def _get_modelspecs(cellids, batch, modelname, multi='mean'):
filepaths = load_batch_modelpaths(batch,
modelname,
cellids,
eval_model=False)
speclists = []
for path in filepaths:
mspaths = []
path = path.replace('http://hyrax.ohsu.edu:3003/',
'/auto/data/nems_db/')
if get_setting('NEMS_RESULTS_DIR').startswith("/Volumes"):
path = path.replace('/auto/', '/Volumes/')
for file in os.listdir(path):
if file.startswith("modelspec"):
mspaths.append(os.path.join(path, file))
speclists.append([load_resource(p) for p in mspaths])
modelspecs = []
for m in speclists:
if len(m) > 1:
if multi == 'first':
this_mspec = m[0]
elif multi == 'all':
this_mspec = m
elif multi == 'mean':
stats = ms.summary_stats(m)
temp_spec = copy.deepcopy(m[0])
phis = [m['phi'] for m in temp_spec]
for p in phis:
for k in p:
for s in stats:
if s.endswith('--' + k):
p[k] = stats[s]['mean']
for m, p in zip(temp_spec, phis):
m['phi'] = p
this_mspec = temp_spec
else:
log.warning(
"Couldn't interpret <multi> parameter. Got: %s,\n"
"Expected one of: 'mean, first, random, all'.\n"
"Using first modelspec instead.", multi)
this_mspec = m[0]
else:
this_mspec = m[0]
modelspecs.append(ms.ModelSpec([this_mspec]))
return modelspecs
# 3. Invert the folding to unwrap the psth back out into a predicted spike_dict by
# simply replacing all epochs in the signal with their psth
respavg = rec['resp'].replace_epochs(per_stim_psth)
respavg.name = 'respavg'
# 4. Now add the signal to the recording
rec.add_signal(respavg)
# Now split into est and val data sets
est, val = rec.split_using_epoch_occurrence_counts(epoch_regex='^STIM_')
# est, val = rec.split_at_time(0.8)
# Load some modelspecs and create their predictions
modelspecs = ms.load_modelspecs(modelspecs_dir, 'TAR010c-18-1')#, regex=('^TAR010c-18-1\.{\d+}\.json'))
# Testing summary statistics:
means, stds = ms.summary_stats(modelspecs)
print("means: {}".format(means))
print("stds: {}".format(stds))
pred = [ms.evaluate(val, m)['pred'] for m in modelspecs]
# Shorthands for unchanging signals
stim = val['stim']
resp = val['resp']
respavg = val['respavg']
def plot_layout(plot_fn_struct):
'''
Accepts a list of lists of functions of 1 argument (ax).
Basically a fancy subplot that lets you lay out functions without
for batch, bs in zip(batches, batstring):
modelspecs = get_batch_modelspecs(batch,
modelname,
multi=multi,
limit=None)
modelspecs_shf = get_batch_modelspecs(batch,
modelnames[1],
multi=multi,
limit=None)
modelspecs_SR = get_batch_modelspecs(batch,
modelnames[0],
multi=multi,
limit=None)
stats = ms.summary_stats(modelspecs,
mod_key=mod_key,
meta_include=meta,
stats_keys=stats_keys)
index = list(stats.keys())
columns = [m[0].get('meta').get('cellid') for m in modelspecs]
midx = 0
fields = ['bg_gain', 'fg_gain']
b = np.array([])
f = np.array([])
tar_id = np.array([])
cellids = []
b_S = np.array([])
f_S = np.array([])
c = np.array([])
cid = []
r_test = np.array([])
def fitted_params_per_cell(cellids,
batch,
modelname,
mod_key='id',
meta=['r_test', 'r_fit', 'se_test'],
multi='mean',
stats_keys=['mean', 'std', 'sem', 'max', 'min']):
'''
Valid meta keys for LBHB (not exhaustive):
r_test, r_fit, r_ceiling, r_floor, cellid, batch ... etc
Valid stats_keys:
mean, std, sem, max, min
Valid 'multi' options (for dealing with multi-modelspec fits):
mean (default), first, all (work in progress)
'''
# query nems_db results to get a list of modelspecs
# (should end up with one modelspec per cell)
modelspecs = _get_modelspecs(cellids, batch, modelname, multi=multi)
if multi == 'all':
raise NotImplementedError
# Flatten sublists of modelspecs
modelspecs = [m for ms in modelspecs for m in ms]
stats = ms.summary_stats(modelspecs,
mod_key=mod_key,
meta_include=meta,
stats_keys=stats_keys)
index = list(stats.keys())
try:
columns = [m[0].get('meta').get('cellid') for m in modelspecs]
except:
log.warning("Couldn't use cellids from modelspecs, using cellids "
"from function parameters instead.")
columns = cellids
data = {}
current = columns[0]
counter = 1
for i, c in enumerate(columns):
# Ensure unique column names, will have duplicates if multi='all'
# and there were multi-fit models included.
if i == 0:
pass
elif c == current:
columns[i] = '%s<%d>' % (c, counter)
counter += 1
else:
current = c
counter = 1
for k in index:
val = ms.try_scalar(stats[k]['values'][i])
if c in data.keys():
data[c].append(val)
else:
data[c] = [val]
if stats_keys:
for s in reversed(stats_keys):
columns.insert(0, s)
data[s] = []
for k in index:
data[s].append(stats[k][s])
return pd.DataFrame(data=data, index=index, columns=columns)