0.98750
30/30 [==============================] - 288s 10s/step - loss: 0.1049 - accuracy: 0.9688 - val_loss: 0.0966 - val_accuracy: 0.9792
Epoch 27/100
30/30 [==============================] - ETA: 0s - loss: 0.0308 - accuracy: 0.9875
Epoch 27: val_accuracy did not improve from 0.98750
30/30 [==============================] - 290s 10s/step - loss: 0.0308 - accuracy: 0.9875 - val_loss: 0.2521 - val_accuracy: 0.9375
Epoch 28/100
30/30 [==============================] - ETA: 0s - loss: 0.0200 - accuracy: 0.9927
Epoch 28: val_accuracy did not improve from 0.98750
30/30 [==============================] - 286s 10s/step - loss: 0.0200 - accuracy: 0.9927 - val_loss: 0.1435 - val_accuracy: 0.9625
Epoch 29/100
30/30 [==============================] - ETA: 0s - loss: 0.0144 - accuracy: 0.9948
Epoch 29: val_accuracy did not improve from 0.98750
30/30 [==============================] - 283s 9s/step - loss: 0.0144 - accuracy: 0.9948 - val_loss: 0.0737 - val_accuracy: 0.9708
Epoch 29: early stopping

可视化结果

from pyecharts.charts import *
import pyecharts.options as opts
from pyecharts.globals import ThemeType
loss = history.history['loss']
val_loss = history.history['val_loss']
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
line_loss = Line()
line_loss.add_xaxis([i for i in range(30)])
line_loss.add_yaxis('loss', loss, label_opts=opts.LabelOpts(is_show=False))
line_loss.add_yaxis('val_loss', val_loss, label_opts=opts.LabelOpts(is_show=False))
line_loss.set_global_opts(legend_opts=opts.LegendOpts(pos_top='5%',pos_left='20%'),
                    tooltip_opts=opts.TooltipOpts(trigger="axis", axis_pointer_type="line"))

line_acc = Line()
line_acc.add_xaxis([i for i in range(30)])
line_acc.add_yaxis('accuracy', acc, label_opts=opts.LabelOpts(is_show=False))
line_acc.add_yaxis('val_accuracy', val_acc, label_opts=opts.LabelOpts(is_show=False))
line_acc.set_global_opts(title_opts=opts.TitleOpts('模型训练过程效果记录', pos_left='center'),
                    legend_opts=opts.LegendOpts(pos_top='5%', pos_left='65%'),
                    yaxis_opts=opts.AxisOpts(is_scale=True),
                    tooltip_opts=opts.TooltipOpts(trigger="axis", axis_pointer_type="line"))

grid = Grid(init_opts=opts.InitOpts(theme=ThemeType.CHALK))
grid.add(line_loss,grid_opts=opts.GridOpts(pos_left='5%', pos_right='55%'))
grid.add(line_acc,grid_opts=opts.GridOpts(pos_left='55%', pos_right='5%'))
grid.render_notebook()

总结

源代码基础上，修改了哪些

设置了早停，epochs从20增加到100

有哪些改进

源代码结果：

修改结果前：val_accuracy: 0.96

修改结果后：val_accuracy: 0.970