def trialbytrial_drive(
mouse,
trace_type='zscore_day',
method='ncp_hals',
cs='',
warp=False,
word='tray',
group_by='all2',
nan_thresh=0.85,
score_threshold=0.8,
drive_type='visual'):
"""
Create a cells x trials array that contains the log inverse p-values
of each trial compared to the distributions of the baseline response
for each cell. Stats us a two tailed KS test.
This is a wrapped function so that a flow Mouse object is correctly
passed to the memoizer and MongoDB database.
"""
drive_mat = trialbytrial_drive_sub(
flow.Mouse(mouse=mouse),
trace_type=trace_type,
method=method,
cs=cs,
warp=warp,
word=word,
group_by=group_by,
nan_thresh=nan_thresh,
score_threshold=score_threshold,
drive_type=drive_type)
return drive_mat
def run_psytrack(mouse, pars=pars_simp_th, include_pavlovian=False):
"""
Simple function for running psytrack behavioral model on a single mouse
"""
# upadate parameters to reflect pavlovian kwarg
pars['include_pavlovian'] = include_pavlovian
psy = flow.Mouse(mouse=mi).psytracker(verbose=True, force=force, pars=pars)
return psy
def groupmouse_fit_disengaged_sated_mean_per_comp(
mice=['OA27', 'OA26', 'OA67', 'VF226', 'CC175'],
trace_type='zscore_day',
method='mncp_hals',
cs='',
warp=False,
words=['orlando', 'already', 'already', 'already', 'already'],
group_by='all',
nan_thresh=0.85,
score_threshold=None,
random_state=None,
init='rand',
rank=18,
verbose=False):
"""
Wrapper function for fit_disengaged_sated_mean_per_comp. Gets a dataframe
of all mice.
"""
mouse_list = []
for m, w in zip(mice, words):
mouse = flow.Mouse(mouse=m)
mouse_list.append(
fit_disengaged_sated_mean_per_comp(mouse,
word=w,
rank=rank,
nan_thresh=nan_thresh,
score_threshold=score_threshold,
method=method,
cs=cs,
warp=warp,
group_by=group_by,
trace_type=trace_type,
random_state=random_state,
init=init,
verbose=verbose))
dfmouse = pd.concat(mouse_list, axis=0)
return dfmouse
def is_center_of_mass_visual(mouse,
trace_type='zscore_day',
method='mncp_hals',
cs='',
warp=False,
word='restaurant',
group_by='all',
nan_thresh=0.85,
score_threshold=0.8,
rank_num=15):
"""
Check if the center of mass of your temporal factors is beyond the
offset of the visual stimulus.
"""
# load your data
cm_kwargs = {
'mouse': flow.Mouse(mouse=mouse),
'method': method,
'cs': cs,
'warp': warp,
'word': word,
'group_by': group_by,
'nan_thresh': nan_thresh,
'score_threshold': score_threshold,
'rank_num': rank_num
}
cm_df = center_of_mass_tempofac(**cm_kwargs)
# set the stimulus offset time (stimulus length + baseline length)
if mouse in ['OA32', 'OA34', 'OA37', 'OA36', 'CB173', 'AS20', 'AS41']:
off_time = 2 + 1 # add 1 for the second before stimulus onset
else:
off_time = 3 + 1
return cm_df['center_of_mass'].values <= 15 * off_time
def varex_norm_bycomp_byday(mouse,
trace_type='zscore_day',
method='mncp_hals',
cs='',
warp=False,
word='orlando',
group_by='all',
nan_thresh=0.85,
rank_num=18,
rectified=True,
verbose=False):
# necessary parameters for determining type of analysis
pars = {'trace_type': trace_type, 'cs': cs, 'warp': warp}
group_pars = {'group_by': group_by}
# save dir
# if cells were removed with too many nan trials
if nan_thresh:
nt_tag = '_nantrial' + str(nan_thresh)
nt_save_tag = ' nantrial ' + str(nan_thresh)
else:
nt_tag = ''
nt_save_tag = ''
# save tag for rectification
if rectified:
r_tag = ' rectified'
r_save_tag = '_rectified'
else:
r_tag = ''
r_save_tag = ''
save_dir = paths.tca_plots(mouse,
'group',
pars=pars,
word=word,
group_pars=group_pars)
save_dir = os.path.join(save_dir, 'varex' + nt_save_tag + r_tag)
if not os.path.isdir(save_dir): os.mkdir(save_dir)
save_dir = os.path.join(save_dir, 'byday_bycomp')
if not os.path.isdir(save_dir): os.mkdir(save_dir)
var_path = os.path.join(
save_dir,
str(mouse) + '_rank' + str(rank_num) + '_norm_varex_by_day' +
r_save_tag + nt_save_tag + '.pdf')
# load dir
load_dir = paths.tca_path(mouse,
'group',
pars=pars,
word=word,
group_pars=group_pars)
tensor_path = os.path.join(
load_dir,
str(mouse) + '_' + str(group_by) + nt_tag + '_group_decomp_' +
str(trace_type) + '.npy')
input_tensor_path = os.path.join(
load_dir,
str(mouse) + '_' + str(group_by) + nt_tag + '_group_tensor_' +
str(trace_type) + '.npy')
ids_tensor_path = os.path.join(
load_dir,
str(mouse) + '_' + str(group_by) + nt_tag + '_group_ids_' +
str(trace_type) + '.npy')
meta_path = os.path.join(
load_dir,
str(mouse) + '_' + str(group_by) + nt_tag + '_df_group_meta.pkl')
# load your data
# ensemble = np.load(tensor_path)
# ensemble = ensemble.item()
# # re-balance your factors ()
# print('Re-balancing factors.')
# for r in ensemble[method].results:
# for i in range(len(ensemble[method].results[r])):
# ensemble[method].results[r][i].factors.rebalance()
# V = ensemble[method]
# X = np.load(input_tensor_path)
# meta = pd.read_pickle(meta_path)
# meta = utils.update_naive_cs(meta)
# orientation = meta.reset_index()['orientation']
# condition = meta.reset_index()['condition']
# speed = meta.reset_index()['speed']
# dates = meta.reset_index()['date']
# # time_in_trial = meta.reset_index()['trial_idx']
# # total_time = pd.DataFrame(data={'total_time': np.arange(len(time_in_trial))}, index=time_in_trial.index)
# learning_state = meta['learning_state']
# trialerror = meta['trialerror']
# ids = np.load(ids_tensor_path)
# create dataframe of dprime values
# dprime_vec = []
# for date in dates:
# date_obj = flow.Date(mouse, date=date)
# dprime_vec.append(pool.calc.performance.dprime(date_obj))
# data = {'dprime': dprime_vec}
# dprime = pd.DataFrame(data=data, index=speed.index)
# dprime = dprime['dprime'] # make indices match to meta
test = calc.var.groupday_varex_byday_bycomp(flow.Mouse(
mouse=mouse, exclude_tags=['bad']),
word=word)
test3 = calc.var.groupday_varex_byday(flow.Mouse(mouse=mouse,
exclude_tags=['bad']),
word=word)
# test = cas.calc.var.groupday_varex_byday_bycomp(flow.Mouse(mouse='VF226'), word='already')
# test3 = cas.calc.var.groupday_varex_byday(flow.Mouse(mouse='VF226'), word='already')
R = rank_num
comp_var_df = test.loc[test['rank'] == R, :]
col_var = deepcopy(comp_var_df['variance_explained_tcamodel'].values)
new_col_var = deepcopy(comp_var_df['variance_explained_tcamodel'].values)
new_col_dates = deepcopy(comp_var_df['date'].values)
daily_var_df = test3.loc[test3['rank'] == R, :]
daily_var_lookup = deepcopy(
daily_var_df['variance_explained_tcamodel'].values)
for c, day in enumerate(np.unique(new_col_dates)):
new_col_dates[new_col_dates == day] = c
new_col_var[comp_var_df['date'].values == day] = (
col_var[comp_var_df['date'].values == day] /
daily_var_lookup[daily_var_df['date'].values == day])
comp_var_df['norm_varex'] = new_col_var
comp_var_df['day_num'] = new_col_dates
g = sns.relplot(x="day_num",
y="norm_varex",
hue="component",
data=comp_var_df.loc[(comp_var_df['day_num'] >= 0)
& (comp_var_df['day_num'] <= 100), :],
legend='full',
kind='line',
alpha=0.8,
palette=sns.color_palette('muted', R),
marker='o')
plt.title(mouse +
': Fraction of total daily variance explained per component')
plt.savefig(var_path, bbox_inches='tight')
def _splice_data_inputs(psydata,
mouse,
trace_type='zscore_day',
method='mncp_hals',
cs='',
warp=False,
word=None,
group_by='all',
nan_thresh=0.85,
score_threshold=0.8,
rank_num=18,
verbose=True):
"""
Create dict used for fitting Pillow model. Main purpose of function
is to align indices from Pillow and TCA since they often sub-select
different trials. This forces Pillow 'y' and 'answer' to have same
trials as TCA and uses TCA trial factors as 'inputs'.
"""
# default TCA params to use
if not word:
if mouse == 'OA27':
word = 'tray'
else:
word = 'obligations' # should be updated to 'obligations'
if verbose:
print('Creating dataframe for ' + mouse + '-' + word)
ms = flow.Mouse(mouse)
psy = ms.psytracker(verbose=True)
dateRuns = psy.data['dateRuns']
trialRuns = psy.data['runLength']
# create your trial indices per day and run
trial_idx = []
for i in trialRuns:
trial_idx.extend(range(i))
# get date and run vectors
date_vec = []
run_vec = []
for c, i in enumerate(dateRuns):
date_vec.extend([i[0]] * trialRuns[c])
run_vec.extend([i[1]] * trialRuns[c])
# create your data dict, transform from log odds to odds ratio
data = {}
for c, i in enumerate(psy.weight_labels):
# adding multiplication step here with binary vector !!!!!!
data[i] = np.exp(psy.fits[c, :]) * psy.inputs[:, c].T
ori_0_in = [i[0] for i in psy.data['inputs']['ori_0']]
ori_135_in = [i[0] for i in psy.data['inputs']['ori_135']]
ori_270_in = [i[0] for i in psy.data['inputs']['ori_270']]
blank_in = [
0 if i == 1 else 1
for i in np.sum((ori_0_in, ori_135_in, ori_270_in), axis=0)
]
# loop through psy data create a binary vectors for trial history
binary_cat = ['ori_0', 'ori_135', 'ori_270', 'prev_reward', 'prev_punish']
for cat in binary_cat:
data[cat + '_th'] = [i[0] for i in psy.data['inputs'][cat]]
data[cat + '_th_prev'] = [i[1] for i in psy.data['inputs'][cat]]
# create a single list of orientations to match format of meta
ori_order = [0, 135, 270, -1]
data['orientation'] = [
ori_order[np.where(np.isin(i, 1))[0][0]]
for i in zip(ori_0_in, ori_135_in, ori_270_in, blank_in)
]
# create your index out of relevant variables
index = pd.MultiIndex.from_arrays(
[[mouse] * len(trial_idx), date_vec, run_vec, trial_idx],
names=['mouse', 'date', 'run', 'trial_idx'])
# make master dataframe
dfr = pd.DataFrame(data, index=index)
# load TCA data
load_kwargs = {
'mouse': mouse,
'method': method,
'cs': cs,
'warp': warp,
'word': word,
'group_by': group_by,
'nan_thresh': nan_thresh,
'score_threshold': score_threshold,
'rank': rank_num
}
tensor, _, _ = load.groupday_tca_model(**load_kwargs)
meta = load.groupday_tca_meta(**load_kwargs)
# add in continuous dprime so psytracker data frame
dp = pool.calc.psytrack.dprime(flow.Mouse(mouse))
dfr['dprime'] = dp
# add in non continuous dprime to meta dataframe
meta = utils.add_dprime_to_meta(meta)
# filter out blank trials
blank_trials_bool = (dfr['orientation'] >= 0)
psy_df = dfr.loc[blank_trials_bool, :]
# check that all runs have matched trial orientations
new_psy_df_list = []
new_meta_df_list = []
drop_trials_bin = np.zeros((len(psy_df)))
dates = meta.reset_index()['date'].unique()
for d in dates:
psy_day_bool = psy_df.reset_index()['date'].isin([d]).values
meta_day_bool = meta.reset_index()['date'].isin([d]).values
psy_day_df = psy_df.iloc[psy_day_bool, :]
meta_day_df = meta.iloc[meta_day_bool, :]
runs = meta_day_df.reset_index()['run'].unique()
drop_pos_day = np.where(psy_day_bool)[0]
for r in runs:
psy_run_bool = psy_day_df.reset_index()['run'].isin([r]).values
meta_run_bool = meta_day_df.reset_index()['run'].isin([r]).values
psy_run_df = psy_day_df.iloc[psy_run_bool, :]
meta_run_df = meta_day_df.iloc[meta_run_bool, :]
psy_run_idx = psy_run_df.reset_index()['trial_idx'].values
meta_run_idx = meta_run_df.reset_index()['trial_idx'].values
# drop extra trials from trace2P that don't have associated imaging
max_trials = np.min([len(psy_run_idx), len(meta_run_idx)])
# get just your orientations for checking that trials are matched
meta_ori = meta_run_df['orientation'].iloc[:max_trials]
psy_ori = psy_run_df['orientation'].iloc[:max_trials]
# make sure all oris match between vectors of the same length each day
assert np.all(psy_ori.values == meta_ori.values)
# check which trials are dropped
drop_pos_run = drop_pos_day[psy_run_bool][:max_trials]
drop_trials_bin[drop_pos_run] = 1
# if everything looks good, copy meta index into psy
meta_new = meta_run_df.iloc[:max_trials]
psy_new = psy_run_df.iloc[:max_trials]
data = {}
for i in psy_new.columns:
data[i] = psy_new[i].values
new_psy_df_list.append(
pd.DataFrame(data=data, index=meta_new.index))
new_meta_df_list.append(meta_new)
meta1 = pd.concat(new_meta_df_list, axis=0)
psy1 = pd.concat(new_psy_df_list, axis=0)
tca_data = {}
for comp_num in range(1, rank_num + 1):
fac = tensor.results[rank_num][0].factors[2][:, comp_num - 1]
tca_data['factor_' + str(comp_num)] = fac[:, None]
# which values were dropped from the psydata. Use this to update psydata
blank_trials_bool[blank_trials_bool] = (drop_trials_bin == 1)
keep_bool = blank_trials_bool
# you don't have any blank trials so drop them.
psydata['y'] = psydata['y'][keep_bool] # 1-2 binary not 0-1
psydata['answer'] = psydata['answer'][keep_bool]
psydata['correct'] = psydata['correct'][keep_bool]
psydata['dateRunTrials'] = psydata['dateRunTrials'][keep_bool]
# recalculate dayLength and runLength and dateRuns
new_runLength = []
new_dayLength = []
new_dateRuns = []
for di in np.unique(psydata['dateRunTrials'][:, 0]):
day_bool = psydata['dateRunTrials'][:, 0] == di
new_dayLength.append(np.sum(day_bool))
day_runs = psydata['dateRunTrials'][day_bool, 1]
for ri in np.unique(day_runs):
run_bool = day_runs == ri
new_runLength.append(np.sum(run_bool))
new_dateRuns.append([di, ri])
psydata['dayLength'] = new_dayLength
psydata['runLength'] = new_runLength
psydata['dateRuns'] = np.array(new_dateRuns)
# update days
clean_days = np.unique(psydata['dateRunTrials'][:, 0])
clean_day_bool = np.isin(psydata['days'], clean_days)
psydata['days'] = psydata['days'][clean_day_bool]
# ensure that you still have the same number of runs
assert len(psydata['runLength']) == len(psydata['dateRuns'])
# ensure that you still have the same number of days
assert len(psydata['dayLength']) == len(psydata['days'])
# reset inputs
psydata['inputs'] = tca_data
if verbose:
print('Successful sync of psytracker and TCA data :)')
print(' Fitting {} days'.format(len(psydata['days'])))
print(' Fitting {} runs'.format(len(psydata['dateRuns'])))
print(' Fitting {} trials'.format(len(psydata['dateRunTrials'])))
print(' Fitting {} total hyper-parameters'.format(rank_num))
return psydata
def _splice_data_y(psydata,
mouse,
trace_type='zscore_day',
method='mncp_hals',
cs='',
warp=False,
word=None,
group_by='all',
nan_thresh=0.85,
score_threshold=0.8,
rank_num=18,
comp_num=1,
verbose=True):
"""
Create a pandas dataframe of trial history modulation indices for one
mouse. Only looks at initial learning stage.
"""
# default TCA params to use
if not word:
if mouse == 'OA27':
word = 'tray'
else:
word = 'obligations' # should be updated to 'obligations'
if verbose:
print('Creating dataframe for ' + mouse + '-' + word)
ms = flow.Mouse(mouse)
psy = ms.psytracker(verbose=True)
dateRuns = psy.data['dateRuns']
trialRuns = psy.data['runLength']
# create your trial indices per day and run
trial_idx = []
for i in trialRuns:
trial_idx.extend(range(i))
# get date and run vectors
date_vec = []
run_vec = []
for c, i in enumerate(dateRuns):
date_vec.extend([i[0]] * trialRuns[c])
run_vec.extend([i[1]] * trialRuns[c])
# create your data dict, transform from log odds to odds ratio
data = {}
for c, i in enumerate(psy.weight_labels):
# adding multiplication step here with binary vector !!!!!!
data[i] = np.exp(psy.fits[c, :]) * psy.inputs[:, c].T
ori_0_in = [i[0] for i in psy.data['inputs']['ori_0']]
ori_135_in = [i[0] for i in psy.data['inputs']['ori_135']]
ori_270_in = [i[0] for i in psy.data['inputs']['ori_270']]
blank_in = [
0 if i == 1 else 1
for i in np.sum((ori_0_in, ori_135_in, ori_270_in), axis=0)
]
# loop through psy data create a binary vectors for trial history
binary_cat = ['ori_0', 'ori_135', 'ori_270', 'prev_reward', 'prev_punish']
for cat in binary_cat:
data[cat + '_th'] = [i[0] for i in psy.data['inputs'][cat]]
data[cat + '_th_prev'] = [i[1] for i in psy.data['inputs'][cat]]
# create a single list of orientations to match format of meta
ori_order = [0, 135, 270, -1]
data['orientation'] = [
ori_order[np.where(np.isin(i, 1))[0][0]]
for i in zip(ori_0_in, ori_135_in, ori_270_in, blank_in)
]
# create your index out of relevant variables
index = pd.MultiIndex.from_arrays(
[[mouse] * len(trial_idx), date_vec, run_vec, trial_idx],
names=['mouse', 'date', 'run', 'trial_idx'])
# make master dataframe
dfr = pd.DataFrame(data, index=index)
# load TCA data
load_kwargs = {
'mouse': mouse,
'method': method,
'cs': cs,
'warp': warp,
'word': word,
'group_by': group_by,
'nan_thresh': nan_thresh,
'score_threshold': score_threshold,
'rank': rank_num
}
tensor, _, _ = load.groupday_tca_model(**load_kwargs)
meta = load.groupday_tca_meta(**load_kwargs)
# add in continuous dprime so psytracker data frame
dp = pool.calc.psytrack.dprime(flow.Mouse(mouse))
dfr['dprime'] = dp
# add in non continuous dprime to meta dataframe
meta = utils.add_dprime_to_meta(meta)
# filter out blank trials
blank_trials_bool = (dfr['orientation'] >= 0)
psy_df = dfr.loc[blank_trials_bool, :]
# check that all runs have matched trial orientations
new_psy_df_list = []
new_meta_df_list = []
drop_trials_bin = np.zeros((len(psy_df)))
dates = meta.reset_index()['date'].unique()
for d in dates:
psy_day_bool = psy_df.reset_index()['date'].isin([d]).values
meta_day_bool = meta.reset_index()['date'].isin([d]).values
psy_day_df = psy_df.iloc[psy_day_bool, :]
meta_day_df = meta.iloc[meta_day_bool, :]
runs = meta_day_df.reset_index()['run'].unique()
drop_pos_day = np.where(psy_day_bool)[0]
for r in runs:
psy_run_bool = psy_day_df.reset_index()['run'].isin([r]).values
meta_run_bool = meta_day_df.reset_index()['run'].isin([r]).values
psy_run_df = psy_day_df.iloc[psy_run_bool, :]
meta_run_df = meta_day_df.iloc[meta_run_bool, :]
psy_run_idx = psy_run_df.reset_index()['trial_idx'].values
meta_run_idx = meta_run_df.reset_index()['trial_idx'].values
# drop extra trials from trace2P that don't have associated imaging
max_trials = np.min([len(psy_run_idx), len(meta_run_idx)])
# get just your orientations for checking that trials are matched
meta_ori = meta_run_df['orientation'].iloc[:max_trials]
psy_ori = psy_run_df['orientation'].iloc[:max_trials]
# make sure all oris match between vectors of the same length each day
assert np.all(psy_ori.values == meta_ori.values)
# check which trials are dropped
drop_pos_run = drop_pos_day[psy_run_bool][:max_trials]
drop_trials_bin[drop_pos_run] = 1
# if everything looks good, copy meta index into psy
meta_new = meta_run_df.iloc[:max_trials]
psy_new = psy_run_df.iloc[:max_trials]
data = {}
for i in psy_new.columns:
data[i] = psy_new[i].values
new_psy_df_list.append(
pd.DataFrame(data=data, index=meta_new.index))
new_meta_df_list.append(meta_new)
meta1 = pd.concat(new_meta_df_list, axis=0)
psy1 = pd.concat(new_psy_df_list, axis=0)
tca_data = {}
fac = tensor.results[rank_num][0].factors[2][:, comp_num - 1]
tca_data['factor_' + str(comp_num)] = fac
fac_df = pd.DataFrame(data=tca_data, index=meta1.index)
# threshold your data in a clever way so that you are not only
# looking at orientation trials
clever_binary = np.ones((len(fac)))
thresh = np.nanstd(fac) * 1
clever_binary[fac > thresh] = 2
# which values were dropped from the psydata. Use this to update psydata
blank_trials_bool[blank_trials_bool] = (drop_trials_bin == 1)
keep_bool = blank_trials_bool
drop_bool = blank_trials_bool == False
# keep_bool = np.logical_and(
# drop_trials_bin == 0, blank_trials_bool.values == True)
# drop_bool = np.logical_or(
# drop_trials_bin == 1, blank_trials_bool.values == False)
# you don't have any blank trials to avoid using them.
psydata['y'][drop_bool] = 1
psydata['answer'][drop_bool] = 1 # 1-2 binary not 0-1
psydata['y'][keep_bool] = clever_binary
psydata['answer'][keep_bool] = clever_binary
print('cleared :)')
return psydata
def groupmouse_correlate_pillow_tca(mice=[
'OA27', 'OA32', 'OA34', 'CC175', 'OA36', 'OA26', 'OA67', 'VF226'
],
words=[
'orlando', 'already', 'already',
'already', 'already', 'already',
'already', 'already'
],
trace_type='zscore_day',
method='mncp_hals',
cs='',
warp=False,
group_by='all',
nan_thresh=0.85,
score_threshold=None):
# preallocate
corr_list = []
pmat_list = []
x_labels = []
# LOADING
for mouse, word in zip(mice, words):
# Load or run your psytracker behavioral model
ms = flow.Mouse(mouse)
psy = ms.psytracker(verbose=True)
dateRuns = psy.data['dateRuns']
trialRuns = psy.data['runLength']
# create your trial indices per day and run
trial_idx = []
for i in trialRuns:
trial_idx.extend(range(i))
# get date and run vectors
date_vec = []
run_vec = []
for c, i in enumerate(dateRuns):
date_vec.extend([i[0]] * trialRuns[c])
run_vec.extend([i[1]] * trialRuns[c])
# create your data dict, transform from log odds to odds ratio
data = {}
for c, i in enumerate(psy.weight_labels):
# adding multiplication step here with binary vector
data[i] = np.exp(psy.fits[c, :]) * psy.inputs[:, c].T
ori_0_in = [i[0] for i in psy.data['inputs']['ori_0']]
ori_135_in = [i[0] for i in psy.data['inputs']['ori_135']]
ori_270_in = [i[0] for i in psy.data['inputs']['ori_270']]
blank_in = [
0 if i == 1 else 1
for i in np.sum((ori_0_in, ori_135_in, ori_270_in), axis=0)
]
# create a single list of orientations to match format of meta
ori_order = [0, 135, 270, -1]
data['orientation'] = [
ori_order[np.where(np.isin(i, 1))[0][0]]
for i in zip(ori_0_in, ori_135_in, ori_270_in, blank_in)
]
# create your index out of relevant variables
index = pd.MultiIndex.from_arrays(
[[mouse] * len(trial_idx), date_vec, run_vec, trial_idx],
names=['mouse', 'date', 'run', 'trial_idx'])
dfr = pd.DataFrame(data, index=index)
# Load TCA results
load_kwargs = {
'mouse': mouse,
'method': method,
'cs': cs,
'warp': warp,
'word': word,
'group_by': group_by,
'nan_thresh': nan_thresh,
'score_threshold': score_threshold
}
tensor, _, _ = load.groupday_tca_model(**load_kwargs)
meta = load.groupday_tca_meta(**load_kwargs)
savepath = paths.tca_plots(mouse,
'group',
word=word,
group_pars={'group_by': group_by})
savepath = os.path.join(savepath, 'psytrack-vs-tca')
if not os.path.isdir(savepath): os.mkdir(savepath)
# Load smooth dprime
dfr['dprime'] = pool.calc.psytrack.dprime(flow.Mouse(mouse))
# CHECK THAT TRIAL INDICES ARE MATCHED AND HAVE MATCHED ORIS
# filter out blank trials
psy_df = dfr.loc[(dfr['orientation'] >= 0), :]
# check that all runs have matched trial orienations
new_psy_df_list = []
new_meta_df_list = []
dates = meta.reset_index()['date'].unique()
for d in dates:
psy_day_bool = psy_df.reset_index()['date'].isin([d]).values
meta_day_bool = meta.reset_index()['date'].isin([d]).values
psy_day_df = psy_df.iloc[psy_day_bool, :]
meta_day_df = meta.iloc[meta_day_bool, :]
runs = meta_day_df.reset_index()['run'].unique()
for r in runs:
psy_run_bool = psy_day_df.reset_index()['run'].isin([r]).values
meta_run_bool = meta_day_df.reset_index()['run'].isin(
[r]).values
psy_run_df = psy_day_df.iloc[psy_run_bool, :]
meta_run_df = meta_day_df.iloc[meta_run_bool, :]
psy_run_idx = psy_run_df.reset_index()['trial_idx'].values
meta_run_idx = meta_run_df.reset_index()['trial_idx'].values
# drop extra trials from trace2P that don't have associated imaging
max_trials = np.min([len(psy_run_idx), len(meta_run_idx)])
# get just your orientations for checking that trials are matched
meta_ori = meta_run_df['orientation'].iloc[:max_trials]
psy_ori = psy_run_df['orientation'].iloc[:max_trials]
# make sure all oris match between vectors of the same length each day
assert np.all(psy_ori.values == meta_ori.values)
# if everything looks good, copy meta index into psy
meta_new = meta_run_df.iloc[:max_trials]
psy_new = psy_run_df.iloc[:max_trials]
data = {}
for i in psy_new.columns:
data[i] = psy_new[i].values
new_psy_df_list.append(
pd.DataFrame(data=data, index=meta_new.index))
new_meta_df_list.append(meta_new)
meta1 = pd.concat(new_meta_df_list, axis=0)
psy1 = pd.concat(new_psy_df_list, axis=0)
# NOW TAKE TCA TRIAL FACTORS AND TRY CORRELATING FOR WITH PILLOW
# put factors for a given rank into a dataframe
ori = 'all'
save_pls = False
iteration = 0
for rank in [18]: # tensor.results
data = {}
for i in range(rank):
fac = tensor.results[rank][iteration].factors[2][:, i]
data['factor_' + str(i + 1)] = fac
fac_df = pd.DataFrame(data=data, index=meta1.index)
# loop over single oris
single_ori = pd.concat([psy1, fac_df],
axis=1).drop(columns='orientation')
corr = np.corrcoef(single_ori.values.T)
single_data = {}
for c, i in enumerate(single_ori.columns):
single_data[i] = corr[:, c]
corr_plt = pd.DataFrame(data=single_data, index=single_ori.columns)
num_corr = np.shape(single_ori)[1]
corrmat = np.zeros((num_corr, num_corr))
pmat = np.zeros((num_corr, num_corr))
for i in range(num_corr):
for k in range(num_corr):
corA, corP = pearsonr(single_ori.values[:, i],
single_ori.values[:, k])
corrmat[i, k] = corA
pmat[i, k] = corP
if mouse == mice[0]:
y_label = single_ori.columns
# stick chunks of corr matrix together
x_labels.extend([mouse + ' ' + s for s in single_ori.columns[0:8]])
corr_list.append(corrmat[:, 0:8])
pmat_list.append(pmat[:, 0:8])
# concatenate final matrix together
corrmat = np.concatenate(corr_list, axis=1)
pmat = np.concatenate(pmat_list, axis=1)
annot = True
figsize = (80, 20)
# create your path for saving
rankpath = os.path.join(savepath, 'rank ' + str(rank))
if not os.path.isdir(rankpath): os.mkdir(rankpath)
var_path_prefix = os.path.join(
rankpath,
mouse + '_psytrack-vs-tca_ori-' + str(ori) + '_rank-' + str(rank))
plt.figure(figsize=figsize)
# plt.figure()
sns.heatmap(corrmat,
annot=annot,
xticklabels=x_labels,
yticklabels=y_label,
square=False,
cbar_kws={'label': 'correlation (R)'})
plt.xticks(rotation=45, ha='right')
plt.title('Pearson-R corrcoef: rank ' + str(rank))
if save_pls:
plt.savefig(var_path_prefix + '_corr.pdf', bbox_inches='tight')
plt.figure(figsize=figsize)
# plt.figure()
sns.heatmap(pmat,
annot=annot,
xticklabels=x_labels,
yticklabels=y_label,
square=False,
cbar_kws={'label': 'p-value'})
plt.xticks(rotation=45, ha='right')
plt.title('Pearson-R p-values: rank ' + str(rank))
if save_pls:
plt.savefig(var_path_prefix + '_pvals.pdf', bbox_inches='tight')
plt.figure(figsize=figsize)
# plt.figure()
logger = np.log10(pmat).flatten()
vmin = np.nanmin(logger[np.isfinite(logger)])
vmax = 0
sns.heatmap(np.log10(pmat),
annot=annot,
xticklabels=x_labels,
yticklabels=y_label,
vmin=vmin,
vmax=vmax,
square=False,
cbar_kws={'label': 'log$_{10}$(p-value)'})
plt.xticks(rotation=45, ha='right')
plt.title('Pearson-R log$_{10}$(p-values): rank ' + str(rank))
if save_pls:
plt.savefig(var_path_prefix + '_log10pvals.pdf', bbox_inches='tight')
# close plots after saving to save memory
if save_pls:
plt.close('all')
