Advanced topics ¶

This file covers some more advanced use cases of nbgrader.

Running nbgrader with JupyterHub ¶

Please see Using nbgrader with JupyterHub.

Advanced “Assignment List” installation ¶

Getting information from the database ¶

nbgrader offers a fairly rich API for interfacing with the database. The API should allow you to access pretty much anything you want, though if you find something that can’t be accessed through the API please open an issue!

In this example, we’ll go through how to create a CSV file of grades for each student and assignment using nbgrader and pandas.

New in version 0.4.0: nbgrader now comes with CSV export functionality out-of-the box using the nbgrader export command. However, this example is still kept for reference as it may be useful for defining your own exporter.

import pandas as pd
from nbgrader.api import Gradebook, MissingEntry

# Create the connection to the database
with Gradebook('sqlite:///gradebook.db') as gb:

    grades = []

    # Loop over each assignment in the database
    for assignment in gb.assignments:

        # Loop over each student in the database
        for student in gb.students:

            # Create a dictionary that will store information about this student's
            # submitted assignment
            score = {}
            score['max_score'] = assignment.max_score
            score['student'] = student.id
            score['assignment'] = assignment.name

            # Try to find the submission in the database. If it doesn't exist, the
            # `MissingEntry` exception will be raised, which means the student
            # didn't submit anything, so we assign them a score of zero.
            try:
                submission = gb.find_submission(assignment.name, student.id)
            except MissingEntry:
                score['score'] = 0.0
            else:
                score['score'] = submission.score

            grades.append(score)

    # Create a pandas dataframe with our grade information, and save it to disk
    grades = pd.DataFrame(grades).set_index(['student', 'assignment']).sortlevel()
    grades.to_csv('grades.csv')

    # Print out what the grades look like
    with open('grades.csv', 'r') as fh:
        print(fh.read())

After running the above code, you should see that grades.csv contains something that looks like:

student,assignment,max_score,score
bitdiddle,ps1,9.0,1.5
hacker,ps1,9.0,3.0

Using nbgrader preprocessors ¶

Several of the nbgrader preprocessors can be used with nbconvert without actually relying on the rest of the nbgrader machinery. In particular, the following preprocessors can be applied to other nbconvert workflows:

ClearOutput – clears outputs of all cells
ClearSolutions – removes solutions between the solution delimeters (see “Autograded answer” cells).
HeaderFooter – concatenates notebooks together, prepending a “header” notebook and/or appending a “footer” notebook to another notebook.
LimitOutput – limits the amount of output any given cell can have. If a cell has too many lines of outputs, they will be truncated.

Using these preprocessors in your own nbconvert workflow is relatively straightforward. In your nbconvert_config.py file, you would add, for example:

c.Exporter.preprocessors = ['nbgrader.preprocessors.ClearSolutions']

See also the nbconvert docs on custom preprocessors.

Calling nbgrader apps from Python ¶

New in version 0.5.0: Much of nbgrader’s high level functionality can now be accessed through an official Python API.

Grading in a docker container ¶

For security reasons, it may be advantageous to do the grading with a kernel running in isolation, e.g. in a docker container. We will assume that docker is already installed and an appropriate image has been downloaded. Otherwise, refer to the docker documentation for information on how to install and run docker.

A convenient way to switch to a kernel running in a docker container is provided by envkernel which serves a double purpose. In a first step, it is writing a new kernelspec file. Later it ensures that the docker container is run and the kernel started.

Presently, envkernel is only available from its Github repository and can be installed directly from there into a virtual environment

pip install https://github.com/NordicHPC/envkernel/archive/master.zip

As an alternative, the script envkernel.py can be put in a different location, e.g. /opt/envkernel, as long as it is accessible there also later during grading.

Now, a new kernel can be installed by means of

./envkernel.py docker --name=NAME --display-name=DNAME DOCKER-IMAGE

Here, NAME should be replaced by the name to be given to the kernel. After installation of the kernel, it will be displayed in the list of kernels when executing jupyter kernelspec list. DNAME should be replaced by the name under which the kernel shall be known in the Jupyter notebook GUI. After installation of the kernel, this name will be listed as a possible kernel when starting a new notebook. Finally, DOCKER-IMAGE should be replaced by the name of the docker image in which the kernel is to be run, e.g. python:3, continuumio/anaconda3, or some other suitable image.

The command given above will install the kernel in the system-wide location for Jupyter data files. If installation in the corresponding user directory is desired, the option --user should be added before the name of the docker image. By default, envkernel will install a Python kernel. For the installation of other kernels, see the README of envkernel.

In order to run the grading process with the new kernel, one can specify its name in nbgrader_config.py

c.ExecutePreprocessor.kernel_name = NAME

where NAME should be replaced by the name chosen when running the envkernel script. Alternatively, the name can be specified when running nbgrader from the command line

nbgrader autograde --ExecutePreprocessor.kernel_name=NAME ASSIGNMENT_NAME

In addition to docker, envkernel also supports singularity as a containerization system. For details on using envkernel with singularity, see the README of envkernel.

Automatic test code generation ¶

New in version 0.9.0.

See also

“Autograder tests” cells with automatically generated test code: General introduction to automatic test code generation.

nbgrader now supports generating test code automatically using ### AUTOTEST and ### HASHED AUTOTEST statements. In this section, you can find more detail on how this works and how to customize the test generation process. Suppose you ask students to create a foo function that adds 5 to an integer. In the source copy of the notebook, you might write something like

### BEGIN SOLUTION
def foo(x):
  return x + 5
### END SOLUTION

In a test cell, you would normally then write test code manually to probe various aspects of the solution. For example, you might check that the function increments 3 to 8 properly, and that the type of the output is an integer.

assert isinstance(foo(3), int), "incrementing an int by 5 should return an int"
assert foo(3) == 8, "3+5 should be 8"

nbgrader now provides functionality to automate this process. Instead of writing tests explicitly, you can instead specify what you want to test, and let nbgrader decide how to test it automatically.

### AUTOTEST foo(3)

This directive indicates that you want to check foo(3) in the student’s notebook, and make sure it aligns with the value of foo(3) in the current source copy. You can write any valid expression (in the language of your notebook) after the ### AUTOTEST directive. For example, you could write

### AUTOTEST (foo(3) - 5 == 3)

to generate test code for the expression foo(3)-5==3 (i.e., a boolean value), and make sure that evaluating the student’s copy of this expression has a result that aligns with the source version (i.e., True). You can write multiple ### AUTOTEST directives in one cell. You can also separate multiple expressions on one line with semicolons:

### AUTOTEST foo(3); foo(4); foo(5) != 8

These directives will insert code into student notebooks where the solution is available in plaintext. If you want to obfuscate the answers in the student copy, you should instead use a ### HASHED AUTOTEST, which will produce a student notebook where the answers are hashed and not viewable by students.

When you generate an assignment containing ### AUTOTEST (or ### HASHED AUTOTEST) statements, nbgrader looks for a file named autotests.yml that contains instructions on how to generate test code. It first looks in the assignment directory itself (in case you want to specify special tests for just that assignment), and if it is not found there, nbgrader searches in the course root directory. The autotests.yml file is a YAML file that looks something like this:

python3:
    setup: "from hashlib import sha1"
    hash: 'sha1({{snippet}}.encode("utf-8")+b"{{salt}}").hexdigest()'
    dispatch: "type({{snippet}})"
    normalize: "str({{snippet}})"
    check: 'assert {{snippet}} == """{{value}}""", """{{message}}"""'
    success: "print('Success!')"

    templates:
        default:
            - test: "type({{snippet}})"
              fail: "type of {{snippet}} is not correct"

            - test: "{{snippet}}"
              fail: "value of {{snippet}} is not correct"

        int:
            - test: "type({{snippet}})"
              fail: "type of {{snippet}} is not int. Please make sure it is int and not np.int64, etc. You can cast your value into an int using int()"

            - test: "{{snippet}}"
              fail: "value of {{snippet}} is not correct"

The outermost level in the YAML file (the example shows an entry for python3) specifies which kernel the configuration applies to. autotests.yml can have separate sections for multiple kernels / languages. The autotests.yml file uses Jinja templates to specify snippets of code that will be executed/inserted into Jupyter notebooks in the process of generating the assignment. You should familiarize yourself with the basics of Jinja templates before proceeding. For each kernel, there are a few configuration settings possible:

dispatch: When you write ### AUTOTEST foo(3), nbgrader needs to know how to test foo(3). It does so by executing foo(3), then checking its type, and then running tests corresponding to that type in the autotests.yml file. Specifically, when generating an assignment, nbgrader substitutes the {{snippet}} template variable with the expression foo(3), and then evaluates the dispatch code based on that. In this case, nbgrader runs type(foo(3)), which will return int, so nbgrader will know to test foo(3) using tests for integer variables.
templates: Once nbgrader determines the type of the expression foo(3), it will look for that type in the list of templates for the kernel. In this case, it will find the int type in the list (it will use the default if the type is not found). Each type will have associated with it a list of test/fail template pairs, which tell nbgrader what tests to run and what messages to print in the event of a failure. Once again, {{snippet}} will be replaced by the foo(3) expression. In autotests.yml above, the int type has two tests: one that checks type of the expression, and one that checks its value. In this case, the student notebook will have two tests: one that checks the value of type(foo(3)), and one that checks the value of foo(3).
normalize: For each test code expression (for example, type(foo(3)) as mentioned previously), nbgrader will execute code using the corresponding Jupyter kernel, which will respond with a result in the form of a string. So nbgrader now knows that if it runs type(foo(3)) at this point in the notebook, and converts the output to a string (i.e., normalizes it), it should obtain "int". However, nbgrader does not know how to convert output to a string; that depends on the kernel! So the normalize code template tells nbgrader how to convert an expression to a string. In the autotests.yml example above, the normalize template suggests that nbgrader should try to compare str(type(foo(3))) to "int".
check: This is the code template that will be inserted into the student notebook to run each test. The template has three variables. {{snippet}} is the normalized test code. The {{value}} is the evaluated version of that test code, based on the source notebook. The {{message}} is text that will be printed in the event of a test failure. In the example above, the check code template tells nbgrader to insert an assert statement to run the test.
hash (optional): This is a code template that is responsible for hashing (i.e., obfuscating) the answers in the student notebok. The template has two variables. {{snippet}} represents the expression that will be hashed, and {{salt}} is used for nbgrader to insert a salt prior to hashing. The salt helps avoid students being able to identify hashes from common question types. For example, a true/false question has only two possible answers; without a salt, students would be able to recognize the hashes of True and False in their notebooks. By adding a salt, nbgrader makes the hashed version of the answer different for each question, preventing identifying answers based on their hashes.
setup (optional): This is a code template that will be run at the beginning of all test cells containing ### AUTOTEST or ### HASHED AUTOTEST directives. It is often used to import special packages that only the test code requires. In the example above, the setup code is used to import the sha1 function from hashlib, which is necessary for hashed test generation.
success (optional): This is a code template that will be added to the end of all test cells containing ### AUTOTEST or ### HASHED AUTOTEST directives. In the generated student version of the notebook, this code will run if all the tests pass. In the example autotests.yml file above, the success code is used to run print('Success!'), i.e., simply print a message to indicate that all tests in the cell passed.

Note

For assignments with ### AUTOTEST and ### HASHED AUTOTEST directives, it is often handy to have an editable copy of the assignment with solutions and test code inserted. You can use nbgrader generate_assignment --source_with_tests to generate this version of an assignment, which will appear in the source_with_tests/ folder in the course repository.

Warning

The default autotests.yml test templates file included with the repository has tests for many common data types (int, dict, list, float, etc). It also has a default test template that it will try to apply to any types that do not have specified tests. If you want to automatically generate your own tests for custom types, you will need to implement those test templates in autotests.yml. That being said, custom object types often have standard Python types as class attributes. Sometimes an easier option is to use nbgrader to test these attributes automatically instead. For example, if obj is a complicated type with no specific test template available, but obj has an int attribute x, you could consider testing that attribute directly, e.g., ### AUTOTEST obj.x.

Warning

The InstantiateTests preprocessor in nbgrader is responsible for generating test code from ### AUTOTEST directives and the autotests.yml file. It has some configuration parameters not yet mentioned here. The most important of these is the InstantiateTests.sanitizers dictionary, which tells nbgrader how to clean up the string output from each kind of Jupyter kernel before using it in the process of generating tests. We have implemented sanitizers for popular kernels in nbgrader already, but you might need to add your own.