"""
import pandas as pd
import numpy as np
import seaborn as sns
from os.path import join
import matplotlib.pyplot as plt
from paper_behavior_functions import seaborn_style, figpath, load_csv, FIGURE_WIDTH, FIGURE_HEIGHT
# Settings
FIG_PATH = figpath()
colors = [[1, 1, 1], [1, 1, 1], [0.6, 0.6, 0.6]]
seaborn_style()
# Load in results from csv file
decoding_result = load_csv('classification_results', 'classification_results_full_bayes.pkl')
# Calculate if decoder performs above chance
chance_level = decoding_result['original_shuffled'].mean()
significance = np.percentile(decoding_result['original'], 2.5)
sig_control = np.percentile(decoding_result['control'], 0.001)
if chance_level > significance:
print('Classification performance not significanlty above chance')
else:
print('Above chance classification performance!')
# %%
f, ax1 = plt.subplots(1, 1, figsize=(FIGURE_WIDTH/5, FIGURE_HEIGHT))
sns.violinplot(data=pd.concat([decoding_result['control'],
decoding_result['original_shuffled'],
trials.fetch(format='frame').join(
subject_info.fetch(format='frame')).sort_values(by=[
'institution_short', 'subject_nickname', 'session_start_time',
'trial_id'
]).reset_index())
behav['institution_code'] = behav.institution_short.map(institution_map)
# split the two types of task protocols (remove the pybpod version number)
behav['task'] = behav['task_protocol'].str[14:20].copy()
# RECODE SOME THINGS JUST FOR PATSY
behav['contrast'] = np.abs(behav.signed_contrast)
behav['stimulus_side'] = np.sign(behav.signed_contrast)
behav['block_id'] = behav['probabilityLeft'].map({80: -1, 50: 0, 20: 1})
else: # load from disk
behav = load_csv('Fig5.csv')
# ========================================== #
#%% 2. DEFINE THE GLM
# ========================================== #
# DEFINE THE MODEL
def fit_glm(behav, prior_blocks=False, folds=5):
# drop trials with contrast-level 50, only rarely present (should not be its own regressor)
behav = behav[np.abs(behav.signed_contrast) != 50]
# use patsy to easily build design matrix
if not prior_blocks:
endog, exog = patsy.dmatrices(
from ibl_pipeline import reference, subject, behavior
use_sessions, _ = query_sessions_around_criterion(criterion='biased',
days_from_criterion=[1, 3])
use_sessions = use_sessions & 'task_protocol LIKE "%biased%"' # only get biased sessions
b = (use_sessions * subject.Subject * subject.SubjectLab * reference.Lab
* behavior.TrialSet.Trial)
b2 = b.proj('institution_short', 'subject_nickname', 'task_protocol',
'trial_stim_contrast_left', 'trial_stim_contrast_right', 'trial_response_choice',
'task_protocol', 'trial_stim_prob_left', 'trial_feedback_type',
'trial_response_time', 'trial_stim_on_time', 'time_zone')
bdat = b2.fetch(order_by='institution_short, subject_nickname, session_start_time, trial_id',
format='frame').reset_index()
behav = dj2pandas(bdat)
behav['institution_code'] = behav.institution_short.map(institution_map()[0])
else:
behav = load_csv('Fig4.csv')
biased_fits = pd.DataFrame()
for i, nickname in enumerate(behav['subject_nickname'].unique()):
if np.mod(i+1, 10) == 0:
print('Processing data of subject %d of %d' % (i+1,
len(behav['subject_nickname'].unique())))
# Get lab and timezone
lab = behav.loc[behav['subject_nickname'] == nickname, 'institution_code'].unique()[0]
time_zone = behav.loc[behav['subject_nickname'] == nickname, 'time_zone'].unique()[0]
if (time_zone == 'Europe/Lisbon') or (time_zone == 'Europe/London'):
time_zone_number = 0
elif time_zone == 'America/New_York':
time_zone_number = -5
elif time_zone == 'America/Los_Angeles':
behavior_analysis.SessionTrainingStatus - subject.Death
& 'training_status = "in_training"'
& 'session_start_time > "%s"' % CUTOFF_DATE)
use_subjects = mice_started_training - still_training
# Get training status and training time in number of sessions and trials
ses = ((use_subjects * behavior_analysis.SessionTrainingStatus *
behavior_analysis.PsychResults).proj(
'subject_nickname', 'training_status', 'n_trials_stim',
'institution_short').fetch(format='frame').reset_index())
ses['n_trials'] = [sum(i) for i in ses['n_trials_stim']]
ses = ses.drop('n_trials_stim', axis=1).dropna()
ses = ses.sort_values(['subject_nickname', 'session_start_time'])
else:
# Load in sessions from csv file
ses = load_csv('Fig2d.csv').dropna()
# Select mice that started training before cut off date
ses = ses.groupby('subject_uuid').filter(
lambda s: s['session_start_time'].min() < CUTOFF_DATE)
# Construct dataframe from query
training_time = pd.DataFrame()
for i, nickname in enumerate(ses['subject_nickname'].unique()):
training_time.loc[i, 'nickname'] = nickname
training_time.loc[i, 'lab'] = ses.loc[ses['subject_nickname'] == nickname,
'institution_short'].values[0]
training_time.loc[i, 'sessions'] = sum(
(ses['subject_nickname'] == nickname)
& ((ses['training_status'] == 'in_training')
| (ses['training_status'] == 'untrainable')))
import numpy as np
import seaborn as sns
from os.path import join
import matplotlib.pyplot as plt
from paper_behavior_functions import seaborn_style, figpath, load_csv, FIGURE_WIDTH, FIGURE_HEIGHT
# Settings
FIG_PATH = figpath()
colors = [[1, 1, 1], [1, 1, 1], [0.6, 0.6, 0.6]]
seaborn_style()
for DECODER in ['bayes', 'forest', 'regression']:
# Load in results from csv file
filename = f'classification_results_basic_{DECODER}.pkl'
decoding_result = load_csv('classification_results', filename)
# Calculate if decoder performs above chance
chance_level = decoding_result['original_shuffled'].mean()
significance = np.percentile(decoding_result['original'], 2.5)
sig_control = np.percentile(decoding_result['control'], 0.001)
if chance_level > significance:
print('\n%s classifier did not perform above chance' % DECODER)
print('Chance level: %.2f (F1 score)' % chance_level)
else:
print('\n%s classifier did not perform above chance' % DECODER)
print('Chance level: %.2f (F1 score)' % chance_level)
print('F1 score: %.2f ± %.3f' % (decoding_result['original'].mean(),
decoding_result['original'].std()))
# %%
subject.Subject * subject.SubjectLab * reference.Lab *
(behavior.TrialSet.Trial & session_keys))
ses = ses.proj(
'institution_short', 'subject_nickname', 'task_protocol',
'trial_stim_contrast_left', 'trial_stim_contrast_right',
'trial_response_choice', 'task_protocol', 'trial_stim_prob_left',
'trial_feedback_type', 'trial_response_time', 'trial_stim_on_time',
'time_zone').fetch(
order_by=
'institution_short, subject_nickname,session_start_time, trial_id',
format='frame').reset_index()
behav = dj2pandas(ses)
behav['institution_code'] = behav.institution_short.map(
institution_map()[0])
else:
behav = load_csv('Fig3.csv')
# Create dataframe with behavioral metrics of all mice
learned = pd.DataFrame(columns=[
'mouse', 'lab', 'perf_easy', 'n_trials', 'threshold', 'bias',
'reaction_time', 'lapse_low', 'lapse_high', 'time_zone', 'UTC'
])
for i, nickname in enumerate(behav['subject_nickname'].unique()):
if np.mod(i + 1, 10) == 0:
print('Processing data of subject %d of %d' %
(i + 1, len(behav['subject_nickname'].unique())))
# Get the trials of the sessions around criterion for this subject
trials = behav[behav['subject_nickname'] == nickname]
trials = trials.reset_index()
FIGURE_WIDTH, FIGURE_HEIGHT,
fit_psychfunc, num_star,
query_session_around_performance)
import scikit_posthocs as sp
from statsmodels.stats.multitest import multipletests
seaborn_style()
figpath = figpath()
pal = group_colors()
institution_map, col_names = institution_map()
col_names = col_names[:-1]
if QUERY == True:
behav = query_session_around_performance(perform_thres=0.8)
else:
behav = load_csv('suppfig_3-4af.pkl')
behav['institution_code'] = behav.lab_name.map(institution_map)
# Create dataframe with behavioral metrics of all mice
learned = pd.DataFrame(columns=[
'mouse', 'institution_short', 'perf_easy', 'n_trials', 'threshold', 'bias',
'reaction_time', 'lapse_low', 'lapse_high', 'trials_per_minute'
])
for i, nickname in enumerate(behav['subject_nickname'].unique()):
if np.mod(i + 1, 10) == 0:
print('Processing data of subject %d of %d' %
(i + 1, len(behav['subject_nickname'].unique())))
# Get the trials of the sessions around criterion for this subject (first
# 90% + next session)
'lapse_high')
query = (not_trained.aggr(status,
session_start_time='max(session_start_time)',
session_n='COUNT(session_start_time)') * status *
subject.Death.proj('death_ts') *
behavior_analysis.PsychResults.proj(*fields) *
behavior.TrialSet.proj('n_trials'))
df = ((query.fetch(format='frame').reset_index().drop('subject_project',
axis=1)))
# Print a breakdown of final training statuses
print(df.training_status.value_counts(), '\n')
# Load the cull reasons from file. These were not available through DJ.
df.subject_uuid = df.subject_uuid.astype(str)
cull_reasons = load_csv('cull_reasons.csv')
df = pd.merge(df, cull_reasons, on='subject_uuid')
# NB: Untrainable training status takes precedence over cull reason
not_trained = len(mice_started_training) - len(trained)
untrainable = df['training_status'] == 'untrainable'
time_limit = (df.cull_reason == 'time limit reached') & ~untrainable
low_trial_n = df['n_trials'] < 400
biased = df['bias'].abs() > 15
low_perf = df['performance_easy'] < 65
# Inspecting deaths
injury = ('acute injury', 'infection or illness', 'issue during surgery')
premature_death = ~untrainable & (df.cull_reason != 'time limit reached')
sick = df.training_status[df.cull_reason.isin(injury)][premature_death]
benign = premature_death & (df.cull_reason
== 'benign experimental impediments')
pal = group_colors()
institution_map, col_names = institution_map()
col_names = col_names[:-1]
# %% ============================== #
# GET DATA FROM TRAINED ANIMALS
# ================================= #
if QUERY is True:
use_subjects = query_subjects()
b = (behavioral_analyses.BehavioralSummaryByDate * use_subjects *
behavioral_analyses.BehavioralSummaryByDate.PsychResults)
behav = b.fetch(
order_by='institution_short, subject_nickname, training_day',
format='frame').reset_index()
behav['institution_code'] = behav.institution_short.map(institution_map)
else:
behav = load_csv('Fig2af.pkl')
# exclude sessions with fewer than 100 trials
behav = behav[behav['n_trials_date'] > 100]
# exclude sessions with less than 3 types of contrast
behav.loc[behav['signed_contrasts'].str.len() < 6, 'threshold'] = np.nan
behav.loc[behav['signed_contrasts'].str.len() < 6, 'bias'] = np.nan
# convolve performance over 3 days
for i, nickname in enumerate(behav['subject_nickname'].unique()):
# 1.Performance
perf = behav.loc[behav['subject_nickname'] == nickname,
'performance_easy'].values
perf_conv = np.convolve(perf, np.ones((3, )) / 3, mode='valid')
# perf_conv = np.append(perf_conv, [np.nan, np.nan])
perf_conv = medfilt(perf, kernel_size=3)
behav.loc[behav['subject_nickname'] == nickname,
'performance_easy'] = perf_conv
group_colors, figpath, load_csv,
FIGURE_WIDTH, FIGURE_HEIGHT, num_star)
# Load some things from paper_behavior_functions
figpath = Path(figpath())
seaborn_style()
institution_map, col_names = institution_map()
pal = group_colors()
cmap = sns.diverging_palette(20, 220, n=3, center="dark")
# ========================================== #
#%% 1. GET GLM FITS FOR ALL MICE
# ========================================== #
print('loading model from disk...')
params_basic = load_csv('model_results', 'params_basic.csv')
params_full = load_csv('model_results', 'params_full.csv')
combined = params_basic.merge(params_full,
on=['institution_code', 'subject_nickname'])
# ========================================== #
# PRINT SUMMARY AND STATS
# ========================================== #
vars = ['6.25', '12.5', '25', '100', 'rewarded', 'unrewarded', 'bias']
for v in vars:
print('basic task, %s: mean %.2f, %f : %f' %
(v, params_basic[v].mean(), params_basic[v].min(),
params_basic[v].max()))
print(
# Construct dataframe
training_time = pd.DataFrame(columns=['sessions'],
data=ses.groupby('subject_nickname').size())
ses['n_trials_date'] = ses['n_trials_date'].astype(int)
training_time['trials'] = ses.groupby(
'subject_nickname').sum()['n_trials_date']
training_time['lab'] = ses.groupby(
'subject_nickname')['institution_short'].apply(list).str[0]
# Change lab name into lab number
training_time['lab_number'] = training_time.lab.map(institution_map)
training_time = training_time.sort_values('lab_number')
training_time = training_time.reset_index()
else:
data = load_csv('Fig2af.pkl').dropna()
use_subjects = data['subject_nickname'].unique(
) # For counting the number of subjects
training_time = pd.DataFrame()
for i, subject in enumerate(use_subjects):
training_time = training_time.append(
pd.DataFrame(
index=[training_time.shape[0] + 1],
data={
'subject_nickname':
subject,
'lab':
data.loc[data['subject_nickname'] == subject,
'institution_short'].unique(),
'sessions':
data.loc[((data['subject_nickname'] == subject)
ses = ((use_sessions & 'task_protocol LIKE "%training%"') *
subject.Subject * subject.SubjectLab * reference.Lab *
(behavior.TrialSet.Trial & session_keys))
ses = ses.proj(
'institution_short', 'subject_nickname', 'task_protocol',
'session_uuid', 'trial_stim_contrast_left',
'trial_stim_contrast_right', 'trial_response_choice', 'task_protocol',
'trial_stim_prob_left', 'trial_feedback_type', 'trial_response_time',
'trial_stim_on_time', 'session_end_time').fetch(
order_by=
'institution_short, subject_nickname,session_start_time, trial_id',
format='frame').reset_index()
behav = dj2pandas(ses)
behav['institution_code'] = behav.institution_short.map(institution_map)
else:
behav = load_csv('Fig3.csv',
parse_dates=['session_start_time', 'session_end_time'])
# Create dataframe with behavioral metrics of all mice
learned = pd.DataFrame(columns=[
'mouse', 'lab', 'perf_easy', 'n_trials', 'threshold', 'bias',
'reaction_time', 'lapse_low', 'lapse_high', 'trials_per_minute'
])
for i, nickname in enumerate(behav['subject_nickname'].unique()):
if np.mod(i + 1, 10) == 0:
print('Processing data of subject %d of %d' %
(i + 1, len(behav['subject_nickname'].unique())))
# Get the trials of the sessions around criterion for this subject
trials = behav[behav['subject_nickname'] == nickname]
trials = trials.reset_index()