def load_model(path=None):
'''
Function to load model
'''
model = tiramisu.FCDenseNet103(n_classes=8).cuda()
weights = torch.load(path)
model.load_state_dict(weights['state_dict'])
return model
def load_model(path=None):
'''
parameters: path to the model weights
return: Model with desired architecture and weights
'''
model = tiramisu.FCDenseNet103(n_classes=8).cuda()
weights = torch.load(path)
model.load_state_dict(weights['state_dict'])
return model
def load_model(path=None): ''' function to load the model parameters : path : the path of the model state dictionary return: model: the model along with the random weights if path is none otherwise initalized with pretrained weight ''' model = tiramisu.FCDenseNet103(n_classes=8).to(device) if path is not None: weights = torch.load(path, map_location = device) model.load_state_dict(weights['state_dict']) return model
val_loader = torch.utils.data.DataLoader(val_cilia,
batch_size=1,
shuffle=True)
print("Loaded training set!")
test_cilia = getCilia.CiliaData(ROOT,
'test',
joint_transform=None,
input_transform=img_transform)
test_loader = torch.utils.data.DataLoader(test_cilia,
batch_size=1,
shuffle=False)
print("Loaded testing set!")
model = tiramisu.FCDenseNet103(n_classes=3, in_channels=1).cuda()
model.apply(training_utils.weights_init)
optimizer = torch.optim.Adam(model.parameters(), lr=LR, weight_decay=1e-4)
criterion = nn.NLLLoss().cuda()
# Training process
for epoch in range(1, N_EPOCHS + 1):
since = time.time()
### Train ###
trn_loss, trn_err = training_utils.train(model, train_loader,
optimizer, criterion, epoch)
print('Epoch {:d}\nTrain - Loss: {:.4f}, Acc: {:.4f}'.format(
epoch, trn_loss, 1 - trn_err))
time_elapsed = time.time() - since
print('Train Time {:.0f}m {:.0f}s'.format(time_elapsed // 60,
time_elapsed % 60))
dropout_data = {
'mode': 'concrete_dropout',
'init_min': 0.01,
'init_max': 0.05,
'weight_reg': 1e-8,
'dropout_reg': 1e0,
'gpu_id': gpu_id
}
elif mode == 'dropout':
dropout_data = {'mode': 'dropout', 'prob': 0.2}
else:
dropout_data = {'mode': 'dropout', 'prob': 0.0}
print(dropout_data)
model = tiramisu.FCDenseNet103(12, dropout_data).cuda(gpu_id)
model.apply(train_utils.weights_init)
optimizer = torch.optim.RMSprop(model.parameters(), lr=LR, weight_decay=1e-4)
criterion = nn.NLLLoss2d(weight=camvid.class_weight.cuda(gpu_id)).cuda(gpu_id)
print('\n######### TRAINING #########\n')
def collect_dropout_probabilities(model):
dropout_probabilities = []
def add_dropout_prob(module):
nonlocal dropout_probabilities
if module.__class__.__name__.endswith('ConcreteDropout'):
#utils.imgs.view_annotated(targets[0])
# ## Train
# In[5]:
LR = 1e-4
LR_DECAY = 0.995
DECAY_EVERY_N_EPOCHS = 1
N_EPOCHS = 100
torch.cuda.manual_seed(0)
# In[6]:
#model = tiramisu.FCDenseNet67(n_classes=12).cuda()
model = tiramisu.FCDenseNet103(n_classes=12).cuda()
model.apply(train_utils.weights_init)
optimizer = torch.optim.RMSprop(model.parameters(), lr=LR, weight_decay=1e-4)
criterion = nn.NLLLoss2d(weight=camvid.class_weight.cuda()).cuda()
# In[7]:
def train(epoch):
since = time.time()
### Train ###
trn_loss, trn_err = train_utils.train(model, train_loader, optimizer,
criterion, epoch)
#print('Epoch {:d}, Train - Loss: {:.4f}, Acc: {:.4f}'.format(epoch, trn_loss, 1-trn_err))
print("Epoch: %d, Train - Loss: %.4f, Acc: %.4f" %
(epoch, trn_loss, 1 - trn_err))
time_elapsed = time.time() - since
def load_model(path=None):
if path == None:
model = tiramisu.FCDenseNet103(n_classes=30).cuda()
model.apply(train_utils.weights_init)
return model
