regress_rr.py

import glob
from os import makedirs, mkdir
from os.path import join, exists
import pandas as pd
import numpy as np
import json
import ipdb
import re
import pickle
import sys
import time
from zipfile import ZipFile

import argparse
from datetime import datetime, timedelta, timezone, timedelta
import pytz

import matplotlib.pyplot as plt
from functools import partial
from collections import Counter
from itertools import repeat, chain, combinations
from multiprocessing import Pool, cpu_count
import tensorflow as tf

from sklearn.preprocessing import StandardScaler, MinMaxScaler, OneHotEncoder
from sklearn.preprocessing import PolynomialFeatures, LabelEncoder
from sklearn.model_selection import KFold, train_test_split
from sklearn.metrics import accuracy_score

from tsfresh.feature_extraction import extract_features
from tsfresh.feature_extraction import settings as tsfresh_settings
from tsfresh.utilities.string_manipulation import get_config_from_string

from modules.datapipeline import get_file_list, load_and_snip, load_data, \
        load_split_data, load_harness_data
from modules.digitalsignalprocessing import vectorized_slide_win as vsw
from modules.digitalsignalprocessing import imu_signal_processing
from modules.digitalsignalprocessing import do_pad_fft,\
        pressure_signal_processing, infer_frequency
from modules.utils import *

from modules.evaluations import Evaluation
from modules.datapipeline import get_windowed_data, DataSynchronizer,\
        parallelize_dataframe
from modules.datapipeline import ProjectFileHandler
from models.ardregression import ARDRegressionClass
from models.knn import KNNClass
from models.svm import SVMClass
from models.lda import LDAClass
from models.svr import SVRClass
from models.logisticregression import LogisticRegressionClass
from models.linearregression import LinearRegressionClass
from models.neuralnet import FNN_HyperModel, LSTM_HyperModel, TunerClass,\
        CNN1D_HyperModel
from models.ridgeclass import RidgeClass
from models.resnet import Regressor_RESNET, Classifier_RESNET
from models.xgboostclass import XGBoostClass

from pprint import PrettyPrinter

from sktime.transformations.panel.rocket import (
        MiniRocket,
        MiniRocketMultivariate,
        MiniRocketMultivariateVariable,
)

from config import WINDOW_SIZE, WINDOW_SHIFT, IMU_FS, DATA_DIR, BR_FS

IMU_COLS =  ['acc_x', 'acc_y', 'acc_z', 'gyr_x', 'gyr_y', 'gyr_z']

def utc_to_local(utc_dt, tz=None):
    return utc_dt.replace(tzinfo=timezone.utc).astimezone(tz=tz)

def datetime_from_utc_to_local(utc_datetime):
    now_timestamp = time.time()
    offset = datetime.fromtimestamp(now_timestamp) - datetime.utcfromtimestamp(now_timestamp)
    return utc_datetime + offset

# Load data
def load_bioharness_file(f:str, skiprows=0, skipfooter=0, **kwargs):
    df_list = []
    method = partial(pd.read_csv, skipinitialspace=True,
                     skiprows=list(range(1, skiprows+1)),
                     skipfooter=skipfooter,
                     header=0,
                     **kwargs
                    )
    df = method(f)
    if 'Time' not in df.columns.values:
        df['Time'] = pd.to_datetime(
            df.rename(columns={'Date':'Day'})[
                ['Day','Month','Year']]) \
                + pd.to_timedelta(df['ms'], unit='ms')
        if pd.isna(df['Time']).any():
            df['Time'].interpolate(inplace=True)
        df['Time'] = pd.to_datetime(df['Time'], format="%d/%m/%Y %H:%M:%S.%f")
        df['Time'] = df['Time'].dt.strftime("%d/%m/%Y %H:%M:%S.%f")
    return df

def load_bioharness_files(f_list:list, skiprows=0, skipfooter=0, **kwargs):
    df_list = []
    method = partial(pd.read_csv, skipinitialspace=True,
                     skiprows=list(range(1, skiprows+1)),
                     skipfooter=skipfooter,
                     header=0, **kwargs)
    for f in f_list:
        df_list.append(load_bioharness_file(f))

    df = pd.concat(df_list, ignore_index=True)
    return df

def bioharness_datetime_to_seconds(val):
    fmt = "%d/%m/%Y %H:%M:%S.%f" 
    dstr = datetime.strptime(val, fmt)
    seconds = dstr.timestamp()
    return seconds

def load_imu_file(imu_file:str):
    hdr_file = imu_file.replace('imudata.gz', 'recording.g3')

    df = pd.read_json(imu_file, lines=True, compression='gzip')
    hdr = pd.read_json(hdr_file, orient='index')
    hdr = hdr.to_dict().pop(0)

    if df.empty: return df, hdr

    data_df = pd.DataFrame(df['data'].tolist())
    df = pd.concat([df.drop('data', axis=1), data_df], axis=1)

    iso_tz = hdr['created']
    tzinfo = pytz.timezone(hdr['timezone'])
    # adjust for UTC
    start_time = datetime.fromisoformat(iso_tz[:-1])
    start_time = utc_to_local(start_time, tz=tzinfo).astimezone(tzinfo)

    na_inds = df.loc[pd.isna(df['accelerometer']), :].index.values
    df.drop(index=na_inds, inplace=True)

    imu_times = df['timestamp'].values
    df['timestamp_interp'] = imu_times
    df['timestamp_interp'] = df['timestamp_interp'].interpolate()
    imu_times = df['timestamp_interp'].values
    imu_datetimes = [start_time + timedelta(seconds=val) \
                     for val in imu_times]
    imu_s = np.array([time.timestamp() for time in imu_datetimes])
    df['sec'] = imu_s

    time_check_thold = df['sec'].min() + 3*3600
    mask = df['sec'] > time_check_thold
    if np.any(mask):
        df = df[np.logical_not(mask)]

    return df, hdr

def load_imu_files(f_list:list):
    data, hdr = [], []
    tmp = []
    for f in f_list:
        tmp.append(load_imu_file(f))
    for l in tmp:
        data.append(l[0])
        hdr.append(l[1])
    data_df = pd.concat(data, axis=0)
    return data_df, hdr

def load_e4_file(e4_file:str):
    ''' First row is the initial time of the session as unix time.
    Second row is the sample rate in Hz'''
    zip_file = ZipFile(e4_file)
    dfs = {csv_file.filename: pd.read_csv(zip_file.open(csv_file.filename)
                                          ,header=None)
           for csv_file in zip_file.infolist()
           if csv_file.filename.endswith('.csv')}
    bvp = dfs["BVP.csv"]
    t0 = bvp.iloc[0].values[0]
    fs = bvp.iloc[1].values[0]
    nsamples = len(bvp) - 2

    t0_datetime = datetime.utcfromtimestamp(t0)
    t0_local = datetime_from_utc_to_local(t0_datetime)
    time = [t0_local.timestamp() + ind*(1/fs) for ind in
            range(nsamples)]
    tmp = [np.nan, np.nan]
    time = tmp + time
    bvp.rename(columns={0: "bvp"}, inplace=True)
    bvp['sec'] = np.array(time)

    head = bvp.iloc[[0, 1]]
    bvp.drop(inplace=True, index=[0, 1])

    hdr = {'start_time': head.iloc[0,0],
           'fs': head.iloc[0,1]}

    return bvp, hdr

def load_e4_files(f_list:list):
    tmp = []
    data = []
    hdr = []
    for f in f_list:
        tmp.append(load_e4_file(f))
    for d, h in tmp:
        data.append(d)
        hdr.append(h)
    data_df = pd.concat(data, axis=0)
    return data_df, hdr

# Synchronising data
def sync_to_ref(df0, df1):
    dsync0 = DataSynchronizer()
    dsync1 = DataSynchronizer()

    time0 = df0['sec'].values
    time1 = df1['sec'].values

    t0 = max((time0[0], time1[0]))
    t1 = min((time0[-1], time1[-1]))
    dsync0.set_bounds(time0, t0, t1)
    dsync1.set_bounds(time1, t0, t1)

    return dsync0.sync_df(df0), dsync1.sync_df(df1)

def pss_br_calculations(win, pss_df=None, br_df=None):
    n_out = 5
    if win[-1] == 0: return [None]*n_out

    dsync = DataSynchronizer()
    pss_fs = BR_FS
    pss_col = [col for col in pss_df.columns.values if\
               'breathing' in col.lower()][0]
    pss_ms = pss_df['ms'].values
    br_ms  = br_df['ms'].values
    t0, t1 = pss_ms[win][0], pss_ms[win][-1]

    diff = pss_ms[win][1:] - pss_ms[win][:-1]
    mask = np.abs(diff/1e3) > 60
    diff_chk = np.any(mask)
    if diff_chk: return [None]*n_out

    # Get pressure estimate for window
    pss_win = pss_df.iloc[win]
    pss_data = pss_win[pss_col]
    pss_filt = pressure_signal_processing(pss_data, fs=pss_fs)
    xf, yf = do_pad_fft(pss_filt, fs=pss_fs)
    pss_est = xf[yf.argmax()]*60

    # Sync and get summary br output
    dsync.set_bounds(br_ms, t0, t1)
    br_win = dsync.sync_df(br_df)

    br_out = np.median(br_win['BR'].values)

    # Get subject and condition
    sbj_out = pss_win['subject'].values[0]
    time_out = np.median(pss_win['sec'].values)
    return time_out, pss_est, br_out, sbj_out, cond_out

def get_pss_br_estimates(pss_df, br_df, window_size=12, window_shift=1):
    pss_fs = BR_FS
    # pss_col = [col for col in pss_df.columns.values if\
    #            'breathing' in col.lower()][0]
    pss_ms = pss_df['sec'].values
    br_ms  = br_df['sec'].values

    inds = np.arange(0, len(pss_ms))
    vsw_out = vsw(inds, len(inds), sub_window_size=int(window_size*pss_fs),
                  stride_size=int(window_shift*pss_fs))

    # dsync = DataSynchronizer()
    pss_est, br_out = [], []
    cond_out, sbj_out = [], []
    func = partial(pss_br_calculations, pss_df=pss_df, br_df=br_df)
    # for i, win in enumerate(vsw_out):
    #     tmp = func(win)

    with Pool(cpu_count()) as p:
        tmp = p.map(func, vsw_out)

    time_out, pss_est, br_out, sbj_out, cond_out = zip(*tmp)

    time_array = np.array(time_out)
    pss_est_array = np.array(pss_est)
    br_out_array = np.array(br_out)
    sbj_out_array = np.array(sbj_out)
    cond_out_array = np.array(cond_out)

    df = pd.DataFrame(
        np.array(
            [time_array, sbj_out_array, cond_out_array,
             pss_est_array, br_out_array]
        ).T,
        columns=['ms', 'subject', 'condition', 'pss', 'br'])
    df.dropna(inplace=True)

    return df

# Multiprocessing task for windowing dataframe
def imu_df_win_task(w_inds, df, i, cols):
    time = df['sec'].values
    if w_inds[-1] == 0: return
    w_df = df.iloc[w_inds]
    t0, t1 = time[w_inds][0], time[w_inds][-1]
    diff = time[w_inds[1:]] - time[w_inds[0:-1]]
    mask = np.abs(diff)>20
    diff_chk = np.any(mask)
    if diff_chk:
        return

    # sbj = w_df['subject'].values.astype(int)
    # sbj_mask = np.any((sbj[1:] - sbj[:-1])>0)
    # if sbj_mask:
    #     return

    if cols is None:
        cols = ['acc_x', 'acc_y', 'acc_z',
                'gyr_x', 'gyr_y', 'gyr_z']

    data = w_df[cols].values

    # DSP
    sd_data = (data - np.mean(data, axis=0))/np.std(data, axis=0)
    # ys = cubic_interp(sd_data, BR_FS, FS_RESAMPLE)
    filt_data = imu_signal_processing(sd_data, IMU_FS)
    x_out = pd.DataFrame(filt_data,
                         columns=[
                             'acc_x', 'acc_y', 'acc_z',
                             'gyro_x', 'gyro_y', 'gyro_z',
                         ])

    sm_out = w_df['BR'].values
    ps_out = w_df['PSS'].values

    x_vec_time = np.median(time[w_inds])

    fs = 1/np.mean(diff)
    ps_freq = int(get_max_frequency(ps_out, fs=fs))

    y_tmp = np.array([x_vec_time, np.nanmedian(sm_out), ps_freq])

    x_out['sec'] = x_vec_time
    x_out['id'] = i
    y_out = pd.DataFrame([y_tmp], columns=['sec', 'br', 'pss'])

    return x_out, y_out

def get_max_frequency(data, fs=IMU_FS):
    data = pressure_signal_processing(data, fs=fs)

    xf, yf = do_pad_fft(data, fs=fs)
    max_freq = xf[yf.argmax()]*60
    return max_freq

def convert_to_float(df):
    cols = df.columns.values
    if 'sec' in cols:
        df['sec'] = df['sec'].astype(float)
    if 'pss' in cols:
        df['pss'] = df['pss'].astype(float)
    if 'br' in cols:
        df['br'] = df['br'].astype(float)
    if 'subject' in cols:
        df['subject'] = df['subject'].astype(float)

def load_and_sync_xsens(subject):
    # load imu
    imu_list = get_file_list('imudata.gz', sbj=subject)
    imu_df_all, imu_hdr_df_all = load_imu_files(imu_list)

    # load bioharness
    pss_list = get_file_list('*Breathing.csv', sbj=subject)
    if len(pss_list) == 0:
        pss_list = get_file_list('BR*.csv', sbj=subject)

    br_list = get_file_list('*Summary*', sbj=subject)

    # load e4 wristband
    e4_list = get_file_list('*.zip', sbj=subject)
    bvp_df_all, bvp_hdr = load_e4_files(e4_list)
    bvp_fs = bvp_hdr[0]['fs']

    xsens_list = []
    # skip the first and last x minute(s)
    minutes_to_skip = .5
    br_skiprows = br_skipfooter = int(minutes_to_skip*60)
    pss_skiprows = pss_skipfooter = int(minutes_to_skip*60*BR_FS)
    # load each bioharness file and sync the imu to it
    for pss_file, br_file in zip(pss_list, br_list):
        pss_df = load_bioharness_file(pss_file, skiprows=pss_skiprows,
                                      skipfooter=pss_skipfooter,
                                      engine='python')
        pss_time = pss_df['Time'].map(bioharness_datetime_to_seconds).values\
                .reshape(-1, 1)
        pss_df['sec'] = pss_time

        br_df = load_bioharness_file(br_file, skiprows=br_skiprows,
                                     skipfooter=br_skipfooter,
                                     engine='python')
        br_time = br_df['Time'].map(bioharness_datetime_to_seconds).values\
                .reshape(-1, 1)
        br_df['sec'] = br_time

        # sync
        br_df, imu_df = sync_to_ref(br_df, imu_df_all.copy())
        pss_df, _ = sync_to_ref(pss_df, imu_df_all.copy())
        bvp_df, _ = sync_to_ref(bvp_df_all.copy(), pss_df.copy())

        # extract relevant data
        acc_data = np.stack(imu_df['accelerometer'].values)
        gyr_data = np.stack(imu_df['gyroscope'].values)
        x_time = imu_df['sec'].values.reshape(-1, 1)

        br_col = [col for col in pss_df.columns.values if\
                  'breathing' in col.lower()][0]
        pss_data = pss_df[br_col].values
        pss_data = np.interp(x_time, pss_df['sec'].values, pss_data)\
                .reshape(-1, 1)

        br_lbl = [col for col in br_df.columns.values if\
                  'br' in col.lower()][0]
        br_data = br_df['BR'].values
        br_data = np.interp(x_time, br_df['sec'].values, br_data)\
                .reshape(-1, 1)

        bvp_data = bvp_df['bvp'].values
        bvp_data = np.interp(x_time, bvp_df['sec'].values, bvp_data)\
                .reshape(-1, 1)

        xsens_data = np.concatenate(
            (x_time, br_data, pss_data, bvp_data, acc_data, gyr_data),
            axis=1)

        columns=['sec'   , 'BR'    , 'PSS'   , 'BVP' ,
                 'acc_x' , 'acc_y' , 'acc_z' ,
                 'gyr_x' , 'gyr_y' , 'gyr_z' , ]
        xsens_df_tmp = pd.DataFrame(xsens_data, columns=columns)

        '''
        print("{:.2f}\t{:.2f}\t{:.2f}".format(br_df.sec.iloc[0],
                                              pss_df.sec.iloc[0],
                                              imu_df.sec.iloc[0]))
        print("{:.2f}\t{:.2f}\t{:.2f}".format(br_df.sec.iloc[-1],
                                              pss_df.sec.iloc[-1],
                                              imu_df.sec.iloc[-1]))
        print(xsens_df_tmp.head())
        '''
        xsens_list.append(xsens_df_tmp)

    if len(xsens_list) > 1:
        xsens_df = pd.concat(xsens_list, axis=0, ignore_index=True)
        xsens_df.reset_index(drop=True, inplace=True)
    else:
        xsens_df = xsens_list[0]

    return xsens_df


def load_tsfresh(subject, project_dir,
                 window_size=12, window_shift=0.2, fs=IMU_FS,
                 overwrite=False):
    cols = ['acc_x', 'acc_y', 'acc_z', 'gyro_x', 'gyro_y', 'gyro_z']
    pkl_file = join(project_dir, 'tsfresh.pkl')
    if exists(pkl_file) and not overwrite:
        return pd.read_pickle(pkl_file)

    xsens_df = load_and_sync_xsens(subject)
    x_df, y_df = get_df_windows(xsens_df,
                                imu_df_win_task,
                                window_size=window_size,
                                window_shift=window_shift,
                                fs=fs,
                               )
    x_features_df = extract_features(
        x_df, column_sort='sec',
        column_id='id',
        # default_fc_parameters=tsfresh_settings.MinimalFCParameters(),
    )
    x_features_df.fillna(0, inplace=True)

    cols = x_features_df.columns.values

    df_out = pd.concat([y_df, x_features_df], axis=1)
    df_out.to_pickle(pkl_file)
    return df_out

def get_activity_log(subject):
    activity_list = get_file_list('activity*.csv', sbj=subject)
    activity_dfs = [pd.read_csv(f) for f in activity_list]
    return pd.concat(activity_dfs, axis=0)

def get_respiration_log(subject):
    log_list = get_file_list('*.json', sbj=subject)
    log_dfs = [pd.read_json(f) for f in log_list]
    return pd.concat(log_dfs, axis=0)

def get_cal_data(event_df, xsens_df):
    fmt ="%Y-%m-%d %H.%M.%S" 
    cal_list = []
    cpms = []
    start_sec = 0
    stop_sec = 0
    for index, row in event_df.iterrows():
        event = row['eventTag']
        timestamp = row['timestamp']
        inhalePeriod = row['inhalePeriod']
        exhalePeriod = row['exhalePeriod']

        cpm = np.round( 60/(inhalePeriod + exhalePeriod) )

        sec = timestamp.to_pydatetime().timestamp()

        if event == 'Start':
            start_sec = sec
            continue
        elif event == 'Stop':
            stop_sec = sec

            dsync = DataSynchronizer()
            dsync.set_bounds(xsens_df['sec'].values, start_sec, stop_sec)

            sync_df = dsync.sync_df(xsens_df.copy())
            cal_data = {'cpm': cpm, 'data': sync_df}
            cal_list.append(cal_data)

            assert np.round(sync_df.sec.iloc[0])==np.round(start_sec), \
            "error with start sync"
            assert np.round(sync_df.sec.iloc[-1])==np.round(stop_sec), \
            "error with stop sync"

    return pd.DataFrame(cal_list)

def get_test_data(cal_df, activity_df, xsens_df, test_standing):
    fmt = "%d/%m/%Y %H:%M:%S"
    start_time = cal_df.iloc[-1]['data'].sec.values[-1]
    data_df = xsens_df[xsens_df.sec > start_time]
    activity_start = 0
    activity_end = 0

    activity_list = []

    for index, row in activity_df.iterrows():
        sec = datetime.strptime(row['Timestamps'], fmt).timestamp()
        if not test_standing and row['Activity'] == 'standing':
            continue
        if row['Event'] == 'start':
            activity_start = sec
        elif row['Event'] == 'end':
            activity_stop = sec

            dsync = DataSynchronizer()
            dsync.set_bounds(data_df['sec'].values, activity_start,
                             activity_stop)

            sync_df = dsync.sync_df(data_df.copy())
            activity_data = {'activity': row['Activity'], 'data': sync_df}
            activity_list.append(activity_data)

    return pd.DataFrame(activity_list)

# save evaluation metrics in single file that handles the models for the
# subject and config
class EvalHandler():
    def __init__(self, y_true, y_pred, subject, pfh, mdl_str, overwrite=False):
        self.subject = subject
        self.config = pfh.config
        self.parent_directory = join(DATA_DIR, 'subject_specific')
        self.fset_id = pfh.fset_id
        self.mdl_str = mdl_str
        self.overwrite = overwrite

        self.evals = Evaluation(y_true, y_pred)

        entry = {'subject': self.subject,
                 'config_id': self.fset_id,
                 'mdl_str': self.mdl_str,
                }
        self.entry = {**entry, **self.config, **self.evals.get_evals()}

        self.eval_history_file = join(self.parent_directory,
                                      'eval_history.csv')
        self.eval_hist = self.load_eval_history()

    def load_eval_history(self):
        if not exists(self.eval_history_file):
            return None
        else:
            return pd.read_csv(self.eval_history_file)

    def update_eval_history(self):
        eval_hist = self.eval_hist
        if eval_hist is None:
            eval_hist = pd.DataFrame([self.entry])
        else:
            index_list = eval_hist[
                (eval_hist['subject'] == self.entry['subject']) &\
                (eval_hist['config_id'] == self.entry['config_id']) &\
                (eval_hist['mdl_str'] == self.entry['mdl_str']) &\
                (eval_hist['cpm'] == self.entry['cpm'])\
            ].index.tolist()
            if len(index_list) == 0:
                print("adding new entry")
                eval_hist = eval_hist._append(self.entry, ignore_index=True)
            elif index_list is not None and self.overwrite:
                eval_hist.loc[index_list[0]] = self.entry
        self.eval_hist = eval_hist

    def save_eval_history(self):
        self.eval_hist.to_csv(self.eval_history_file, index=False)

# Train IMU - RR models across subjects
def imu_rr_model(subject,
                 window_size=12,
                 window_shift=0.2,
                 lbl_str='pss',
                 mdl_str='knn',
                 overwrite=False,
                 feature_method='tsfresh',
                 train_len:int=3,
                 test_standing=False,
                ):
    # window_size, window_shift, intra, inter
    cal_str = 'cpm'
    fs = IMU_FS
    tmp = []
    imu_cols = ['acc_x', 'acc_y', 'acc_z', 'gyro_x', 'gyro_y', 'gyro_z']
    bvp_cols= ['bvp']

    # TODO: 
        # implement and input args config by data cols
        # implement and input args config with test_standing
    data_cols = imu_cols + bvp_cols

    do_minirocket = False
    use_tsfresh   = False
    overwrite_tsfresh = True
    train_size = int(train_len)

    if feature_method == 'tsfresh':
        use_tsfresh = True
    elif feature_method == 'minirocket':
        do_minirocket = True

    config = {'window_size'   : window_size,
              'window_shift'  : window_shift,
              'lbl_str'       : lbl_str,
              'do_minirocket' : do_minirocket,
              'use_tsfresh'   : use_tsfresh,
              'train_len'     : train_len,
              'data_cols'     : data_cols,
              'test_standing' : test_standing,
             }

    pfh = ProjectFileHandler(config)
    pfh.set_home_directory(join(DATA_DIR, 'subject_specific', subject))
    pfh.set_parent_directory('imu_rr')
    id_check = pfh.get_id_from_config()
    if id_check is None:
        pfh.set_project_directory()
        pfh.save_metafile()
    else:
        pfh.set_id(int(id_check))
        pfh.set_project_directory()
        print('Using pre-set data id: ', pfh.fset_id)
    project_dir = pfh.project_directory

    if not use_tsfresh:
        xsens_df = load_and_sync_xsens(subject)
    else:
        xsens_df = load_tsfresh(subject,
                                project_dir,
                                window_size=window_size,
                                window_shift=window_shift,
                                fs=IMU_FS,
                                overwrite=overwrite_tsfresh)

    activity_df = get_activity_log(subject)
    event_df = get_respiration_log(subject)

    cal_df = get_cal_data(event_df, xsens_df)

    # include standing or not
    test_df_tmp = get_test_data(cal_df, activity_df, xsens_df, test_standing)
    test_df = pd.concat([df for df in test_df_tmp['data']], axis=0)
    
    for combi in combinations(cal_df[cal_str].values, train_len):
        combi_str = "-".join([str(x) for x in combi])
        pfh.config[cal_str] = combi_str
        marker = f'imu_rr_{subject}_id{pfh.fset_id}_combi{combi_str}'
        print(marker)

        train_df = pd.concat(
            [cal_df[cal_df[cal_str] == cpm]['data'].iloc[0] for cpm in combi],
            axis=0
        )

        assert np.isin(train_df.index.values, test_df.index.values).any()==False,\
                "overlapping test and train data"

        print("train")
        print(train_df.shape)
        print("test")
        print(test_df.shape)

        if do_minirocket:
            x_train_df, y_train_df = get_df_windows(train_df,
                                                    imu_df_win_task,
                                                    window_size=window_size,
                                                    window_shift=window_shift,
                                                    fs=fs,
                                                   )
            x_test_df, y_test_df = get_df_windows(test_df, 
                                                  imu_df_win_task,
                                                  window_size=window_size,
                                                  window_shift=window_shift,
                                                  fs=fs,
                                                 )

            x_train = make_windows_from_id(x_train_df, imu_cols)
            x_test  = make_windows_from_id(x_test_df, imu_cols)
            y_train = y_train_df[lbl_str].values.reshape(-1, 1)
            y_test  = y_test_df[lbl_str].values.reshape(-1, 1)

            print("minirocket transforming...")
            x_train = np.swapaxes(x_train, 1, 2)
            x_test = np.swapaxes(x_test, 1, 2)
            minirocket = MiniRocketMultivariate()
            x_train    = minirocket.fit_transform(x_train)
            x_test     = minirocket.transform(x_test)
        elif use_tsfresh:
            x_train = train_df.iloc[:, 3:].values
            y_train = train_df[lbl_str].values.reshape(-1, 1)
            x_test  = test_df.iloc[:, 3:].values
            y_test  = test_df[lbl_str].values.reshape(-1, 1)
        else:
            x_train_df, y_train_df = get_df_windows(train_df,
                                                    imu_df_win_task,
                                                    window_size=window_size,
                                                    window_shift=window_shift,
                                                    fs=fs,
                                                   )
            x_test_df, y_test_df = get_df_windows(test_df, 
                                                  imu_df_win_task,
                                                  window_size=window_size,
                                                  window_shift=window_shift,
                                                  fs=fs,
                                                 )

            x_train = make_windows_from_id(x_train_df, imu_cols)
            x_test  = make_windows_from_id(x_test_df, imu_cols)
            y_train = y_train_df[lbl_str].values.reshape(-1, 1)
            y_test  = y_test_df[lbl_str].values.reshape(-1, 1)

        transforms, model = model_training(mdl_str, x_train, y_train,
                                           marker, validation_data=None,
                                           overwrite=overwrite,
                                           is_regression=True,
                                           project_directory=project_dir,
                                           window_size=int(window_size*fs),
                                           extra_train=200,
                                          )

        if transforms is not None:
            x_test = transforms.transform(x_test)

        preds = model.predict(x_test)

        eval_handle = EvalHandler(y_test.flatten(), preds.flatten(), subject,
                                  pfh, mdl_str, overwrite=overwrite)
        eval_handle.update_eval_history()
        eval_handle.save_eval_history()

        pp = PrettyPrinter()
        pp.pprint(eval_handle.load_eval_history())

        fig, ax = plt.subplots()
        fig_title = '_'.join([mdl_str, subject]+[combi_str])
        ax.plot(y_test)
        ax.plot(preds)
        ax.set_title(fig_title)
        ax.legend([lbl_str, 'pred'])
        fig_dir = join(project_dir, 'figures')
        if not exists(fig_dir): mkdir(fig_dir)
        fig.savefig(join(fig_dir, fig_title+".png"))
        plt.close()

def arg_parser():
    parser = argparse.ArgumentParser()
    parser.add_argument("-m", '--model', type=str,
                        default='linreg',
                        choices=['linreg', 'ard', 'xgboost', 'knn',
                                 'svr', 'cnn1d', 'fnn', 'lstm', 'ridge',
                                 'elastic'],
                       )
    parser.add_argument("-s", '--subject', type=int,
                        default=2,
                        choices=list(range(1,3))+[-1],
                       )
    parser.add_argument("-f", '--feature_method', type=str,
                        default='minirocket',
                        choices=['tsfresh', 'minirocket', 'None']
                       )
    parser.add_argument("-o", '--overwrite', type=int,
                        default=0,
                       )
    parser.add_argument('--win_size', type=int,
                        default=12,
                       )
    parser.add_argument('--win_shift', type=float,
                        default=0.2,
                       )
    parser.add_argument('-l', '--lbl_str', type=str,
                        default='pss',
                       )
    parser.add_argument('-tl', '--train_len', type=int,
                        default=3,
                        help='minutes of data to use for calibration'
                       )
    args = parser.parse_args()
    return args

if __name__ == '__main__':
    # choose either intra or inter subject features to use for model training
    # '[!M]*'
    np.random.seed(100)
    n_subject_max = 2
    args = arg_parser()

    mdl_str        = args.model
    subject        = args.subject
    feature_method = args.feature_method
    window_size    = args.win_size
    window_shift   = args.win_shift
    lbl_str        = args.lbl_str
    train_len      = args.train_len
    overwrite      = args.overwrite

    print(args)
    assert train_len>0,"--train_len must be an integer greater than 0"

    subject_pre_string = 'Pilot'

    if subject > 0:
        subject = subject_pre_string+str(subject).zfill(2)

        imu_rr_model(subject, window_size=window_size, window_shift=window_shift,
                     lbl_str=lbl_str, mdl_str=mdl_str, overwrite=overwrite,
                     feature_method=feature_method, train_len=train_len
                    )
    else:
        subjects = [subject_pre_string+str(i).zfill(2) for i in \
                    range(1, n_subject_max+1) if i not in imu_issues]
        imu_rr_func = partial(imu_rr_model,
                              window_size=window_size,
                              window_shift=window_shift,
                              lbl_str=lbl_str,
                              mdl_str=mdl_str,
                              overwrite=overwrite,
                              feature_method=feature_method,
                              train_len=train_len
                             )

        if mdl_str in ['fnn', 'lstm', 'cnn1d', 'elastic', 'ard', 'xgboost']:
            for subject in subjects:
                imu_rr_func(subject)
        else:
            ncpu = min(len(subjects), cpu_count())
            with Pool(ncpu) as p:
                p.map(imu_rr_func, subjects)

    print(args)