Tag: ML

Jupyter Notebook: Getting Started and Installation

Jupyter Notebook provides a simple way to run code in Python, R, Scala and more. While it’s mainly used in research related fields, Jupyter can be applied to a wide variety of applications.

Jupyter is especially useful for two groups in particular. Data scientists and machine learning experts benefit from modular execution. They can load a large data set once and try many different experiments/code changes; saving an enormous amount of time in the process. Academics make up the other group. Documenting work as you go could not be easier. Professors can auto grade students work using tools like Vocareum, built to work with Jupyter.

Above is a screenshot directly from a notebook. It runs in a web interface and is quite easy to use. Markdown syntax can be added to create documentation blocks — much more expressive than the brief inline comments we programmers often get used to.

Installing Jupyter Notebook

Some prefer the native installation while others like to keep everything in a self-contained Docker container. I will outline both methods — choose the one that works best for you.

Native

Jupyter Notebook is easy to install with pip. I assume you already have Python and pip installed. If that’s the case, simply run

python3 -m pip install jupyter #for Python3
python -m pip install jupyter  #for Python2

Congratulations, that’s it! To run Jupyter, simply open up a new terminal in the directory you want the notebooks to be saved. Then type:

jupyter notebook

Your default browser should automatically open to the Jupyter instance.

Docker

This guide assumes Docker is already installed. If you’re unfamiliar with Docker, please check out their guide.

Using Jupyter with Docker is easy, a container is already maintained. Simply run the container with the following command:

docker run -it -v /path/to/jupyter/directory:/work --net=host --rm jupyter/all-spark-notebook

The -v flag is used to share a local directory with Docker. /path/to/jupyter/directory should point to a local directory where you want the notebooks to be saved. When inside the Jupyter instance, be sure to save everything inside the /work directory, or it will not be saved.

Once the Docker container is launched, a unique URL will be printed to the console. Copy and paste that into a web browser and you’re good to go! You may notice the example below says 127.0.0.1 or a seemingly random string of numbers. If this is the case for you, be sure to substitute 127.0.0.1. For instance, open a web browser and go directly to the URL http://127.0.0.1:8888/?token=f19d2097c0e3455e3589c985b182b93a6c9f022612d9a2cc. Your token will be different!

Your first Jupyter project

With Jupyter successfully installed, your screen should be similar to the one pictured below.

Once Jupyter Notebook is installed and launched, we can create our first actual notebook. If you installed with Docker, be sure to click on the work directory first. Then click the New dropdown, then select Python 3 (or a different language if you prefer).

All that remains is filling it with content. Each content block is a cell. There are two main types of cells: code and Markdown.

Let’s create a new code cell.

Run the code in a cell by clicking the play button with the cell selected, or by hitting ctrl + enter on the keyboard.

Now, let’s demonstrate one of the main benefits of Jupyter. Say we need to load data, which takes a lot of time. If this was a standard Python script, the data would have to be loaded during each subsequent run. This is not the case with Jupyter. We can simply put the data loading code into its own cell.

Saving & Checkpoints

To save your work, simply click the Save & Checkpoint button.

As the name suggests, manually saving also creates a checkpoint. Checkpoints are a form of basic version control — you can easily roll back to any checkpoint later on. Work in a notebook will also be periodically auto-saved, but checkpoints must be created manually.

To share your work with someone else, simply send them the .ipynb file. They can launch the file using their own Jupyter installation and pick up right where you left off. While you can use Jupyter with a version control system (like Git of Mercurial) by checking in the .ipynb files, it’s not easy to see the individual code changes later. This is my single biggest complaint when using Jupyter. If anyone has found a solution, I would love to hear it!

Posted on by Edward Atter 2

A Hands-On Introduction to Machine Learning

First, let me begin by setting some expectations. This is not a guide for the hardcore ML researchers out there. This is meant to be a practical introduction to machine learning that any computer scientist can follow, without much prior knowledge of the ML domain. I feel that there are many guides that focus solely on the academics of ML, but neglect to mention how simple it is to apply towards real life applications. Even naive approaches are often surprisingly effective.

To get started, we need to install scikit and other dependencies.

pip install numpy scipy scikit-learn

For simple implementations like we’ll see today, most of the challenge revolves around data prep. Our first step is to download the JSON training data from Kaggle. If you don’t already have an account, you’ll need to create one now. Once the account is created, visit the “What’s Cooking?” competition page and select the data tab to access the downloads.

Next, we need to format the data. A CSV format will be used, with each column representing a different ingredient and each row a single recipe. In ML lingo, the ingredients are features, the data we use as a basis for the predictions. Labels, the answer to each recipe, will be generated similarly. There’s nothing too novel with the code here, just some data wrangling.

Create a file called parser.py with the following code:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
#!/usr/bin/env python3

# Enable python2 compatability
from __future__ import print_function

import json

def main():
    # Define input/output file names
    train_file = "train.json"
    test_file = "test.json"
    train_file_out = "train.csv"
    test_file_out = "test.csv"
    train_file_out_labels = "train-labels.txt"
    json_data = None
    with open(train_file, 'r') as f:
        json_data = f.read()
    train_obj = json.loads(json_data)

    # Empty arrays to hold information
    labels_train = []
    labels_test = []
    # ingredients is defined as a set to prevent duplicates
    ingredients = set()

    # Generate corresponding labels and simultaneously make
    # exhaustive set of all posible cuisines (labels)
    with open(train_file_out_labels, 'w') as f:
        for recipe in train_obj:
            label = recipe["cuisine"]
            print(label, file=f)
            labels_train.append(label)
            for ingredient in recipe["ingredients"]:
                ingredients.add(ingredient)

    with open(test_file, 'r') as f:
        json_data = f.read()
    test_obj = json.loads(json_data)

    # The test file may introduce ingredients not included in training set
    # This ensures they're included
    for recipe in test_obj:
        for ingredient in recipe["ingredients"]:
            ingredients.add(ingredient)

    # Transform set to list to ensure iteration order is constant
    ingredients_list = list(ingredients)

    # Generate the CSV files
    generate_csv_for_each_recipe(ingredients_list, train_obj, train_file_out)
    generate_csv_for_each_recipe(ingredients_list, test_obj, test_file_out)

def generate_csv_for_each_recipe(ingredients_list, json_obj, output_file):
    """
    Creates an output csv file with each ingredient being a column
    and each recipe a row. 1 will represent the recipe contains the
    given ingredient if the recipe includes that incredient, else 0

    ingredients_list -- the full list of ingredients (without duplicates)
    json_obj -- the json object with recipes returned by json.loads
    output_file -- the name of the generated CSV file
    """
    # Loop thru each recipe
    with open(output_file, 'w') as f:
        for recipe in json_obj:
            rl = set()
            first = True
            s = ""
            for ingredient in recipe["ingredients"]:
                rl.add(ingredient)
            # This builds the csv row of ingredients for current recipe
            # Add 1 for ingredient if included in recipe; else 0
            for j in ingredients_list:
                # Don't prepend "," for first item
                if first != True:
                    s += ","  
                else:
                    first = False

                # Add 1 or 0 as explained above
                if j not in rl:
                    s += "0"
                else:
                    s += "1"
            print(s, file=f)

if __name__ == "__main__":
    main()

Create a new python script called train.py. Add some simple imports and variables that will prove useful later.

import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split

filename = "train.csv"
label_file = "train-labels.txt"
test_file = "test.csv"
prediction_output = "predictions.txt"

Load the data files generated previously with the parser script.

# Load the training features into a np array
features = np.loadtxt(filename, delimiter=',', dtype=np.uint8)
# Load the labels
with open(label_file) as f:
    labels = f.readlines()
# Strip any new line characters or extra spaces
labels = [x.strip() for x in labels]
# Convert to np array
labels = np.asarray(labels)

The next step is to split the training data into a training set and a testing set. This will allow us to estimate how well the classifier does on the “real” test data. Think about it, the testing data from Kaggle does not include the answers (labels). To have an easy way to see how well we’re doing, it’s necessary to split the data we do have answers for. It is not okay to test with the same features used in training — the accuracy will be artificially high. Scikit includes a handy feature to split the data for us.

# Split data up into training and test data
X_train, X_test, y_train, y_test = train_test_split(features, labels)

Instantiate and train the classifier with the split data set created from the last step. Selecting the best classifier is beyond the scope of this article, so we’ll just use the Logistic Regression classifier in this example, which performs pretty well. Scikit has an example testing different classifiers, if you want to explore.

print("Starting training...")
clf = LogisticRegression()
clf.fit(X_train, y_train)
score = clf.score(X_test, y_test)
print("Model has accuracy of " + str(score * 100) + "%")

Let’s use the same classifier to make predictions over the Kaggle test set, the one we don’t know the answers to. We’ll format this as simply one prediction per line.

print("Predicting over the Kaggle test set")
test_data = np.loadtxt(test_file, delimiter=',', dtype=np.uint8)
predictions = clf.predict(test_data)

with open(prediction_output, "w") as f:
    for prediction in predictions:
        print(prediction, file=f)

The last script we’ll write takes the predictions created in the last step and formats it in the specific way Kaggle expects. This will let us see how we performed against other solutions to the What’s Cooking Challenge.

Create a new script called kaggle.py. As usual, import the required modules and define a few helpful variables.

#!/usr/bin/env python3

from __future__ import print_function
import json

predict_file = "predictions.txt"
test_file = "test.json"
output_file = "kaggle.csv"

Read the prediction file

with open(predict_file) as f:
    labels = f.readlines()
labels = [x.strip() for x in labels]

Open the Kaggle test file and parse as JSON

with open(test_file, 'r') as f:
    json_data = f.read()
obj = json.loads(json_data)

Open the output file for writing and format as the Kaggle spec requires.

with open(output_file, "w") as out:
    # Print CSV headers
    print("id,cuisine", file=out)

    i = 0
    # Iterate through each recipe in the test file
    # Follow the spec in CSV format,
    # the recipe id followed by the cuisine prediction
    for recipe in obj:
        idx = recipe["id"]
        ingredient = labels[i]
        print(str(idx) + "," + ingredient, file=out)
        i += 1

To see how well you did, submit the generated kaggle.csv file to the Kaggle competition.

The complete code is available on GitHub.

For such a naive solution, we did pretty well here — successfully classifying over 77% of the recipes. There is, of course, room for improvement. It’s unlikely you’ll top the leader board with ready-made classifiers, but it’s close enough for many real-life problems and an excellent start to a future in machine learning.

Posted on by Edward Atter 0