run_inversion.py

"""
Main code for inverting for the ambient noise source distribution

:copyright:
    noisi development team
:license:
    GNU Lesser General Public License, Version 3 and later
    (https://www.gnu.org/copyleft/lesser.html)
"""

# # Gradient-based iterative Inversion
import yaml
import os
import pprint
import shutil
import time
from pandas import read_csv
import numpy as np
import h5py
import sys
from glob import glob
import matplotlib
import obspy
import csv
from obspy import UTCDateTime
from datetime import date
matplotlib.use('agg')

import noisi
from noisi.util.setup_new import setup_proj
from noisi.scripts.source_grid import setup_sourcegrid
from noisi.scripts.run_sourcesetup import source_setup
from noisi.scripts.correlation import run_corr
from noisi.util.corr_obs_copy import copy_corr
from noisi.scripts.run_measurement import run_measurement
from noisi.scripts.kernel import run_kern
from noisi.scripts.assemble_gradient import assemble_grad
from noisi.util.smoothing import smooth
from noisi.util.inv_step_test import steplengthtest
from noisi.util.output_copy import output_copy
from noisi.util.add_metadata import assign_geographic_metadata
from noisi.util.corr_add_noise import corr_add_noise
from noisi.util.output_plot import output_plot
#from noisi.util.obspy_data_download import download_data_inv
from noisi.util.obspy_mass_download import obspy_mass_downloader
from noisi.util.ants_crosscorrelate import ants_preprocess,ants_crosscorrelate
from noisi.util.compress_files import sac_to_asdf
from noisi.util.compress_files import npy_to_h5
from noisi.scripts.run_wavefieldprep import precomp_wavefield

def precompute_wavefield(args, comm, size, rank):
    pw = precomp_wavefield(args, comm, size, rank)
    pw.precompute()

import functools
print = functools.partial(print, flush=True)


pp = pprint.PrettyPrinter()


from mpi4py import MPI
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()

########################################################################
# argument class
########################################################################
class args(object):
    def __init__(self):
        pass
        
        
inv_args = args()
start_iter = 0


t_0 = time.time()


########################################################################
# read in inversion_config.yml and load attributes
########################################################################

with open(sys.argv[1]) as f:
    inv_config = yaml.safe_load(f)
    
# set attributes for inv_args (ignore source_setup_parameters)
for conf in ['main', 'data_download', 'inversion_config', 'project_config', 'grid_config', 'svp_grid_config', 'auto_data_grid_config', 'wavefield_config', 'source_config', 'measr_config']:
    for key in inv_config[conf]:
        setattr(inv_args,key,inv_config[conf][key]) 
        

inv_args.comm = comm
inv_args.size = size
inv_args.rank = rank

config = inv_config["project_config"]

if inv_args.svp_grid:
    only_ocean = inv_args.svp_only_ocean
else:
    only_ocean = False
    
    
# stationlist absolute path

if inv_args.stationlist is not None:
    inv_args.stationlist = os.path.abspath(inv_args.stationlist)
    inv_args.stationlist_init = os.path.abspath(inv_args.stationlist)

elif inv_args.stationlist is None:
    inv_args.stationlist_init = None


# get noisi folder
noisi_inv_path = os.path.dirname(os.path.dirname(noisi.__file__))

#if rank == 0:
#    print(f"Current directory: {noisi_inv_path}")

    
noisi_path = os.path.join(noisi_inv_path,'noisi')
setattr(inv_args,'noisi_path',noisi_path)

output_path = inv_args.output_folder

if rank == 0:
    if not os.path.isdir(output_path):
        os.makedirs(output_path)

comm.barrier()

os.chdir(output_path)

comm.barrier()

########################################################################
# change to output directory and create project folder
########################################################################

if rank == 0:
    
    print("===="*20)
    print(f"Changing directory to {os.getcwd()}")
    print(f"Setting up project {inv_args.project_name}")
    setup_proj(inv_args,comm,size,rank)
    print("===="*20)


comm.barrier()

# read in config.yml file
configfile_proj = f"./{inv_args.project_name}/config.yml"
with open(configfile_proj,"r") as f:
    config_proj = yaml.safe_load(f)

    
setattr(inv_args,'project_path',config_proj["project_path"])

comm.barrier()

if rank == 0:

    # copy inversion_config.yml to project
    inv_conf_path = os.path.join(noisi_inv_path,sys.argv[1])
    print(f"Copying {inv_conf_path} to project folder")
    src = inv_conf_path
    dst = os.path.join(config_proj["project_path"],"inversion_config.yml")
    shutil.copy(src,dst)
    
    # try to copy batch file
    for file in glob(os.path.join(noisi_inv_path,"run_inversion_batch.sbatch")):
        shutil.copy2(file,config_proj["project_path"])
    
    # update project config file
    config_proj.update(inv_config['auto_data_grid_config'])
    config_proj.update(inv_config['svp_grid_config'])
    config_proj.update(inv_config['wavefield_config'])
    config_proj.update(inv_config['grid_config'])
    config_proj.update(inv_config['project_config'])

    with open(configfile_proj,"w") as f:
        yaml.safe_dump(config_proj,f,sort_keys=False)

    #print("Project config file:")
    #pp.pprint(config_proj)
        
comm.barrier()

# time for project
if rank==0:
    run_time = open(os.path.join(inv_args.project_path,'runtime.txt'),'w+')
    run_time.write(f"Number of cores: {size} \n")
    t_1 = time.time()
    run_time.write(f"Project setup: {np.around((t_1-t_0)/60,4)} \n")

########################################################################
# download (on only one rank) and process data 
########################################################################

if inv_args.download_data:
    
    inv_args.add_metadata = False
    
    # choose date for which data should be downloaded
    if inv_args.download_data_date == "yesterday":
        t_start = UTCDateTime(date.today())-60*60*24*inv_args.download_data_days
        t_end = UTCDateTime(date.today())
    else:
        t_start = UTCDateTime(inv_args.download_data_date)
        t_end = t_start + 60*60*24*inv_args.download_data_days
    
    inv_args.t_start = t_start
    inv_args.t_end = t_end
    
    if rank == 0:
        print("="*25)
        print("Downloading and processing data...")
        print("="*25)

        stationlist_new = obspy_mass_downloader(inv_args)

        # change stationlist to new stationlist
        print(f"Copying new stationlist {stationlist_new} to project folder")
        src = stationlist_new
        dst = os.path.join(config_proj["project_path"],"stationlist.csv")
        inv_args.stationlist = dst
        shutil.copy(src,dst)

    
    comm.barrier()

    # time for data download
    if rank==0:

        t_100 = time.time()
        run_time.write(f"Data Download: {np.around((t_100-t_1)/60,4)} \n")

    comm.barrier()
    
    ants_preprocess(inv_args,comm,size,rank)
    
    comm.barrier()

    if rank==0:

        
        stationlist_var = read_csv(inv_args.stationlist,keep_default_na=False)
        station_dict = dict()

        for sta_i in stationlist_var.iterrows():
            sta = sta_i[1]
            station_dict.update({f"{sta['net']}.{sta['sta']}":[sta['lat'],sta['lon']]})

        
        ###### CHECK PREPROCESSED DATA AND DELETE HIGH AMPLITUDES ########
        proc_path = os.path.join(inv_args.project_path,'data','processed')
        proc_files = glob(os.path.join(proc_path,"*.MSEED"))
        
        proc_delete_count = 0
        
        for file in proc_files:
            
            file_name = os.path.basename(file)
            
            net_1 = file_name.split('.')[0]
            sta_1 = file_name.split('.')[1]
                        
            #print(data)
            st = obspy.read(file)
            #st.filter('bandpass',freqmin=0.01,freqmax=0.05,corners=5)
            st.merge()
            
            data_var = st[0].data
            
            # delete files with amplitudes above 5e-6
            # threshold chosen by look at some data but should be checked
            if np.max(data_var) > 5e-6:
                os.remove(file)
                proc_delete_count += 1
                    
            elif f'{net_1}.{sta_1}' not in list(station_dict.keys()):
                os.remove(file)
                proc_delete_count += 1
                    
                    
        print(f"Deleted {proc_delete_count} of {np.size(proc_files)} processed files with amplitude above 5e-6 or not station in stationlist.")
                    
        # time for data preprocessing
        t_101 = time.time()
        run_time.write(f"Data Pre-processing: {np.around((t_101-t_100)/60,4)} \n")

    comm.barrier()
        
    ants_crosscorrelate(inv_args,comm,size,rank)
    
    comm.barrier()
    
    if rank==0:
        

        ##### CHANGE STATIONLIST TO ONLY HAVE STATIONS WITH CROSS-CORRELATIONS #######
        # open stationlist
        
        stationlist_var = read_csv(inv_args.stationlist,keep_default_na=False)
        station_dict = dict()

        for sta_i in stationlist_var.iterrows():
            sta = sta_i[1]
            station_dict.update({f"{sta['net']}.{sta['sta']}":[sta['lat'],sta['lon']]})

        # get cross-correlation net.sta
        # would probably be better with a set
        corr_net_sta = []
        corr_files = glob(os.path.join(inv_args.project_path,'data/correlations/*SAC'))
        
        stations_csv = [['net','sta','lat','lon']]

                          
        for file in corr_files:

            file_name = os.path.basename(file)

            net_1 = file_name.split('--')[0].split('.')[0]
            sta_1 = file_name.split('--')[0].split('.')[1]
            net_2 = file_name.split('--')[1].split('.')[0]
            sta_2 = file_name.split('--')[1].split('.')[1]

            if f'{net_1}.{sta_1}' not in corr_net_sta:
                
                try:
                    corr_net_sta.append(f'{net_1}.{sta_1}')       
                    stations_csv.append([net_1,sta_1,station_dict[f'{net_1}.{sta_1}'][0],station_dict[f'{net_1}.{sta_1}'][1]])
                except:
                    print(f"Could not add {net_1}.{sta_1} to stationlist. Deleting correlation {file}..")
                    # delete correlation files that are not in stationlist
                    os.remove(file)
                    
            if f'{net_2}.{sta_2}' not in corr_net_sta:
            
                try:
                    corr_net_sta.append(f'{net_2}.{sta_2}')
                    stations_csv.append([net_2,sta_2,station_dict[f'{net_2}.{sta_2}'][0],station_dict[f'{net_2}.{sta_2}'][1]])
                except:
                    print(f"Could not add {net_2}.{sta_2} to stationlist. Deleting correlation {file}..")
                    # delete correlation files that are not in stationlist
                    os.remove(file)


        # write new stationlist
        with open(inv_args.stationlist, 'w') as csvFile:
            writer = csv.writer(csvFile)
            writer.writerows(stations_csv)
        csvFile.close()
                          

        # time for data cross-correlating
        t_102 = time.time()
        run_time.write(f"Data Cross-correlating: {np.around((t_102-t_101)/60,4)} \n")

    # check if there are any stations in stationlist after data download
    stationlist_check = read_csv(inv_args.stationlist)

    # change path to observed cross-correlations
    inv_args.observed_corr = os.path.join(inv_args.project_path,'data','correlations')
    inv_config['inversion_config']['observed_corr'] = inv_args.observed_corr
    
    comm.barrier()

    # Stop the inversion if there is no data
    if stationlist_check.shape[0]<=2:
        if rank == 0:
            print("Not enough data available for inversion.")
        # exit
        sys.exit()

    t_1 = time.time()

    if rank == 0:
        print("="*25)
        print("Downloading and processing data done.")
        print("="*25)
        
    
else:
    if rank == 0 and inv_args.stationlist != None:
        # copy stationlist
        print(f"Copying {inv_args.stationlist} to project folder")
        src = inv_args.stationlist
        dst = os.path.join(config_proj["project_path"],"stationlist.csv")
        
        inv_args.stationlist = dst
        
        shutil.copy(src,dst)

        
comm.barrier()
    
########################################################################
# setup sourcegrid
########################################################################

if rank == 0:
    print("Setting up sourcegrid..")
    # create the source grid
    setup_sourcegrid(inv_args, comm, size, rank)

comm.barrier()

if rank==0:

    # time for sourcegrid
    t_2 = time.time()
    run_time.write(f"Sourcegrid: {np.around((t_2-t_1)/60,4)} \n")

########################################################################
# Convert wavefield
########################################################################


if rank == 0:
    print("Converting wavefield..")
    
if inv_args.wavefield_type == 'greens':

    if hasattr(inv_args, 'wavefield_greens_copy') and not inv_args.wavefield_greens_copy:
        if rank == 0:
            print(f"Not copying already prepared wavefield. Path: {inv_args.wavefield_path}")

    else:
        # copy greens folder to project
        wf_path = os.path.join(inv_args.project_path, 'greens')
        
        if rank == 0:
            print(f"Copying already prepared wavefield from {inv_config['wavefield_config']['wavefield_path']}")
            if os.path.exists(wf_path):
                shutil.rmtree(wf_path)
            shutil.copytree(inv_args.wavefield_path,wf_path)
            print("Copying done.")

        inv_args.wavefield_path = wf_path


    comm.barrier()

    # should check here if wavefield are the same
else:
    precompute_wavefield(inv_args, comm, size, rank)

    wf_path = os.path.join(inv_args.project_path, 'greens')
    inv_args.wavefield_path = wf_path
    #pass
    
comm.barrier()



if rank==0:
    # time for wavefield
    t_3 = time.time()
    run_time.write(f"Wavefield: {np.around((t_3-t_2)/60,4)} \n")


########################################################################
# Setup source directory
########################################################################

### change model observed only = True
source_name = 'source_1'
    
setattr(inv_args,'new_model',True)
inv_args.source_model = os.path.join(inv_args.project_path,source_name)
    
if rank == 0:

    print("Creating source directory..")
    # setup source and change config
    # create the source grid
    

    source_setup(inv_args,comm,size,rank)
    
source_configfile = os.path.join(inv_args.project_path,source_name,'source_config.yml')
measr_configfile = os.path.join(inv_args.project_path,source_name,'measr_config.yml')
source_setup_configfile = os.path.join(inv_args.project_path,source_name,'source_setup_parameters.yml')

comm.barrier()

if rank == 0:
    with open(source_configfile) as f:
        config_source = yaml.safe_load(f)

    with open(measr_configfile) as f:
        config_measr = yaml.safe_load(f)
        
    config_measr.update(inv_config['measr_config'])
    config_source.update(inv_config['source_config'])
    config_source.update({'source_setup_file':source_setup_configfile})
    config_sourcesetup = inv_config['source_setup_config']
    
    
    if inv_args.observed_corr == None:
        config_source['model_observed_only'] = False

    if inv_args.verbose:
        print("Source config files:")
        pp.pprint(config_source)
        pp.pprint(config_sourcesetup)

        print("Measurment config file:")
        pp.pprint(config_measr)
    
    with open(source_configfile,"w") as f:
        yaml.safe_dump(config_source,f,sort_keys=False)

    with open(source_setup_configfile,"w") as f:
        yaml.safe_dump(config_sourcesetup,f,sort_keys=False)
        
    with open(measr_configfile,"w") as f:
        yaml.safe_dump(config_measr,f,sort_keys=False)
        
comm.barrier()

setattr(inv_args,'new_model',False)



########################################################################
# Copy observed cross-correlations
########################################################################

if not os.path.isdir(os.path.join(inv_args.source_model,"observed_correlations_slt")):
    if rank == 0:
        os.makedirs(os.path.join(inv_args.source_model,"observed_correlations_slt"))

if not inv_args.observed_corr == None:
    if rank == 0:
        print("Copying observed correlations")

    n_corr_copy = copy_corr(inv_args,comm,size,rank)

    if n_corr_copy == 0:
        if rank == 0:
            print("Did not copy any observed cross-correlations. Exiting..")

        sys.exit()

comm.barrier()

if rank==0:
    # time for corr copy
    t_4 = time.time()
    run_time.write(f"Copy correlations: {np.around((t_4-t_3)/60,4)} \n")
    run_time.write(f"Number of correlations: {np.size(os.listdir(os.path.join(inv_args.source_model,'observed_correlations')))} \n")
    
    
    
########################################################################
# Setup source distribution
########################################################################


###### need to do initial weight test here #######

config_sourcesetup = inv_config['source_setup_config'][0]

if 'weight_test' in list(config_sourcesetup.keys()) and config_sourcesetup['weight_test']:
    
    if rank == 0:
        print("Performing weight test for initial model..")
    
    # set to steplengthrun to only use subset of correlations
    setattr(inv_args,'steplengthrun',True)
    setattr(inv_args,'step',0)
    
    # make starting model with weight 1
    if rank == 0:
        config_sourcesetup['weight'] = float(1)

        with open(source_setup_configfile,"w+") as f:
            yaml.safe_dump([config_sourcesetup],f,sort_keys=False)
                
    source_setup(inv_args,comm,size,rank)
    
    comm.barrier()
    # copy it and re-use it so starting model isn't recalculate every time
   
    if rank == 0: 
        shutil.copy2(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}','starting_model.h5'),os.path.join(inv_args.source_model,f'iteration_{inv_args.step}','starting_model_1.h5'))
    
    comm.barrier()
    
    model_ones = h5py.File(os.path.join(inv_args.source_model,f'iteration_0','starting_model_1.h5'),'r')['model'][()]
    
    
    # make array of weights
    weights_arr = 10**np.linspace(-7,1,9)
    
    weights_misfit_arr = []
    
    for i,weight in enumerate(weights_arr[::-1]):
        
        # create starting model by multipling starting_model_1.h5 with value
        model_weight = model_ones * weight
        
        if rank == 0:
            
            print(f'Working on {i+1} of {np.size(weights_arr)} weights with {weight:.3e}')

            with h5py.File(os.path.join(inv_args.source_model,f'iteration_0','starting_model.h5')) as fh:
                del fh['model']
                fh.create_dataset('model',data=model_weight.astype(np.float32))
                fh.close() 
        
        comm.barrier()
        
        # compute correlations
        run_corr(inv_args,comm,size,rank)
        
        comm.barrier()
        
        # compute measurements
        run_measurement(inv_args,comm,size,rank)
        
        comm.barrier()
        # get misfit
        measr_step_var = read_csv(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',f'{inv_args.mtype}.0.measurement.csv'))

        l2_norm_all = np.asarray(measr_step_var['l2_norm'])
        l2_norm = l2_norm_all[~np.isnan(l2_norm_all)]
        mf_step_var = np.mean(l2_norm)
                
        weights_misfit_arr.append([weight,mf_step_var])
        
        # delete correlation files
        if rank == 0:
            print(f"Misfit for {i+1}: {mf_step_var:.3e}")

            for file in glob(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}/corr/*.sac')):
                os.remove(file)
                
        comm.barrier()
            
        # stop if the misfit of new one is higher than the previous one
        # assumes we're going from worse to better model, i.e. starting at a strong model
                
        if i > 0  and mf_step_var > np.asarray(weights_misfit_arr).T[1][-2]:
            if rank == 0:
                print("Misfit increased. Stopping weight test.")
            break
            
    # pick the weight
    weights_misfit_arr = np.asarray(sorted(weights_misfit_arr,key = lambda x:x[0],reverse=True)).T
    weights_min = weights_misfit_arr[0][np.argmin(weights_misfit_arr[1])]
        
    # change the weight in the source setup config file and inversion config file
    if rank == 0:
    
        print(f"Final weight: {weights_min:.3e}")

        model_weight = model_ones * weights_min
        
        with h5py.File(os.path.join(inv_args.source_model,f'iteration_0','starting_model.h5'),'r+') as fh:
            del fh['model']
            fh.create_dataset('model',data=model_weight.astype(np.float32))
            fh.close() 
    
        config_sourcesetup['weight'] = float(weights_min)

        with open(source_setup_configfile,"w+") as f:
            yaml.safe_dump([config_sourcesetup],f,sort_keys=False)
            
        with open(os.path.join(inv_args.project_path,'inversion_config.yml')) as f:
            inv_config = yaml.safe_load(f)

        inv_config['source_setup_config'][0]['weight'] = float(weights_min)

        with open(os.path.join(inv_args.project_path,'inversion_config.yml'), 'w') as f:
            yaml.dump(inv_config, f, sort_keys=False)
            
    comm.barrier()
            
    # change back
    setattr(inv_args,'steplengthrun',False)
    
else:
    # initial source distribution has to be given in yaml file
    if rank == 0:        
        print("Setting up source distribution..")

    source_setup(inv_args,comm,size,rank)

    comm.barrier()


if rank==0:
    # time for source setup
    t_5 = time.time()
    run_time.write(f"Source setup: {np.around((t_5-t_4)/60,4)} \n")
    

#sys.exit()

    
########################################################################
# Begin inversion, first check for already computed iterations
########################################################################

# check for already calculated iterations
models = glob(os.path.join(inv_args.source_model,'iteration*.h5'))
steps = [int(os.path.basename(file).split("_")[1][:-3]) for file in models]

# set attribute
setattr(inv_args,'steplengthrun',False)


if steps != []:
    
    if np.max(steps) == inv_args.nr_iterations:
        if rank == 0:
            print("All iterations already calculated. Exiting..")
        sys.exit()
        
    else:
        if rank == 0:
            print("\n")
            print(f"---------- STARTING AT ITERATION {np.max(steps)} ----------")
            print("\n")

        mf_dict = dict()
        
        setattr(inv_args,'step',np.max(steps))
        start_iter = np.max(steps)
        t_9 = time.time()
        
        # Compute misfit
        measr_step_var = read_csv(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',f'{inv_args.mtype}.0.measurement.csv'))

        l2_norm_all = np.asarray(measr_step_var['l2_norm'])
        l2_norm = l2_norm_all[~np.isnan(l2_norm_all)]
        mf_step_var = np.mean(l2_norm)

        mf_dict.update({f'iteration_{inv_args.step}':mf_step_var})
        
        if rank == 0:
            print('Misfit dictionary: ',mf_dict)      
else:
    
    ########################################################################
    # Compute iteration 0. Exits if only forward simulation wanted
    ########################################################################
    setattr(inv_args,'step',0)
    
    
    if rank == 0:
        print("\n")
        print(f"---------- ITERATION 0 ----------")
        print("\n")


    if rank == 0:
        print("Computing cross-correlations..")
        
    run_corr(inv_args,comm,size,rank)
    
    comm.barrier()
    
    if rank==0:
        #time for corr
        t_6 = time.time()
        run_time.write(f"Correlations iteration_0: {np.around((t_6-t_5)/60,4)} \n")

    if inv_args.add_metadata:
        if rank == 0:
            print("Adding metadata to correlations..")

        corr_path = os.path.join(inv_args.source_model,"iteration_0","corr")
        stationlist_path = inv_args.stationlist
        assign_geographic_metadata(corr_path,stationlist_path,comm,size,rank)

    comm.barrier()

    
    if inv_args.add_noise not in [False, None]:
        if rank == 0:
            print("Adding noise to cross-correlations..")
        
        corr_path = os.path.join(inv_args.source_model,"iteration_0","corr")
        corr_add_noise(inv_args,comm,size,rank,corr_path, perc=inv_args.add_noise,method="amp")
            
    comm.barrier()
    


    ########################################################################
    # Correlations
    ########################################################################
    
    if inv_args.observed_corr == None:
        if rank == 0:
            print("Cross-correlations done.")
            
            if inv_args.compress_output_files == True:

                print("Converting SAC files to ASDF..")
                corr_path = os.path.join(inv_args.source_model,f"iteration_0","corr")
                corr_files = glob(os.path.join(corr_path,'*.sac'))
                corr_asdf_file = os.path.join(corr_path,f'corr_iter_0.h5')
                corr_asdf_path = sac_to_asdf(corr_files,corr_asdf_file,n=4)

                print("Removing SAC files..")

                for file in corr_files:
                    os.remove(file)

                print(f"Correlations in {corr_asdf_path}")
            
            
            print("No observed correlations, exiting..")
        sys.exit()

    ########################################################################
    # Measurement
    ########################################################################
        
    # take measurements 
    if rank == 0:
        print("Cross-correlations done.")
        print("Taking measurement..")

    run_measurement(inv_args,comm,size,rank)

    comm.barrier()
    
    if rank==0:
        # time for measurement
        t_7 = time.time()
        run_time.write(f"Measurement iteration_0: {np.around((t_7-t_6)/60,4)} \n")

    # check if there are adjoint sources. If not exit since kernels no kernels will be computed
    adjt_path = os.path.join(inv_args.source_model,'iteration_0/adjt')
    if os.listdir(adjt_path) == []:
        if rank == 0:
            print("!"*30)
            print("No adjoint sources found. Can't compute any sensitivity kernels.")
            
            if inv_args.compress_output_files == True:

                print("Converting SAC files to ASDF..")
                corr_path = os.path.join(inv_args.source_model,f"iteration_0","corr")
                corr_files = glob(os.path.join(corr_path,'*.sac'))
                corr_asdf_file = os.path.join(corr_path,f'corr_iter_0.h5')
                corr_asdf_path = sac_to_asdf(corr_files,corr_asdf_file,n=4)

                print("Removing SAC files..")

                for file in corr_files:
                    os.remove(file)

                print(f"Correlations in {corr_asdf_path}")
            
            
            
            print("Exiting..")
            
        sys.exit()

    ########################################################################
    # Kernels
    ########################################################################
        
    # compute kernels
    if rank == 0:
        print("Measurements taken.")
        print("Computing sensitivity kernels..")

    run_kern(inv_args,comm,size,rank)

    comm.barrier()
    
    if rank==0:
        # time for kern
        t_8 = time.time()
        run_time.write(f"Kernels iteration_0: {np.around((t_8-t_7)/60,4)} \n")

    
    ########################################################################
    # Gradient
    ########################################################################
    
    if rank == 0:
        print("Sensitivity kernels done.")
        print("Assembling gradient..")

    assemble_grad(inv_args,comm,size,rank)

    comm.barrier()

    if rank==0:
        # time for gradient
        t_9 = time.time()
        run_time.write(f"Gradient iteration_0: {np.around((t_9-t_8)/60,4)} \n")

    if rank == 0:
        print("Gradient assembled.")


    if inv_args.nr_iterations == 0:
        if rank == 0:
            
            if inv_args.compress_output_files == True:

                print("Converting SAC to ASDF and NPY to H5..")
                corr_path = os.path.join(inv_args.source_model,"iteration_0","corr")
                corr_files = glob(os.path.join(corr_path,'*.sac'))
                corr_asdf_file = os.path.join(corr_path,f'corr_iter_{inv_args.step}.h5')
                corr_asdf_path = sac_to_asdf(corr_files,corr_asdf_file,n=4)
                
                adjt_path = os.path.join(inv_args.source_model,"iteration_0","adjt")
                adjt_files = glob(os.path.join(adjt_path,'*.sac'))
                adjt_asdf_file = os.path.join(adjt_path,f'adjt_iter_{inv_args.step}.h5')
                adjt_asdf_path = sac_to_asdf(adjt_files,adjt_asdf_file,n=6)
                
                kern_path = os.path.join(inv_args.source_model,"iteration_0","kern")
                kern_files = glob(os.path.join(kern_path,'*.npy'))
                kern_h5_file = os.path.join(kern_path,f'kern_iter_{inv_args.step}.h5')
                kern_h5_path = npy_to_h5(kern_files,kern_h5_file,n=4)
                
                print("Removing SAC/NPY files..")

                for file in corr_files:
                    os.remove(file)
                
                for file in adjt_files:
                    os.remove(file)
                    
                for file in kern_files:
                    os.remove(file)
                    
                print(f"Correlations in {corr_asdf_path}")
                print(f"Adjoint sources in {adjt_asdf_path}")
                print(f"Sensitivity Kernels in {kern_h5_path}")

            print("0 iterations, exiting..")
        
        sys.exit()
                
    else:
        pass

    
    comm.barrier()
            

    # Compute initial misfit
    ########### initial misfit #########
    mf_dict = dict()


    measr_step_var = read_csv(os.path.join(inv_args.source_model,f'iteration_0',f'{inv_config["measr_config"]["mtype"]}.0.measurement.csv'))

    l2_norm_all = np.asarray(measr_step_var['l2_norm'])
    l2_norm = l2_norm_all[~np.isnan(l2_norm_all)]
    mf_step_0 = np.mean(l2_norm)

    mf_dict.update({f'iteration_0':mf_step_0})


    if rank == 0:
        print(f'Misfit for iteration 0: {mf_step_0:.4e}')
        print('Misfit dictionary: ',mf_dict)        

    comm.barrier()
    
########################################################################
# Begin inversion loop
########################################################################
    
    
for iter_nr in range(start_iter, inv_args.nr_iterations):

    setattr(inv_args,'step',iter_nr)

    t_9900 = time.time()
    
    # get source distribution and gradient
    source_distr_path = os.path.join(inv_args.source_model,f'iteration_{iter_nr}/starting_model.h5')
    sourcegrid_path=os.path.join(inv_args.project_path,"sourcegrid.npy")

    source_distr = h5py.File(source_distr_path,'r')
    source_grid = np.asarray(source_distr['coordinates'])

    source_distr_data = np.asarray(source_distr['model'])
    source_distr_max = np.max(source_distr_data)
    
    if np.all(source_distr_data==0):
        source_distr_data_norm = source_distr_data
    else:
        source_distr_data_norm = source_distr_data/np.max(abs(source_distr_data))

    gradient_path = os.path.join(inv_args.source_model,f'iteration_{iter_nr}','grad','grad_all.npy')
    grad_data = np.load(gradient_path)
    grad_data_norm = grad_data[0]/np.max(abs(grad_data[0]))

    grid_full = source_grid


    ########################################################################
    # Smoothing and step length test
    ########################################################################
    
    # smooth the gradient using noisi function
    grad_smooth_path = os.path.join(inv_args.source_model,f'iteration_{iter_nr}','grad','grad_all_smooth.npy')

    smooth_arr = inv_args.step_smooth
    step = np.asarray([i[0] for i in smooth_arr])
    smooth_var = np.asarray([i[1] for i in smooth_arr])

    step_bool = iter_nr < step

    for j in range(np.size(step_bool)):
        if step_bool[j]:
            sigma_degrees = smooth_var[j]
            break
        elif not any(step_bool):
            sigma_degrees = smooth_var[-1]
            break
        else:
            continue    

    sigma=[sigma_degrees*111000]
    
    if rank == 0:
        print(f'Smoothing iteration {iter_nr} with {sigma_degrees} degree smoothing...')
        np.save(os.path.join(inv_args.source_model,f'iteration_{iter_nr}','grad',f'smoothing_iter_{iter_nr}.npy'),sigma_degrees)

    comm.barrier()

    smooth(gradient_path,grad_smooth_path,sourcegrid_path,sigma=sigma,cap=inv_args.smoothing_cap,thresh=1e-10,comm=comm,size=size,rank=rank)
    
    comm.barrier()

    if rank==0:

        t_9901 = time.time()
        run_time.write(f"Smoothing iteration_{inv_args.step}: {np.around((t_9901-t_9900)/60,4)} \n")

    if rank == 0:
        print(f'Smoothing iteration {iter_nr} done.')
        print("Performing step length test..")
        

    # step length test
    step_length,slt_success = steplengthtest(inv_args,comm,size,rank,mf_dict,gradient_path,grad_smooth_path,sourcegrid_path,sigma=sigma,cap=inv_args.smoothing_cap,thresh=1e-10)

    
    comm.barrier()
    
    if not slt_success:
        if rank == 0:
            print('!'*60)
            print(f'!!! Misfit cannot be reduced for iteration {inv_args.step} !!!')
            print(f'!!!!!! Stopping Inversion !!!!!')
            print('!'*60)        
            
        break
    
    
    ########################################################################
    # Update model
    ########################################################################
    if rank==0:

        t_9902 = time.time()
        run_time.write(f"Steplengthtest iteration_{inv_args.step}: {np.around((t_9902-t_9901)/60,4)} \n")

    inv_args.step += 1
    setattr(inv_args,'steplengthrun',False)

    # update noise distribution
    if rank == 0:
        print("Updating noise source distribution..")
        
        grad_smooth = np.load(grad_smooth_path).T
        
        # take minimum value and normalise 
        #grad_smooth_norm = grad_smooth/np.abs(np.min(grad_smooth))
        grad_smooth_norm = grad_smooth/np.max(np.abs(grad_smooth))

        distr_update_norm = source_distr_data_norm - grad_smooth_norm*step_length

        # set all negative values to 0
        distr_update_norm[distr_update_norm < 0] = 0
        
        # scale back up again 
        if source_distr_max == 0:
            distr_update = np.asarray([distr_update_norm][0])
        else:
            distr_update = np.asarray([distr_update_norm*source_distr_max][0])
            
        # square the distribution then normalise and rescale
        #distr_update_2 = np.power(distr_update,2)
        #distr_update_2_norm = distr_update_2/np.max(np.abs(distr_update_2))
        #distr_update = distr_update_2_norm * source_distr_max
        

        # save new distribution in new .h5 file
        sourcemodel_new = os.path.join(inv_args.source_model,f'iteration_{inv_args.step}.h5')
        shutil.copy2(source_distr_path,sourcemodel_new)
            
        
        with h5py.File(sourcemodel_new,'r+') as fh:
            del fh['model']
            fh.create_dataset('model',data=distr_update.astype(np.float32))
                
        print("\n")
        print(f"---------- ITERATION {iter_nr+1} ----------")
        print("\n")
                
        print(f"New sourcemodel for iteration {inv_args.step}: ", os.path.join(inv_args.source_model,f'iteration_{inv_args.step}.h5')) 

        # make step folder
        if not os.path.exists(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}')):
            os.makedirs(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}'))

            for d in ['adjt','grad','corr','kern']:
                os.mkdir(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',d))

        # copy sourcemodel to starting_model.h5 and base_model.h5
        shutil.copy2(sourcemodel_new,os.path.join(inv_args.source_model,f'iteration_{inv_args.step}','starting_model.h5'))

        print(f'Made iteration_{inv_args.step} folder.')
        
    comm.barrier()




    if inv_args.compress_output_files == True and rank == 0:

        print("Converting SAC to ASDF and NPY to H5..")
        corr_path = os.path.join(inv_args.source_model,f'iteration_{inv_args.step-1}',"corr")
        corr_files = glob(os.path.join(corr_path,'*.sac'))
        corr_asdf_file = os.path.join(corr_path,f'corr_iter_{inv_args.step-1}.h5')
        corr_asdf_path = sac_to_asdf(corr_files,corr_asdf_file,n=4)

        adjt_path = os.path.join(inv_args.source_model,f'iteration_{inv_args.step-1}',"adjt")
        adjt_files = glob(os.path.join(adjt_path,'*.sac'))
        adjt_asdf_file = os.path.join(adjt_path,f'adjt_iter_{inv_args.step-1}.h5')
        adjt_asdf_path = sac_to_asdf(adjt_files,adjt_asdf_file,n=6)

        kern_path = os.path.join(inv_args.source_model,f'iteration_{inv_args.step-1}',"kern")
        kern_files = glob(os.path.join(kern_path,'*.npy'))
        kern_h5_file = os.path.join(kern_path,f'kern_iter_{inv_args.step-1}.h5')
        kern_h5_path = npy_to_h5(kern_files,kern_h5_file,n=4)

        print("Removing SAC/NPY files..")

        for file in corr_files:
            os.remove(file)

        for file in adjt_files:
            os.remove(file)

        for file in kern_files:
            os.remove(file)

        print(f"Correlations in {corr_asdf_path}")
        print(f"Adjoint sources in {adjt_asdf_path}")
        print(f"Sensitivity Kernels in {kern_h5_path}")
        
        
    if rank==0:

        t_9903 = time.time()
        run_time.write(f"Noisesource Update iteration_{inv_args.step}: {np.around((t_9903-t_9902)/60,4)} \n")

    ########################################################################
    # Correlations
    ########################################################################
    
    if rank == 0:
        print(f"Computing cross-correlations for iteration {inv_args.step}..")
    
    run_corr(inv_args,comm,size,rank)
    
    comm.barrier()
   
    if rank==0:

        t_9904 = time.time()
        run_time.write(f"Correlations iteration_{inv_args.step}: {np.around((t_9904-t_9903)/60,4)} \n")

    if inv_args.add_noise not in [False, None]:
        if rank == 0:
            print("Adding noise to cross-correlations..")
            
        corr_path = os.path.join(inv_args.source_model,f"iteration_{inv_args.step}","corr")
        corr_add_noise(inv_args,comm,size,rank,corr_path,perc=inv_args.add_noise,method="amp")
            
    comm.barrier()
    
    ########################################################################
    # Measurement
    ########################################################################
 
    if rank == 0:
        print(f"Cross-correlations for iteration {inv_args.step} done.")
        print(f"Taking measurement for iteration {inv_args.step}..")
        
    run_measurement(inv_args,comm,size,rank)

    comm.barrier()
   
    if rank==0:
        t_9905 = time.time()
        run_time.write(f"Measurement iteration_{inv_args.step}: {np.around((t_9905-t_9904)/60,4)} \n")

    ########################################################################
    # Kernel
    ########################################################################
    
    if rank == 0:
        print(f"Measurements for iteration {inv_args.step} taken.")
        print(f"Computing sensitivity kernels for iteration {inv_args.step}..")

    run_kern(inv_args,comm,size,rank)
    
    comm.barrier()
    
    if rank==0:
        t_9906 = time.time()
        run_time.write(f"Kernels iteration_{inv_args.step}: {np.around((t_9906-t_9905)/60,4)} \n")   

    ########################################################################
    # Gradient 
    ########################################################################.
    
    if rank == 0:
        print(f"Sensitivity kernels for iteration {inv_args.step} done.")
        print(f"Assembling gradient for iteration {inv_args.step}..")
    
    assemble_grad(inv_args,comm,size,rank)
    
    comm.barrier()

    if rank==0:
        t_9907 = time.time()
        run_time.write(f"Gradient iteration_{inv_args.step}: {np.around((t_9907-t_9906)/60,4)} \n")

    
    # Compute misfit
    measr_step_var = read_csv(os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',f'{inv_config["measr_config"]["mtype"]}.0.measurement.csv'))

    l2_norm_all = np.asarray(measr_step_var['l2_norm'])
    l2_norm = l2_norm_all[~np.isnan(l2_norm_all)]
    mf_step_var = np.mean(l2_norm)

    mf_dict.update({f'iteration_{inv_args.step}':mf_step_var})

    if rank == 0:
        print(f'Misfit for iteration {inv_args.step}: {mf_step_0:.4e}')
        print('Misfit dictionary: ',mf_dict)    
        

if inv_args.compress_output_files == True and rank == 0:

    print("Converting SAC to ASDF and NPY to H5..")
    corr_path = os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',"corr")
    corr_files = glob(os.path.join(corr_path,'*.sac'))
    corr_asdf_file = os.path.join(corr_path,f'corr_iter_{inv_args.step}.h5')
    corr_asdf_path = sac_to_asdf(corr_files,corr_asdf_file,n=4)

    adjt_path = os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',"adjt")
    adjt_files = glob(os.path.join(adjt_path,'*.sac'))
    adjt_asdf_file = os.path.join(adjt_path,f'adjt_iter_{inv_args.step}.h5')
    adjt_asdf_path = sac_to_asdf(adjt_files,adjt_asdf_file,n=6)

    kern_path = os.path.join(inv_args.source_model,f'iteration_{inv_args.step}',"kern")
    kern_files = glob(os.path.join(kern_path,'*.npy'))
    kern_h5_file = os.path.join(kern_path,f'kern_iter_{inv_args.step}.h5')
    kern_h5_path = npy_to_h5(kern_files,kern_h5_file,n=4)

    print("Removing SAC/NPY files..")

    for file in corr_files:
        os.remove(file)

    for file in adjt_files:
        os.remove(file)

    for file in kern_files:
        os.remove(file)

    print(f"Correlations in {corr_asdf_path}")
    print(f"Adjoint sources in {adjt_asdf_path}")
    print(f"Sensitivity Kernels in {kern_h5_path}")

        
        
if rank == 0:
    print("\n")
    print(f"---------- ITERATIONS DONE ----------")
    print("\n")
    
    print(f"Inversion done. Misfit reduced from {list(mf_dict.values())[0]:.4e} to {list(mf_dict.values())[-1]:.4e}")

        # inversion time
    t_10 = time.time()
    run_time.write(f"Inversion: {np.around((t_10-t_9)/60,4)} \n")
    run_time.write(f"Total runtime: {np.around((t_10-t_0)/60,4)} \n")
    run_time.close()

    print("Copying files..")
    output_copy(inv_args.project_path)
    print("Files copied to output folder.")

comm.barrier()

if inv_args.output_plot:
    output_plot(inv_args,os.path.join(inv_args.project_path,"output"),only_ocean=only_ocean,triangulation=True)
    
comm.barrier()
        
if rank == 0:

    # make a csv file with names of observed correlations
    corr_name_file = os.path.join(inv_args.project_path,"output/used_obs_corr_list.csv")
    
    corr_name_list = [[file] for file in os.listdir(os.path.join(inv_args.source_model,'observed_correlations'))]
    
    # write new stationlist
    with open(corr_name_file, 'w') as csvFile:
        writer = csv.writer(csvFile)
        writer.writerows(corr_name_list)
    csvFile.close()

    

    if inv_args.compress_output_files == 'delete':
        print("Deleting correlation and kernel files..")

        for file in glob(os.path.join(inv_args.source_model,'iteration_*/corr/*.sac')):
            os.remove(file)

        for file in glob(os.path.join(inv_args.source_model,'iteration_*/adjt/*.sac')):
            os.remove(file)

        for file in glob(os.path.join(inv_args.source_model,'iteration_*/kern/*.0.npy')):
            os.remove(file)

        print("Files removed.")


        
comm.barrier()

if rank == 0:
    print("================ INVERSION SCRIPT DONE ===================")