How to use tensorflow on a single node?

I want to use tensorflow on SDIL: what do I have to do?

assigned to @ai5116

By till.riedel on 2022-05-18T10:47:56 (imported from GitLab)

The first steps covers the creation of a new virtual environment. Please use the following lines:

python3 -m venv hpc_tensorflow_sn source hpc_tensorflow_sn/bin/activate

The first lines creates a new virtual environment, the second activates it.

Now you can install Tensorflow via:

pip install --upgrade tensorflow

Let us install Matplotlib as well to run the example:

pip install matplotlib

Install now Jupyter in your virtual environment and add the virtual environment as Jupyter Kernel:

ipython kernel install --name "hpc_tensorflow_sn" --user

You can now run the following code (download: HPC_Tensorflow_SN.ipynb) in Jupyter

import tensorflow as tf
image_data = tf.keras.datasets.fashion_mnist

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

import matplotlib.pyplot as plt
plt.imshow(train_images[0])
train_labels[0]

train_images = train_images / 255.0
test_images = test_images / 255.0

model = tf.keras.Sequential([
                             tf.keras.layers.Flatten(input_shape=(28,28)),
                             tf.keras.layers.Dense(128,activation="relu"),
                             tf.keras.layers.Dense(10)
])
model.compile(optimizer="adam",
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=["accuracy"]
              )
model.fit(train_images,train_labels,epochs=10)

model.predict(test_images)[0]

test_labels[0]

By using NBConvert you can use

jupyter nbconvert --to notebook --execute HPC_Tensorflow_SN.ipynb

to run your Jupyter Notebook on the commandline

Usually the login node doesn´t have a NVIDIA GPU, so you can´t run the notebook but you can gain some ressources for a single run on one node by using:

salloc -p sdil --time=01:00:00 --gres=gpu:1

You specify basically the SDIL partition, wall-time and that you want a GPU.

After granting access to a node you can use the command:

jupyter nbconvert --to notebook --execute HPC_Tensorflow_SN.ipynb

The Jupyter notebook will be executed. There will be a new notebook with executed cells which you can investigate. To run it on slurm see #51

By rainer.duda on 2022-06-01T16:22:13 (imported from GitLab)

added answered label

changed the description

By till.riedel on 2022-06-01T16:25:57 (imported from GitLab)

added rookie woes label

unassigned @till.riedel