【tensorflow】T5.运动鞋识别

最新推荐文章于 2024-07-18 21:04:33 发布

矮山

最新推荐文章于 2024-07-18 21:04:33 发布

阅读量580

点赞数 10

文章标签： tensorflow 人工智能 python

本文链接：https://blog.csdn.net/pung0302/article/details/139524903

版权

>- **🍨 本文为[🔗365天深度学习训练营](https://mp.weixin.qq.com/s/0dvHCaOoFnW8SCp3JpzKxg) 中的学习记录博客**
>- **🍖 原作者：[K同学啊](https://mtyjkh.blog.csdn.net/)**

一、前期工作和数据预处理

#包含导包和数据，数据处理、检查和配置数据集

from tensorflow.keras import models, layers
import tensorflow as tf
import pathlib, PIL
import matplotlib.pyplot as plt

data_dir = "/Volumes/T7 Shield/code/shoes"
data_dir = pathlib.Path(data_dir)

image_count = len(list(data_dir.glob("*/*/*.jpg")))
print(image_count)

roses = list(data_dir.glob("train/nike/*.jpg"))
PIL.Image.open(roses[0])

out[1]：578 #img_count

batch_size = 32
image_height = 224
image_width = 224

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "/Volumes/T7 Shield/code/shoes/train",
    image_size = (image_height, image_width),
    batch_size = batch_size,
    seed = 123
    
    )

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "/Volumes/T7 Shield/code/shoes/test",
    image_size = (image_height, image_width),
    batch_size = batch_size,
    seed = 123
    
    )

class_names = val_ds.class_names
print(class_names)

out[2]: Found 502 files belonging to 2 classes.
Found 76 files belonging to 2 classes.
['adidas', 'nike'] #class_names

plt.figure(figsize = (20, 10))

for image, labels in val_ds.take(1):
    for i in range(20):
        plt.subplot(5, 10, i+1)
        
        plt.imshow(image[i].numpy().astype("uint8"))
        plt.title(class_names[labels[i]])
        plt.axis("off")

out[3]:

AUTOTUNE = tf.data.AUTOTUNE
train_ds = train_ds.cache().shuffle(1000).prefetch(AUTOTUNE)
val_ds = val_ds.cache().prefetch(AUTOTUNE)

二、搭建cnn网络：

#三个3*3卷积，两个平均池化，外加两个dropout层

model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1. / 255, input_shape = (image_height, image_width, 3)),
    
    layers.Conv2D(16, (3, 3), activation = "relu", input_shape = (image_height, image_width, 3)),
    layers.AveragePooling2D(2, 2),
    layers.Conv2D(32, (3, 3), activation = "relu"),
    layers.AveragePooling2D(2, 2),
    layers.Dropout(0.3),
    layers.Conv2D(64, (3,3), activation = "relu"),
    layers.Dropout(0.3),
    
    layers.Flatten(),
    layers.Dense(128, activation = "relu"),
    layers.Dense(2)
    
    ])

model.summary()

summary一下模型

Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
rescaling (Rescaling) (None, 224, 224, 3) 0

conv2d (Conv2D) (None, 222, 222, 16) 448

average_pooling2d (AverageP (None, 111, 111, 16) 0
ooling2D)

conv2d_1 (Conv2D) (None, 109, 109, 32) 4640

average_pooling2d_1 (Averag (None, 54, 54, 32) 0
ePooling2D)

dropout (Dropout) (None, 54, 54, 32) 0

conv2d_2 (Conv2D) (None, 52, 52, 64) 18496

dropout_1 (Dropout) (None, 52, 52, 64) 0

flatten (Flatten) (None, 173056) 0

dense (Dense) (None, 128) 22151296

dense_1 (Dense) (None, 2) 258

=================================================================
Total params: 22,175,138
Trainable params: 22,175,138
Non-trainable params: 0
_________________________________________________________________

三、编译

1. 学习率

这里的问题出在原代码里改了学习率，学习率太大了一直在跳，这里把初始lr改到了0.001，正确率达到了原代码的0.89474：

initial_learning_rate = 0.001

lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate, 
    decay_steps = 10, 
    decay_rate = 0.92,
    staircase = True
    
    )

2. 优化器：

opt = tf.keras.optimizers.Adam(learning_rate=lr_schedule)

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

3. 加点小组件（checkpointer和earlystooping）

checkpointer = tf.keras.callbacks.ModelCheckpoint(
    "best_model.h5",
    monitor = "val_accuracy",
    verbose = 1,
    save_best_only = True,
    save_weights_only = True
    
    )

# earlystopping = tf.keras.callbacks.EarlyStopping(
#     monitor = "val_accuracy",
#     min_delta = 0.001,
#     patience = 30,
#     verbose = 1
    
#     )

#因为在看最后能跑到多少准确率，就把earlystopping注解了

四、模型训练：

epochs = 50

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

acc = history.history["accuracy"]
val_acc = history.history["val_accuracy"]

loss = history.history["loss"]
val_loss = history.history["val_loss"]

Epoch 1/50
15/16 ━━━━━━━━━━━━━━━━━━━━ 0s 611ms/step - accuracy: 0.5082 - loss: 7.3733
Epoch 1: val_accuracy improved from -inf to 0.72368, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 15s 659ms/step - accuracy: 0.5110 - loss: 6.9457 - val_accuracy: 0.7237 - val_loss: 0.6176
Epoch 2/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.6835 - loss: 0.5930
Epoch 2: val_accuracy did not improve from 0.72368
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.6692 - loss: 0.6071 - val_accuracy: 0.6974 - val_loss: 0.6057
Epoch 3/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.7108 - loss: 0.5699
Epoch 3: val_accuracy did not improve from 0.72368
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.7147 - loss: 0.5641 - val_accuracy: 0.7105 - val_loss: 0.5437
Epoch 4/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.7161 - loss: 0.5444
Epoch 4: val_accuracy did not improve from 0.72368
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.7129 - loss: 0.5471 - val_accuracy: 0.6316 - val_loss: 0.5698
Epoch 5/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.7426 - loss: 0.5090
Epoch 5: val_accuracy improved from 0.72368 to 0.75000, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 1s 63ms/step - accuracy: 0.7414 - loss: 0.5107 - val_accuracy: 0.7500 - val_loss: 0.5678
Epoch 6/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.7393 - loss: 0.4984
Epoch 6: val_accuracy improved from 0.75000 to 0.77632, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 1s 62ms/step - accuracy: 0.7450 - loss: 0.4941 - val_accuracy: 0.7763 - val_loss: 0.4796
Epoch 7/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.8263 - loss: 0.3997
Epoch 7: val_accuracy did not improve from 0.77632
16/16 ━━━━━━━━━━━━━━━━━━━━ 1s 17ms/step - accuracy: 0.8310 - loss: 0.3977 - val_accuracy: 0.7632 - val_loss: 0.4676
Epoch 8/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.8544 - loss: 0.3663
Epoch 8: val_accuracy did not improve from 0.77632
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.8558 - loss: 0.3665 - val_accuracy: 0.7763 - val_loss: 0.4454
Epoch 9/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.8896 - loss: 0.3128
Epoch 9: val_accuracy improved from 0.77632 to 0.81579, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 4s 262ms/step - accuracy: 0.8902 - loss: 0.3136 - val_accuracy: 0.8158 - val_loss: 0.4194
Epoch 10/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9004 - loss: 0.2866
Epoch 10: val_accuracy did not improve from 0.81579
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.8968 - loss: 0.2890 - val_accuracy: 0.7895 - val_loss: 0.5074
Epoch 11/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.8945 - loss: 0.3077
Epoch 11: val_accuracy did not improve from 0.81579
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.8930 - loss: 0.3085 - val_accuracy: 0.7632 - val_loss: 0.4555
Epoch 12/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9072 - loss: 0.2591
Epoch 12: val_accuracy improved from 0.81579 to 0.82895, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 1s 62ms/step - accuracy: 0.9078 - loss: 0.2585 - val_accuracy: 0.8289 - val_loss: 0.3995
Epoch 13/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9364 - loss: 0.2387
Epoch 13: val_accuracy did not improve from 0.82895
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9364 - loss: 0.2367 - val_accuracy: 0.8026 - val_loss: 0.3749
Epoch 14/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9595 - loss: 0.1921
Epoch 14: val_accuracy did not improve from 0.82895
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9559 - loss: 0.1958 - val_accuracy: 0.7895 - val_loss: 0.3649
Epoch 15/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9691 - loss: 0.1668
Epoch 15: val_accuracy did not improve from 0.82895
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9628 - loss: 0.1728 - val_accuracy: 0.8158 - val_loss: 0.3955
Epoch 16/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9615 - loss: 0.1708
Epoch 16: val_accuracy improved from 0.82895 to 0.84211, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 2s 121ms/step - accuracy: 0.9574 - loss: 0.1742 - val_accuracy: 0.8421 - val_loss: 0.3746
Epoch 17/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9718 - loss: 0.1319
Epoch 17: val_accuracy did not improve from 0.84211
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9705 - loss: 0.1348 - val_accuracy: 0.7895 - val_loss: 0.3599
Epoch 18/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9560 - loss: 0.1589
Epoch 18: val_accuracy did not improve from 0.84211
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9568 - loss: 0.1577 - val_accuracy: 0.7763 - val_loss: 0.3542
Epoch 19/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9448 - loss: 0.1468
Epoch 19: val_accuracy did not improve from 0.84211
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9499 - loss: 0.1435 - val_accuracy: 0.8289 - val_loss: 0.3404
Epoch 20/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9627 - loss: 0.1271
Epoch 20: val_accuracy did not improve from 0.84211
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9632 - loss: 0.1276 - val_accuracy: 0.8421 - val_loss: 0.3270
Epoch 21/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9788 - loss: 0.1303
Epoch 21: val_accuracy improved from 0.84211 to 0.86842, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 1s 62ms/step - accuracy: 0.9779 - loss: 0.1287 - val_accuracy: 0.8684 - val_loss: 0.3635
Epoch 22/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9822 - loss: 0.1072
Epoch 22: val_accuracy did not improve from 0.86842
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9793 - loss: 0.1093 - val_accuracy: 0.8421 - val_loss: 0.3243
Epoch 23/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9845 - loss: 0.0964
Epoch 23: val_accuracy improved from 0.86842 to 0.89474, saving model to best_model.weights.h5
16/16 ━━━━━━━━━━━━━━━━━━━━ 1s 62ms/step - accuracy: 0.9830 - loss: 0.0978 - val_accuracy: 0.8947 - val_loss: 0.3190
Epoch 24/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9755 - loss: 0.1034
Epoch 24: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9756 - loss: 0.1040 - val_accuracy: 0.8289 - val_loss: 0.3195
Epoch 25/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9782 - loss: 0.0975
Epoch 25: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9779 - loss: 0.0966 - val_accuracy: 0.8816 - val_loss: 0.3066
Epoch 26/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9755 - loss: 0.1012
Epoch 26: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9758 - loss: 0.1002 - val_accuracy: 0.8684 - val_loss: 0.3077
Epoch 27/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9771 - loss: 0.0921
Epoch 27: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9771 - loss: 0.0917 - val_accuracy: 0.8947 - val_loss: 0.3060
Epoch 28/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9871 - loss: 0.0825
Epoch 28: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9866 - loss: 0.0844 - val_accuracy: 0.8553 - val_loss: 0.3174
Epoch 29/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9521 - loss: 0.1154
Epoch 29: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9561 - loss: 0.1104 - val_accuracy: 0.8421 - val_loss: 0.3100
Epoch 30/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9904 - loss: 0.0769
Epoch 30: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9881 - loss: 0.0783 - val_accuracy: 0.8947 - val_loss: 0.3013
Epoch 31/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9829 - loss: 0.0853
Epoch 31: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9824 - loss: 0.0851 - val_accuracy: 0.8947 - val_loss: 0.2996
Epoch 32/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9746 - loss: 0.0935
Epoch 32: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9752 - loss: 0.0915 - val_accuracy: 0.8947 - val_loss: 0.3012
Epoch 33/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9932 - loss: 0.0719
Epoch 33: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9910 - loss: 0.0744 - val_accuracy: 0.8947 - val_loss: 0.3006
Epoch 34/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9742 - loss: 0.0802
Epoch 34: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9739 - loss: 0.0806 - val_accuracy: 0.8947 - val_loss: 0.3002
Epoch 35/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9876 - loss: 0.0812
Epoch 35: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9862 - loss: 0.0803 - val_accuracy: 0.8947 - val_loss: 0.3010
Epoch 36/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9809 - loss: 0.0893
Epoch 36: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9824 - loss: 0.0862 - val_accuracy: 0.8816 - val_loss: 0.3063
Epoch 37/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9823 - loss: 0.0873
Epoch 37: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 18ms/step - accuracy: 0.9835 - loss: 0.0866 - val_accuracy: 0.8947 - val_loss: 0.3014
Epoch 38/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9727 - loss: 0.0840
Epoch 38: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9737 - loss: 0.0835 - val_accuracy: 0.8947 - val_loss: 0.3000
Epoch 39/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9924 - loss: 0.0641
Epoch 39: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9899 - loss: 0.0670 - val_accuracy: 0.8947 - val_loss: 0.2997
Epoch 40/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9857 - loss: 0.0771
Epoch 40: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 18ms/step - accuracy: 0.9853 - loss: 0.0760 - val_accuracy: 0.8947 - val_loss: 0.3000
Epoch 41/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9958 - loss: 0.0703
Epoch 41: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9938 - loss: 0.0708 - val_accuracy: 0.8947 - val_loss: 0.3001
Epoch 42/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9864 - loss: 0.0670
Epoch 42: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9858 - loss: 0.0679 - val_accuracy: 0.8947 - val_loss: 0.2987
Epoch 43/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9808 - loss: 0.0822
Epoch 43: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9809 - loss: 0.0815 - val_accuracy: 0.8947 - val_loss: 0.2984
Epoch 44/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9865 - loss: 0.0760
Epoch 44: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9858 - loss: 0.0762 - val_accuracy: 0.8947 - val_loss: 0.2981
Epoch 45/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9807 - loss: 0.0799
Epoch 45: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9820 - loss: 0.0779 - val_accuracy: 0.8947 - val_loss: 0.2977
Epoch 46/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9740 - loss: 0.0852
Epoch 46: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9762 - loss: 0.0829 - val_accuracy: 0.8947 - val_loss: 0.2979
Epoch 47/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9756 - loss: 0.0744
Epoch 47: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9771 - loss: 0.0740 - val_accuracy: 0.8947 - val_loss: 0.2984
Epoch 48/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9830 - loss: 0.0688
Epoch 48: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9823 - loss: 0.0704 - val_accuracy: 0.8947 - val_loss: 0.2986
Epoch 49/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9787 - loss: 0.0873
Epoch 49: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9793 - loss: 0.0850 - val_accuracy: 0.8947 - val_loss: 0.2982
Epoch 50/50
13/16 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.9885 - loss: 0.0714
Epoch 50: val_accuracy did not improve from 0.89474
16/16 ━━━━━━━━━━━━━━━━━━━━ 0s 17ms/step - accuracy: 0.9883 - loss: 0.0723 - val_accuracy: 0.8947 - val_loss: 0.2979

五、模型评估：

acc = history.history["accuracy"]
val_acc = history.history["val_accuracy"]

loss = history.history["loss"]
val_loss = history.history["val_loss"]

epochs = range(len(loss))

plt.figure(figsize = (12, 4))

plt.subplot(1, 2, 1)
plt.plot(epochs, acc, label = "train accuracy")
plt.plot(epochs, val_acc, label = "validation accuracy")
plt.legend(loc = "lower right")
plt.title("Accuracy")

plt.subplot(1, 2, 2)
plt.plot(epochs, loss, label = "train loss")
plt.plot(epochs, val_loss, label = "validation loss")
plt.legend(loc = "lower right")
plt.title("Loss")

plt.show()

六、总结：

很神奇的是，半个月前选学习率的时候，在本地用cpu跑的0.001只能达到0.7几左右，到0.0005才能到0.85，0.0001的话就50轮完全不够用，收敛不到家（无论如何都跑不到原代码的0.89），遂放弃。半个月后为了写这周的笔记，同样的参数只能达到0.7几的准确率，完全和之前的结果不一样......于是放到gpu上跑，lr在0.001是和原代码跑出来的结果基本一样的（曲线还是有点不大相同），可以知道lr的正确答案应该就是0.001，但是差距这么大是我没想到的...

矮山

关注

10
点赞
踩
12

收藏

觉得还不错? 一键收藏
0
评论
【tensorflow】T5.运动鞋识别

很神奇的是，半个月前选学习率的时候，在本地用cpu跑的0.001只能达到0.7几左右，到0.0005才能到0.85，0.0001的话就50轮完全不够用，收敛不到家（无论如何都跑不到原代码的0.89），遂放弃。半个月后为了写这周的笔记，同样的参数只能达到0.7几的准确率，完全和之前的结果不一样......于是放到gpu上跑，lr在0.001是和原代码跑出来的结果基本一样的（曲线还是有点不大相同），可以知道lr的正确答案应该就是0.001，但是差距这么大是我没想到的...
复制链接

扫一扫