spacekit - single visit mosaic alignment modeling

How-To

  • Generate Predictions (unlabeled data)

  • Train and Evaluate Model (labeled data)

  • Generate Synthetic Corruption Images (artificial corruption data)

Background Summary

Observations taken within a single visit are intended to represent a multi-wavelength view of objects in a given field-of-view. Most of the time, such observations are taken using pairs of guide stars to provide a very stable field-of-view throughout the entire visit resulting in images which overlap almost perfectly. Unfortunately, due to increasing limitations of the aging telescope systems, this is not always possible. The result is an increasing number of visits where observations:

  • drift and/or roll during the course of the visit

  • re-acquire at slightly different pointings from one orbit to the next

Therefore, data across each visit cannot automatically be assumed to align. Single-visit mosaic (SVM) processing attempts to:

  • correct these relative alignment errors

  • align the data to an absolute astrometric frame

In order to ensure that data across all filters used can be drizzled onto the same pixel grid.

Model Architecture

The SVM Classifier is a neural network trained on both “good” and “bad” (or valid and compromised) final drizzle products (images), along with statistical information produced during SVM processing quality analysis tests. The model’s objective is to identify SVM total detection images that may have experienced an alignment error.

Ensemble Classifier

  • Multi-Layer Perceptron (MLP)

  • 3D Image CNN

Dataset

  • SVM Regression test results

  • SVM total detection images

—

## Setup

Install with pip

$ pip install spacekit

$ git clone https://github.com/spacetelescope/spacekit
$ cd spacekit
$ pip install -e .

## Run

Example: HST Single Visit Mosaic Alignment Classification

### Classify new data using pre-trained model:

  1. Preprocess data (scrape from regression test json and fits files, scrub/preprocess dataframe, generate png images for ML)

*from the command line*

$ python -m spacekit.skopes.hst.svm.prep path/to/svmdata -f=svm_data.csv

*from python*

from spacekit.skopes.hst.svm.prep import run_preprocessing
input_path = "/path/to/svm/datasets"
run_preprocessing(input_path)

# This is equivalent to using the default kwargs:
run_preprocessing(input_path, h5=None, fname="svm_data", output_path=None, json_pattern="*_total*_svm_*.json", crpt=0)

Outputs: * svm_data.csv * raw_svm_data.csv * svm_data.h5 * img/

  1. Generate predictions

*from the command line*

$ python -m spacekit.skopes.hst.svm.predict svm_data.csv img

*from python*

from spacekit.skopes.hst.svm.predict import predict_alignment
data_file = "svm_data.csv"
img_path = "img"
predict_alignment(data_file, img_path)

# This is equivalent to using the default kwargs:
predict_alignment(data_file, img_path, model_path=None, output_path=None, size=None)

Outputs: * predictions/clf_report.txt * predictions/compromised.txt * predictions/predictions.csv

### Build, train, evaluate new classifier from labeled data

Run step 1 (prep) above, then:

*from the command line*

# Note: there are several option flags you can also include in this command
$ python -m spacekit.skopes.hst.svm.train svm_data.csv img

*from Python*

# import spacekit training submodule
from spacekit.skopes.hst.svm.train import run_training

training_data = "svm_data.csv" # preprocessed dataframe (see step 1 above)
img_path = "img" # preprocessed PNG image files (see step 1 above)

run_training(training_data, img_path)

# This is the same as using the default kwargs
com, val = run_training(training_data, img_path, synth_data=None, norm=0, model_name=None, params=None, output_path=None)

# Optional: view plots
com.draw_plots()
val.draw_plots()