Streaming: Processing Unlimited Frames On-Disk
==============================================
A key feature of trackpy is the ability to process an unlimited number
of frames.
For feature-finding, this is straightforward: a frame is loaded,
features are located, the locations are saved the disk, and the memory
is cleared for the next frame. For linking, the problem is more
challenging, but trackpy handles all this complexity for you, using as
little memory as possible throughout.
When data sets become large, beginning-friendly file formats like CSV or
Excel become impractical. We recommend using the HDF5 file format, which
is trackpy can read and write out of the box. (HDF5 is `widely
used <http://en.wikipedia.org/wiki/Hierarchical_Data_Format>`__; you can
be sure it will be around for many, many years to come.)
If you have some other format in mind, see the end of this tutorial,
where we explain how to extend trackpy's interface to support other
formats.
Install PyTables
----------------
You need pytables, which you can easily install using conda. (Type this
command into a Terminal or Command Prompt.)
::
conda install pytables
Locate Features, Streaming Results into an HDF5 File
----------------------------------------------------
.. code:: python
import trackpy as tp
import pims
.. code:: python
def gray(image):
return image[:, :, 0]
images = pims.ImageSequence('../sample_data/bulk_water/*.png', process_func=gray)
images = images[:10] # We'll take just the first 10 frames for demo purposes.
.. code:: python
# For this demo, we'll first remove the file if it already exists.
!rm -f data.h5
We can use ``locate`` inside a loop:
.. code:: python
with tp.PandasHDFStore('data.h5') as s: # This opens an HDF5 file. Data will be stored and retrieved by frame number.
for image in images:
features = tp.locate(image, 11, invert=True) # Find the features in a given frame.
s.put(features) # Save the features to the file before continuing to the next frame.
or, equivalently, we can use ``batch``, which accepts the storage file
as ``output``.
.. code:: python
with tp.PandasHDFStore('data.h5') as s:
tp.batch(images, 11, invert=True, output=s)
.. parsed-literal::
Frame 9: 510 features
We can get the data for a given frame:
.. code:: python
with tp.PandasHDFStore('data.h5') as s:
frame_2_results = s.get(2)
frame_2_results.head() # Display the first few rows.
.. raw:: html
<div>
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>x</th>
<th>y</th>
<th>mass</th>
<th>size</th>
<th>ecc</th>
<th>signal</th>
<th>raw_mass</th>
<th>ep</th>
<th>frame</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>295.886941</td>
<td>5.624598</td>
<td>372.183679</td>
<td>2.580235</td>
<td>0.189920</td>
<td>17.396842</td>
<td>8915</td>
<td>0.108849</td>
<td>2</td>
</tr>
<tr>
<th>1</th>
<td>68.203651</td>
<td>6.471969</td>
<td>416.980545</td>
<td>2.888723</td>
<td>0.088487</td>
<td>13.265092</td>
<td>9267</td>
<td>0.069054</td>
<td>2</td>
</tr>
<tr>
<th>2</th>
<td>336.888818</td>
<td>6.445367</td>
<td>340.325712</td>
<td>2.565562</td>
<td>0.028504</td>
<td>16.418269</td>
<td>8815</td>
<td>0.130158</td>
<td>2</td>
</tr>
<tr>
<th>3</th>
<td>432.423210</td>
<td>6.799269</td>
<td>564.962429</td>
<td>3.048429</td>
<td>0.302762</td>
<td>17.614302</td>
<td>9131</td>
<td>0.080412</td>
<td>2</td>
</tr>
<tr>
<th>4</th>
<td>460.570088</td>
<td>7.785952</td>
<td>359.136047</td>
<td>2.924729</td>
<td>0.124871</td>
<td>12.503980</td>
<td>8907</td>
<td>0.110293</td>
<td>2</td>
</tr>
</tbody>
</table>
</div>
Or dump all the data, if your machine has enough memory to hold it:
.. code:: python
with tp.PandasHDFStore('data.h5') as s:
all_results = s.dump()
all_results.head() # Display the first few rows.
.. raw:: html
<div>
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>x</th>
<th>y</th>
<th>mass</th>
<th>size</th>
<th>ecc</th>
<th>signal</th>
<th>raw_mass</th>
<th>ep</th>
<th>frame</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>103.430478</td>
<td>5.247191</td>
<td>308.202079</td>
<td>2.738100</td>
<td>0.039502</td>
<td>11.795655</td>
<td>8983</td>
<td>0.098772</td>
<td>0</td>
</tr>
<tr>
<th>1</th>
<td>294.831759</td>
<td>5.692167</td>
<td>355.060049</td>
<td>2.574877</td>
<td>0.162698</td>
<td>17.422941</td>
<td>8917</td>
<td>0.109846</td>
<td>0</td>
</tr>
<tr>
<th>2</th>
<td>311.069767</td>
<td>7.223679</td>
<td>255.933257</td>
<td>3.321975</td>
<td>0.007893</td>
<td>5.627285</td>
<td>8644</td>
<td>0.204852</td>
<td>0</td>
</tr>
<tr>
<th>3</th>
<td>431.496378</td>
<td>7.273025</td>
<td>627.442294</td>
<td>2.872567</td>
<td>0.273653</td>
<td>19.695498</td>
<td>9199</td>
<td>0.074267</td>
<td>0</td>
</tr>
<tr>
<th>4</th>
<td>36.061983</td>
<td>8.255091</td>
<td>483.621872</td>
<td>2.973328</td>
<td>0.123753</td>
<td>13.635345</td>
<td>9531</td>
<td>0.053765</td>
<td>0</td>
</tr>
</tbody>
</table>
</div>
You can dump the first N frames using ``s.dump(N)``.
Link Trajectories, Streaming From and Updating the HDF5 File
------------------------------------------------------------
.. code:: python
with tp.PandasHDFStore('data.h5') as s:
for linked in tp.link_df_iter(s, 3, neighbor_strategy='KDTree'):
s.put(linked)
.. parsed-literal::
Frame 9: 510 trajectories present
The original data is overwritten.
.. code:: python
with tp.PandasHDFStore('data.h5') as s:
frame_2_results = s.get(2)
frame_2_results.head() # Display the first few rows.
.. raw:: html
<div>
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>x</th>
<th>y</th>
<th>mass</th>
<th>size</th>
<th>ecc</th>
<th>signal</th>
<th>raw_mass</th>
<th>ep</th>
<th>frame</th>
<th>particle</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>295.886941</td>
<td>5.624598</td>
<td>372.183679</td>
<td>2.580235</td>
<td>0.189920</td>
<td>17.396842</td>
<td>8915</td>
<td>0.108849</td>
<td>2</td>
<td>1</td>
</tr>
<tr>
<th>1</th>
<td>432.423210</td>
<td>6.799269</td>
<td>564.962429</td>
<td>3.048429</td>
<td>0.302762</td>
<td>17.614302</td>
<td>9131</td>
<td>0.080412</td>
<td>2</td>
<td>3</td>
</tr>
<tr>
<th>2</th>
<td>36.560049</td>
<td>8.409618</td>
<td>440.901203</td>
<td>3.004805</td>
<td>0.132634</td>
<td>12.938901</td>
<td>9508</td>
<td>0.055229</td>
<td>2</td>
<td>4</td>
</tr>
<tr>
<th>3</th>
<td>68.203651</td>
<td>6.471969</td>
<td>416.980545</td>
<td>2.888723</td>
<td>0.088487</td>
<td>13.265092</td>
<td>9267</td>
<td>0.069054</td>
<td>2</td>
<td>5</td>
</tr>
<tr>
<th>4</th>
<td>629.016761</td>
<td>7.998041</td>
<td>499.506812</td>
<td>3.298593</td>
<td>0.183847</td>
<td>9.350802</td>
<td>9374</td>
<td>0.062147</td>
<td>2</td>
<td>7</td>
</tr>
</tbody>
</table>
</div>
Framewise Data Interfaces
-------------------------
Built-in interfaces
~~~~~~~~~~~~~~~~~~~
There are three different interfaces. You can use them interchangeably.
They offer different performance advantages.
- ``PandasHDFStore`` -- fastest for a small (~100) number of frames
- ``PandasHDFStoreBig`` -- fastest for a medium or large number of
frames
- ``PandasHDFStoreSingleNode`` -- optimizes HDF queries that access
multiple frames (advanced)
Writing your own interface
~~~~~~~~~~~~~~~~~~~~~~~~~~
Trackpy implements a generic interface that could be used to store and
retrieve particle tracking data in any file format. We hope that it can
make it easier for researchers who use different file formats to
exchange data. Any in-house format could be accessed using the same
simple interface demonstrated above.
At present, the interface is implemented only for HDF5 files. To extend
it to any format, write a class subclassing ``trackpy.FramewiseData``.
This custom class must implement the methods ``put``, ``get``,
``close``, and ``__iter__`` and the properties ``max_frame`` and
``t_column``. Refer to the built-in classes in
`framewise\_data.py <https://github.com/soft-matter/trackpy/blob/master/trackpy/framewise_data.py>`__
for examples to work from.