Multiple environments have been installed in ModelArts notebook instances, including TensorFlow. You can use pip install to install external libraries from a Jupyter notebook or terminal to facilitate use.

Installing an External Library from a Jupyter Notebook

Assume that you want to install Shapely from a notebook instance. Follow the following instructions:

In the left navigation pane of the ModelArts management console, choose DevEnviron > Notebooks. Open a notebook instance in the displayed notebook instance list.
In the Jupyter Notebook page that is displayed, click New and select the required AI engine from the drop-down list.
In the displayed window, type the following command in the code input bar to install Shapely:
pip install shapely

Installing an External Library from a Terminal

Assume that you want to install Shapely from the terminal of a notebook instance by using pip. Follow the following instructions:

In the left navigation pane of the ModelArts management console, choose DevEnviron > Notebooks. Open a notebook instance in the displayed notebook instance list.
In the displayed Jupyter dashboard, click New and choose Terminal from the shortcut menu.
For a notebook instance that does not use the AI engine of the Multi-Engine type, enter the following command in the code input bar to install Shapely:
/opt/conda/envs/python27_tf/bin/pip install Shapely
The Multi-Engine notebook instance can use multiple engines. By referring to the README file in the /home/ma-user/ path, switch to the installation package of the corresponding engine environment and install Shapely. For example, you can install Shapely from TensorFlow-1.13.1 with the following code:
```
source /home/ma-user/anaconda3/bin/activate TensorFlow-1.13.1
pip install shapely
```

Table 1 lists the Python paths of the TensorFlow, MXNet, PyTorch, Caffe, Scikit-learn & XGBoost, and Spark engines in the terminal. The pip system is also installed in the same directory as the related engine. For details about the engines used by Multi-Engine notebook instances, refer to the README file.

**Table 1** AI engines and their installation paths
AI Engine	Version	Python Path
TensorFlow	TF-1.8.0-python2.7	/opt/conda/envs/python27_tf/bin/python
TensorFlow	TF-1.8.0-python3.6	/opt/conda/envs/python36_tf/bin/python
MXNet	MXNet-1.2.1-python2.7	/opt/conda/envs/python27_mxnet/bin/python
MXNet	MXNet-1.2.1-python3.6	/opt/conda/envs/python36_mxnet/bin/python
PyTorch	PyTorch-1.0.0-python2.7	/opt/conda/envs/python27_pytorch/bin/python
PyTorch	PyTorch-1.0.0-python3.6	/opt/conda/envs/python36_pytorch/bin/python
Caffe	Caffe-1.0.0-python2.7	/opt/conda/envs/python27_caffe/bin/python
Scikit-learn & XGBoost	ML-1.0.0-python2.7	/opt/notebook/anaconda2/bin/python
Spark	Spark-2.2.0-python2.7	/opt/notebook/anaconda2/bin/python
Scikit-learn & XGBoost	ML-1.0.0-python3.6	/opt/notebook/anaconda3/bin/python
Spark	Spark-2.2.0-python3.6	/opt/notebook/anaconda3/bin/python

When you create a ModelArts training job, a new independent running environment is started, which is not associated with the packages installed in the Notebook environment. Therefore, add os.system('pip install xxx') to the startup code before importing the installation package.

For example, if you need to use the Shapely dependency in the training job, add the following code to the startup code:

os.system('pip install Shapely')
import Shapely

Installing External Libraries and Kernels in Notebook Instances

Installing an External Library from a Jupyter Notebook

Installing an External Library from a Terminal