print(train.keras_model.summary())
# **2 #this will be an upper limit on vertices per batch
verbosity = 2
import os
samplepath = train.val_data.getSamplePath(train.val_data.samples[0])
callbacks = []
for i in range(10):
plotoutdir = train.outputDir + "/event_" + str(i + 2)
os.system('mkdir -p ' + plotoutdir)
callbacks.append(
plotEventDuringTraining(outputfile=plotoutdir + "/sn",
samplefile=samplepath,
after_n_batches=4000,
batchsize=100000,
on_epoch_end=False,
use_event=2 + i))
from configSaver import copyModules
copyModules(train.outputDir)
from betaLosses import config as loss_config
loss_config.energy_loss_weight = 0.0001
loss_config.use_energy_weights = False
loss_config.q_min = 0.5
loss_config.no_beta_norm = False
loss_config.potential_scaling = 1.
loss_config.s_b = 1.
loss_config.position_loss_weight = 0.00001
samplepath = train.val_data.getSamplePath(train.val_data.samples[0])
print("using sample for plotting ",samplepath)
callbacks = []
import os
publishpath = '[email protected]:/eos/home-j/jkiesele/www/HGCalML_trainings/'+os.path.basename(os.path.normpath(train.outputDir))
for i in range(6,10):
ev = i
plotoutdir = train.outputDir + "/event_" + str(ev)
os.system('mkdir -p ' + plotoutdir)
callbacks.append(
plotEventDuringTraining(
outputfile=plotoutdir + "/sn",
samplefile=samplepath,
after_n_batches=200,
batchsize=100000,
on_epoch_end=False,
publish = publishpath+"_event_"+ str(ev),
use_event=ev)
)
model, history = train.trainModel(nepochs=1,
run_eagerly=True,
batchsize=nbatch,
batchsize_use_sum_of_squares=False,
checkperiod=1, # saves a checkpoint model every N epochs
verbose=verbosity,
backup_after_batches=100,
additional_callbacks=callbacks+
[CyclicLR (base_lr = learningrate,
max_lr = learningrate*5.,