def Validation(model, optimizer, lossfunction, targetvar, train_graphs,
val_graphs, train_queue, val_queue, val_queue_fill,
train_queue_fill, workers, batch_size, val_batches, epoch,
epochs):
deadcheck = False
traincheck = False
val_loss = torch.tensor([0], dtype=float).to(device)
for k in range(val_batches):
with torch.no_grad():
batch = Processing.GrabBatch(val_queue, device)
output = model(batch)[:, 0]
loss = lossfunction(output, batch.y[:, targetvar])
val_loss += loss
deadcheck = Processing.Process_check(val_queue_fill, deadcheck, k,
val_batches)
if deadcheck == True and traincheck == False and epoch < epochs - 1:
train_queue_fill = Processing.Spawn_Processes(
workers, train_graphs, train_queue, batch_size)
traincheck = True
deadcheck = Processing.Process_check(val_queue_fill, deadcheck, k,
val_batches)
if deadcheck == False:
print('Validation Processes Still Alive. Terminating...')
for process in val_queue_fill:
process.terminate()
if epoch < epochs - 1:
train_queue_fill = Processing.Spawn_Processes(
workers, train_graphs, train_queue, batch_size)
return val_loss, train_queue_fill
def Predict(model,prediction_graphs,workers,pred_mini_batches,batch_size,currfolder,device,targetvar):
# print('PREDICTING: \n \
# model : %s \n \
# n_events: %s' %(baseline,pred_mini_batches*batch_size))
predictions = []
truths = []
pred_events = []
manager = torch.multiprocessing.Manager()
q = manager.Queue()
slaves = Processing.Spawn_Processes(workers, prediction_graphs, q,batch_size)
dead_check = False
model.eval()
with torch.no_grad():
for mini_batch in range(0,pred_mini_batches):
data = Processing.GrabBatch(q,device)
prediction = model(data)
truth = Pidclass(data.y[:,targetvar]).unsqueeze(1).detach().cpu().numpy()
pred_events.extend(data.event_no.detach().cpu().numpy())
predictions.extend(prediction.detach().cpu().numpy())
truths.extend(truth)
dead_check =Processing.Process_check(slaves, dead_check, mini_batch, pred_mini_batches)
#print (" predict batch ",mini_batch,"out of " ,pred_mini_batches)
if( dead_check == False):
for slave in slaves:
slave.terminate()
print('Saving results...')
truths = pd.DataFrame(truths)
predictions = pd.DataFrame(predictions)
pred_events = pd.DataFrame(pred_events)
result = pd.concat([pred_events,truths, predictions],axis = 1)
result.columns = ['event_no','Pid','Antineutrino','Neutrino']
result.to_csv(currfolder + 'predictions.csv',index=False)
def Trainloop(model,optimizer,lossfunction,targetvar,device,train_graphs,val_graphs,workers,batch_size,train_batches,val_batches,epochs):
average_train_loss_per_epoch = list()
average_val_loss_per_epoch=list()
for epoch in tqdm(range(epochs)):
deadcheck=False
valcheck=False
model.train()
train_loss = torch.tensor([0],dtype = float).to(device)
if epoch==0:
Manager=torch.multiprocessing.Manager()
train_queue=Manager.Queue()
val_queue=Manager.Queue()
train_queue_fill=Processing.Spawn_Processes(workers, train_graphs, train_queue, batch_size)
for k in range(train_batches):
with torch.enable_grad():
model.train()
batch=Processing.GrabBatch(train_queue,device)
optimizer.zero_grad()
output=model(batch)
# print (output)
# print (batch.x,output.size(),batch)
# print(output.size())
loss=lossfunction(output,batch.y[:,targetvar])
# print (loss)
loss.backward()
optimizer.step()
# print (loss,output,batch.y[:,targetvar])
deadcheck=Processing.Process_check(train_queue_fill,deadcheck,k,train_batches)
train_loss +=loss
if deadcheck==True and valcheck==False:
val_queue_fill=Processing.Spawn_Processes(workers,val_graphs,val_queue,batch_size)
valcheck=True
# print (k)
if(torch.sum(torch.isnan(output)) != 0):
raise TypeError('NAN ENCOUNTERED AT : %s / %s'%(k,train_batches))
if( deadcheck == False):
print('Training Processes Still Alive. Terminating...')
for process in train_queue_fill:
process.terminate()
val_queue_fill=Processing.Spawn_Processes(workers,val_graphs,val_queue,batch_size)
if deadcheck==True and valcheck==False:
valcheck=True
with torch.no_grad():
val_loss,train_queue_fill=Validation(model,optimizer,lossfunction,targetvar,train_graphs,val_graphs,train_queue,val_queue,val_queue_fill,train_queue_fill,workers,batch_size,val_batches,epoch,epochs)
average_train_loss_per_epoch.append(train_loss.item()/(train_batches*batch_size))
average_val_loss_per_epoch.append(val_loss.item()/(val_batches*batch_size))
print (train_loss.item()/(train_batches*batch_size))
deadcheck=Processing.Process_check(train_queue_fill,deadcheck,k,train_batches)
if( deadcheck == False):
print('Training done. Terminating slaves...')
for process in train_queue_fill:
process.terminate()
del batch,loss,train_loss,val_loss
return model,average_train_loss_per_epoch,average_val_loss_per_epoch
train_graphs.append(train_graphs_job)
Manager=torch.multiprocessing.Manager()
train_queue=Manager.Queue()
val_queue=Manager.Queue()
lossfunction=CrossEntropyLoss().to(device)
lossarr=np.zeros(epochs)
for epoch in tqdm(range(epochs)):
train_queue_fill=Processing.Spawn_Processes(workers, train_graphs, train_queue, batch_size)
deadcheck=False
valcheck=False
model.train()
losstemp=np.zeros(valbatches)
for k in range(minibatches):
batch=Processing.GrabBatch(train_queue,device)
optimizer.zero_grad()
output=model(batch)
loss=lossfunction(output,batch.y)
loss.backward()
optimizer.step()
deadcheck=Processing.Process_check(train_queue_fill,deadcheck,k,minibatches)
if deadcheck==True and valcheck==False:
print (deadcheck)
val_queue_fill=Processing.Spawn_Processes(workers,val_graphs,val_queue,batch_size)
valcheck=True
model.eval()
for k in range(valbatches):
batch=Processing.GrabBatch(val_queue,device)
output=model(batch)