Transform Data

Amazon SageMaker Data Wrangler provides numerous ML data transforms to streamline cleaning, transforming, and featurizing your data. When you add a transform, it adds a step to the data flow. Each transform you add modifies your dataset and produces a new dataframe. All subsequent transforms apply to the resulting dataframe.

Data Wrangler includes built-in transforms, which you can use to transform columns without any code. You can also add custom transformations using PySpark, Python (User-Defined Function), pandas, and PySpark SQL. Some transforms operate in place, while others create a new output column in your dataset.

You can apply transforms to multiple columns at once. For example, you can delete multiple columns in a single step.

You can apply the Process numeric and Handle missing transforms only to a single column.

Use this page to learn more about these built-in and custom transforms.

Transform UI

Most of the built-in transforms are located in the Prepare tab of the Data Wrangler UI. You can access the join and concatenate transforms through the data flow view. Use the following table to preview these two views.

Join Datasets

You join dataframes directly in your data flow. When you join two datasets, the resulting joined dataset appears in your flow. The following join types are supported by Data Wrangler.

Use the following procedure to join two dataframes.

Concatenate Datasets

Balance Data

You can balance the data for datasets with an underrepresented category. Balancing a dataset can help you create better models for binary classification.

You can use the Balance data operation to balance your data using one of the following operators:

You can use all transforms for datasets containing both numeric and non-numeric features. SMOTE interpolates values by using neighboring samples. Data Wrangler uses the R-squared distance to determine the neighborhood to interpolate the additional samples. Data Wrangler only uses numeric features to calculate the distances between samples in the underrepresented group.

For two real samples in the underrepresented group, Data Wrangler interpolates the numeric features by using a weighted average. It randomly assigns weights to those samples in the range of [0, 1]. For numeric features, Data Wrangler interpolates samples using a weighted average of the samples. For samples A and B, Data Wrangler could randomly assign a weight of 0.7 to A and 0.3 to B. The interpolated sample has a value of 0.7A + 0.3B.

Data Wrangler interpolates non-numeric features by copying from either of the interpolated real samples. It copies the samples with a probability that it randomly assigns to each sample. For samples A and B, it can assign probabilities 0.8 to A and 0.2 to B. For the probabilities it assigned, it copies A 80% of the time.

Custom Transforms

The Custom Transforms group allows you to use Python (User-Defined Function), PySpark, pandas, or PySpark (SQL) to define custom transformations. For all three options, you use the variable df to access the dataframe to which you want to apply the transform. To apply your custom code to your dataframe, assign the dataframe with the transformations that you've made to the df variable. If you're not using Python (User-Defined Function), you don't need to include a return statement. Choose Preview to preview the result of the custom transform. Choose Add to add the custom transform to your list of Previous steps.

You can import the popular libraries with an import statement in the custom transform code block, such as the following:

df.rename(columns={"A column": "A_column", "B column": "B_column"})

If you include print statements in the code block, the result appears when you select Preview. You can resize the custom code transformer panel. Resizing the panel provides more space to write code. The following image shows the resizing of the panel.

The following sections provide additional context and examples for writing custom transform code.

Python (User-Defined Function)

The Python function gives you the ability to write custom transformations without needing to know Apache Spark or pandas. Data Wrangler is optimized to run your custom code quickly. You get similar performance using custom Python code and an Apache Spark plugin.

To use the Python (User-Defined Function) code block, you specify the following:

Using the pandas mode gives better performance. The Python mode makes it easier for you to write transformations by using pure Python functions.

The following video shows an example of how to use custom code to create a transformation. It uses the Titanic dataset to create a column with the person's salutation.

PySpark

The following example extracts date and time from a timestamp.

from pyspark.sql.functions import from_unixtime, to_date, date_format
df = df.withColumn('DATE_TIME', from_unixtime('TIMESTAMP'))
df = df.withColumn( 'EVENT_DATE', to_date('DATE_TIME')).withColumn(
'EVENT_TIME', date_format('DATE_TIME', 'HH:mm:ss'))

pandas

The following example provides an overview of the dataframe to which you are adding transforms.

df.info()

PySpark (SQL)

The following example creates a new dataframe with four columns: name, fare, pclass, survived.

SELECT name, fare, pclass, survived FROM df

If you don’t know how to use PySpark, you can use custom code snippets to help you get started.

Data Wrangler has a searchable collection of code snippets. You can use to code snippets to perform tasks such as dropping columns, grouping by columns, or modelling.

To use a code snippet, choose Search example snippets and specify a query in the search bar. The text you specify in the query doesn’t have to match the name of the code snippet exactly.

The following example shows a Drop duplicate rows code snippet that can delete rows with similar data in your dataset. You can find the code snippet by searching for one of the following:

The following snippet has comments to help you understand the changes that you need to make. For most snippets, you must specify the column names of your dataset in the code.

# Specify the subset of columns
# all rows having identical values in these columns will be dropped

subset = ["col1", "col2", "col3"]
df = df.dropDuplicates(subset)  

# to drop the full-duplicate rows run
# df = df.dropDuplicates()            
                        
        

To use a snippet, copy and paste its content into the Custom transform field. You can copy and paste multiple code snippets into the custom transform field.

Custom Formula

Use Custom formula to define a new column using a Spark SQL expression to query data in the current dataframe. The query must use the conventions of Spark SQL expressions.

df.rename(columns={"A column": "A_column", "B column": "B_column"})

You can use this transform to perform operations on columns, referencing the columns by name. For example, assuming the current dataframe contains columns named col_a and col_b, you can use the following operation to produce an Output column that is the product of these two columns with the following code:

col_a * col_b

Other common operations include the following, assuming a dataframe contains col_a and col_b columns:

For more information, see the Spark documentation on selecting data.

Reduce Dimensionality within a Dataset

Reduce the dimensionality in your data by using Principal Component Analysis (PCA). The dimensionality of your dataset corresponds to the number of features. When you use dimensionality reduction in Data Wrangler, you get a new set of features called components. Each component accounts for some variability in the data.

The first component accounts for the largest amount of variation in the data. The second component accounts for the second largest amount of variation in the data, and so on.

You can use dimensionality reduction to reduce the size of the data sets that you use to train models. Instead of using the features in your dataset, you can use the principal components instead.

To perform PCA, Data Wrangler creates axes for your data. An axis is an affine combination of columns in your dataset. The first principal component is the value on the axis that has the largest amount of variance. The second principal component is the value on the axis that has the second largest amount of variance. The nth principal component is the value on the axis that has the nth largest amount of variance.

You can configure the number of principal components that Data Wrangler returns. You can either specify the number of principal components directly or you can specify the variance threshold percentage. Each principal component explains an amount of variance in the data. For example, you might have a principal component with a value of 0.5. The component would explain 50% of the variation in the data. When you specify a variance threshold percentage, Data Wrangler returns the smallest number of components that meet the percentage that you specify.

The following are example principal components with the amount of variance that they explain in the data.

If you specify a variance threshold percentage of 94 or 95, Data Wrangler returns Component 1 and Component 2. If you specify a variance threshold percentage of 96, Data Wrangler returns all three principal components.

You can use the following procedure to run PCA on your dataset.

Encode Categorical

Categorical data is usually composed of a finite number of categories, where each category is represented with a string. For example, if you have a table of customer data, a column that indicates the country a person lives in is categorical. The categories would be Afghanistan, Albania, Algeria, and so on. Categorical data can be nominal or ordinal. Ordinal categories have an inherent order, and nominal categories do not. The highest degree obtained (High school, Bachelors, Masters, and so on) is an example of ordinal categories.

Encoding categorical data is the process of creating a numerical representation for categories. For example, if your categories are Dog and Cat, you may encode this information into two vectors, [1,0] to represent Dog, and [0,1] to represent Cat.

When you encode ordinal categories, you may need to translate the natural order of categories into your encoding. For example, you can represent the highest degree obtained with the following map: {"High school": 1, "Bachelors": 2, "Masters":3}.

Use categorical encoding to encode categorical data that is in string format into arrays of integers.

The Data Wrangler categorical encoders create encodings for all categories that exist in a column at the time the step is defined. If new categories have been added to a column when you start a Data Wrangler job to process your dataset at time t, and this column was the input for a Data Wrangler categorical encoding transform at time t-1, these new categories are considered missing in the Data Wrangler job. The option you select for Invalid handling strategy is applied to these missing values. Examples of when this can occur are:

In these situations, these new categories are considered missing values in the Data Wrangler job.

You can choose from and configure an ordinal and a one-hot encode. Use the following sections to learn more about these options.

Both transforms create a new column named Output column name. You specify the output format of this column with Output style:

Ordinal Encode

Select Ordinal encode to encode categories into an integer between 0 and the total number of categories in the Input column you select.

Invalid handing strategy: Select a method to handle invalid or missing values.

One-Hot Encode

Select One-hot encode for Transform to use one-hot encoding. Configure this transform using the following:

Similarity encode

Use similarity encoding when you have the following:

The similarity encoder creates embeddings for columns with categorical data. An embedding is a mapping of discrete objects, such as words, to vectors of real numbers. It encodes similar strings to vectors containing similar values. For example, it creates very similar encodings for "California" and "Calfornia".

Data Wrangler converts each category in your dataset into a set of tokens using a 3-gram tokenizer. It converts the tokens into an embedding using min-hash encoding.

The following example shows how the similarity encoder creates vectors from strings.

The similarity encodings that Data Wrangler creates:

For the preceding reasons, similarity encoding is more versatile than one-hot encoding.

To add the similarity encoding transform to your dataset, use the following procedure.

Featurize Text

Use the Featurize Text transform group to inspect string-typed columns and use text embedding to featurize these columns.

This feature group contains two features, Character statistics and Vectorize. Use the following sections to learn more about these transforms. For both options, the Input column must contain text data (string type).

Character Statistics

Use Character statistics to generate statistics for each row in a column containing text data.

This transform computes the following ratios and counts for each row, and creates a new column to report the result. The new column is named using the input column name as a prefix and a suffix that is specific to the ratio or count.

Vectorize

Text embedding involves mapping words or phrases from a vocabulary to vectors of real numbers. Use the Data Wrangler text embedding transform to tokenize and vectorize text data into term frequency–inverse document frequency (TF-IDF) vectors.

When TF-IDF is calculated for a column of text data, each word in each sentence is converted to a real number that represents its semantic importance. Higher numbers are associated with less frequent words, which tend to be more meaningful.

When you define a Vectorize transform step, Data Wrangler uses the data in your dataset to define the count vectorizer and TF-IDF methods . Running a Data Wrangler job uses these same methods.

You configure this transform using the following:

Transform Time Series

In Data Wrangler, you can transform time series data. The values in a time series dataset are indexed to specific time. For example, a dataset that shows the number of customers in a store for each hour in a day is a time series dataset. The following table shows an example of a time series dataset.

For the preceding table, the Number of Customers column contains the time series data. The time series data is indexed on the hourly data in the Time (hour) column.

You might need to perform a series of transformations on your data to get it in a format that you can use for your analysis. Use the Time series transform group to transform your time series data. For more information about the transformations that you can perform, see the following sections.

Group by a Time Series

You can use the group by operation to group time series data for specific values in a column.

For example, you have the following table that tracks the average daily electricity usage in a household.

If you choose to group by ID, you get the following table.

Each entry in the time series sequence is ordered by the corresponding timestamp. The first element of the sequence corresponds to the first timestamp of the series. For household_0, 30 is the first value of the Electricity Usage Series. The value of 30 corresponds to the first timestamp of 1/1/2020.

You can include the starting timestamp and ending timestamp. The following table shows how that information appears.

You can use the following procedure to group by a time series column.

Resample Time Series Data

Time series data usually has observations that aren't taken at regular intervals. For example, a dataset could have some observations that are recorded hourly and other observations that are recorded every two hours.

Many analyses, such as forecasting algorithms, require the observations to be taken at regular intervals. Resampling gives you the ability to establish regular intervals for the observations in your dataset.

You can either upsample or downsample a time series. Downsampling increases the interval between observations in the dataset. For example, if you downsample observations that are taken either every hour or every two hours, each observation in your dataset is taken every two hours. The hourly observations are aggregated into a single value using an aggregation method such as the mean or median.

Upsampling reduces the interval between observations in the dataset. For example, if you upsample observations that are taken every two hours into hourly observations, you can use an interpolation method to infer hourly observations from the ones that have been taken every two hours. For information on interpolation methods, see pandas.DataFrame.interpolate.

You can resample both numeric and non-numeric data.

Use the Resample operation to resample your time series data. If you have multiple time series in your dataset, Data Wrangler standardizes the time interval for each time series.

The following table shows an example of downsampling time series data by using the mean as the aggregation method. The data is downsampled from every two hours to every hour.

You can use the following procedure to resample time series data.

Handle Missing Time Series Data

If you have missing values in your dataset, you can do one of the following:

Imputing a missing value involves replacing the data by either specifying a value or by using an inferential method. The following are the methods that you can use for imputation:

Some of the imputation methods might not be able to impute of all the missing value in your dataset. For example, a Forward fill can't impute a missing value that appears at the beginning of the time series. You can impute the values by using either a forward fill or a backward fill.

You can either impute missing values within a cell or within a column.

The following example shows how values are imputed within a cell.

The following example shows how values are imputed within a column.

You can use the following procedure to handle missing values.

Validate the Timestamp of Your Time Series Data

You might have time stamp data that is invalid. You can use the Validate time stamp function to determine whether the timestamps in your dataset are valid. Your timestamp can be invalid for one or more of the following reasons:

If you have invalid timestamps in your dataset, you can't perform your analysis successfully. You can use Data Wrangler to identify invalid timestamps and understand where you need to clean your data.

The time series validation works in one of the two ways:

You can configure Data Wrangler to do one of the following if it encounters missing values in your dataset:

You can validate the timestamps on columns that either have the timestamp type or the string type. If the column has the string type, Data Wrangler converts the type of the column to timestamp and performs the validation.

You can use the following procedure to validate the timestamps in your dataset.

Standardizing the Length of the Time Series

If you have time series data stored as arrays, you can standardize each time series to the same length. Standardizing the length of the time series array might make it easier for you to perform your analysis on the data.

You can standardize your time series for data transformations that require the length of your data to be fixed.

Many ML algorithms require you to flatten your time series data before you use them. Flattening time series data is separating each value of the time series into its own column in a dataset. The number of columns in a dataset can't change, so the lengths of the time series need to be standardized between you flatten each array into a set of features.

Each time series is set to the length that you specify as a quantile or percentile of the time series set. For example, you can have three sequences that have the following lengths:

You can set the length of all of the sequences as the length of the sequence that has the 50th percentile length.

Time series arrays that are shorter than the length you've specified have missing values added. The following is an example format of standardizing the time series to a longer length: [2, 4, 5, NaN, NaN, NaN].

You can use different approaches to handle the missing values. For information on those approaches, see Handle Missing Time Series Data.

The time series arrays that are longer than the length that you specify are truncated.

You can use the following procedure to standardize the length of the time series.

Extract Features from Your Time Series Data

If you're running a classification or a regression algorithm on your time series data, we recommend extracting features from the time series before running the algorithm. Extracting features might improve the performance of your algorithm.

Use the following options to choose how you want to extract features from your data:

Use the following the procedure to extract features from your time series data.

Use Lagged Features from Your Time Series Data

For many use cases, the best way to predict the future behavior of your time series is to use its most recent behavior.

The most common uses of lagged features are the following:

Create a Datetime Range In Your Time Series

You might have time series data that don't have timestamps. If you know that the observations were taken at regular intervals, you can generate timestamps for the time series in a separate column. To generate timestamps, you specify the value for the start timestamp and the frequency of the timestamps.

For example, you might have the following time series data for the number of customers at a restaurant.

If you know that the restaurant opened at 5:00 PM and that the observations are taken hourly, you can add a timestamp column that corresponds to the time series data. You can see the timestamp column in the following table.

Use the following procedure to add a datetime range to your data.

Use a Rolling Window In Your Time Series

You can extract features over a time period. For example, for time, t, and a time window length of 3, and for the row that indicates the tth timestamp, we append the features that are extracted from the time series at times t - 3, t -2, and t - 1. For information on extracting features, see Extract Features from Your Time Series Data.

You can use the following procedure to extract features over a time period.

Featurize Datetime

Use Featurize date/time to create a vector embedding representing a datetime field. To use this transform, your datetime data must be in one of the following formats:

You can choose to Infer datetime format and provide a Datetime format. If you provide a datetime format, you must use the codes described in the Python documentation. The options you select for these two configurations have implications for the speed of the operation and the final results.

When you use this transform, you specify an Input column which contains datetime data in one of the formats listed above. The transform creates an output column named Output column name. The format of the output column depends on your configuration using the following:

Additionally, you must choose an Embedding mode. For linear models and deep networks, we recommend choosing cyclic. For tree-based algorithms, we recommend choosing ordinal.

Format String

The Format string transforms contain standard string formatting operations. For example, you can use these operations to remove special characters, normalize string lengths, and update string casing.

This feature group contains the following transforms. All transforms return copies of the strings in the Input column and add the result to a new, output column.

Handle Outliers

Machine learning models are sensitive to the distribution and range of your feature values. Outliers, or rare values, can negatively impact model accuracy and lead to longer training times. Use this feature group to detect and update outliers in your dataset.

When you define a Handle outliers transform step, the statistics used to detect outliers are generated on the data available in Data Wrangler when defining this step. These same statistics are used when running a Data Wrangler job.

Use the following sections to learn more about the transforms this group contains. You specify an Output name and each of these transforms produces an output column with the resulting data.

Robust standard deviation numeric outliers

This transform detects and fixes outliers in numeric features using statistics that are robust to outliers.

You must define an Upper quantile and a Lower quantile for the statistics used to calculate outliers. You must also specify the number of Standard deviations from which a value must vary from the mean to be considered an outlier. For example, if you specify 3 for Standard deviations, a value must fall more than 3 standard deviations from the mean to be considered an outlier.

The Fix method is the method used to handle outliers when they are detected. You can choose from the following:

Standard Deviation Numeric Outliers

This transform detects and fixes outliers in numeric features using the mean and standard deviation.

You specify the number of Standard deviations a value must vary from the mean to be considered an outlier. For example, if you specify 3 for Standard deviations, a value must fall more than 3 standard deviations from the mean to be considered an outlier.

The Fix method is the method used to handle outliers when they are detected. You can choose from the following:

Quantile Numeric Outliers

Use this transform to detect and fix outliers in numeric features using quantiles. You can define an Upper quantile and a Lower quantile. All values that fall above the upper quantile or below the lower quantile are considered outliers.

The Fix method is the method used to handle outliers when they are detected. You can choose from the following:

Min-Max Numeric Outliers

This transform detects and fixes outliers in numeric features using upper and lower thresholds. Use this method if you know threshold values that demark outliers.

You specify a Upper threshold and a Lower threshold, and if values fall above or below those thresholds respectively, they are considered outliers.

The Fix method is the method used to handle outliers when they are detected. You can choose from the following:

Replace Rare

When you use the Replace rare transform, you specify a threshold and Data Wrangler finds all values that meet that threshold and replaces them with a string that you specify. For example, you may want to use this transform to categorize all outliers in a column into an "Others" category.

Handle Missing Values

Missing values are a common occurrence in machine learning datasets. In some situations, it is appropriate to impute missing data with a calculated value, such as an average or categorically common value. You can process missing values using the Handle missing values transform group. This group contains the following transforms.

Fill Missing

Use the Fill missing transform to replace missing values with a Fill value you define.

Impute Missing

Use the Impute missing transform to create a new column that contains imputed values where missing values were found in input categorical and numerical data. The configuration depends on your data type.

For numeric data, choose an imputing strategy, the strategy used to determine the new value to impute. You can choose to impute the mean or the median over the values that are present in your dataset. Data Wrangler uses the value that it computes to impute the missing values.

For categorical data, Data Wrangler imputes missing values using the most frequent value in the column. To impute a custom string, use the Fill missing transform instead.

Add Indicator for Missing

Use the Add indicator for missing transform to create a new indicator column, which contains a Boolean "false" if a row contains a value, and "true" if a row contains a missing value.

Drop Missing

Use the Drop missing option to drop rows that contain missing values from the Input column.

Manage Columns

You can use the following transforms to quickly update and manage columns in your dataset:

Manage Rows

Use this transform group to quickly perform sort and shuffle operations on rows. This group contains the following:

Manage Vectors

Use this transform group to combine or flatten vector columns. This group contains the following transforms.

Process Numeric

Use the Process Numeric feature group to process numeric data. Each scalar in this group is defined using the Spark library. The following scalars are supported:

Sampling

After you've imported your data, you can use the Sampling transformer to take one or more samples of it. When you use the sampling transformer, Data Wrangler samples your original dataset.

You can choose one of the following sample methods:

You can stratify a randomized sample to make sure that it represents the original distribution of the dataset.

You might be performing data preparation for multiple use cases. For each use case, you can take a different sample and apply a different set of transformations.

The following procedure describes the process of creating a random sample.

The following procedure describes the process of creating a stratified sample.

Search and Edit

Use this section to search for and edit specific patterns within strings. For example, you can find and update strings within sentences or documents, split strings by delimiters, and find occurrences of specific strings.

The following transforms are supported under Search and edit. All transforms return copies of the strings in the Input column and add the result to a new output column.

Split data

Use the Split data transform to split your dataset into two or three datasets. For example, you can split your dataset into a dataset used to train your model and a dataset used to test it. You can determine the proportion of the dataset that goes into each split. For example, if you’re splitting one dataset into two datasets, the training dataset can have 80% of the data while the testing dataset has 20%.

Splitting your data into three datasets gives you the ability to create training, validation, and test datasets. You can see how well the model performs on the test dataset by dropping the target column.

Your use case determines how much of the original dataset each of your datasets get and the method you use to split the data. For example, you might want to use a stratified split to make sure that the distribution of the observations in the target column are the same across datasets. You can use the following split transforms:

After you split the data, you can apply additional transformations to each dataset. For most use cases, they aren't necessary.

Data Wrangler calculates the proportions of the splits for performance. You can choose an error threshold to set the accuracy of the splits. Lower error thresholds more accurately reflect the proportions that you specify for the splits. If you set a higher error threshold, you get better performance, but lower accuracy.

For perfectly split data, set the error threshold to 0. You can specify a threshold between 0 and 1 for better performance. If you specify a value greater than 1, Data Wrangler interprets that value as 1.

If you have 10000 rows in your dataset and you specify an 80/20 split with an error of 0.001, you would get observations approximating one of the following results:

The number of observations for the testing set in the preceding example is in the interval between 8010 and 7990.

By default, Data Wrangler uses a random seed to make the splits reproducible. You can specify a different value for the seed to create a different reproducible split.

Parse Value as Type

Use this transform to cast a column to a new type. The supported Data Wrangler data types are:

Validate String

Use the Validate string transforms to create a new column that indicates that a row of text data meets a specified condition. For example, you can use a Validate string transform to verify that a string only contains lowercase characters. The following transforms are supported under Validate string.

The following transforms are included in this transform group. If a transform outputs a Boolean value, True is represented with a 1 and False is represented with a 0.

Unnest JSON Data

If you have a .csv file, you might have values in your dataset that are JSON strings. Similarly, you might have nested data in columns of either a Parquet file or a JSON document.

Use the Flatten structured operator to separate the first level keys into separate columns. A first level key is a key that isn't nested within a value.

For example, you might have a dataset that has a person column with demographic information on each person stored as JSON strings. A JSON string might look like the following.

 "{"seq": 1,"name": {"first": "Nathaniel","last": "Ferguson"},"age": 59,"city": "Posbotno","state": "WV"}"
            

The Flatten structured operator converts the following first level keys into additional columns in your dataset:

Data Wrangler puts the values of the keys as values under the columns. The following shows the column names and values of the JSON.

seq, name,                                    age, city, state
1, {"first": "Nathaniel","last": "Ferguson"}, 59, Posbotno, WV
            

For each value in your dataset containing JSON, the Flatten structured operator creates columns for the first-level keys. To create columns for nested keys, call the operator again. For the preceding example, calling the operator creates the columns:

The following example shows the dataset that results from calling the operation again.

seq, name,                                    age, city, state, name_first, name_last
1, {"first": "Nathaniel","last": "Ferguson"}, 59, Posbotno, WV, Nathaniel, Ferguson
            

Choose Keys to flatten on to specify the first-level keys that want to extract as separate columns. If you don't specify any keys, Data Wrangler extracts all the keys by default.

Explode Array

Use Explode array to expand the values of the array into separate output rows. For example, the operation can take each value in the array, [[1, 2, 3,], [4, 5, 6], [7, 8, 9]] and create a new column with the following rows:

                [1, 2, 3]
                [4, 5, 6]
                [7, 8, 9]
            

Data Wrangler names the new column, input_column_name_flatten.

You can call the Explode array operation multiple times to get the nested values of the array into separate output columns. The following example shows the result of calling the operation multiple times on a dataset with a nested array.

Transform Image Data

Use Data Wrangler to import and transform the images that you're using for your machine learning (ML) pipelines. After you've prepared your image data, you can export it from your Data Wrangler flow to your ML pipeline.

You can use the information provided here to familiarize yourself with importing and transforming image data in Data Wrangler. Data Wrangler uses OpenCV to import images. For more information about supported image formats, see Image file reading and writing.

After you've familiarized yourself with the concepts of transforming your image data, go through the following tutorial, Prepare image data with Amazon SageMaker Data Wrangler.

The following industries and use cases are examples where applying machine learning to transformed image data can be useful:

When you work with image data in Data Wrangler, you go through the following process:

Use the following procedure to import your image data.

Data Wrangler uses the open-source imgaug library for its built-in image transformations. You can use the following built-in transformations:

Use the following procedure to transform your images without writing code.

In addition to using the transformations that Data Wrangler provides, you can also use your own custom code snippets. For more information about using custom code snippets, see Custom Transforms. You can import the OpenCV and imgaug libraries within your code snippets and use the transforms associated with them. The following is an example of a code snippet that detects edges within the images.

# A table with your image data is stored in the `df` variable
import cv2
import numpy as np
from pyspark.sql.functions import column

from sagemaker_dataprep.compute.operators.transforms.image.constants import DEFAULT_IMAGE_COLUMN, IMAGE_COLUMN_TYPE
from sagemaker_dataprep.compute.operators.transforms.image.decorators import BasicImageOperationDecorator, PandasUDFOperationDecorator


@BasicImageOperationDecorator
def my_transform(image: np.ndarray) -> np.ndarray:
  # To use the code snippet on your image data, modify the following lines within the function
    HYST_THRLD_1, HYST_THRLD_2 = 100, 200
    edges = cv2.Canny(image,HYST_THRLD_1,HYST_THRLD_2)
    return edges
    

@PandasUDFOperationDecorator(IMAGE_COLUMN_TYPE)
def custom_image_udf(image_row):
    return my_transform(image_row)
    

df = df.withColumn(DEFAULT_IMAGE_COLUMN, custom_image_udf(column(DEFAULT_IMAGE_COLUMN)))
        

When apply transformations in your Data Wrangler flow, Data Wrangler only applies them to a sample of the images in your dataset. To optimize your experience with the application, Data Wrangler doesn't apply the transforms to all of your images.

To apply the transformations to all of your images, export your Data Wrangler flow to an Amazon S3 location. You can use the images that you've exported in your training or inference pipelines. Use a destination node or a Jupyter Notebook to export your data. You can access either method for exporting your data from the Data Wrangler flow. For information about using these methods, see Export to Amazon S3.

Filter data

Use Data Wrangler to filter the data in your columns. When you filter the data in a column, you specify the following fields:

You can filter on the following conditions:

For a column that has the categories, male and female, you can filter out all the male values. You could also filter for all the female values. Because there are only male and female values in the column, the filter returns a column that only has female values.

You can also add multiple filters. The filters can be applied across multiple columns or the same column. For example, if you're creating a column that only has values within a certain range, you add two different filters. One filter specifies that the column must have values greater than the value that you provide. The other filter specifies that the column must have values less than the value that you provide.

Use the following procedure to add the filter transform to your data.

Map Columns for Amazon Personalize

Data Wrangler integrates with Amazon Personalize, a fully managed machine learning service that generates item recommendations and user segments. You can use the Map columns for Amazon Personalize transform to get your data into a format that Amazon Personalize can interpret. For more information about the transforms specific to Amazon Personalize, see Importing data using Amazon SageMaker Data Wrangler. For more information about Amazon Personalize see What is Amazon Personalize?