如何在 Tensorflow-2.0 中绘制 tf.keras 模型？

Question

我升级到 Tensorflow 2.0，没有tf.summary.FileWriter("tf_graphs", sess.graph)。我正在查看有关此问题的其他一些 StackOverflow 问题，他们说使用tf.compat.v1.summary etc。在 Tensorflow 版本 2 中肯定有一种方法可以绘制和可视化 tf.keras 模型。它是什么？我正在寻找像下面这样的张量板输出。谢谢！

Answer 1

您可以可视化任何tf.function 修饰函数的图形，但首先，您必须跟踪其执行情况。

可视化 Keras 模型的图形意味着可视化它的 call 方法。

默认情况下，此方法没有tf.function 修饰，因此您必须将模型调用包装在正确修饰的函数中并执行它。

import tensorflow as tf

model = tf.keras.Sequential(
    [
        tf.keras.layers.Flatten(input_shape=(28, 28)),
        tf.keras.layers.Dense(32, activation="relu"),
        tf.keras.layers.Dropout(0.2),
        tf.keras.layers.Dense(10, activation="softmax"),
    ]
)


@tf.function
def traceme(x):
    return model(x)


logdir = "log"
writer = tf.summary.create_file_writer(logdir)
tf.summary.trace_on(graph=True, profiler=True)
# Forward pass
traceme(tf.zeros((1, 28, 28, 1)))
with writer.as_default():
    tf.summary.trace_export(name="model_trace", step=0, profiler_outdir=logdir)

Answer 2

根据docs，一旦您的模型经过训练，您就可以使用 Tensorboard 来可视化图表。

首先，定义您的模型并运行它。然后，打开 Tensorboard 并切换到 Graph 选项卡。

---

最小可编译示例

此示例取自文档。首先，定义您的模型和数据。

# Relevant imports.
%load_ext tensorboard

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from datetime import datetime
from packaging import version

import tensorflow as tf
from tensorflow import keras

# Define the model.
model = keras.models.Sequential([
    keras.layers.Flatten(input_shape=(28, 28)),
    keras.layers.Dense(32, activation='relu'),
    keras.layers.Dropout(0.2),
    keras.layers.Dense(10, activation='softmax')
])

model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy'])

(train_images, train_labels), _ = keras.datasets.fashion_mnist.load_data()
train_images = train_images / 255.0

接下来，训练您的模型。在这里，您需要为 Tensorboard 定义一个回调，以用于可视化统计数据和图表。

# Define the Keras TensorBoard callback.
logdir="logs/fit/" + datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = keras.callbacks.TensorBoard(log_dir=logdir)

# Train the model.
model.fit(
    train_images,
    train_labels, 
    batch_size=64,
    epochs=5, 
    callbacks=[tensorboard_callback])

训练结束后，在你的笔记本中运行

%tensorboard --logdir logs

并切换到导航栏中的 Graph 选项卡：

你会看到一个看起来很像这样的图表：

Answer 3

这是目前对我有用的东西 (TF 2.0.0)，基于 tf.keras.callbacks.TensorBoard 代码：

# After model has been compiled
from tensorflow.python.ops import summary_ops_v2
from tensorflow.python.keras.backend import get_graph
tb_path = '/tmp/tensorboard/'
tb_writer = tf.summary.create_file_writer(tb_path)
with tb_writer.as_default():
    if not model.run_eagerly:
        summary_ops_v2.graph(get_graph(), step=0)

Answer 4

另一个选择是使用这个网站：https://lutzroeder.github.io/netron/

使用 .h5 或 .tflite 文件生成图表。

它所基于的 github repo 可以在这里找到： https://github.com/lutzroeder/netron

Answer 5

这是 tf2.x 的解决方案，Graph visualization of subclassed model/layer

import tensorflow as tf
print("TensorFlow version:", tf.__version__)

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


class MyModel(Model):
    def __init__(self, dim):
        super(MyModel, self).__init__()
        self.conv1 = Conv2D(16, 3, activation='relu')
        self.conv2 = Conv2D(32, 3, activation='relu')
        self.conv3 = Conv2D(8, 3, activation='relu')
        self.flatten = Flatten()
        self.d1 = Dense(128, activation='relu')
        self.d2 = Dense(1)

    def call(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.flatten(x)
        x = self.d1(x)
        return self.d2(x)
    def build_graph(self):
        x = Input(shape=(dim))
        return Model(inputs=[x], outputs=self.call(x))


dim = (28, 28, 1)

# Create an instance of the model
model = MyModel((dim))
model.build((None, *dim))
model.build_graph().summary()


tf.keras.utils.plot_model(model.build_graph(), to_file="model.png",
           expand_nested=True, show_shapes=True)

输出是

TensorFlow version: 2.5.0

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 28, 28, 1)]       0         
_________________________________________________________________
conv2d (Conv2D)              (None, 26, 26, 16)        160       
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 24, 24, 32)        4640      
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 22, 22, 8)         2312      
_________________________________________________________________
flatten (Flatten)            (None, 3872)              0         
_________________________________________________________________
dense (Dense)                (None, 128)               495744    
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 129       
=================================================================
Total params: 502,985
Trainable params: 502,985
Non-trainable params: 0

这里也是图形可视化