keras模型的构建、训练、加载(模型不刷屏)
目录
设置模型不刷屏
让模型训练时不刷屏，但是要动态显示训练过程
- 推荐方法
import ipykernel # 设置keras训练、预测的verbose=1 - 尝试方法
输出结果：# 使用print("\r ……") for i in range(20): if i % 10 == 0: print("") time.sleep(0.2) print("\r 数值为: " + str(i), end="")
keras模型构建与训练
搭建深度学习模型
keras搭建模型使用import ipykernel。而使得模型的输出进度条不会刷屏！
以手写数字识别为例，搭建深度学习识别模型！
@staticmethod
def get_h5():
x = keras.Input([28, 28])
reshape1 = keras.layers.Reshape([28, 28, 1])(x)
Conv2D_1 = keras.layers.Conv2D(32, (3, 3), activation="relu")(reshape1)
MaxPooling_1 = keras.layers.MaxPool2D((2, 2))(Conv2D_1)
Conv2D_2 = keras.layers.Conv2D(64, (3, 3), activation="relu")(MaxPooling_1)
MaxPooling_2 = keras.layers.MaxPool2D((2, 2))(Conv2D_2)
Conv2D_3 = keras.layers.Conv2D(64, (3, 3), activation="relu")(MaxPooling_2)
flatten_1 = keras.layers.Flatten()(Conv2D_3)
Dense_1 = keras.layers.Dense(64, activation="relu")(flatten_1)
Dense_2 = keras.layers.Dense(10, activation="softmax")(Dense_1)
model = keras.Model(x, Dense_2)
print(model.summary())
return model
if __name__ == '__main__':
model_1 = get_h5()
print(model_1.summary())
打印模型架构信息：
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 28, 28) 0
_________________________________________________________________
reshape_1 (Reshape) (None, 28, 28, 1) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 26, 26, 32) 320
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 13, 13, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 11, 11, 64) 18496
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 5, 5, 64) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 3, 3, 64) 36928
_________________________________________________________________
flatten_1 (Flatten) (None, 576) 0
_________________________________________________________________
dense_1 (Dense) (None, 64) 36928
_________________________________________________________________
dense_2 (Dense) (None, 10) 650
=================================================================
Total params: 93,322
Trainable params: 93,322
Non-trainable params: 0
_________________________________________________________________
None
数据准备
手写数字识别数据集是一个经典的数据集，这里我们直接获取或使用keras.datasets.mnist.load_data()获取。第一次执行，本地没有会下载，第二次会直接加载本地的数据！
@staticmethod
def load_data():
# 先下载数据（http://yann.lecun.com/exdb/mnist/）到一个文件夹;若使用下面的命令
# 直接下载，数据默认会保存到用户目录下的 .keras/datasets/ 里面;当然你也可以指定缓
# 存的路径，如 "./mnist_data"
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data("./mnist_data")
return x_train, y_train, x_test, y_test
模型编译、训练
def compile(self):
model = self.model
model.compile(keras.optimizers.Adam(),
# loss=tf.keras.losses.categorical_crossentropy(),
loss="binary_crossentropy",
metrics=["acc", "poisson"])
return model
# 注：metrics里你可以添加自己需要的监控指标！
# ------------------------------------------------------------------------------
def fit(self):
model = self.compile()
checkpoint = keras.callbacks.ModelCheckpoint("checkpoint/best_model.h5", save_best_only=True)
model.fit(self.train_x, keras.utils.np_utils.to_categorical(self.data[1], 10),
batch_size=8, epochs=20, verbose=1,
callbacks=[checkpoint], validation_split=0.3,
shuffle=True)
model.save("./checkpoint/mnist_model.h5")
return model
模型预测
使用plot_model会有一些问题，建议参考传送门！ 我这边查遍了所有可查的资源也没有解决可视化的bug，先留着吧，以后研究。
def inference(self):
if os.path.exists(os.getcwd() + "/checkpoint/best_model.h5"):
model = keras.models.load_model(os.getcwd() + "/checkpoint/best_model.h5")
else:
model = self.fit()
# keras.utils.plot_model(model, to_file="./model.png")
pred_x = model.predict(self.test_x)
return pred_x
使用笔记本电脑测试输出模型架构：
# 测试代码：
model = keras.models.load_model(os.getcwd() + "/checkpoint/best_model.h5")
keras.utils.plot_model(model, to_file="./model.png")
测试结果：
输出预测的数值
h5_c = KerasH5()
test_pred = h5_c.inference()
# 获取预测的test图片数字
test_pred = np.argmax(test_pred, axis=1)
print(test_pred - h5_c.data[3])
完整代码
class KerasH5:
def __init__(self):
self.model = self.get_h5()
self.data = self.load_data()
self.handle_data()
@staticmethod
def get_h5():
x = keras.Input([28, 28])
reshape1 = keras.layers.Reshape([28, 28, 1])(x)
Conv2D_1 = keras.layers.Conv2D(32, (3, 3), activation="relu")(reshape1)
MaxPooling_1 = keras.layers.MaxPool2D((2, 2))(Conv2D_1)
Conv2D_2 = keras.layers.Conv2D(64, (3, 3), activation="relu")(MaxPooling_1)
MaxPooling_2 = keras.layers.MaxPool2D((2, 2))(Conv2D_2)
Conv2D_3 = keras.layers.Conv2D(64, (3, 3), activation="relu")(MaxPooling_2)
flatten_1 = keras.layers.Flatten()(Conv2D_3)
Dense_1 = keras.layers.Dense(64, activation="relu")(flatten_1)
Dense_2 = keras.layers.Dense(10, activation="softmax")(Dense_1)
model = keras.Model(x, Dense_2)
print(model.summary())
return model
@staticmethod
def load_data():
# 先下载数据（http://yann.lecun.com/exdb/mnist/）到一个文件夹;若使用下面的命令
# 直接下载，数据默认会保存到用户目录下的 .keras/datasets/ 里面;当然你也可以指定缓
# 存的路径，如 "./mnist_data"
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data("./mnist_data")
return x_train, y_train, x_test, y_test
def handle_data(self):
self.train_x = self.data[0] / 255
self.test_x = self.data[2] / 255
pass
def compile(self):
model = self.model
model.compile(keras.optimizers.Adam(),
# loss=tf.keras.losses.categorical_crossentropy(),
loss="binary_crossentropy",
metrics=["acc", "poisson"])
return model
def fit(self):
model = self.compile()
checkpoint = keras.callbacks.ModelCheckpoint("checkpoint/best_model.h5", save_best_only=True)
model.fit(self.train_x, keras.utils.np_utils.to_categorical(self.data[1], 10),
batch_size=8, epochs=20, verbose=1,
callbacks=[checkpoint], validation_split=0.3,
shuffle=True)
model.save("./checkpoint/mnist_model.h5")
return model
def inference(self):
if os.path.exists(os.getcwd() + "/checkpoint/best_model.h5"):
model = keras.models.load_model(os.getcwd() + "/checkpoint/best_model.h5")
else:
model = self.fit()
# keras.utils.plot_model(model, to_file="./model.png")
pred_x = model.predict(self.test_x)
return pred_x
def evaluate(self):
model = keras.models.load_model(os.getcwd() + "/checkpoint/best_model.h5")
res = model.evaluate(self.test_x, keras.utils.np_utils.to_categorical(self.data[3], 10),
batch_size=100, verbose=1)
# "Returns the loss value & metrics values for the model in test mode."
# res = ["loss", "acc", "poisson"] , for my metrics.
return res
if __name__ == '__main__':
h5_c = KerasH5()
test_pred = h5_c.inference()
test_pred = np.argmax(test_pred, axis=1)
print(test_pred - h5_c.data[3])
print(h5_c.evaluate())
h5模型文件加载与分析
我们生成的模型文件主要有两个关键字：'model_weights' 与 'optimizer_weights'，你感兴趣的东西主要在第一个里面！'model_weights'里面主要是node与weight。测试'model_weights'关键字里面的'conv2d_1'，"conv2d_1"关键字里面还是嵌套了关键字"conv2d_1"，下一级关键字为：bias:0与kernel:0，即模型的权值！这里的0标识了相应的层级输出单元的索引，如某单元有两个输出，那么可能会出现kernel:1。h5文件的结构如下：
-best_model.h5
--model_weights
---conv2d_1
----conv2d_1
-----bias
-----kernel
……
--optimizer_weights
best_model的主要node
左边的顺序是打印的顺序，可以发现其并不是我们模型架构的本身顺序，那么如何还原我们的模型呢？
我们的顺序是"input_1"------>"reshape_1"------>"conv2d_1"，如果知道了这些顺序，实际上我们自定义计算路径实现预测。h5文件中的每个过滤器的权值是分开的，HDFView可视化工具只能可视化第一个filter。如图所示：
Conv2d_2有64个3x3x32的filter，也只显示了3x3x32的！
# 输出"conv2d_1"的kernel:0
import h5py
with h5py.File("./checkpoint/best_model.h5", "r") as f:
i = 0
for node in f['model_weights']['conv2d_1']['conv2d_1']["kernel:0"]:
print(f"第{i}个数组：\n", node)
i += 1
输出结果为：
第0个数组：
[[[ 0.10564 -0.16097234 0.09915016 0.24613042 -0.08456963
-0.00454617 -0.05960701 0.06925311 -0.06459317 0.22562805
0.01301621 0.01064035 -0.20790304 0.04066228 -0.11448191
0.18316612 -0.17505792 0.1834265 -0.19409508 0.1662196
0.24422747 0.00287141 0.05400374 0.14116095 -0.4437251
-0.19318981 -0.1868813 -0.07384571 0.04616514 -0.09265891
-0.06916037 -0.28320184]]
[[ 0.24024534 0.03670109 -0.00417878 0.04924528 -0.00590578
0.06448627 0.13575661 0.07192774 -0.06872132 -0.01627582
-0.02804667 -0.15541585 0.0873699 0.11507094 -0.04232902
0.17835136 0.06311885 0.0213728 -0.3960938 0.12034985
-0.09786782 -0.14218147 -0.1110921 0.1760532 0.04455886
0.06668992 -0.20053528 0.01761849 -0.23805025 0.0733271
0.1568982 -0.12684295]]
[[ 0.0986921 0.00448399 0.10003785 -0.18682715 0.01353289
-0.11072443 -0.18143338 -0.07571129 -0.00445067 -0.25813603
0.00260918 -0.080878 0.15965232 -0.23025012 0.3168462
0.00272685 -0.0179387 -0.24480435 -0.2689363 0.26252237
-0.29930016 0.08890419 -0.2922434 -0.07410804 0.15559785
-0.03862172 -0.17935905 -0.33666146 -0.13682601 -0.06384996
-0.06077206 0.10360008]]]
第1个数组：
[[[-1.89561710e-01 8.70923400e-02 -1.47527248e-01 -2.86200404e-01
-2.20933463e-02 4.41786088e-02 -1.72818810e-01 -4.97002229e-02
1.34274061e-03 1.04595050e-01 1.01386808e-01 -3.04525405e-01
1.77128479e-01 -6.56002238e-02 -2.25848466e-01 -1.99412674e-01
-5.70780300e-02 1.16424084e-01 -1.70288030e-02 2.00436041e-01
2.16925085e-01 -1.99801117e-01 5.98984323e-02 1.81850672e-01
-1.89085931e-01 -2.12413073e-02 8.51334855e-02 1.15921244e-01
6.72528148e-02 1.37529252e-02 -1.16040766e-01 -9.02416930e-02]]
[[ 4.85295467e-02 -8.73176605e-02 1.86447635e-01 -4.45128798e-01
1.05745822e-01 7.90806860e-02 1.87297508e-01 2.26620194e-02
-5.34360297e-02 7.66759068e-02 -1.40923887e-01 -6.28831536e-02
9.31783319e-02 2.10334659e-01 -3.17964345e-01 -1.68596301e-02
1.50268286e-01 5.58755957e-02 -5.04233837e-02 -2.05894206e-02
-4.43578288e-02 1.26300499e-01 -6.44076392e-02 5.97948246e-02
1.02995068e-01 9.61822867e-02 1.38428584e-01 1.52803034e-01
1.28563449e-01 5.74405119e-02 2.58390248e-01 7.32380375e-02]]
[[-1.20110594e-01 6.92583397e-02 2.58188009e-01 -1.25160798e-01
-1.95767395e-02 4.51400131e-02 4.41203974e-02 4.18249443e-02
7.98251852e-02 -2.23809376e-01 1.39213994e-01 2.15288758e-01
-1.33189708e-01 -2.56082207e-01 2.64832348e-01 -1.06671266e-01
1.16286250e-02 -2.82497197e-01 -1.07641041e-01 3.55815107e-04
-4.32559460e-01 1.26489013e-01 -5.14838053e-03 -5.63367665e-01
1.28047854e-01 -8.79025981e-02 2.73709178e-01 -2.41094410e-01
2.20919736e-02 -4.37319390e-02 1.70607254e-01 2.54879922e-01]]]
第2个数组：
[[[-0.28743413 -0.05039541 -0.32322106 -0.28576022 -0.01186745
-0.16136153 -0.40770677 -0.03401943 0.04251064 0.17044757
-0.06370702 -0.25491872 -0.05032514 -0.05304395 0.10895949
-0.23599699 -0.08287958 0.09605569 0.33950958 -0.3004691
0.06335799 0.08101434 0.04215294 -0.23539479 -0.30663756
-0.20703389 0.06114375 -0.23180345 -0.10535791 -0.00877953
-0.6093956 0.08673523]]
[[-0.4594167 0.08436025 -0.19094844 0.24358904 -0.07383776
0.06015397 0.08437141 -0.0696937 -0.01023775 -0.13062379
-0.11832908 0.15471423 -0.09247598 0.15709819 -0.27268386
-0.25660586 0.05452353 0.18700227 0.28142843 -0.41201025
-0.26123935 0.16155057 0.04419636 -0.6761446 0.11539403
-0.01430293 0.10560115 0.24880193 -0.04546466 -0.06051769
-0.3192044 0.12822656]]
[[-0.3266369 0.07745046 0.00325512 0.31925917 -0.20738624
-0.09920612 0.09558655 -0.00091549 -0.01535833 -0.2739862
-0.21334378 0.04550457 -0.24382149 -0.0177168 -0.01421373
-0.3137606 -0.17557812 -0.07298016 0.29398265 -0.18772653
-0.2527745 -0.01731345 -0.02425755 0.08157137 0.12236273
0.0889151 -0.11918794 0.15772383 0.01994926 0.04337122
0.02604716 -0.05431315]]]
conv2d_1的卷积权值主要由3个数组组成，每个数组的shape为 3x1x32 ，即kernel:0的shape大小为: 3x3x1x32 。这与理论值是符合的，但是这里没有按过滤器进行切片分层，而是按shape的顺序进行切片的！相信有不少的同学在刚刚接触shape时，都知道卷积层的shape维度分别为 "卷积核的高"、"卷积核的宽"、输入通道数、过滤器的数量(pytorch不一样，注意区分)。
这里我为什么不直接输出kernel:0？因为直接输出看不见值，输出是一个封装的对象！如果你一定要直接输出可以采样命令：print(f['model_weights']['conv2d_1']['conv2d_1']["kernel:0"].value)，但是很有可能会给你下面的警告：
H5pyDeprecationWarning: dataset.value has been deprecated. Use dataset[()] instead.
下面我们看一下weight的bias权值：
import h5py
with h5py.File("./checkpoint/best_model.h5", "r") as f:
print(f['model_weights']['conv2d_1']['conv2d_1']["bias:0"])
print(f['model_weights']['conv2d_1']['conv2d_1']["bias:0"].value)
输出结果为：
# H5pyDeprecationWarning: dataset.value has been deprecated. Use dataset[()] instead.
# print(f['model_weights']['conv2d_1']['conv2d_1']["bias:0"].value)
<HDF5 dataset "bias:0": shape (32,), type "<f4">
[-0.01121292 -0.23676746 -0.07182489 0.0063485 -0.1473221 -0.20649968
-0.13457218 -0.18226808 -0.16741337 -0.05513388 -0.15839782 -0.05088074
-0.10574292 -0.17863815 -0.03026129 0.02034929 -0.22656491 -0.06224731
-0.00224195 -0.01806859 -0.04112944 -0.16773042 -0.20201148 -0.006715
-0.05799676 -0.20864251 -0.08409153 -0.11148676 -0.14344956 -0.15199727
-0.08022476 -0.17667659]
从node与weight还原模型
回顾建立mnist数据集的建模过程，我们在预测、评估时，是直接加载的"./checkpoint/best_model.h5"模型，然后就进行了相应的预测，那么keras是怎么知道先将数据接入Input、再将数据传到reshape后传到conv2d_1呢？难道它有魔法吗？如果我们知道其原理，那么完全可以摆脱架构的束缚，直接基于权值文件建立一个模型，有什么用？降低环境要求，轻量级，计算提速，可控！理论上只要模型文件保存了模型结构与权值，那么呢都能够还原模型，不然要它有何用！就为了让keras这些库加载一下吗？
keras是怎么读取文件的？
# load_model的代码块
if h5py is None:
raise ImportError('`load_model` requires h5py.')
model = None
opened_new_file = not isinstance(filepath, h5py.Group)
f = h5dict(filepath, 'r')
try:
model = _deserialize_model(f, custom_objects, compile)
finally:
if opened_new_file:
f.close()
目光如炬的你应该发现了，模型加载实际上最辛苦的崽是 _deserialize_model ，这个家伙是个内嵌函数，下面先分析一下它！
-
这个崽主要是进行模型结构复现与编译！
-
第一步，配置模型config文件：
model_config = f['model_config'] if model_config is None: raise ValueError('No model found in config.')从上面我们不难发现，'model_config'这个key是很关键的，没有它我们就没有办法重建模型结构！在上一小节中，我们只输出了'model_weights' 与 'optimizer_weights'关键字，实际上你可以在 "f.attrs.items()" 里找到相关的配置！
-
根据config生成模型结构
model_config = json.loads(model_config.decode('utf-8')) model = model_from_config(model_config, custom_objects=custom_objects)model_config 加载为json数据，model_from_config函数将json格式的配置文件转为keras的模型！
可以看见layers是按模型的顺序罗列的，且"inbound_nodes"记录了当前节点的输入来自于哪里！这个非常重要，有这个信息才能将模型复现！
后面就是一堆的判断与构造，你可以查看源代码理解其如何构造！
清澈的爱，只为中国
