SVD Codec

SVDCodec compresses a 2D patch with a truncated singular value decomposition.

For background on the method, see low-rank approximation.

Configuration

Set exactly one of rank or energy.

Option	Type	Meaning
rank	int	Number of singular values to keep. Higher ranks keep more detail and compress less. Gives a predictable compression ratio.
energy	float in (0, 1]	Fraction of the signal energy to keep. Gives a predictable error.

Options given to the constructor act as defaults. They can be overridden for a single call by passing them to compress_patch() or write_compressed():

codec = SVDCodec(rank=20)
compress_patch(patch, codec)             # rank 20
compress_patch(patch, codec, rank=5)     # rank 5 for this call only
compress_patch(patch, codec, energy=0.9) # use energy criterion for this call

Example

import numpy as np
import dascore as dc
from comdas import SVDCodec, compress_patch

patch = dc.get_example_patch("example_event_2")

codec = SVDCodec(energy=0.9)
compressed = compress_patch(patch, codec)
payload = compressed.data.payload
error = np.linalg.norm(patch.data - np.asarray(compressed.data))
error /= np.linalg.norm(patch.data)
print(
    f"rank={payload.rank:4d}  "
    f"ratio={codec.compression_ratio(payload):6.1f}x  "
    f"relative error={error:.3f}"
)

Limitations

• Only 2D data is supported.

• Encoding computes a full SVD, so very large patches may be slow to compress. Decoding is fast.

API

See the module reference.