DAMASK_EICMD/processing/post/ascii2hdf5.py

143 lines
5.7 KiB
Python
Executable File

#!/usr/bin/env python2.7
# -*- coding: UTF-8 no BOM -*-
# ------------------------------------------------------------------- #
# NOTE: #
# 1. Not all output is defined in the DS_HDF5.xml, please add new #
# new one to the system wide definition file #
# <DAMASK_ROOT>/lib/damask/DS_HDF5.xml #
# or specify your own when initializing HDF5 class #
# 2. Somehow the point cloud structure cannot be properly handled #
# by Xdmf, which is a descriptive wrapper for visualizing HDF5 #
# using Paraview. The current solution is using cell structured #
# HDF5 so that Xdmf can describe the data shape as a rectangular #
# mesh rather than polyvertex. #
# TODO: #
# 1. remove the <ASCII_TABLE>._tmp file, basically need a way to #
# just load data from ASCII table. #
# 2. a progress monitor when transferring data from ASCII table #
# to HDF5. #
# 3. a more flexible way handle the data structure rather than a #
# xml file. #
# ------------------------------------------------------------------- #
import os
import damask
import numpy as np
from optparse import OptionParser
scriptName = os.path.splitext(os.path.basename(__file__))[0]
scriptID = ' '.join([scriptName, damask.version])
# ----- helper function ----- #
def get_rectMshVectors(xyz_array, posNum):
"""Get Vx, Vy, Vz for rectLinear grid"""
# need some improvement, and only works for rectangular grid
v = sorted(list(set(xyz_array[:, posNum])))
v_interval = (v[2]+v[1])/2.0 - (v[1]+v[0])/2.0
v_start = (v[1]+v[0])/2.0 - v_interval
v_end = (v[-1]+v[-2])/2.0 + v_interval
V = np.linspace(v_start, v_end, len(v)+1)
return V
# ----- MAIN ---- #
desp_msg = "Convert DAMASK ascii table to HDF5 file"
parser = OptionParser(option_class=damask.extendableOption,
usage='%prog options [file[s]]',
description=desp_msg,
version=scriptID)
parser.add_option('-D', '--DefinitionFile',
dest='storage definition file',
type='string',
metavar='string',
help='definition file for H5 data storage')
parser.add_option('-p', '--pos', '--position',
dest='pos',
type='string', metavar='string',
help='label of coordinates [%default]')
parser.set_defaults(DefinitionFile='default',
pos='pos')
(options, filenames) = parser.parse_args()
filename = filenames[0]
if options.DefinitionFile == 'default':
defFile = None
else:
defFile = options.DefinitionFile
# ----- read in data using DAMASK ASCII table class ----- #
asciiTable = damask.ASCIItable(name=filename, buffered=False)
asciiTable.head_read()
asciiTable.data_readArray()
incNum = int(asciiTable.data[asciiTable.label_index('inc'), 0])
fullTable = np.copy(asciiTable.data) # deep copy all data, just to be safe
labels = asciiTable.labels()
labels_idx = [asciiTable.label_index(label) for label in labels]
featuresDim = [labels_idx[i+1] - labels_idx[i] for i in xrange(len(labels)-1)]
featuresDim.append(fullTable.shape[1] - labels_idx[-1])
# ----- figure out size and grid ----- #
pos_idx = asciiTable.label_index('pos')
xyz_array = asciiTable.data[:, pos_idx:pos_idx+3]
Vx = get_rectMshVectors(xyz_array, 0)
Vy = get_rectMshVectors(xyz_array, 1)
Vz = get_rectMshVectors(xyz_array, 2)
# use the dimension of the rectangular grid to reshape all other data
mshGridDim = [len(Vx)-1, len(Vy)-1, len(Vz)-1]
# ----- compose cmd log ----- #
cmd_log = " ".join([scriptID, filename])
# ----- create a new HDF5 file and save the data -----#
# force remove existing HDF5 file
h5fName = filename.replace(".txt", ".h5")
try:
os.remove(h5fName)
except OSError:
pass
h5f = damask.H5Table(h5fName,
new_file=True,
dsXMLFile=defFile)
# adding increment number as root level attributes
h5f.add_attr('inc', incNum)
# add the mesh grid data now
h5f.add_data("Vx", Vx, cmd_log=cmd_log)
h5f.add_data("Vy", Vy, cmd_log=cmd_log)
h5f.add_data("Vz", Vz, cmd_log=cmd_log)
# add the rest of data from table
labelsProcessed = ['inc']
for fi in xrange(len(labels)):
featureName = labels[fi]
# remove trouble maker "("" and ")" from label/feature name
if "(" in featureName:
featureName = featureName.replace("(", "")
if ")" in featureName:
featureName = featureName.replace(")", "")
# skip increment and duplicated columns in the ASCII table
if featureName in labelsProcessed:
continue
featureIdx = labels_idx[fi]
featureDim = featuresDim[fi]
# grab the data hook
dataset = fullTable[:, featureIdx:featureIdx+featureDim]
# mapping 2D data onto a 3D rectangular mesh to get 4D data
# WARNING: In paraview, the data for a recmesh is mapped as:
# --> len(z), len(y), len(x), size(data)
# dataset = dataset.reshape((mshGridDim[0],
# mshGridDim[1],
# mshGridDim[2],
# dataset.shape[1]))
# write out data
print "adding {}...".format(featureName)
h5f.add_data(featureName, dataset, cmd_log=cmd_log)
# write down the processed label
labelsProcessed.append(featureName)