ba-lab / code-snippets

Code Snippets - Python, R, Bash, etc.

GPU Code Snippets

---

How to use a specific GPU (in prayog10 server)?

wget https://raw.githubusercontent.com/ba-lab/code-snippets/master/train-test.py

# Use 1st GPU
CUDA_VISIBLE_DEVICES=0 python3 train-test.py

# Use 2nd GPU
CUDA_VISIBLE_DEVICES=1 python3 train-test.py

# Use CPU
CUDA_VISIBLE_DEVICES=-1 python3 train-test.py

---

How to allow GPU memory growth?

Add the following code at the beginning of your Python script or Notebook:

Option 0:

for gpu in tf.config.experimental.list_physical_devices('GPU'):
	print('Setting gpu growth for', gpu)
	tf.config.experimental.set_memory_growth(gpu, True)

Option 1:

import keras.backend as K
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
K.tensorflow_backend.set_session(sess)

Option 2:

config = tf.ConfigProto()
config.gpu_options.allow_growth=True
sess = tf.Session(config=config)

---

How to remotely access the Jupyter Notebook in prayog02?

In 1st Terminal start Jupyter service:

ssh [e-mail]
pip3 install jupyter (if not installed)
CUDA_VISIBLE_DEVICES=0 jupyter notebook --no-browser --port=8892
OR
CUDA_VISIBLE_DEVICES=0 /home/user/.local/bin/jupyter-notebook --no-browser --port=8892

This will give a URL path; leave the terminal open.

In 2nd terminal Port forward (this is a local terminal; not server terminal)

• If you are using Windows follow this instead

$ ssh -L 8892:127.0.0.1:8892 -N -f -l user prayog02.umsl.edu

Open the URL path in your local browser.

---

How to remotely access the Tensorboard in prayog02?

Install Tensorboard:

pip3 list
pip3 uninstall tensorboard
pip3 uninstall tensorflow
pip3 uninstall tensorflow-gpu
pip3 install tf-nightly-gpu-2.0-preview

In 1st Terminal start tensorboard service:

ssh [e-mail]
cd project-directory (this is where your logs will be written)
rm -r tb-logs
tensorboard --logdir ./tb-logs/

This will give a URL path (along with a port number, say 6007); leave the terminal open.

In 2nd terminal Port forward (this is a local terminal; not server terminal)

• If you are using Windows follow this instead

$ ssh -L 6007:127.0.0.1:6007 -N -f -l user prayog02.umsl.edu

Open the URL path in your local browser http://localhost:6007/

Test with the following Python code:

import tensorflow as tf
import datetime

# Clear any logs from previous runs
!rm -rf ./tb-logs/ 

mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

def create_model():
  return tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(512, activation='relu'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation='softmax')
  ])

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

log_dir="./tb-logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir, histogram_freq=1)

model.fit(x=x_train, 
          y=y_train, 
          epochs=3, 
          validation_data=(x_test, y_test), 
          callbacks=[tensorboard_callback])

---

How to rsync?

rsync -av --progress source/ destination/ --exclude dir2

---

How to increase the display decimals (precision) of metrics in Keras?

• Change “1e-4” to “1e-9” in generic_utils.py (at two places) to increase the output precision
• Not sure if this impacts only the display but also the actual accuracy calculations.

vim /home/notebook/anaconda3/lib/python3.6/site-packages/keras/utils/generic_utils.py
EDIT: info += ' %.4f' % avg

---

How to test GPU speed?

• Run this code

---

How to find out which device GPU device tensorflow is using?

from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())

---

How to make remote folders accessible locally in Mac?

brew cask install osxfuse
brew install sshfs
sshfs [e-mail]:/home/badri/ /Users/badriadhikari/prayog10.umsl.edu -ovolname=prayog10