Profile TensorFlow using TensorBoard

This guide will show you how to use the TensorFlow Profiler to profile the execution of your TensorFlow code.

Example Code

Copy and paste the following code into tf-profile.py.

from datetime import datetime
import os
import tensorflow

from tensorflow.keras.datasets import mnist
from tensorflow import keras
from tensorflow.keras import layers

(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

model = keras.Sequential([
    layers.Dense(512, activation="relu"),
    layers.Dense(10, activation="softmax")
])

model.compile(optimizer="rmsprop",
              loss="sparse_categorical_crossentropy",
              metrics=["accuracy"])

train_images = train_images.reshape((60000, 28 \* 28))
train_images = train_images.astype("float32") / 255
test_images = test_images.reshape((10000, 28 \* 28))
test_images = test_images.astype("float32") / 255

# Create a TensorBoard callback
logs = "logs/" + datetime.now().strftime("%Y%m%d-%H%M%S")

tboard_callback = tensorflow.keras.callbacks.TensorBoard(log_dir = logs,
                                                 histogram_freq = 1,
                                                 profile_batch = '10,20')

model.fit(train_images,
          train_labels,
          epochs=10,
          batch_size=128,
          callbacks = [tboard_callback])

The tensorflow.keras.callbacks.TensorBoard command will create a tensorboard callback and profile_batch will pick batch number 10 to batch number 20.

Local profiling on your own computer

Run the code:
```
python tf-profile.py
```
Compress the logs folder:
```
bashtar -zcvf ./logs.tar.gz ./logs
```
Download the tarball file with sftp and/or hal-ondemand.
Decompress the tarball file:
```
tar -zxvf ./logs.tar.gz
```
Install the TensorBoard profile plugin in your python environment:
```
pip install tensorboard_plugin_profile
```
Launch the TensorBoard with profiler installed:
```
tensorboard --logdir ./logs
```
Open the TensorBoard dashboard in your web browser. (Google Chrome is strongly recommended.)

Remote Profiling on HAL

Coming soon!