User Guide#

The demo notebook is the best place to start:

This guide provides some more details on the API and concepts for using mesmerize-core.

Mesmerize-core interfaces with CaImAn algorithms, helps with data organization, and provides useful functions for evaluation and visualization. It is a collection of “pandas extensions” – functions that operate on pandas DataFrames. This enables you to create a “psuedo-database” of your calcium imaging data and CaImAn generated output files. No database setup or experience is required, it operates purely on pandas and standard file systems.

Since this framework uses pandas extensions, you should be relatively comfortable with basic pandas operations. If you’re familiar with numpy then pandas will be easy, here’s a quick start guide from the pandas docs:

Accessors and Extensions#

There are 3 accessors that the mesmerize-core API provides, caiman, mcorr and cnmf. These allow you to perform operations on a pandas.DataFrame or invidual DataFrame rows, which are called pandas.Series. In mesmerize-core the individual rows contain data that pertain to a single batch item.

A batch item is the combination of:

  • input data (input movie)

  • parameters

  • algorithm

  • output data (depends on algorithm)

  • a user defined name for your convenience, multiple batch items can have the same name

  • UUID (universally unique identifier), a 128 bit integer that uniquely identifies a batch item and is used to organize the output data. You must never modify the UUID, they are computer generated.


Some common caiman extensions are:

  • caiman.add_item() - adds a new batch item to the dataframe

  • - runs the batch item

  • caiman.get_corr_image() - gets the correlation image for the batch item

Some motion correction extensions:

  • mcorr.get_output() - get the motion corrected memmaped numpy array

  • mcorr.get_shifts() - get the x, y shifts per frame

Some CNMF extensions:

  • cnmf.get_contours() - get the spatial contours and centers of mass

  • cnmf.get_rcm() - get the reconstructed movie, i.e. (A * C)

  • cnmf.get_residuals() - get the residuals, i.e. Y - (A * C) - (b * f)

  • cnmf.run_eval() - runs component quality evaluation

You must use the appropriate accessor on a DataFrame or Series (row) to access the appropriate extension functions. Accessors that operate at the level of the DataFrame can only be referenced using the DataFrame instance.

For example the caiman.add_item() extension operates on a DataFrame, so you can use it like this:

# imports
from mesmerize_core import *

# load an existing DataFrame
df = load_batch("/path/to/batch.pickle")

# in this case `df` is a DataFrame instance
# we can use the `caiman` accessor to utilize
# common caiman extensions that operate at
# the level of the DataFrame

# for example, ``add_item()`` works at the level of a DataFrame

In contrast some common extensions, such as cnmf.get_contours() operate on pandas.Series, i.e. individual DataFrame rows. All the motion correction and CNMF extensions are Series extensions. You will need to using indexing on the DataFrame to get the pandas.Series (row) that you want.

# imports
from mesmerize_core import *

# load an existing DataFrame
df = load_batch("/path/to/batch.pickle")

# df.iloc[n] will return the pandas.Series, i.e. row at the `nth` index

# let's assume the item at index `0` is an mcorr item
# we can get the memmaped output movie
mcorr_movie_memmaped_array = df.iloc[0].mcorr.get_output()

# let's assume the item at index `1` is a cnmf(e) item
# we can get the contours
contours, coms = df.iloc[1].cnmf.get_contours()

Use of the mcorr and cnmf accessors isn’t limited to indexing through iloc[n], you can use any combination of pandas indexing that results in a pandas.Series.

Some common extensions are valid for getting outputs from motion correction and CNMF. For example the correlation image can be obtained regardless of motion correction or CNMF using the common caiman accessor on a dataframe row.

from mesmerize_core import *

# the 0th index, i.e. first row, in the dataframe
corr_img = df.iloc[0].caiman.get_corr_image()

# plot with matplotlib
from matplotlib import pyplot as plt



Using the wrong accessor and extension on a batch item (row/pandas Series) will raise an exception. For example, you cannot use cnmf.get_contours() on a motion correction batch item.

More examples

We can get motion corrected outputs as a memmaped numpy array using the mcorr accessor and get_output() function. We can also get CNMF outputs from another batch item, such as temporal components, using the cnmf accessor and get_temporal() function.

# get the output memmap after motion correction

memmap([[[ 1.09921265e+01,  5.52584839e+00,  2.44244690e+01, ...,
       2.74850464e+00,  5.92257690e+00,  3.67776489e+00],
     [ 8.48319397e+01,  4.00158539e+01,  6.09210205e+00, ...,
       3.89350281e+01,  5.72113037e+01,  2.35960083e+01],
     [ 1.09254852e+02,  8.75248413e+01,  1.91671143e+01, ...,
       2.50050354e+01,  7.38364258e+01,  1.21587524e+01],

df.iloc[0].mcorr.get_output().shape # returns [n_frames, x_pix, y_pix]

(3000, 170, 170)

# get temporal and spacial components

# this will return the [n_neurons, n_frames] array
array([[-22.34959017, -22.34959017, -22.34959017, ..., -22.34959017,
    -22.34959017, -22.34959017],
   [-24.06055624, -24.06055624,   0.73800929, ..., -24.03839339,
    -24.04034401, -24.04212251],
   [-20.06077687, -20.06077687, -20.06077687, ..., -20.06077687,
    -20.06077687, -20.06077687],

Common Extensions#

API reference for common extensions

These extensions with the accessor caiman contain functions that are common to both motion correction and CNMF. The most frequent common extension you will probably use is add_item() which adds a new batch item (row) to the DataFrame.

Basic structure of using add_item():

    algo=<name of algorithm, mcorr, cnmf, or cnmfe>,
    item_name=<a name for you to keep track of this item>,
    params=<params dict for algo>,

See the API reference for more details.


from mesmerize_core import *
# create a new batch
df = create_batch("/path/to/batch.pickle")

# params, exactly the same as what you'd directly use with CaImAn
mcorr_params =\
'main': # this key is required to specify that these are the "main" params for the algorithm
        'max_shifts': [24, 24],
        'strides': [48, 48],
        'overlaps': [24, 24],
        'max_deviation_rigid': 3,
        'border_nan': 'copy',
        'pw_rigid': True,
        'gSig_filt': None


You can add multiple “batch items” using the same input movie and set the same item_name but use different params. This enables you to perform a gridsearch to find the optimal params for your input movie.

You can run a batch item using the run() extension on an individual DataFrame row (i.e. Series). At the moment the only supported backend is subprocess, the “batch item” is run using the corresponding algorithm in an external subprocess so you can continue using your notebook, i.e. calling run() is non-blocking. run() returns a subprocess.Popen instance.

You can set the maximum number of processes to spawn using the MESMERIZE_N_PROCESSES environment variable. By default it will use n_cpus - 1.


# assuming a batch dataframe is already loaded
# runs the item at the 0th index

You can run an entire DataFrame from the 0th index (i.e. first row) to the last index (-1), or run certain ranges just by using for loops. I would recommend a pandas tutorial if this sounds complicated (pandas concepts and syntax are similar to numpy).


You MUST call wait() on the subprocess.Popen instance after the run() call, otherwise you will spawn hundreds of processes for multiple batch items simultaneously!

from tqdm import tqdm # for a progress bar

# run an entire dataframe
for ix, r in tqdm(df.iterrows(), total=df.index.size):
    process =
    process.wait()  # this line is VERY IMPORTANT!!

# or run only certain rows
for ix, r in tqdm(df.iterrows(), total=df.index.size):
    if ix < 30:  # skip the first 29 items
    if ix > 100:  # skip items after index 99

    process =

Data management#

See the API reference for more details on these extensions.

There are some extensions under the common caiman accessor that help with data management, they operate on the DataFrame (not Series/rows).


This will return the row, i.e. pandas.Series for the given UUID.


row = df.caiman.uloc("fd2b3734-96b1-4656-945e-6860df9b711e")


Removes the batch item i.e. row within the DataFrame (a.k.a pandas.Series), from the DataFrame. Also delete corresponding output files from disk if remove_data=True (it is True by default). safe_removal (default True) is useful to make sure you do not delete an mcorr item if this mcorr output is used later in the dataframe for cnmf.

The batch item to remove is indicated by an int index or UUID (either as a str or UUID object).


Get the list of UUIDs of all batch items that use the output of the batch item passed to get_children(). For example, you can get the UUIDs of all downstream CNMF batch items that use the output from a given mcorr batch item.

Note: This feature is experimental and its behavior may change in future releases.


Get the UUID of the parent batch item. For example, you can pass the UUID of a CNMF batch item to get_parent() to get the UUID of the mcorr batch item whose output was used as the input for the CNMF batch item.

Note: This feature is experimental and its behavior may change in future releases.

Motion Correction Extensions#

API reference for motion correction extensions

These extensions with the accessor mcorr contain functions that are exclusive to motion correction.


This returns the memmaped numpy array of the motion corrected movie. It allows fast random access scrolling which is useful for fast random-access scrolling during visualization. See the Visuzalition page for details on visualization, we recommend mesmerize-viz and fastplotlib.


This returns the Path to the memmaped numpy array. The most common use for this extension is for using the motion corrected movie as the input movie for CNMF(E). You can use the returned path from mcorr.get_output_path() to set the input_movie_path argument for caiman.add_item()

CNMF Extensions#

These extensions with the accessor cnmf contain functions that are exclusive to CNMF, such as getting the contours and centers of mass for spatial components, getting the temporal components and dF/F0, running component evaluation, getting the reconstructed movie, residuals, etc. See the API reference for CNMF extensions which extensively documents these extensions along with several examples.

Input movie files#

mesmerize-core will work with any input files that caiman accepts, input movie paths are directly passed to caiman when you use caiman.add_item(...input_movie_path="/path/to/movie"). However if you use caiman.get_input_movie() only tiff files and files that can be opened with pims are currently supported, BUT it is easy even for a Python beginner to use caiman.get_input_movie() with your own file formats. You just need to create function that returns an indexable array for your file format and pass it: caiman.get_input_movie(my_func).

For example:

from mesmerize_core import *

df = load_batch("/path/to/batch.pickle")

def my_func(path):
    # your file handling code here
    # must return an indexable array
    return a  # shape must be [n_frames, x, y]

movie = df.iloc[0].caiman.get_input_movie(my_func)