DeepEcho Synthetic Data Generator

0.8.1 · active · verified Fri Apr 17

DeepEcho is a Python library within the SDV ecosystem for generating sequential synthetic data from real-world datasets using Generative Adversarial Networks (GANs). It's designed for data that has a temporal or sequential component, such as time series or event logs. Currently at version 0.8.1, the library maintains an active development pace with frequent updates and bug fixes, often aligning with the broader SDV ecosystem's release cycle.

Common errors

```
ModuleNotFoundError: No module named 'deepecho.synthesizers'
```
cause Attempting to import the `DeepEchoSynthesizer` class, which was removed in DeepEcho v0.7.0.

fix Change your import statement to `from deepecho.models import DeepEcho`.
```
TypeError: DeepEcho.__init__() missing 1 required positional argument: 'metadata'
```
cause As of v0.7.0, `DeepEcho` requires a `metadata` object during initialization to understand the data structure, especially for sequential data.

fix Obtain a `sdv.metadata.Metadata` object for your data (e.g., using `sdv.metadata.SingleTableMetadata.load_from_dataframe()`) and pass it to the DeepEcho constructor: `model = DeepEcho(metadata=my_metadata)`.
```
ValueError: The input data must contain the primary key column 'column_name' defined in the metadata.
```
cause The dataframe provided to `DeepEcho.fit()` is missing a column that is declared as a primary key in the `metadata` object, or the column name in the data does not match the metadata.

fix Verify that your input dataframe contains all columns specified as keys in the `metadata`. Ensure column names in the dataframe exactly match those in the metadata.

Warnings

breaking The main class `DeepEchoSynthesizer` was removed in v0.7.0. It was replaced by `DeepEcho` in `deepecho.models`.
Fix: Update your import statement from `from deepecho.synthesizers import DeepEchoSynthesizer` to `from deepecho.models import DeepEcho`.
breaking DeepEcho models no longer handle data preprocessing internally as of v0.7.0. All preprocessing, data loading, and metadata generation must now be handled externally, primarily using the `sdv` library.
Fix: Ensure you are using `sdv` to load your data and generate a `sdv.metadata.Metadata` object. Pass this metadata object explicitly to the `DeepEcho` model constructor: `model = DeepEcho(metadata=my_metadata)`.
gotcha When working with sequential data, it is critical that the `metadata` object correctly defines the primary key, parent key (if nested), and sequence key for each table. Incorrect metadata can lead to errors during fitting or generation of nonsensical synthetic data.
Fix: Thoroughly review and, if necessary, manually adjust the `sdv` metadata to accurately reflect your sequential data's structure. Use `metadata.set_table_meta()` to specify keys if auto-detection is insufficient.

Install

pip install deepecho Install DeepEcho

Imports

DeepEcho
wrong:
```
from deepecho.synthesizers import DeepEchoSynthesizer
```
correct:
```
from deepecho.models import DeepEcho
```
As of DeepEcho v0.7.0, `DeepEchoSynthesizer` was removed. The main entry point is now `DeepEcho` in `deepecho.models`.

Quickstart

This quickstart demonstrates how to use DeepEcho to generate synthetic sequential data. It uses `sdv.datasets.demo.get_sequential_demo` to obtain sample sequential data and its corresponding metadata, which is essential for DeepEcho to correctly model the data's structure. The `DeepEcho` model is then initialized with the metadata, fitted to the real data, and finally used to sample new synthetic sequences.

from deepecho.models import DeepEcho
from sdv.datasets.demo import get_sequential_demo

# Get demo data for sequential modeling from SDV
metadata, data = get_sequential_demo()

# Initialize and fit the DeepEcho model
# Metadata is crucial for DeepEcho to understand the sequential structure
model = DeepEcho(metadata=metadata)
model.fit(data)

# Generate 100 rows of synthetic sequential data
synthetic_data = model.sample(num_rows=100)

print(synthetic_data.head())

view raw JSON →