Unexpected acceleration with sigpy.mri.poisson

[ad12]

Describe the bug
Output masks produced by sigpy.mri.poisson are not within the expected +/-0.1 bounds of the expected acceleration (passed in as an argument). We also see this in cases where crop_corner=False and in the trivial case acceleration=1.0.

To Reproduce

# desired acceleration
desired_accel = 1.0  # also tried with 4.0 and 6.0

mask = sigpy.mri.poisson(
            (512, 80), 
            accel=desired_accel,
            calib=(16, 16),
            dtype=np.bool,
            crop_corner=False,
        )

acc = 512 * 80 / mask.sum()   # between 7.2-7.6 depending on the seed

Expected behavior
The output mask should correspond to an acceleration closer to the desired_accel

Desktop (please complete the following information):

• OS: Ubuntu 20.04
• SigPy Version: 0.1.20

[ad12]

Results from a more exhaustive search. The Expected Acceleration is the acceleration passed into the the poisson function.

Results

	Mask Shape	Calibration Size	Crop Corner?	Expected Acceleration	Acceleration
0	(512, 80)	(16, 16)	False	1	7.38284
1	(512, 80)	(16, 16)	True	1	8.9218
2	(512, 80)	(16, 16)	False	4	7.38284
3	(512, 80)	(16, 16)	True	4	8.9218
4	(512, 80)	(16, 16)	False	6	7.38284
5	(512, 80)	(16, 16)	True	6	8.9218
6	(512, 80)	(16, 16)	False	12	11.9104
7	(512, 80)	(16, 16)	True	12	12.0719
8	(320, 256)	(20, 20)	False	1	1.44702
9	(320, 256)	(20, 20)	True	1	1.7772
10	(320, 256)	(20, 20)	False	4	4.074
11	(320, 256)	(20, 20)	True	4	4.04344
12	(320, 256)	(20, 20)	False	6	6.06276
13	(320, 256)	(20, 20)	True	6	5.97128
14	(320, 256)	(20, 20)	False	12	12.0987
15	(320, 256)	(20, 20)	True	12	11.9434

To Reproduce

import numpy as np
import pandas as pd
from sigpy.mri import poisson
from tabulate import tabulate

N = 1
data = []

for img_shape, calib_size in zip([(512, 80), (320, 256)], [(16, 16), (20, 20)]):
    for accel in [1.0, 4.0, 6.0, 12.0]:
        for crop_corner in [False, True]:
            outputs = []
            for i in range(N):
                mask = poisson(img_shape, accel, calib=calib_size, dtype=np.bool, crop_corner=crop_corner)
                outputs.append(np.product(img_shape) / mask.sum())
            avg_accel = np.mean(outputs)
            data.append({"Image Shape": str(img_shape), "Calibration Size": str(calib_size), "Crop Corner?": crop_corner, "Expected Acceleration": accel, "Acceleration": avg_accel})
            print(data[-1])

import pdb; pdb.set_trace()
data = pd.DataFrame(data)
print(tabulate(data, headers=data.columns, tablefmt="github"))

[frankong]

This is fixed in 08bef3c.
accel=1 is still not supported, and the function will now throw an exception (cd27418)

Here's an instance of running the above code:

	Image Shape	Calibration Size	Crop Corner?	Expected Acceleration	Acceleration
0	(512, 80)	(16, 16)	False	4	3.97516
1	(512, 80)	(16, 16)	True	4	3.94605
2	(512, 80)	(16, 16)	False	6	6.06635
3	(512, 80)	(16, 16)	True	6	6.04576
4	(512, 80)	(16, 16)	False	12	11.9556
5	(512, 80)	(16, 16)	True	12	12.04
6	(320, 256)	(20, 20)	False	4	3.94472
7	(320, 256)	(20, 20)	True	4	3.93354
8	(320, 256)	(20, 20)	False	6	5.91181
9	(320, 256)	(20, 20)	True	6	6.00234
10	(320, 256)	(20, 20)	False	12	12.0012
11	(320, 256)	(20, 20)	True	12	11.9347