[DataFile] Add parallel dataset compression

Description

For a while now, we are using the parallelized compression offered by write_compressed_frames. This MR adds this capability to the DataFrame API to write such compressed datasets side-by-side with uncompressed ones.

How Has This Been Tested?

Running LPD_Correct_Fast with default arguments.

Types of changes

• Refactor (refactoring code with no functionality changes)

Reviewers

@ahmedk

notebooks/Jungfrau/Jungfrau_Gain_Correct_and_Verify_NBC.ipynb

 %% Cell type:markdown id: tags:
 
 # Jungfrau Offline Correction #
 
 Author: European XFEL Detector Group, Version: 2.0
 
 Offline Calibration for the Jungfrau Detector
 
 %% Cell type:code id: tags:
 
 ``` python
 in_folder = "/gpfs/exfel/exp/SPB/202130/p900204/raw"  # the folder to read data from, required
 out_folder =  "/gpfs/exfel/data/scratch/ahmedk/test/remove"  # the folder to output to, required
 run = 91  # run to process, required
 metadata_folder = ""  # Directory containing calibration_metadata.yml when run by xfel-calibrate
 sequences = [-1]  # sequences to correct, set to [-1] for all, range allowed
 sequences_per_node = 1  # number of sequence files per cluster node if run as slurm job, set to 0 to not run SLURM parallel
 
 # Parameters used to access raw data.
 karabo_id = "SPB_IRDA_JF4M"  # karabo prefix of Jungfrau devices
 karabo_da = ['JNGFR01', 'JNGFR02', 'JNGFR03', 'JNGFR04', 'JNGFR05', 'JNGFR06', 'JNGFR07', 'JNGFR08']  # data aggregators
 receiver_template = "JNGFR{:02d}"  # Detector receiver template for accessing raw data files. e.g. "JNGFR{:02d}"
 instrument_source_template = '{}/DET/{}:daqOutput'  # template for source name (filled with karabo_id & receiver_id). e.g. 'SPB_IRDA_JF4M/DET/JNGFR01:daqOutput'
 ctrl_source_template = '{}/DET/CONTROL'  # template for control source name (filled with karabo_id_control)
 karabo_id_control = ""  # if control is on a different ID, set to empty string if it is the same a karabo-id
 
 # Parameters for calibration database.
 use_dir_creation_date = True  # use the creation data of the input dir for database queries
 cal_db_interface = "tcp://max-exfl016:8017#8025" # the database interface to use
 cal_db_timeout = 180000  # timeout on caldb requests
 
 # Parameters affecting corrected data.
 relative_gain = True  # do relative gain correction.
 strixel_sensor = False  # reordering for strixel detector layout.
 strixel_double_norm = 2.0  # normalization to use for double-size pixels, only applied for strixel sensors.
 limit_trains = 0  # ONLY FOR TESTING. process only first N trains, Use 0 to process all.
 chunks_ids = 32  # HDF chunk size for memoryCell and frameNumber.
 chunks_data = 1  # HDF chunk size for pixel data in number of frames.
 
 # Parameters for retrieving calibration constants
 manual_slow_data = False  # if true, use manually entered bias_voltage, integration_time, gain_setting, and gain_mode values
 integration_time = 4.96  # integration time in us, will be overwritten by value in file
 gain_setting = 0  # 0 for dynamic gain, 1 for dynamic HG0, will be overwritten by value in file
 gain_mode = 0  # 0 for runs with dynamic gain setting, 1 for fixgain. It will be overwritten by value in file, if manual_slow_data is set to True.
 mem_cells = -1  # Set mem_cells to -1 to automatically use the value stored in RAW data.
 bias_voltage = 180  # will be overwritten by value in file
 
 # Parameters for plotting
 skip_plots = False  # exit after writing corrected files
 plot_trains = 500  # Number of trains to plot for RAW and CORRECTED plots. Set to -1 to automatically plot all trains.
 cell_id_preview = 15  # cell Id used for preview in single-shot plots
 
 # Parameters for ROI selection and reduction
 roi_definitions = [-1]  # List with groups of 6 values defining ROIs, e.g. [3, 120, 180, 200, 550, -2] for module 3 (JNGFR03), slice 120:180, 200:550, average along axis -2 (slow scan, or -1 for fast scan)
 
 def balance_sequences(in_folder, run, sequences, sequences_per_node, karabo_da):
     from xfel_calibrate.calibrate import balance_sequences as bs
     return bs(in_folder, run, sequences, sequences_per_node, karabo_da)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 import multiprocessing
 import sys
 import warnings
 from functools import partial
 from pathlib import Path
 
 import h5py
 import matplotlib
 import matplotlib.pyplot as plt
 import numpy as np
 import pasha as psh
 import tabulate
 from IPython.display import Latex, Markdown, display
 from extra_data import H5File, RunDirectory, by_id, components
 from extra_geom import JUNGFRAUGeometry
 from matplotlib.colors import LogNorm
 
 from cal_tools import h5_copy_except
 from cal_tools.jungfraulib import JungfrauCtrl
 from cal_tools.enums import BadPixels
 from cal_tools.files import DataFile
 from cal_tools.step_timing import StepTimer
 from cal_tools.tools import (
     get_constant_from_db_and_time,
     get_dir_creation_date,
     get_pdu_from_db,
     map_seq_files,
-    write_compressed_frames,
     CalibrationMetadata,
 )
 from iCalibrationDB import Conditions, Constants
 
 warnings.filterwarnings('ignore')
 
 matplotlib.use('agg')
 %matplotlib inline
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 in_folder = Path(in_folder)
 out_folder = Path(out_folder)
 run_folder = in_folder / f'r{run:04d}'
 run_dc = RunDirectory(run_folder)
 instrument_src = instrument_source_template.format(karabo_id, receiver_template)
 
 out_folder.mkdir(parents=True, exist_ok=True)
 
 print(f"Run is: {run}")
 print(f"Instrument H5File source: {instrument_src}")
 print(f"Process modules: {karabo_da}")
 
 creation_time = None
 if use_dir_creation_date:
     creation_time = get_dir_creation_date(in_folder, run)
     print(f"Using {creation_time} as creation time")
 
 if karabo_id_control == "":
     karabo_id_control = karabo_id
 
 if any(axis_no not in {-2, -1, 2, 3} for axis_no in roi_definitions[5::6]):
     print("ROI averaging must be on axis 2/3 (or equivalently -2/-1). "
           f"Axis numbers given: {roi_definitions[5::6]}")
     sys.exit(1)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 # Read available sequence files to correct.
 mapped_files, num_seq_files = map_seq_files(
     run_folder, karabo_da, sequences)
 
 if not len(mapped_files):
     raise IndexError(
         "No sequence files available to correct for the selected sequences and karabo_da.")
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 print(f"Processing a total of {num_seq_files} sequence files")
 table = []
 fi = 0
 for kda, sfiles in mapped_files.items():
     for k, f in enumerate(sfiles):
         if k == 0:
             table.append((fi, kda, k, f))
         else:
             table.append((fi, "", k,  f))
         fi += 1
 md = display(Latex(tabulate.tabulate(
     table, tablefmt='latex',
     headers=["#", "module", "# module", "file"])))
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 ctrl_src = ctrl_source_template.format(karabo_id_control)
 ctrl_data = JungfrauCtrl(run_dc, ctrl_src)
 
 if mem_cells < 0:
     memory_cells, sc_start = ctrl_data.get_memory_cells()
 
     mem_cells_name = "single cell" if memory_cells == 1 else "burst"
     print(f"Run is in {mem_cells_name} mode.\nStorage cell start: {sc_start:02d}")
 else:
     memory_cells = mem_cells
     mem_cells_name = "single cell" if memory_cells == 1 else "burst"
     print(f"Run is in manually set to {mem_cells_name} mode. With {memory_cells} memory cells")
 
 if not manual_slow_data:
     integration_time = ctrl_data.get_integration_time()
     bias_voltage = ctrl_data.get_bias_voltage()
     gain_setting = ctrl_data.get_gain_setting()
     gain_mode = ctrl_data.get_gain_mode()
 
 print(f"Integration time is {integration_time} us")
 print(f"Gain setting is {gain_setting} (run settings: {ctrl_data.run_settings})")
 print(f"Gain mode is {gain_mode} ({ctrl_data.run_mode})")
 print(f"Bias voltage is {bias_voltage} V")
 print(f"Number of memory cells are {memory_cells}")
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 strixel_transform = None
 output_frame_shape = None
 
 if strixel_sensor:
     from cal_tools.jfalgs import strixel_transform, strixel_shape, strixel_double_pixels
     output_frame_shape = strixel_shape()
     Ydouble, Xdouble = strixel_double_pixels()
     print('Strixel sensor transformation enabled')
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Retrieving calibration constants ###
 
 %% Cell type:code id: tags:
 
 ``` python
 condition = Conditions.Dark.jungfrau(
     memory_cells=memory_cells,
     bias_voltage=bias_voltage,
     integration_time=integration_time,
     gain_setting=gain_setting,
     gain_mode=gain_mode,
 )
 
 empty_constants = {
     "Offset": np.zeros((512, 1024, memory_cells, 3), dtype=np.float32),
     "BadPixelsDark": np.zeros((512, 1024, memory_cells, 3), dtype=np.uint32),
     "RelativeGain": None,
     "BadPixelsFF": None,
 }
 metadata = CalibrationMetadata(metadata_folder or out_folder)
 # NOTE: this notebook will not overwrite calibration metadata file
 const_yaml = metadata.get("retrieved-constants", {})
 
 def get_constants_for_module(karabo_da: str):
     """ Get calibration constants for given module of Jungfrau
 
     :return:
         offset_map (offset map),
         mask (mask of bad pixels),
         gain_map (map of relative gain factors),
         db_module (name of DB module),
         when (dictionary: constant - creation time)
     """
 
     when = dict()
     const_data = dict()
 
     if const_yaml:
         for cname, mdata in const_yaml[karabo_da]["constants"].items():
             const_data[cname] = dict()
             when[cname] = mdata["creation-time"]
             if when[cname]:
                 with h5py.File(mdata["file-path"], "r") as cf:
                     const_data[cname] = np.copy(
                         cf[f"{mdata['dataset-name']}/data"])
             else:
                 const_data[cname] = empty_constants[cname]
     else:
         retrieval_function = partial(
             get_constant_from_db_and_time,
             karabo_id=karabo_id,
             karabo_da=karabo_da,
             cal_db_interface=cal_db_interface,
             creation_time=creation_time,
             timeout=cal_db_timeout,
             print_once=False,
         )
 
         for cname, cempty in empty_constants.items():
             const_data[cname], when[cname] = retrieval_function(
                 condition=condition,
                 constant=getattr(Constants.jungfrau, cname)(),
                 empty_constant=cempty,
             )
 
     offset_map = const_data["Offset"]
     mask = const_data["BadPixelsDark"]
     gain_map = const_data["RelativeGain"]
     mask_ff = const_data["BadPixelsFF"]
 
     # Combine masks
     if mask_ff is not None:
         mask |= np.moveaxis(mask_ff, 0, 1)
 
     if memory_cells > 1:
         # move from x, y, cell, gain to cell, x, y, gain
         offset_map = np.moveaxis(offset_map, [0, 1], [1, 2])
         mask = np.moveaxis(mask, [0, 1], [1, 2])
     else:
         offset_map = np.squeeze(offset_map)
         mask = np.squeeze(mask)
 
     # masking double size pixels
     mask[..., [255, 256], :, :] |= BadPixels.NON_STANDARD_SIZE
     mask[..., [255, 256, 511, 512, 767, 768], :] |= BadPixels.NON_STANDARD_SIZE
 
     if gain_map is not None:
         if memory_cells > 1:
             gain_map = np.moveaxis(gain_map, [0, 2], [2, 0])
             # add extra empty cell constant
             b = np.ones(((1,)+gain_map.shape[1:]))
             gain_map = np.concatenate((gain_map, b), axis=0)
         else:
             gain_map = np.moveaxis(np.squeeze(gain_map), 1, 0)
 
     return offset_map, mask, gain_map, karabo_da, when
 
 with multiprocessing.Pool() as pool:
     r = pool.map(get_constants_for_module, karabo_da)
 
 # Print timestamps for the retrieved constants.
 constants = {}
 for offset_map, mask, gain_map, k_da, when in r:
     print(f'Constants for module {k_da}:')
     for const in when:
         print(f'  {const} injected at {when[const]}')
 
     if gain_map is None:
         print("No gain map found")
         relative_gain = False
 
     constants[k_da] = (offset_map, mask, gain_map)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 # Correct a chunk of images for offset and gain
 def correct_train(wid, index, d):
     d = d.astype(np.float32)  # [cells, x, y]
     g = gain[index]
 
     # Copy gain over first to keep it at the original 3 for low gain.
     if strixel_transform is not None:
         strixel_transform(g, out=gain_corr[index, ...])
     else:
         gain_corr[index, ...] = g
 
     # Jungfrau gains 0[00], 1[01], 3[11]
     # Change low gain to 2 for indexing purposes.
     g[g==3] = 2
 
     # Select memory cells
     if memory_cells > 1:
         """
         Even though it is correct to assume that memory cells pattern
         can be the same across all trains (for one correction run
         taken with one acquisition), it is preferred to not assume
         this to account for exceptions that can happen.
         """
         m = memcells[index].copy()
         # 255 is a cell value pointing to no cell image data (image of 0 pixels).
         # Corresponding image will be corrected with constant of cell 0. To avoid values of 0.
         # This line is depending on not storing the modified memory cells in the corrected data.
         m[m==255] = 0
 
         offset_map_cell = offset_map[m, ...]  # [16 + empty cell, x, y]
         mask_cell = mask[m, ...]
     else:
         offset_map_cell = offset_map
         mask_cell = mask
 
     # Offset correction
     offset = np.choose(g, np.moveaxis(offset_map_cell, -1, 0))
 
     d -= offset
 
     # Gain correction
     if relative_gain:
         if memory_cells > 1:
             gain_map_cell = gain_map[m, ...]
         else:
             gain_map_cell = gain_map
         cal = np.choose(g, np.moveaxis(gain_map_cell, -1, 0))
         d /= cal
 
     msk = np.choose(g, np.moveaxis(mask_cell, -1, 0))
 
     if strixel_transform is not None:
         strixel_transform(d, out=data_corr[index, ...])
         data_corr[index, :, Ydouble, Xdouble] /= strixel_double_norm
 
         strixel_transform(msk, out=mask_corr[index, ...])
     else:
         data_corr[index, ...] = d
         mask_corr[index, ...] = msk
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 step_timer = StepTimer()
 
 n_cpus = multiprocessing.cpu_count()
 context = psh.context.ProcessContext(num_workers=n_cpus)
 print(f"Using {n_cpus} workers for correction.")
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 def save_reduced_rois(ofile, data_corr, mask_corr, karabo_da):
     """If ROIs are defined for this karabo_da, reduce them and save to the output file"""
     rois_defined = 0
     module_no = int(karabo_da[-2:])
     params_source = f'{karabo_id}/ROIPROC/{karabo_da}'
     rois_source = f'{params_source}:output'
 
     for i in range(len(roi_definitions) // 6):
         roi_module, a1, a2, b1, b2, mean_axis = roi_definitions[i*6 : (i+1)*6]
         if roi_module == module_no:
             rois_defined += 1
             # Apply the mask and average remaining pixels to 1D
             roi_data = data_corr[..., a1:a2, b1:b2].mean(
                 axis=mean_axis, where=(mask_corr[..., a1:a2, b1:b2] == 0)
             )
 
             # Create Instrument section to later add corrected datasets.
             outp_source = ofile.create_instrument_source(rois_source)
             outp_source.create_key(f'data.roi{rois_defined}.data', data=roi_data)
 
             ctrl_source = ofile.create_control_source(params_source)
             ctrl_source.create_run_key('region', np.array([[a1, a2, b1, b2]]))
             ctrl_source.create_run_key('reduce_axis', np.array([mean_axis]))
 
     if rois_defined:
         # Copy the index for the new source
         # Create count/first datasets at INDEX source.
         ofile.copy(f'INDEX/{karabo_id}/DET/{karabo_da}:daqOutput/data',
                    f'INDEX/{rois_source}')
         ntrains = ofile['INDEX/trainId'].shape[0]
         ctrl_source.create_index(ntrains)
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Correcting RAW data ###
 
 %% Cell type:code id: tags:
 
 ``` python
 # Loop over modules
 for local_karabo_da, mapped_files_module in mapped_files.items():
     instrument_src_kda = instrument_src.format(int(local_karabo_da[-2:]))
 
     for sequence_file in mapped_files_module:
         # Save corrected data in an output file with name
         # of corresponding raw sequence file.
         ofile_name = sequence_file.name.replace("RAW", "CORR")
         out_file = out_folder / ofile_name
 
         # Load sequence file data collection, data.adc keydata,
         # the shape for data to later created arrays of the same shape,
         # and number of available trains to correct.
         seq_dc = H5File(sequence_file)
         seq_dc_adc = seq_dc[instrument_src_kda, "data.adc"]
         ishape = seq_dc_adc.shape  # input shape.
         corr_ntrains = ishape[0]  # number of available trains to correct.
         all_train_ids = seq_dc_adc.train_ids
 
         # Raise a WARNING if this sequence has no trains to correct.
         # Otherwise, print number of trains with no data.
         if corr_ntrains == 0:
             print(f"\t- WARNING: No trains to correct for {sequence_file.name}: "
                   f"Skipping the processing of this file.")
             continue
         elif len(all_train_ids) != corr_ntrains:
             print(f"{sequence_file.name} has {len(seq_dc_adc.train_ids) - corr_ntrains} "
                   "trains with missing data.")
 
         # For testing, limit corrected trains. i.e. Getting output faster.
         if limit_trains > 0:
             print(f"\nCorrected trains are limited to: {limit_trains} trains")
             corr_ntrains = min(corr_ntrains, limit_trains)
 
         print(f"\nNumber of corrected trains are: {corr_ntrains} for {ofile_name}")
 
         # Load constants from the constants dictionary.
         # These arrays are used by `correct_train()` function
         offset_map, mask, gain_map = constants[local_karabo_da]
 
         # Determine total output shape.
         if output_frame_shape is not None:
             oshape = (*ishape[:-2], *output_frame_shape)
         else:
             oshape = ishape
 
         # Allocate shared arrays for corrected data. Used in `correct_train()`
         data_corr = context.alloc(shape=oshape, dtype=np.float32)
         gain_corr = context.alloc(shape=oshape, dtype=np.uint8)
         mask_corr = context.alloc(shape=oshape, dtype=np.uint32)
 
         step_timer.start()
         # Overwrite seq_dc after eliminating empty trains or/and applying limited images.
         seq_dc = seq_dc.select(
             instrument_src_kda, "*", require_all=True).select_trains(np.s_[:corr_ntrains])
 
         # Load raw images(adc), gain, memcells, and frame numbers.
         data = seq_dc[instrument_src_kda, "data.adc"].ndarray()
         gain = seq_dc[instrument_src_kda, "data.gain"].ndarray()
         memcells = seq_dc[instrument_src_kda, "data.memoryCell"].ndarray()
         frame_number = seq_dc[instrument_src_kda, "data.frameNumber"].ndarray()
 
         # Validate that the selected cell id to preview is available in raw data.
         if memory_cells > 1:
             # For plotting, assuming that memory cells are sorted the same for all trains.
             found_cells = memcells[0] == cell_id_preview
             if any(found_cells):
                 cell_idx_preview = np.where(found_cells)[0][0]
             else:
                 print(f"The selected cell_id_preview {cell_id_preview} is not available in burst mode. "
                       f"Previewing cell `{memcells[0]}`.")
                 cell_idx_preview = 0
         else:
             cell_idx_preview = 0
 
         # Correct data per train
         context.map(correct_train, data)
         step_timer.done_step(f"Correction time.")
 
         step_timer.start()
 
         # Create CORR files and add corrected data sections.
         sel_trains = np.isin(all_train_ids, seq_dc.train_ids)
         image_counts = seq_dc[instrument_src_kda, "data.adc"].data_counts(labelled=False)
 
         with DataFile(out_file, 'w') as outp_file:
             # Create INDEX datasets.
             outp_file.create_index(
                 train_ids=seq_dc.train_ids,
                 timestamps=seq_dc.files[0].file["INDEX/timestamp"][sel_trains],
                 flags=seq_dc.files[0].validity_flag[sel_trains])
 
             # Create Instrument section to later add corrected datasets.
             outp_source = outp_file.create_instrument_source(instrument_src_kda)
 
             # Create count/first datasets at INDEX source.
             outp_source.create_index(data=image_counts)
 
             # RAW memoryCell and frameNumber are not corrected. But we are storing only
             # the values for the corrected trains.
             outp_source.create_key(
                 "data.memoryCell", data=memcells,
                 chunks=(min(chunks_ids, memcells.shape[0]), 1))
             outp_source.create_key(
                 "data.frameNumber", data=frame_number,
                 chunks=(min(chunks_ids, frame_number.shape[0]), 1))
             # Add main corrected `data.adc`` dataset and store corrected data.
             outp_source.create_key(
                 "data.adc", data=data_corr,
                 chunks=(min(chunks_data, data_corr.shape[0]), *oshape[1:]))
-
-            write_compressed_frames(
-                gain_corr, outp_file, f"{outp_source.name}/data/gain", comp_threads=8)
-            write_compressed_frames(
-                mask_corr, outp_file, f"{outp_source.name}/data/mask", comp_threads=8)
+            outp_source.create_compressed_key(
+                "data.gain", data=gain_corr)
+            outp_source.create_compressed_key(
+                "data.mask", data=mask_corr)
 
             save_reduced_rois(outp_file, data_corr, mask_corr, local_karabo_da)
 
             # Create METDATA datasets
             outp_file.create_metadata(like=seq_dc)
 
         step_timer.done_step(f'Saving data time.')
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Processing time summary ###
 
 %% Cell type:code id: tags:
 
 ``` python
 print(f"Total processing time {step_timer.timespan():.01f} s")
 step_timer.print_summary()
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 if skip_plots:
     print('Skipping plots')
     import sys
     sys.exit(0)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 # Positions are given in pixels
 mod_width = (256 * 4) + (2 * 3)  # inc. 2px gaps between tiles
 mod_height = (256 * 2) + 2
 if karabo_id == "SPB_IRDA_JF4M":
     # The first 4 modules are rotated 180 degrees relative to the others.
     # We pass the bottom, beam-right corner of the module regardless of its
     # orientation, requiring a subtraction from the symmetric positions we'd
     # otherwise calculate.
     x_start, y_start = 1125, 1078
     module_pos = [
         (x_start - mod_width, y_start - mod_height - (i * (mod_height + 33)))
         for i in range(4)
     ] + [
         (-x_start, -y_start + (i * (mod_height + 33))) for i in range(4)
     ]
     orientations = [(-1, -1) for _ in range(4)] + [(1, 1) for _ in range(4)]
 elif karabo_id == "FXE_XAD_JF1M":
     module_pos = ((-mod_width//2, 33),(-mod_width//2, -mod_height -33))
     orientations = [(-1,-1), (1,1)]
 else:
     module_pos = ((-mod_width//2,-mod_height//2),)
     orientations = None
 
 geom = JUNGFRAUGeometry.from_module_positions(module_pos, orientations=orientations, asic_gap=0)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 first_seq = 0 if sequences == [-1] else sequences[0]
 
 with RunDirectory(out_folder, f"*{run}*S{first_seq:05d}*") as corr_dc:
     # Reading CORR data for plotting.
     jf_corr = components.JUNGFRAU(
         corr_dc,
         detector_name=karabo_id,
     ).select_trains(np.s_[:plot_trains])
     tid, jf_corr_data = next(iter(jf_corr.trains(require_all=True)))
 
 # Shape = [modules, trains, cells, x, y]
 # TODO: Fix the case if not all modules were requested to be corrected.
 # For example if only one modules was corrected. An assertion error is expected
 # at `geom.plot_data_fast`, while plotting corrected images.
 corrected = jf_corr.get_array("data.adc")[:, :, cell_idx_preview, ...].values
 corrected_train = jf_corr_data["data.adc"][
     :, cell_idx_preview, ...
 ].values  # loose the train axis.
 
 mask = jf_corr.get_array("data.mask")[:, :, cell_idx_preview, ...].values
 mask_train = jf_corr_data["data.mask"][:, cell_idx_preview, ...].values
 
 with RunDirectory(f"{in_folder}/r{run:04d}/", f"*S{first_seq:05d}*") as raw_dc:
 
     # Reading RAW data for plotting.
     jf_raw = components.JUNGFRAU(raw_dc, detector_name=karabo_id).select_trains(
         np.s_[:plot_trains]
     )
 
 raw = jf_raw.get_array("data.adc")[:, :, cell_idx_preview, ...].values
 raw_train = (
     jf_raw.select_trains(by_id[[tid]])
     .get_array("data.adc")[:, 0, cell_idx_preview, ...]
     .values
 )
 
 gain = jf_raw.get_array("data.gain")[:, :, cell_idx_preview, ...].values
 gain_train_cells = (
     jf_raw.select_trains(by_id[[tid]]).get_array("data.gain")[:, :, :, ...].values
 )
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 db_modules = get_pdu_from_db(
     karabo_id=karabo_id,
     karabo_da=karabo_da,
     constant=Constants.jungfrau.Offset(),
     condition=condition,
     cal_db_interface=cal_db_interface,
     snapshot_at=creation_time,
 )
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Mean RAW Preview
 
 %% Cell type:code id: tags:
 
 ``` python
 print(f"The per pixel mean of the first {raw.shape[1]} trains of the first sequence file")
 
 fig, ax = plt.subplots(figsize=(18, 10))
 raw_mean = np.mean(raw, axis=1)
 geom.plot_data_fast(
     raw_mean,
     ax=ax,
     vmin=min(0.75*np.median(raw_mean[raw_mean > 0]), 2000),
     vmax=max(1.5*np.median(raw_mean[raw_mean > 0]), 16000),
     cmap="jet",
     colorbar={'shrink': 1, 'pad': 0.01},
 )
 ax.set_title(f'{karabo_id} - Mean RAW', size=18)
 plt.show()
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Mean CORRECTED Preview
 
 %% Cell type:code id: tags:
 
 ``` python
 print(f"The per pixel mean of the first {corrected.shape[1]} trains of the first sequence file")
 
 fig, ax = plt.subplots(figsize=(18, 10))
 corrected_mean = np.mean(corrected, axis=1)
 _corrected_vmin = min(0.75*np.median(corrected_mean[corrected_mean > 0]), -0.5)
 _corrected_vmax = max(2.*np.median(corrected_mean[corrected_mean > 0]), 100)
 
 mean_plot_kwargs = dict(
     vmin=_corrected_vmin, vmax=_corrected_vmax, cmap="jet"
 )
 
 if not strixel_sensor:
     geom.plot_data_fast(
         corrected_mean,
         ax=ax,
         colorbar={'shrink': 1, 'pad': 0.01},
         **mean_plot_kwargs
     )
 else:
     ax.imshow(corrected_mean.squeeze(), aspect=10, **mean_plot_kwargs)
 
 ax.set_title(f'{karabo_id} - Mean CORRECTED', size=18)
 
 plt.show()
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 fig, ax = plt.subplots(figsize=(18, 10))
 corrected_masked = corrected.copy()
 corrected_masked[mask != 0] = np.nan
 corrected_masked_mean = np.nanmean(corrected_masked, axis=1)
 del corrected_masked
 
 if not strixel_sensor:
     geom.plot_data_fast(
         corrected_masked_mean,
         ax=ax,
         colorbar={'shrink': 1, 'pad': 0.01},
         **mean_plot_kwargs
     )
 else:
     ax.imshow(corrected_mean.squeeze(), aspect=10, **mean_plot_kwargs)
 
 ax.set_title(f'{karabo_id} - Mean CORRECTED with mask', size=18)
 
 plt.show()
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 display(Markdown((f"#### A single image from train {tid}")))
 
 fig, ax = plt.subplots(figsize=(18, 10))
 
 single_plot_kwargs = dict(
     vmin=min(0.75 * np.median(corrected_train[corrected_train > 0]), -0.5),
     vmax=max(2.0 * np.median(corrected_train[corrected_train > 0]), 100),
     cmap="jet"
 )
 
 if not strixel_sensor:
     geom.plot_data_fast(
         corrected_train,
         ax=ax,
         colorbar={"shrink": 1, "pad": 0.01},
         **single_plot_kwargs
     )
 else:
     ax.imshow(corrected_train.squeeze(), aspect=10, **single_plot_kwargs)
 
 ax.set_title(f"{karabo_id} - CORRECTED train: {tid}", size=18)
 
 plt.show()
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 def do_2d_plot(data, edges, y_axis, x_axis, title):
     fig = plt.figure(figsize=(10, 10))
     ax = fig.add_subplot(111)
     extent = [
         np.min(edges[1]),
         np.max(edges[1]),
         np.min(edges[0]),
         np.max(edges[0]),
     ]
 
     im = ax.imshow(
         data[::-1, :],
         extent=extent,
         aspect="auto",
         norm=LogNorm(vmin=1, vmax=np.max(data))
     )
     ax.set_xlabel(x_axis)
     ax.set_ylabel(y_axis)
     ax.set_title(title)
     cb = fig.colorbar(im)
     cb.set_label("Counts")
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Gain Bit Value
 
 %% Cell type:code id: tags:
 
 ``` python
 for i, (pdu, mod) in enumerate(zip(db_modules, karabo_da)):
     h, ex, ey = np.histogram2d(
         raw[i].flatten(),
         gain[i].flatten(),
         bins=[100, 4],
         range=[[0, 10000], [0, 4]],
     )
     do_2d_plot(
         h,
         (ex, ey),
         "Signal (ADU)",
         "Gain Bit Value (high gain=0[00], medium gain=1[01], low gain=3[11])",
         f"Module {mod} ({pdu})",
     )
 ```
 
 %% Cell type:markdown id: tags:
 
 ## Signal Distribution ##
 
 %% Cell type:code id: tags:
 
 ``` python
 for i, (pdu, mod) in enumerate(zip(db_modules, karabo_da)):
     fig, axs = plt.subplots(nrows=2, ncols=1, figsize=(18, 10))
     corrected_flatten = corrected[i].flatten()
     for ax, hist_range in zip(axs, [(-100, 1000), (-1000, 10000)]):
         h = ax.hist(
             corrected_flatten,
             bins=1000,
             range=hist_range,
             log=True,
         )
         l = ax.set_xlabel("Signal (keV)")
         l = ax.set_ylabel("Counts")
         _ = ax.set_title(f'Module {mod} ({pdu})')
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Maximum GAIN Preview
 
 %% Cell type:code id: tags:
 
 ``` python
 display(Markdown((f"#### The per pixel maximum of train {tid} of the GAIN data")))
 
 fig, ax = plt.subplots(figsize=(18, 10))
 gain_max = np.max(gain_train_cells, axis=(1, 2))
 geom.plot_data_fast(
     gain_max,
     ax=ax,
     cmap="jet",
     colorbar={'shrink': 1, 'pad': 0.01},
 )
 plt.show()
 ```
 
 %% Cell type:markdown id: tags:
 
 ## Bad Pixels ##
 The mask contains dedicated entries for all pixels and memory cells as well as all three gains stages. Each mask entry is encoded in 32 bits as:
 
 %% Cell type:code id: tags:
 
 ``` python
 table = []
 for item in BadPixels:
     table.append(
         (item.name, f"{item.value:016b}"))
 md = display(Latex(tabulate.tabulate(
     table, tablefmt='latex',
     headers=["Bad pixel type", "Bit mask"])))
 ```
 
 %% Cell type:markdown id: tags:
 
 ### Single Image Bad Pixels ###
 
 A single image bad pixel map for the first image of the first train
 
 %% Cell type:code id: tags:
 
 ``` python
 display(Markdown(f"#### Bad pixels image for train {tid}"))
 
 fig, ax = plt.subplots(figsize=(18, 10))
 if not strixel_sensor:
     geom.plot_data_fast(
         np.log2(mask_train),
         ax=ax,
         vmin=0, vmax=32, cmap="jet",
         colorbar={'shrink': 1, 'pad': 0.01},
     )
 else:
     ax.imshow(np.log2(mask_train).squeeze(), vmin=0, vmax=32, cmap='jet', aspect=10)
 
 plt.show()
 ```