def print_percentages(what, df_sel, df, color='red'):
if color == 'red':
lu.print_red(f"{what} ",
f"{len(df_sel)} = {round(100*len(df_sel)/len(df),1)}%")
elif color == 'blue':
lu.print_blue(f"{what} ",
f"{len(df_sel)} = {round(100*len(df_sel)/len(df),1)}%")
def do_classification(skim_dir, list_models, sntypes):
""" SNN classification
"""
import SuperNNova.supernnova.conf as conf
from SuperNNova.supernnova.data import make_dataset
from SuperNNova.supernnova.visualization import early_prediction
from SuperNNova.supernnova.validation import validate_rnn, metrics
lu.print_blue(f"Classifying {skim_dir}")
# get config args
snn_args = conf.get_args()
# create database
snn_args.data = True
snn_args.data_testing = True
snn_args.dump_dir = f"{skim_dir}/"
snn_args.raw_dir = f"{skim_dir}/"
snn_args.fits_dir = "./"
snn_args.sntypes = sntypes
settings = conf.get_settings(snn_args)
# # make dataset
# make_dataset.make_dataset(settings)
for model in list_models:
# add model file to settings
snn_args.model_files = model
settings = conf.get_settings(snn_args)
model_files = [model]
# # # classify
# snn_args.validate_rnn = True
model_settings = conf.get_settings_from_dump(
settings,
snn_args.model_files,
override_source_data=settings.override_source_data,
)
# fetch predictions
prediction_file = validate_rnn.get_predictions(
model_settings, model_file=snn_args.model_files)
# Compute metrics
metrics.get_metrics_singlemodel(model_settings,
prediction_file=prediction_file,
model_type="rnn")
# plot lcs
model_settings.model_files = snn_args.model_files
early_prediction.make_early_prediction(model_settings, nb_lcs=20)
# evaluate classifications
df = eu.fetch_prediction_info(settings, model_settings, skim_dir)
# plots init
path_plot = f"{snn_args.dump_dir}/figures/"
eu.plot_efficiency(df, skim_dir, path_plot)
# "the classified sample"
eu.pair_plots(df, path_plot)
def skim_data(raw_dir, dump_dir, fits_file, timevar, debug=False):
""" Skim PHOT and HEAD.FITS
"""
list_files = glob.glob(os.path.join(f"{raw_dir}", "*PHOT.FITS"))
if debug:
lu.print_yellow('Debugging mode')
list_files = list_files[:1]
lu.print_green(
f"Starting data skimming, found {len(list_files)} to operate on")
# load Bazin
df_fits = None
if Path(fits_file).exists():
df_fits = du.load_fits(fits_file)
tmp_type_list = []
filenames = []
# skim each FITS file
for fname in list_files:
# fetch data year as prefix
dump_prefix = Path(fname).name.split("_")[0]
lu.print_blue(f"Processing: {dump_prefix}")
df_header, df_phot = du.read_fits(fname)
if df_fits is not None:
df_header = pd.merge(df_header, df_fits, on='SNID')
df_header = df_header[[
k for k in df_header.keys() if 'Unnamed' not in k
]]
# apply cuts
unique_types, filename = apply_cut_save(df_header,
df_phot,
timevar=timevar,
dump_dir=dump_dir,
dump_prefix=dump_prefix)
tmp_type_list += unique_types
filenames.append(filename)
def get_sample_stats_and_plots(df_pred,
photo_Ia,
photo_nonIa,
skim_dir,
model_files=None,
out_dir=None,
plot=False):
# inspect sample
path_plots = f"{skim_dir}/figures/"
Path(path_plots).mkdir(parents=True, exist_ok=True)
vars_to_plot = [
k for k in [
'REDSHIFT_FINAL', 'PRIVATE(DES_numepochs_ml)', 'all_class0',
'PRIVATE(DES_cand_type)', 'TYPE', 'PRIVATE(DES_mjd_trigger)',
'PKMJDINI'
] if k in photo_Ia['all'].keys()
]
if 'fake' in skim_dir:
df_dic = {'all_lcs': df_pred, 'photo Ia sample': photo_Ia['all']}
vars_to_plot += [
'PRIVATE(DES_fake_salt2x1)', 'PRIVATE(DES_fake_salt2c)'
]
else:
df_dic = {
'all_lcs': df_pred,
'photo Ia sample': photo_Ia['all'],
'contaminants': photo_Ia['spec_nonIa'],
'photo other but spec Ia ': photo_nonIa['spec_Ia']
}
for var in [k for k in vars_to_plot]:
vu.plot_superimposed_hist(df_dic,
var,
nameout=f"{path_plots}/hist_{var}_dist.png",
log=True)
for var in ['FLUXCAL_max', 'SNRMAX1']:
# photo sample zoom
vu.plot_superimposed_hist(df_dic,
var,
nameout=f"{path_plots}/hist_{var}_dist.png",
log=True,
limits_from_photo_sample=True)
# Stats
lu.print_green(cut_type)
eu.print_percentages(f"photo Ias ",
photo_Ia['all'],
df_pred,
color='blue')
if dtype == 'real':
lu.print_blue(f" are spec Ias ", len(photo_Ia['spec_Ia']))
lu.print_blue(f' are spec other', len(photo_Ia['spec_nonIa']))
lu.print_blue(
f' gals ',
len(photo_Ia['spec_nonIa'][photo_Ia['spec_nonIa']['TYPE'] == 81]))
lu.print_red(f"missed Ias ", len(photo_nonIa['spec_Ia']))
# dump sample
if not out_dir:
out_dir = f"{skim_dir}/sample/"
Path(out_dir).mkdir(parents=True, exist_ok=True)
# filla na
photo_Ia['all_no_spec_nonIa'] = photo_Ia['all_no_spec_nonIa'].fillna(0)
photo_Ia['spec_nonIa'] = photo_Ia['spec_nonIa'].fillna(0)
photo_Ia['all_no_spec_nonIa'][[
'SNID', 'HOSTGAL_OBJID', 'DEC', 'RA', 'TYPE', 'REDSHIFT_FINAL',
'HOSTGAL_PHOTOZ', 'HOSTGAL_SPECZ', 'all_class0', 'c', 'x1'
]].to_csv(f'{out_dir}/photo_Ia.csv')
# dump contaminants
photo_Ia['spec_nonIa'][[
'SNID', 'HOSTGAL_OBJID', 'DEC', 'RA', 'TYPE', 'REDSHIFT_FINAL',
'HOSTGAL_PHOTOZ', 'HOSTGAL_SPECZ', 'all_class0', 'c', 'x1'
]].to_csv(f'{out_dir}/photo_Ia_spec_contamination.csv')
# dump missed Ias
photo_nonIa['spec_Ia'][[
'SNID', 'HOSTGAL_OBJID', 'DEC', 'RA', 'TYPE', 'REDSHIFT_FINAL',
'HOSTGAL_PHOTOZ', 'HOSTGAL_SPECZ', 'all_class0', 'c', 'x1'
]].to_csv(f'{out_dir}/photo_nonIa_spec_Ia.csv')
# plot lcs
if plot:
vu.plot_early_classification(skim_dir,
prefix='photo_Ia_',
df=photo_Ia['all'],
model_files=model_files,
out_dir=out_dir)
vu.plot_early_classification(skim_dir,
prefix='photo_Ia_spec_contamination',
df=photo_Ia['spec_nonIa'],
model_files=model_files,
out_dir=out_dir)
vu.plot_early_classification(skim_dir,
prefix='photo_nonIa_',
df=photo_nonIa['all'],
model_files=model_files,
out_dir=out_dir)
vu.plot_early_classification(skim_dir,
prefix='photo_nonIa_spec_Ia_',
df=photo_nonIa['spec_Ia'],
model_files=model_files,
out_dir=out_dir)
# sample histograms of type
vu.plot_hist(photo_Ia['all'],
'TYPE',
nameout=f"{path_plots}/photo_Ia_hist_type.png",
log=True)
vu.plot_hist(photo_nonIa['all'],
'TYPE',
nameout=f"{path_plots}/photo_nonIa_hist_type.png",
log=True)
df_dic = {
'photo Ia sample': photo_Ia['all'],
'photo & spec Ia ': photo_Ia['spec_Ia'],
'photo other but spec Ia ': photo_nonIa['spec_Ia']
}
for var in [
'REDSHIFT_FINAL', 'HOSTGAL_PHOTOZ', 'HOSTGAL_SPECZ',
'all_class0'
]:
vu.plot_superimposed_hist(
df_dic,
var,
nameout=f"{path_plots}/hist_{var}_dist_spec.png",
log=True,
only_positive_x=True)
for var in ['FLUXCAL_max', 'c', 'x1']:
vu.plot_superimposed_hist(
df_dic,
var,
nameout=f"{path_plots}/hist_{var}_dist_spec.png",
log=True,
only_positive_x=False,
bins=20)
path_des_data = os.environ.get("DES_DATA")
model_name = "vanilla_S_0_CLF_2_R_None_photometry_DF_1.0_N_global_lstm_32x2_0.05_128_True_mean_C"
list_models = glob.glob("../SuperNNova_general/trained_models_mutant/*/*.pt")
for model in list_models:
model_files = [model]
for dtype in ["real", "fake"]:
df_pred = {}
photo_Ia = {}
photo_nonIa = {}
for cut_type in ['clump']: #['bazin','clump','trigger']:
print()
lu.print_blue(
f'_____STATS FOR {dtype} with window {cut_type} model {Path(model).name.split("_")[0]}_____'
)
print()
skim_dir = f"./dumps/{dtype}/{cut_type}/"
# fetch predictions
df_pred[cut_type] = du.load_predictions_and_info(
skim_dir, model_name)
# add salt2 fit parameters
raw_dir = f"{path_des_data}/DESALL_forcePhoto_{dtype}_snana_fits/"
saltfit = du.load_fitres(raw_dir)
saltfit = saltfit[['SNID'] + [
k for k in saltfit.keys() if k not in df_pred[cut_type].keys()
]]
# load preds & enrich
df_pred = du.load_predictions(fname_preds)
df_pred = du.enrich_predictions(df_pred, path_dtype_data)
# save the preds
dic_pred[dtype][name_model] = df_pred
# Select a default "photometric" sample
photo_sample = df_pred[df_pred['predicted_target'] == 0]
dic_pred[dtype][name_model]['photo_sample'] = np.array([df_pred['predicted_target'] == 0])[0]
# save photo sample
cols_to_save = ['SNID','HOSTGAL_OBJID', 'DEC', 'RA', 'SNTYPE', 'REDSHIFT_FINAL', 'HOSTGAL_PHOTOZ', 'HOSTGAL_SPECZ']
cols_to_save += ['all_class0'] if 'vanilla' in name_model else ['all_class0_median','all_class0_std']
photo_sample[cols_to_save].to_csv(f"{out_dir}/photo_sample.csv")
lu.print_blue(name_model)
print(f'photo sample {len(photo_sample)} representing {int(len(photo_sample)/len(df_pred)*100)}%')
# metrics
if dtype == 'real':
# Ias
dic_pred[dtype][name_model]['photo_spec_Ia'] = np.array([(df_pred['predicted_target'] == 0) & ((df_pred['SNTYPE'] == 1) | (
df_pred['SNTYPE'] == 101))])[0]
spec_Ia = df_pred[(df_pred['SNTYPE'] == 1) |
(df_pred['SNTYPE'] == 101)]
print('spec Ia', len(dic_pred[dtype][name_model][dic_pred[dtype][name_model]['photo_spec_Ia']==True]), f"from {len(spec_Ia)}")
# non Ias
dic_pred[dtype][name_model]['photo_spec_nonIa'] = np.array([(df_pred['predicted_target'] == 0) & (df_pred['SNTYPE'] != 0) & (df_pred['SNTYPE'] != 1) & (
df_pred['SNTYPE'] != 101)])[0]
spec_non_Ia = df_pred[(df_pred['SNTYPE'] != 0) & (df_pred['SNTYPE'] != 1) &
(df_pred['SNTYPE'] != 101)]