Merge pull request #2 from jramapuram/hotfix/pytorch0.4

layers need int and model needs relative module

models/HVAE_2level.py

@@ -139,7 +139,7 @@ def calculate_likelihood(self, X, dir, mode='test', S=5000, MB=500):
  x_single = X[j].unsqueeze(0)
 
  a = []
- for r in range(0, R):
+ for r in range(0, int(R)):
  # Repeat it for all training points
  x = x_single.expand(S, x_single.size(1))
 

models/PixelHVAE_2level.py

@@ -157,7 +157,7 @@ def calculate_likelihood(self, X, dir, mode='test', S=5000, MB=500):
  x_single = X[j].unsqueeze(0)
 
  a = []
- for r in range(0, R):
+ for r in range(0, int(R)):
  # Repeat it for all training points
  x = x_single.expand(S, x_single.size(1)).contiguous()
 

models/VAE.py

@@ -17,7 +17,7 @@
 from utils.visual_evaluation import plot_histogram
 from utils.nn import he_init, GatedDense, NonLinear
 
-from Model import Model
+from models.Model import Model
 # -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
 
 #=======================================================================================================================
@@ -108,7 +108,7 @@ def calculate_likelihood(self, X, dir, mode='test', S=5000, MB=100):
  x_single = X[j].unsqueeze(0)
 
  a = []
- for r in range(0, R):
+ for r in range(0, int(R)):
  # Repeat it for all training points
  x = x_single.expand(S, x_single.size(1))
 

models/convHVAE_2level.py

@@ -160,7 +160,7 @@ def calculate_likelihood(self, X, dir, mode='test', S=5000, MB=500):
  x_single = X[j].unsqueeze(0)
 
  a = []
- for r in range(0, R):
+ for r in range(0, int(R)):
  # Repeat it for all training points
  x = x_single.expand(S, x_single.size(1)).contiguous()
 

utils/evaluation.py

@@ -48,9 +48,22 @@ def evaluate_vae(args, model, train_loader, data_loader, epoch, dir, mode):
 
  if mode == 'test':
  # load all data
- test_data = Variable(data_loader.dataset.data_tensor)
- test_target = Variable(data_loader.dataset.target_tensor)
- full_data = Variable(train_loader.dataset.data_tensor)
+ # grab the test data by iterating over the loader
+ # there is no standardized tensor_dataset member across pytorch datasets
+ test_data, test_target = [], []
+ for data, lbls in data_loader:
+ test_data.append(data)
+ test_target.append(lbls)
+
+ test_data, test_target = [torch.cat(test_data, 0), torch.cat(test_target, 0).squeeze()]
+
+ # grab the train data by iterating over the loader
+ # there is no standardized tensor_dataset member across pytorch datasets
+ full_data = []
+ for data, _ in train_loader:
+ full_data.append(data)
+
+ full_data = torch.cat(full_data, 0)
 
  if args.cuda:
  test_data, test_target, full_data = test_data.cuda(), test_target.cuda(), full_data.cuda()
@@ -110,4 +123,4 @@ def evaluate_vae(args, model, train_loader, data_loader, epoch, dir, mode):
  if mode == 'test':
  return evaluate_loss, evaluate_re, evaluate_kl, log_likelihood_test, log_likelihood_train, elbo_test, elbo_train
  else:
- return evaluate_loss, evaluate_re, evaluate_kl
+ return evaluate_loss, evaluate_re, evaluate_kl

utils/nn.py

@@ -42,7 +42,7 @@ def __init__(self, input_size, output_size, bias=True, activation=None):
  super(NonLinear, self).__init__()
 
  self.activation = activation
- self.linear = nn.Linear(input_size, output_size, bias=bias)
+ self.linear = nn.Linear(int(input_size), int(output_size), bias=bias)
 
  def forward(self, x):
  h = self.linear(x)
@@ -301,4 +301,4 @@ def forward(self, x):
  h2 = self.bn2(h1)
  h2 = self.act(h2)
  h2 = self.h2(h2)
- return x + h2
+ return x + h2