batch_size=batch_size,
shuffle=False,
drop_last=True)
print("Validation batches: {}".format(len(valiLoader)))
# setup training
epochs = int(iterations / len(trainLoader) + 0.5)
netG = TurbNetG(channelExponent=expo, dropout=dropout)
print(netG) # print full net
model_parameters = filter(lambda p: p.requires_grad, netG.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print("Initialized TurbNet with {} trainable params ".format(params))
netG.apply(weights_init)
if len(doLoad) > 0:
netG.load_state_dict(torch.load(doLoad))
print("Loaded model " + doLoad)
netG.cuda()
criterionL1 = nn.L1Loss()
criterionL1.cuda()
optimizerG = optim.Adam(netG.parameters(),
lr=lrG,
betas=(0.5, 0.999),
weight_decay=0.0)
targets = Variable(torch.FloatTensor(batch_size, 3, 128, 128))
inputs = Variable(torch.FloatTensor(batch_size, 3, 128, 128))
targets = targets.cuda()
inputs = inputs.cuda()
losses = []
models = []
# for si in range(25):
for si in range(1):
s = chr(96 + si)
if (si == 0):
s = "" # check modelG, and modelG + char
# modelFn = "./" + prefix + "modelG{}{}".format(suffix,s)
modelFn = path
if not os.path.isfile(modelFn):
continue
models.append(modelFn)
log(lf, "Loading " + modelFn)
netG.load_state_dict(torch.load(modelFn, map_location=device))
log(lf, "Loaded " + modelFn)
netG.to(device)
criterionL1 = nn.L1Loss()
criterionL1.to(device)
L1val_accum = 0.0
L1val_dn_accum = 0.0
lossPer_p_accum = 0
lossPer_v_accum = 0
lossPer_accum = 0
netG.eval()
for i, data in enumerate(testLoader, 0):
inputs_cpu, targets_cpu = data
models = []
loss_p_list = []
loss_v_list = []
accum_list = []
for si in range(25):
s = chr(96 + si)
if (si == 0):
s = "" # check modelG, and modelG + char
modelFn = "./" + prefix + "modelG{}{}".format(suffix, s)
if not os.path.isfile(modelFn):
continue
models.append(modelFn)
log(lf, "Loading " + modelFn)
netG.load_state_dict(torch.load(modelFn))
log(lf, "Loaded " + modelFn)
netG.cuda()
criterionLoss = nn.L1Loss()
#criterionLoss = nn.MSELoss()
#criterionLoss = nn.SmoothL1Loss()
criterionLoss.cuda()
Lossval_accum = 0.0
Lossval_dn_accum = 0.0
lossPer_p_accum = 0
lossPer_v_accum = 0
lossPer_accum = 0
netG.eval()
def getModel(expo):
netG = TurbNetG(channelExponent=expo).to(device)
netG.load_state_dict(torch.load(f'models/model_w_{expo}', map_location=device))
netG.eval()
return netG