Sinter is a software tool/library for doing fast monte carlo sampling of quantum error correction circuits.
- How it works
- How to install
- How to use: Python API
- How to use: Linux Command Line
- The csv format for sample statistics
Sinter takes Stim circuits annotated with noise, detectors, and logical observables. It uses stim to sample the circuits and a decoder such as pymatching to predict whether the logical observables were flipped or not, given the detector data. It records how often this succeeds, and how often it fails (the error rate).
Sinter uses python multiprocessing to do parallel sampling across multiple CPU cores, dynamically decides which circuits need more samples based on parameters specified by the user (such as a target number of errors), saves the results to as simple CSV format, and has some basic plotting functionality for viewing the results.
Sinter doesn't support cloud compute, but it does scale well on a single machine. I've tested it on 2 core machines, 4 core machines, and 96 core machines. Although there are potential pitfalls (e.g. setting batch sizes too large causes thrashing), sinter generally achieves good resource utilization of the processes you assign to it.
Sinter is available as a pypi package. It can be installed using pip:
pip install sinter
When you are in a python virtual environment with sinter installed, you have
access to a command line command sinter which can be used to perform tasks
from the command line. You can also import sinter in a python program in order
to use sinter's python API.
This example assumes you are in a python environment with sinter and
pymatching installed.
import stim
import sinter
import matplotlib.pyplot as plt
# Generates surface code circuit tasks using Stim's circuit generation.
def generate_example_tasks():
for p in [0.001, 0.005, 0.01]:
for d in [3, 5]:
yield sinter.Task(
circuit=stim.Circuit.generated(
rounds=d,
distance=d,
after_clifford_depolarization=p,
code_task=f'surface_code:rotated_memory_x',
),
json_metadata={
'p': p,
'd': d,
},
)
def main():
# Collect the samples (takes a few minutes).
samples = sinter.collect(
num_workers=4,
max_shots=1_000_000,
max_errors=1000,
tasks=generate_example_tasks(),
decoders=['pymatching'],
)
# Print samples as CSV data.
print(sinter.CSV_HEADER)
for sample in samples:
print(sample.to_csv_line())
# Render a matplotlib plot of the data.
fig, ax = plt.subplots(1, 1)
sinter.plot_error_rate(
ax=ax,
stats=samples,
group_func=lambda stat: f"Rotated Surface Code d={stat.json_metadata['d']}",
x_func=lambda stat: stat.json_metadata['p'],
)
ax.loglog()
ax.set_ylim(1e-5, 1)
ax.grid()
ax.set_title('Logical Error Rate vs Physical Error Rate')
ax.set_ylabel('Logical Error Probability (per shot)')
ax.set_xlabel('Physical Error Rate')
ax.legend()
# Save to file and also open in a window.
fig.savefig('plot.png')
plt.show()
# NOTE: This is actually necessary! If the code inside 'main()' was at the
# module level, the multiprocessing children spawned by sinter.collect would
# also attempt to run that code.
if __name__ == '__main__':
main()Example output to stdout:
shots, errors, discards, seconds,decoder,strong_id,json_metadata
1000000, 837, 0, 36.6,pymatching,9f7e20c54fec45b6aef7491b774dd5c0a3b9a005aa82faf5b9c051d6e40d60a9,"{""d"":3,""p"":0.001}"
53498, 1099, 0, 6.52,pymatching,3f40432443a99b933fb548b831fb54e7e245d9d73a35c03ea5a2fb2ce270f8c8,"{""d"":3,""p"":0.005}"
16269, 1023, 0, 3.23,pymatching,17b2e0c99560d20307204494ac50e31b33e50721b4ebae99d9e3577ae7248874,"{""d"":3,""p"":0.01}"
1000000, 151, 0, 77.3,pymatching,e179a18739201250371ffaae0197d8fa19d26b58dfc2942f9f1c85568645387a,"{""d"":5,""p"":0.001}"
11363, 1068, 0, 12.5,pymatching,a4dec28934a033215ff1389651a26114ecc22016a6e122008830cf7dd04ba5ad,"{""d"":5,""p"":0.01}"
61569, 1001, 0, 24.5,pymatching,2fefcc356752482fb4c6d912c228f6d18762f5752796c668b6abeb7775f5de92,"{""d"":5,""p"":0.005}"
and the corresponding image saved to plot.png:
Sinter's python module exposes a variety of methods that are handy for plotting or analyzing QEC data. See the sinter API reference.
This example assumes you are using a linux command line in a python virtualenv with sinter installed.
For this example, we will use Stim's circuit generation functionality to produce
circuits to benchmark.
We will make rotated surface code circuits with various physical error rates,
with filenames like rotated_d5_p0.001_surface_code.stim.
mkdir -p circuits
python -c "
import stim
for p in [0.001, 0.005, 0.01]:
for d in [3, 5]:
with open(f'circuits/d={d},p={p},b=X,type=rotated_surface_memory.stim', 'w') as f:
c = stim.Circuit.generated(
rounds=d,
distance=d,
after_clifford_depolarization=p,
after_reset_flip_probability=p,
before_measure_flip_probability=p,
before_round_data_depolarization=p,
code_task=f'surface_code:rotated_memory_x')
print(c, file=f)
"Normally, making the circuit files is the hardest step, because they are what specifies the problem you are sampling from. Almost all of the work you do will generally involve creating the exact perfect circuit file for your needs. But this is just an example, so we'll use normal surface code circuits.
You can use sinter to collect statistics on each circuit by using the sinter collect command.
This command takes options specifying how much data to collect, how to do decoding, etc.
The processes argument decides how many workers to use. Set it to auto to set
it to the number of CPUs on your machine.
The metadata_func argument can be used to specify custom python expression that turns the path
into a dictionary or other JSON object associated with the circuit.
If you set metadata_func to auto then will use the method
sinter.comma_separated_key_values(path) which parses
stim circuit paths like folder/a=2,b=test.stim into a dictionary like {'a': 2, 'b': 'test'}.
By default, sinter writes the collected statistics to stdout as CSV data.
One particularly important option that changes this behavior is --save_resume_filepath,
which allows the command to be interrupted and restarted without losing data.
Any data already at the file specified by --save_resume_filepath will count towards the
amount of statistics asked to be collected, and sinter will append new statistics to this file
instead of overwriting it.
sinter collect \
--processes auto \
--circuits circuits/*.stim \
--metadata_func auto \
--decoders pymatching \
--max_shots 1_000_000 \
--max_errors 1000 \
--save_resume_filepath stats.csvBeware that if you SIGKILL or SIGTEM sinter, instead of just using SIGINT, it's possible (though unlikely) that you are killing it just as it writes a row of CSV data. This truncates the data, which requires manual intervention on your part to fix (e.g. by deleting the partial row using a text editor).
Note that the CSV data written by sinter will contain multiple rows for each case, because sinter starts by running small batches to see roughly what the error rate is before moving to larger batch sizes.
You can get a single-row-per-case CSV file by using sinter combine:
sinter combine stats.csv shots, errors, discards, seconds,decoder,strong_id,json_metadata
58591, 1067, 0, 5.50,pymatching,bb46c8fca4d9fd9d4d27a5039686332ac5e24011a7f2aea5a65f6040445567c0,"{""b"":""X"",""d"":3,""p"":0.005,""type"":""rotated_surface_memory""}"
1000000, 901, 0, 73.4,pymatching,4c0780830fe1747ab22767b69d1178f803943c83dd4afa6d241acf02e6dfa71f,"{""b"":""X"",""d"":3,""p"":0.001,""type"":""rotated_surface_memory""}"
16315, 1026, 0, 2.39,pymatching,64d81b177ef1a455644ac3e03f374394cd8ad385ba2ee0ac147b2405107564fc,"{""b"":""X"",""d"":3,""p"":0.01,""type"":""rotated_surface_memory""}"
1000000, 157, 0, 116.5,pymatching,100855c078af0936d098cecbd8bfb7591c0951ae69527c002c9c5f4c79bde129,"{""b"":""X"",""d"":5,""p"":0.001,""type"":""rotated_surface_memory""}"
61677, 1005, 0, 21.2,pymatching,6d7b8b312a5460c7fe08119d3c7a040daa25bd34d524611160e4aac6196293fe,"{""b"":""X"",""d"":5,""p"":0.005,""type"":""rotated_surface_memory""}"
10891, 1021, 0, 7.43,pymatching,477252e968f0f22f64ccb058c0e1e9c77b765f60f74df8b6707de7ec65ed13b7,"{""b"":""X"",""d"":5,""p"":0.01,""type"":""rotated_surface_memory""}"
You can use sinter plot to view the results you've collected.
This command takes a CSV file, an argument --group_func indicating how to
group the statistics into curves, an argument --x_func indicating how to
pick the X coordinate of each point, and various other arguments. Each *_func
argument takes a string that will be evaluated as a python expression, with
various useful values in scope such as a metadata value containing the
json metadata for the various points being evaluated. There is also a special
m value where m.key is shorthand for metadata.get('key', None).
Here is an example of a sinter plot command:
sinter plot \
--in stats.csv \
--group_func "f'''Rotated Surface Code d={m.d}'''" \
--x_func m.p \
--xaxis "[log]Physical Error Rate" \
--fig_size 1024 1024 \
--out surface_code_figure.png \
--showWhich will save a png image of, and also open a window showing, a plot like this one:
Sinter saves samples as a table using a Comma Separated Value format. For example:
shots,errors,discards,seconds,decoder,strong_id,json_metadata
1000000, 837, 0, 36.6,pymatching,9f7e20c54fec45b6aef7491b774dd5c0a3b9a005aa82faf5b9c051d6e40d60a9,"{""d"":3,""p"":0.001}"
53498, 1099, 0, 6.52,pymatching,3f40432443a99b933fb548b831fb54e7e245d9d73a35c03ea5a2fb2ce270f8c8,"{""d"":3,""p"":0.005}"
16269, 1023, 0, 3.23,pymatching,17b2e0c99560d20307204494ac50e31b33e50721b4ebae99d9e3577ae7248874,"{""d"":3,""p"":0.01}"
1000000, 151, 0, 77.3,pymatching,e179a18739201250371ffaae0197d8fa19d26b58dfc2942f9f1c85568645387a,"{""d"":5,""p"":0.001}"
11363, 1068, 0, 12.5,pymatching,a4dec28934a033215ff1389651a26114ecc22016a6e122008830cf7dd04ba5ad,"{""d"":5,""p"":0.01}"
61569, 1001, 0, 24.5,pymatching,2fefcc356752482fb4c6d912c228f6d18762f5752796c668b6abeb7775f5de92,"{""d"":5,""p"":0.005}"
The columns are:
shots(unsigned int): How many times the circuit was sampled.errors(unsigned int): How many times the decoder failed to predict any logical observable.discards(unsigned int): How many times a shot was discarded because a postselected detector fired or because the decoder incorrectly predicted the value of a postselected observable. Discarded shots never count as errors.seconds(non-negative float): How many CPU core seconds it took to simulate and decode these shots.decoder(str): Which decoder was used.strong_id(str): Hex representation of a cryptographic hash of the problem being sampled from. The hashed data includes the exact circuit that was simulated, the decoder that was used, the exact detector error model that was given to the decoder, the postselection rules that were applied, and the metadata associated with the circuit. The purpose of the strong id is to make it impossible to accidentally combine shots that were from separate circuits or separate versions of a circuit.json_metadata(json): A free form field that can store any value representable in Java Script Object Notation. For example, this could be a dictionary with helpful keys like "noise_level" or "circuit_name". The json value is serialized into JSON and then escaped so that it can be put into the CSV data (e.g. quotes get doubled up).custom_counts(json[Dict[str, int]]): An optional field that can store a dictionary from string keys to integer counts represented in Java Script Object Notation. The counts can be a huge variety of things, ranging from per-observable error counts to detection event counts. In general, any value that should be added when merging rows could be in these counters.
Note shots may be spread across multiple rows. For example, this data:
shots,errors,discards,seconds,decoder,strong_id,json_metadata
500000, 437, 0, 20.5,pymatching,9f7e20c54fec45b6aef7491b774dd5c0a3b9a005aa82faf5b9c051d6e40d60a9,"{""d"":3,""p"":0.001}"
500000, 400, 0, 16.1,pymatching,9f7e20c54fec45b6aef7491b774dd5c0a3b9a005aa82faf5b9c051d6e40d60a9,"{""d"":3,""p"":0.001}"
has the same total statistics as this data:
shots,errors,discards,seconds,decoder,strong_id,json_metadata
1000000, 837, 0, 36.6,pymatching,9f7e20c54fec45b6aef7491b774dd5c0a3b9a005aa82faf5b9c051d6e40d60a9,"{""d"":3,""p"":0.001}"
just split over two rows instead of combined into one.