debug_outdir,
trainable=True,
debugplots_after=1500,
record_metrics=True,
eweighted=True,
)
# this will create issues with the output and is only needed if used in a full dim model.
# so it's ok to pop it here for training
presel.pop('scatterids')
return DictModel(inputs=Inputs, outputs=presel)
import training_base_hgcal
train = training_base_hgcal.HGCalTraining()
if not train.modelSet():
train.setModel(pretrain_model,
td = train.train_data.dataclass(),
debug_outdir=train.outputDir+'/intplots')
train.setCustomOptimizer(tf.keras.optimizers.Adam())
#
train.compileModel(learningrate=1e-4)
train.keras_model.summary()
#start somewhere
#from model_tools import apply_weights_from_path
model_outputs = [('pred_beta', pred_beta), ('pred_ccoords',pred_ccoords),
('pred_energy',pred_energy),
('pred_pos',pred_pos),
('pred_time',pred_time),
('pred_id',pred_id),
('pred_dist',pred_dist),
('row_splits',rs)]
for i, (x, y) in enumerate(zip(backgatheredids, backgathered_coords)):
model_outputs.append(('backgatheredids_'+str(i), x))
model_outputs.append(('backgathered_coords_'+str(i), y))
return RobustModel(model_inputs=Inputs, model_outputs=model_outputs)
import training_base_hgcal
train = training_base_hgcal.HGCalTraining(testrun=False, resumeSilently=True, renewtokens=False)
if not train.modelSet():
train.setModel(gravnet_model)
train.setCustomOptimizer(tf.keras.optimizers.Nadam(
clipnorm=0.001
))
train.compileModel(learningrate=1e-4,
loss=None)
verbosity = 2
import os
samplepath=train.val_data.getSamplePath(train.val_data.samples[0])