utils/eval_baselines.py

"""
Runs several baseline compression algorithms and stores results for each FITS file in a csv.
This code is written functionality-only and cleaning it up is a TODO.
"""


import os
import re
from pathlib import Path
import argparse
import os.path
from astropy.io import fits
import numpy as np
from time import time
import pandas as pd
from tqdm import tqdm

from astropy.io.fits import CompImageHDU
from imagecodecs import (
    jpeg2k_encode, 
    jpeg2k_decode, 
    jpegls_encode, 
    jpegls_decode, 
    jpegxl_encode,
    jpegxl_decode,
    rcomp_encode,
    rcomp_decode,
)

# Functions that require some preset parameters. All others default to lossless.

jpegxl_encode_max_effort_preset = lambda x: jpegxl_encode(x, lossless=True, effort=9)
jpegxl_encode_preset = lambda x: jpegxl_encode(x, lossless=True)


def split_uint16_to_uint8(arr):
    # Ensure the input is of the correct type
    assert arr.dtype == np.uint16, "Input array must be of type np.uint16"
    
    # Compute the top 8 bits and the bottom 8 bits
    top_bits = (arr >> 8).astype(np.uint8)
    bottom_bits = (arr & 0xFF).astype(np.uint8)
    
    return top_bits, bottom_bits

def find_matching_files():
    """
    Returns list of test set file paths.
    """
    df = pd.read_json("./splits/full_test.jsonl", lines=True)
    return list(df['image'])

def benchmark_imagecodecs_compression_algos(arr, compression_type):

    encoder, decoder = ALL_CODECS[compression_type]

    write_start_time = time()
    encoded = encoder(arr)
    write_time = time() - write_start_time

    read_start_time = time()
    if compression_type == "RICE":
        decoded = decoder(encoded, shape=arr.shape, dtype=np.uint16)
    else:
        decoded = decoder(encoded)
    read_time = time() - read_start_time

    assert np.array_equal(arr, decoded)

    buflength = len(encoded)

    return {compression_type + "_BPD": buflength / arr.size,
            compression_type + "_WRITE_RUNTIME": write_time,
            compression_type + "_READ_RUNTIME": read_time,
            #compression_type + "_TILE_DIVISOR": np.nan,
           }

def main(dim):

    save_path = f"baseline_results_{dim}.csv"

    file_paths = find_matching_files()
    
    df = pd.DataFrame(columns=columns, index=[str(p) for p in file_paths])
    
    print(f"Number of files to be tested: {len(file_paths)}")
    
    ct = 0

    for path in tqdm(file_paths):
        with fits.open(path) as hdul:
            if dim == '2d': # compress the first timestep frame, R wavelength band (index 2)
                arr = hdul[0].data[0][2]
                arrs = [arr]
            elif dim == '2d-top': # same as 2d, but only top 8 bits. This is to compare with similarly preprocessed neural approaches.
                arr = hdul[0].data[0][2]
                arr = split_uint16_to_uint8(arr)[0]
                arrs = [arr]
            elif dim == '2d-bottom': # same as 2d, but only bottom 8 bits. This is to compare with similarly preprocessed neural approaches.
                arr = hdul[0].data[0][2]
                arr = split_uint16_to_uint8(arr)[1]
                arrs = [arr]
            elif dim == '3dt' and len(hdul[0].data) > 2: # 3D tensor with first 3 timestep frames of wavelength band index 2
                arr = hdul[0].data[0:3][2]
                arrs = [arr]
            elif dim == '3dw' and len(hdul[0].data[0]) > 2: # 3D tensor with first timestep frame on wavelength bands of indices 1,2,3 (G, R, I bands)
                arr = hdul[0].data[0][0:3]
                arrs = [arr]
            elif dim == '3dt_reshape' and len(hdul[0].data) > 2: # Same as 3dt but reshape into 2D array, for compatibility with JPEG-LS and RICE
                arr = hdul[0].data[0:3][2].reshape((800, -1))
                arrs = [arr]
            elif dim == '3dw_reshape' and len(hdul[0].data[0]) > 2: # Same as 3dw but reshape into 2D array, for compatibility with JPEG-LS and RICE
                arr = hdul[0].data[0][0:3].reshape((800, -1))
                arrs = [arr]
            elif dim == 'tw': # Iterate through all possible arrays where the x,y spatial location is fixed, and the remaining 2D array consists of ALL timesteps, ALL wavelengths. 
                init_arr = hdul[0].data
                def arrs_gen():
                    for i in range(init_arr.shape[-2]):
                        for j in range(init_arr.shape[-1]):
                            yield init_arr[:, :, i, j]
                            
                arrs = arrs_gen()
            else:
                continue

        ct += 1
        if ct % 10 == 0:
            print(df.mean())
            df.to_csv(save_path)

        for arr_idx, arr in enumerate(arrs):
            for algo in ALL_CODECS.keys():
                try:
                    if algo == "JPEG_2K" and (dim == '3dt' or dim == '3dw'):
                        test_results = benchmark_imagecodecs_compression_algos(arr.transpose(1, 2, 0), algo)
                    else:
                        test_results = benchmark_imagecodecs_compression_algos(arr, algo)

                    for column, value in test_results.items():
                        if column in df.columns:
                            df.at[path + f"_arr_{arr_idx}", column] = value

                except Exception as e:
                    print(f"Failed at {path} under exception {e}.")
                    
                    
        print(df.mean())
        df.to_csv(save_path)


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Process some 2D or 3D data.")
    parser.add_argument(
        "dimension", 
        choices=['2d', '2d-top', '2d-bottom', '3dt', '3dw', 'tw', '3dt_reshape', '3dw_reshape'],
        help="Specify whether the data is 2d, 3dt (3d time dimension), 3dw (3d wavelength dimension), 2d-top (only top 8 bits), 2d-bottom (only bottom 8 bits), tw (only a single x,y spatial location but all timesteps and wavelengths), 3dt_reshape or 3dw_reshape for the 2D flattened 3D evals, for use on JPEG-LS or RICE."
    )
    args = parser.parse_args()
    dim = args.dimension.lower()
    
    # RICE REQUIRES UNIQUE INPUT OF ARR SHAPE AND DTYPE INTO DECODER
    
    if dim == '3dw' or dim == '3dt' or dim == 'tw':
        ALL_CODECS = {
        "JPEG_XL_MAX_EFFORT": [jpegxl_encode_max_effort_preset, jpegxl_decode],
        "JPEG_XL": [jpegxl_encode_preset, jpegxl_decode],
        "JPEG_2K": [jpeg2k_encode, jpeg2k_decode],
        }
    else:
        ALL_CODECS = {
            "JPEG_XL_MAX_EFFORT": [jpegxl_encode_max_effort_preset, jpegxl_decode],
            "JPEG_XL": [jpegxl_encode_preset, jpegxl_decode],
            "JPEG_2K": [jpeg2k_encode, jpeg2k_decode], 
            "JPEG_LS": [jpegls_encode, jpegls_decode],
            "RICE": [rcomp_encode, rcomp_decode],
        }

    columns = []
    for algo in ALL_CODECS.keys():
        columns.append(algo + "_BPD")
        columns.append(algo + "_WRITE_RUNTIME")
        columns.append(algo + "_READ_RUNTIME")
        #columns.append(algo + "_TILE_DIVISOR")
        
    main(dim)