if opts.evalMode & (not opts.loadSVM): #to train an SVM for eval
svm = train_svm(dae=dae,
svm=svm,
trainLoader=trainLoader,
testLoader=testLoader,
exDir=opts.load_DAE_from,
lr=opts.svmLR)
if opts.evalMode:
assert opts.loadDAE == True
eval_mode(dae, opts.load_DAE_from, opts.M, testLoader, svm=svm)
exit()
else:
#Create a folder for this experiment
exDir = make_new_folder(opts.outDir)
print 'Outputs will be saved to:', exDir
save_input_args(exDir, opts) #save training opts
#Create optimizers
optimDAE = optim.RMSprop(dae.parameters(), lr=opts.lr)
optimDIS = optim.RMSprop(dis.parameters(),
lr=opts.lr,
momentum=opts.momentum)
#Keeping track of training
losses = {'enc': [], 'rec': [], 'dis': [], 'test rec': []}
#Start training
for e in range(opts.maxEpochs):
if vae.useCUDA:
print 'using CUDA'
vae.cuda()
classer.cuda()
else:
print '\n *** NOT USING CUDA ***\n'
#eval or train
if opts.evalMode:
opts.loadCLASSER = True
opts.loadDELTAZ = True
opts.loadVAE = True
else:
#create a folder to save the exp
exDir = make_new_folder(
opts.outDir) # Create a new folder to save results and model info
print 'Outputs will be saved to:', exDir
save_input_args(exDir, opts)
####### Load pre-trained model or train a model #######
#Load pre-trained nets if available
if opts.loadVAE:
print 'loadding vae...'
vae.load_params(opts.load_VAE_from)
if opts.loadCLASSER:
print 'loadding classer...'
classer.load_params(opts.load_CLASSER_from)
if opts.loadDELTAZ:
print 'loading deltaZ...'
deltaZ = torch.load(join(opts.load_DELTAZ_from, 'deltaZ'))
else:
