def combined_plots(n_test,
n_top,
output_dir,
plot_var,
n_dims=4,
n_constituents=20):
sample_topo = make_sample(n_constituents,
'qcd-topo',
'top-topo',
n_test,
n_top,
n_dims,
adjust_weights=True)
sample_UFO = make_sample(n_constituents,
'qcd-UFO',
'BSM',
n_test,
n_top,
n_dims,
adjust_weights=True)
samples = [sample_UFO, sample_topo]
plot_distributions(samples,
output_dir,
plot_var,
sig_bins=200,
bkg_bins=400,
sig_tag='top')
sys.exit()
if args.valid_cuts == '': args.valid_cuts = valid_cuts
else: args.valid_cuts = valid_cuts + '& (' + args.valid_cuts + ')'
inputs = {'scalars': scalars, 'images': [], 'others': others}
validation(args.output_dir, args.results_in, args.plotting, args.n_valid,
data_files, inputs, args.valid_cuts, args.sep_bkg,
args.runDiffPlots)
elif args.results_in != '':
print('\nOption --results_in not matching any file --> aborting\n')
if args.results_in != '': sys.exit()
# MODEL CREATION AND MULTI-GPU DISTRIBUTION
n_gpus = min(args.n_gpus,
len(tf.config.experimental.list_physical_devices('GPU')))
train_batch_size = max(1, n_gpus) * args.batch_size
valid_batch_size = max(1, n_gpus) * max(args.batch_size, int(5e3))
sample = make_sample(data_files[0], [0, 1], input_data, args.n_tracks,
args.n_classes)[0]
model = create_model(args.n_classes, sample, args.NN_type, args.FCN_neurons,
CNN, args.l2, args.dropout, train_data, n_gpus)
# ARGUMENTS AND VARIABLES SUMMARY
args.scaling = args.scaling == 'ON' and scalars != []
args.t_scaling = args.t_scaling == 'ON' and 'tracks' in scalars + images
if args.NN_type == 'CNN':
print('\nCNN ARCHITECTURE:')
for shape in [
shape for shape in CNN
if shape in [sample[key].shape[1:] for key in sample]
]:
print(format(str(shape), '>8s') + ':', str(CNN[shape]))
print('\nPROGRAM ARGUMENTS:')
print(tabulate(vars(args).items(), tablefmt='psql'))
num_workers=world.workers)
else:
data_test = dataloader(mode="test",
transform=tran,
folder=args.evalFolder)
data_test = DataLoader(data_test,
batch_size=world.batch_size,
shuffle=False,
num_workers=world.workers)
print(">EVAL FOLDER:", args.evalFolder)
if world.verbose:
print("loaded data")
# sample data and show configs
data_sample = data[0]
img = utils.make_sample(data_sample)
utils.show_config()
# initialize model and restore training process
m = ITrackerModel()
if world.verbose:
print("initialized model")
epoch_before = 0
best = 1e10
if world.doLoad:
saved = utils.load_checkpoint(world.weight_file)
if saved:
print('>Loading checkpoint for epoch %05d with loss %.5f ...' %
(saved['epoch'], saved['best_prec1']))
state = saved['state_dict']
try:
from keras.models import load_model
from utils import files_for_train, generate_train_file, make_sample
epochs = 3
input_length = 128
batch_size = 256
input_path = "/home/biot/projects/Audio_generator/data/"
output_path = "/home/biot/projects/Audio_generator/generated_music/"
# input_path = "/root/inspiron/Audio_generator/data/"
# output_path = "/root/inspiron/Audio_generator/generated_music/"
sample_length = 120000
model = load_model('music_gen_model_5.h5')
make_sample(model=model, input_path=input_path, output_path=output_path, sample_length=sample_length, epoch=epochs,
input_length=input_length, extension='mp3')
'images': images if args.images == 'ON' else [],
'tracks': [],
'scalars': scalars if args.scalars == 'ON' else []
}
all_var = {
**train_var, 'others': others
}
scalars = train_var['scalars']
# ARCHITECTURE SELECTION AND MULTI-GPU DISTRIBUTION
n_gpus = min(args.n_gpus,
len(tf.config.experimental.list_physical_devices('GPU')))
devices = ['/gpu:0', '/gpu:1', '/gpu:2', '/gpu:3']
tf.debugging.set_log_device_placement(False)
strategy = tf.distribute.MirroredStrategy(devices=devices[:n_gpus])
sample, _ = make_sample(data_file, all_var, [0, 1], args.n_tracks,
args.n_classes)
with strategy.scope():
if tf.__version__ >= '2.1.0':
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
model = multi_CNN(args.n_classes, args.NN_type, sample, args.l2,
args.dropout, args.alpha, args.CNN_neurons,
args.FCN_neurons, **train_var)
model.summary()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
# SAMPLES SIZES AND APPLIED CUTS ON PHYSICS VARIABLES
sample_size = len(h5py.File(data_file, 'r')['eventNumber'])
args.n_train = [0, min(sample_size, args.n_train)
] if args.cross_valid == 'OFF' else [0, 0]
train_var = {'images' :images if args.images =='ON' else [], 'tracks':[],
'scalars':scalars if args.scalars=='ON' else []}
other_var = ['eventNumber', 'p_TruthType', 'p_iffTruth', 'p_LHTight', 'p_LHMedium', 'p_LHLoose',
'p_eta', 'p_et_calo','p_LHValue']
total_var = {**train_var, 'others':other_var}; scalars = train_var['scalars']
# MULTI-GPU DISTRIBUTION
n_gpus = min(args.n_gpus, len(tf.config.experimental.list_physical_devices('GPU')))
devices = ['/gpu:0', '/gpu:1', '/gpu:2', '/gpu:3']
tf.debugging.set_log_device_placement(False)
strategy = tf.distribute.MirroredStrategy(devices=devices[:n_gpus])
with strategy.scope():
if tf.__version__ >= '2.1.0' and len(total_var['images']) > 1:
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
sample, _ = make_sample(data_file, total_var, [0,1], args.n_tracks, args.n_classes)
func_args = (args.n_classes, args.NN_type, sample, args.l2, args.dropout, CNN, args.FCN_neurons)
model = multi_CNN(*func_args, **train_var)
print('\nNEURAL NETWORK ARCHITECTURE'); model.summary()
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
# SAMPLES SIZES AND APPLIED CUTS ON PHYSICS VARIABLES
sample_size = len(h5py.File(data_file, 'r')['eventNumber'])
args.n_train = [0, min(sample_size, args.n_train)] if args.cross_valid == 'OFF' else [0,0]
args.n_valid = [args.n_train[-1], min(args.n_train[-1]+args.n_valid, sample_size )]
if args.cross_valid == 'OFF' and args.n_folds > 1: args.n_valid = args.n_train
#args.valid_cuts += ' & (sample["p_et_calo"] >= 20)'
#args.valid_cuts += ' & (sample["p_et_calo"] > 4.5) & (sample["p_et_calo"] < 20)'
#args.valid_cuts += ' & (abs(sample["p_eta"] > 0.6)'
tf.compat.v1.logging.ERROR) #Suppressing Tensorflow warnings
seed = None if args.n_epochs > 0 else 0
model = VariationalAutoEncoder(args.FC_layers,
args.n_dims * args.n_constituents,
seed=seed)
args.model_in = args.output_dir + '/' + args.model_in
args.model_out = args.output_dir + '/' + args.model_out
args.scaler_in = args.output_dir + '/' + args.scaler_in
args.scaler_out = args.output_dir + '/' + args.scaler_out
if args.model_in != args.output_dir + '/':
if not os.path.isfile(args.model_in):
sys.exit()
print('\nLoading pre-trained weights from:', args.model_in)
sys.stdout = open(os.devnull, 'w') #Suppressing screen output
sample = make_sample(args.n_constituents,
bkg_data,
OoD_data,
n_dims=args.n_dims)
dataset = make_datasets(sample, sample)
train_model(model, dataset.take(1), dataset.take(1))
sys.stdout = sys.__stdout__ #Resuming screen output
model.load_weights(args.model_in)
if args.scaling == 'ON' and os.path.isfile(args.scaler_in):
print('\nLoading scaler transform from:', args.scaler_in)
scaler = pickle.load(open(args.scaler_in, 'rb'))
args.output_dir += '/plots'
for path in list(
accumulate([folder + '/' for folder in args.output_dir.split('/')])):
try:
os.mkdir(path)
except FileExistsError:
pass
print(
"\noption [--results_in] was given but no matching file found in the right path, aborting.."
)
print("results_in file =", args.output_dir + '/' + args.results_in, '\n')
if args.results_in != '': sys.exit()
# MULTI-GPU DISTRIBUTION
n_gpus = min(args.n_gpus,
len(tf.config.experimental.list_physical_devices('GPU')))
devices = ['/gpu:0', '/gpu:1', '/gpu:2', '/gpu:3']
tf.debugging.set_log_device_placement(False)
strategy = tf.distribute.MirroredStrategy(devices=devices[:n_gpus])
with strategy.scope():
if tf.__version__ >= '2.1.0' and len(variables['images']) >= 1:
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
sample, _ = make_sample(args.data_file, variables, [0, 1], args.n_tracks,
args.n_classes)
func_args = (args.n_classes, args.NN_type, sample, args.l2, args.dropout,
CNN, args.FCN_neurons)
model = multi_CNN(*func_args, **train_var)
print('\nNEURAL NETWORK ARCHITECTURE')
model.summary()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
# ARGUMENTS AND VARIABLES TABLES
args.scaler_in = args.scaler_in if '.pkl' in args.scaler_in else ''
args.model_in = args.model_in if '.h5' in args.model_in else ''
args.results_in = args.results_in if '.h5' in args.results_in else ''
args.NN_type = 'FCN' if train_var['images'] == [] else args.NN_type
args.scaling = (args.scaling == 'ON' and scalars != [])
tf.compat.v1.logging.ERROR) #suppress Tensorflow warnings
for path in list(
accumulate([folder + '/'
for folder in args.output_dir.split('/')])):
try:
os.mkdir(path)
except FileExistsError:
pass
print('\nTRAINING SAMPLE:')
sig_bins, bkg_bins = 200, 600
log = args.weight_type != 'flat_pt'
train_sample = make_sample(args.n_train,
args.n_W,
args.n_dims,
metrics,
train_var,
args.n_constituents,
bkg_cuts,
sig_cuts,
bkg='qcd',
sig='W')
train_sample = {
key: utils.shuffle(train_sample[key], random_state=0)
for key in train_sample
}
train_sample = reweight_sample(train_sample, sig_bins, bkg_bins,
args.weight_type)
var_distributions(train_sample,
args.output_dir,
sig_bins,
bkg_bins,
var='pt',
