Fixes some issues

cascadenet/network/layers/conv.py

@@ -1,7 +1,7 @@
 import theano
 import lasagne
 from lasagne.layers import Layer, prelu
-from helper import ensure_set_name
+from .helper import ensure_set_name
 
 if theano.config.device == 'cuda':
     from lasagne.layers.dnn import Conv2DDNNLayer as ConvLayer

cascadenet/network/layers/data_consistency.py

@@ -1,6 +1,6 @@
 import theano.tensor as T
 from lasagne.layers import MergeLayer, get_output
-from fourier import FFT2Layer, FFTCLayer
+from .fourier import FFT2Layer, FFTCLayer
 from cascadenet.network.theano_extensions.fft_helper import fftshift, ifftshift
 from cascadenet.network.theano_extensions.fft import fft, ifft 
 

cascadenet/network/layers/fourier.py

@@ -18,11 +18,11 @@
     from cascadenet.network.theano_extensions.gpuarray.fft2 import cufft2 as fft2g
     from cascadenet.network.theano_extensions.gpuarray.fft2 import cuifft2 as ifft2g
     use_cuda = True
-    print "Using GPU version of fft layers"
+    print ("Using GPU version of fft layers")
 else:
     from cascadenet.network.theano_extensions.fft2_lasagne import fft2, ifft2
     use_cuda = False
-    print "Using CPU version of fft layers"
+    print ("Using CPU version of fft layers")
 
 
 from cascadenet.network.theano_extensions.fft_helper import fftshift, ifftshift
@@ -215,7 +215,7 @@ def get_output_for(self, input, **kwargs):
             # nt = self.data_shape[-1]
             # out, updates = theano.scan(loop_over_n,
             #                            non_sequences=input,
-            #                            sequences=xrange(nt))
+            #                            sequences=range(nt))
             # return out
 
         out, updates = theano.scan(self.transform, sequences=input)

cascadenet/network/layers/pool.py

@@ -84,7 +84,7 @@ def get_output_shape_for(self, input_shape):
 
         tr = len(output_shape) - self.n
 
-        for i in xrange(self.n):
+        for i in range(self.n):
             output_shape[tr+i] = pool.pool_output_length(input_shape[tr+i],
                                                          pool_size=self.pool_size[i],
                                                          stride=self.stride[i],

cascadenet/network/layers/shape.py

@@ -24,9 +24,9 @@ def get_output_for(self, input, deterministic=False, **kwargs):
         out = T.zeros((input.shape[0], self.output_shape[1],
                        self.output_shape[2], self.output_shape[3]))
 
-        for x in xrange(self.r):
+        for x in range(self.r):
             # loop across all feature maps belonging to this channel
-            for y in xrange(self.r):
+            for y in range(self.r):
                 out = T.inc_subtensor(out[:, :, x::self.r, y::self.r],
                                       input[:, self.r*x+y::self.r*self.r, :, :])
         return out
@@ -51,10 +51,10 @@ def get_output_for(self, input, deterministic=False, **kwargs):
         out = T.zeros((input.shape[0], self.output_shape[1],
                        self.output_shape[2], self.output_shape[3], self.output_shape[4]))
 
-        for x in xrange(r):
+        for x in range(r):
             # loop across all feature maps belonging to this channel
-            for y in xrange(r):
-                for z in xrange(r):
+            for y in range(r):
+                for z in range(r):
                     out = T.inc_subtensor(out[..., x::r, y::r, z::r],
                                           input[:, (r**2)*x+(r*y)+z::r*r*r, ...])
         return out

cascadenet/network/layers/simple.py

@@ -1,7 +1,7 @@
 import theano.tensor as T
 import lasagne
 from lasagne.layers import Layer
-from helper import ensure_set_name
+from .helper import ensure_set_name
 
 
 class IdLayer(Layer):

cascadenet/network/model.py

@@ -8,7 +8,7 @@ def cascade_resnet(pr, net, input_layer, n=5, nf=64, b=lasagne.init.Constant, **
     n_channel = shape[1]
     net[pr+'conv1'] = l.Conv(input_layer, nf, 3, b=b(), name=pr+'conv1')
 
-    for i in xrange(2, n):
+    for i in range(2, n):
         net[pr+'conv%d'%i] = l.Conv(net[pr+'conv%d'%(i-1)], nf, 3, b=b(),
                                     name=pr+'conv%d'%i)
 
@@ -37,7 +37,7 @@ def cascade_resnet_3d_avg(pr, net, input_layer, n=5, nf=64,
     # Conv layers
     net[pr+'conv1'] = l.Conv3D(net[pr+'kavg'], nf, k, b=b(), name=pr+'conv1')
 
-    for i in xrange(2, n):
+    for i in range(2, n):
         net[pr+'conv%d'%i] = l.Conv3D(net[pr+'conv%d'%(i-1)], nf, k, b=b(),
                                       name=pr+'conv%d'%i)
 
@@ -67,7 +67,7 @@ def build_cascade_cnn_from_list(shape, net_meta, lmda=None):
     j = 0
     for cascade_net, cascade_n in net_meta:
         # Cascade layer
-        for i in xrange(cascade_n):
+        for i in range(cascade_n):
             pr = 'c%d_' % j
             net, output_layer = cascade_net(pr, net, input_layer,
                                             **{'cascade_i': j})

main_2d.py

@@ -45,7 +45,7 @@ def iterate_minibatch(data, batch_size, shuffle=True):
     if shuffle:
         data = np.random.permutation(data)
 
-    for i in xrange(0, n, batch_size):
+    for i in range(0, n, batch_size):
         yield data[i:i+batch_size]
 
 
@@ -171,7 +171,7 @@ def compile_fn(network, net_config, args):
     train, validate, test = create_dummy_data()
 
     print('Start Training...')
-    for epoch in xrange(num_epoch):
+    for epoch in range(num_epoch):
         t_start = time.time()
         # Training
         train_err = 0

main_3d.py

@@ -45,7 +45,7 @@ def iterate_minibatch(data, batch_size, shuffle=True):
     if shuffle:
         data = np.random.permutation(data)
 
-    for i in xrange(0, n, batch_size):
+    for i in range(0, n, batch_size):
         yield data[i:i+batch_size]
 
 
@@ -173,7 +173,7 @@ def compile_fn(network, net_config, args):
     # Compile function
     train_fn, val_fn = compile_fn(net, net_config, args)
 
-    for epoch in xrange(num_epoch):
+    for epoch in range(num_epoch):
         t_start = time.time()
         # Training
         train_err = 0

utils/compressed_sensing.py

@@ -104,7 +104,7 @@ def shear_grid_mask(shape, acceleration_rate, sample_low_freq=True,
     Nt, Nx, Ny = shape
     start = np.random.randint(0, acceleration_rate)
     mask = np.zeros((Nt, Nx))
-    for t in xrange(Nt):
+    for t in range(Nt):
         mask[t, (start+t)%acceleration_rate::acceleration_rate] = 1
 
     xc = Nx / 2
@@ -135,15 +135,15 @@ def perturbed_shear_grid_mask(shape, acceleration_rate, sample_low_freq=True,
     Nt, Nx, Ny = shape
     start = np.random.randint(0, acceleration_rate)
     mask = np.zeros((Nt, Nx))
-    for t in xrange(Nt):
+    for t in range(Nt):
         mask[t, (start+t)%acceleration_rate::acceleration_rate] = 1
 
     # brute force
     rand_code = np.random.randint(0, 3, size=Nt*Nx)
     shift = np.array([-1, 0, 1])[rand_code]
     new_mask = np.zeros_like(mask)
-    for t in xrange(Nt):
-        for x in xrange(Nx):
+    for t in range(Nt):
+        for x in range(Nx):
             if mask[t, x]:
                 new_mask[t, (x + shift[t*x])%Nx] = 1