optimizer=int(x[6]),
verbose=True)
print("loading model 3")
mymodel.load_state_dict(
torch.load("%s/IN%s_bestmodel3.params" %
(loc, '_sumO' if mymodel.sum_O else '')))
inputTestFiles = glob.glob(
"/mnt/ccnas2/bdp/mzl20/proj_conda/lhc_jet_tagging_NN/JEDInet_pytorch/150p_dataset/val/jetImage*_%sp*.h5"
% nParticles)
loss = nn.CrossEntropyLoss(reduction='mean')
test_set = InEventLoader(file_names=inputTestFiles,
nP=nParticles,
feature_name='jetConstituentList',
label_name='jets',
verbose=False)
test_loader = torch.utils.data.DataLoader(test_set,
batch_size=batch_size,
shuffle=False)
# PRINTING DATA
import sys
import numpy
"""
for mydict in test_loader:
data = mydict['jetConstituentList']
target = mydict['jets']
np.set_printoptions(suppress=True)
if mymodel.optimizer == 1:
optimizer = optim.Adadelta(mymodel.parameters(), lr = 0.0001)
else:
optimizer = optim.Adam(mymodel.parameters(), lr = 0.0001)
loss_train = np.zeros(n_epochs)
loss_val = np.zeros(n_epochs)
nBatches_per_training_epoch = len(inputTrainFiles)*10000/batch_size
nBatches_per_validation_epoch = len(inputValFiles)*10000/batch_size
print("nBatches_per_training_epoch: %i" %nBatches_per_training_epoch)
print("nBatches_per_validation_epoch: %i" %nBatches_per_validation_epoch)
for i in range(n_epochs):
if mymodel.verbose: print("Epoch %s" % i)
# Define the data generators from the training set and validation set.
random.shuffle(inputTrainFiles)
random.shuffle(inputValFiles)
train_set = InEventLoader(file_names=inputTrainFiles, nP=nParticles,
feature_name ='jetConstituentList',label_name = 'jets', verbose=False)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=False)
val_set = InEventLoader(file_names=inputValFiles, nP=nParticles,
feature_name ='jetConstituentList',label_name = 'jets', verbose=False)
val_loader = torch.utils.data.DataLoader(val_set, batch_size=batch_size, shuffle=False)
####
# train
for batch_idx, mydict in enumerate(train_loader):
data = mydict['jetConstituentList']
target = mydict['jets']
if args_cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data), Variable(target)
optimizer.zero_grad()
out = mymodel(data)
l = loss(out, target)
def model_evaluate(mymodel):
#loss = nn.CrossEntropyLoss(reduction='sum')
loss = nn.CrossEntropyLoss(reduction='mean')
if mymodel.optimizer == 1:
optimizer = optim.Adadelta(mymodel.parameters(), lr=0.0001)
else:
optimizer = optim.Adam(mymodel.parameters(), lr=0.0001)
loss_train = np.zeros(n_epochs)
loss_val = np.zeros(n_epochs)
# Define the data generators from the training set and validation set. Let's try a batch size of 12.
import glob
#inputTrainFiles = glob.glob("/data/ML/mpierini/hls-fml/jetImage*_%sp*.h5" %nParticles)
#inputValFiles = glob.glob("/data/ML/mpierini/hls-fml/VALIDATION/jetImage*_%sp*.h5" %nParticles)
import os
if os.path.isdir('/imdata/NEWDATA'):
inputTrainFiles = glob.glob("/imdata/NEWDATA/jetImage*_%sp*.h5" %
nParticles)
inputValFiles = glob.glob(
"/imdata/NEWDATA/VALIDATION/jetImage*_%sp*.h5" % nParticles)
elif os.path.isdir('/data/shared/hls-fml/NEWDATA'):
inputTrainFiles = glob.glob(
"/data/shared/hls-fml/NEWDATA/jetImage*_%sp*.h5" % nParticles)
inputValFiles = glob.glob(
"/data/shared/hls-fml/NEWDATA/VALIDATION/jetImage*_%sp*.h5" %
nParticles)
elif os.path.isdir('/home/jduarte/NEWDATA'):
inputTrainFiles = glob.glob("/home/jduarte/NEWDATA/jetImage*_%sp*.h5" %
nParticles)
inputValFiles = glob.glob(
"/home/jduarte/NEWDATA/VALIDATION/jetImage*_%sp*.h5" % nParticles)
elif os.path.isdir('/bigdata/shared/hls-fml/NEWDATA'):
inputTrainFiles = glob.glob(
"/bigdata/shared/hls-fml/NEWDATA/jetImage*_%sp*.h5" % nParticles)
inputValFiles = glob.glob(
"/bigdata/shared/hls-fml/NEWDATA/VALIDATION/jetImage*_%sp*.h5" %
nParticles)
random.shuffle(inputTrainFiles)
random.shuffle(inputValFiles)
print("example file", inputTrainFiles[0])
nBatches_per_training_epoch = len(inputTrainFiles) * 10000 / batch_size
nBatches_per_validation_epoch = len(inputValFiles) * 10000 / batch_size
train_set = InEventLoader(file_names=inputTrainFiles,
nP=nParticles,
feature_name='jetConstituentList',
label_name='jets',
verbose=False)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=batch_size,
shuffle=False)
val_set = InEventLoader(file_names=inputValFiles,
nP=nParticles,
feature_name='jetConstituentList',
label_name='jets',
verbose=False)
val_loader = torch.utils.data.DataLoader(val_set,
batch_size=batch_size,
shuffle=False)
for i in range(n_epochs):
if mymodel.verbose: print("Epoch %s" % i)
#for j in range(0, xtrain.size()[0], batch_size):
#for (batch_idx, mydict) in tqdm(enumerate(train_loader),total=nBatches_per_training_epoch):
for (batch_idx, mydict) in enumerate(train_loader):
data = mydict['jetConstituentList']
target = mydict['jets']
if args_cuda:
data, target = data.cuda(), target.cuda()
#### ?????
data, target = Variable(data), Variable(target)
optimizer.zero_grad()
out = mymodel(data)
l = loss(out, target)
l.backward()
optimizer.step()
loss_train[i] += l.cpu().data.numpy() / nBatches_per_training_epoch
#loss_train[i] = loss_train[i]/float(xtrain.size()[0])
#for j in range(0, xval.size()[0], batch_size):
#for (batch_idx, mydict) in tqdm(enumerate(val_loader),total=nBatches_per_validation_epoch):
for (batch_idx, mydict) in enumerate(val_loader):
data = mydict['jetConstituentList']
target = mydict['jets']
if args_cuda:
data, target = data.cuda(), target.cuda()
#### ?????
data, target = Variable(data, volatile=True), Variable(target)
out_val = mymodel(data)
l_val = loss(out_val, target)
loss_val[i] += l_val.cpu().data.numpy(
) / nBatches_per_validation_epoch
#loss_val[i] = loss_val[i]/float(xval.size()[0])
if mymodel.verbose: print("Training Loss: %f" % loss_train[i])
if mymodel.verbose: print("Validation Loss: %f" % loss_val[i])
#that below does not trigger soon enough
if all(loss_val[max(0, i - patience):i] > min(
np.append(loss_val[0:max(0, i -
patience)], 200))) and i > patience:
print("Early Stopping at", i)
break
#that above does not trigger soon enough
if i > (2 * patience):
last_avg = np.mean(loss_val[i - patience:i])
previous_avg = np.mean(loss_val[i - 2 * patience:i - patience])
#print ("last",last_avg,"previous",previous_avg)
if last_avg > previous_avg:
print("Early Avg Stopping at", i)
break
if i > patience:
last_min = min(loss_val[i - patience:i])
overall_min = min(loss_val[:i - patience])
#print ("last",last_min,"overall",overall_min)
if last_min > overall_min:
print("Early min Stopping at", i)
break
loss_val = loss_val[loss_val > 0]
return loss_val[-1]