Update codes.py

codes.py

@@ -89,79 +89,7 @@ def load_image_pil_accelerated(image_path, dim=128):
 
 
 def preprocess_image(image_path, dim = 128):
-    img = load_image_pil_accelerated(image_path, dim)
-    return img.unsqueeze(0)
-
-'''
-def load_image_from_url(image_path, dim = 128):
-    img = Image.open(image_path).convert("RGB")
-    img = img.resize((dim, dim))
-    return img
-
-def preprocess_image(image_path, dim = 128):
-    img = load_img(image_path, target_size=(dim, dim))
-    img = img_to_array(img)
-    img = np.expand_dims(img, axis=0)
+    img = torch.zeros([1,3,dim,dim])
+    img[0] = load_image_pil_accelerated(image_path, dim)
     return img
 
-
-def create_model(dim = 128):
-  # configure unet input shape (concatenation of moving and fixed images)
-  volshape = (dim,dim,3)
-  unet_input_features = 2*volshape[:-1]
-  inshape = (*volshape[:-1],unet_input_features)
-  nb_conv_per_level=1
-  enc_nf = [dim, dim, dim, dim]
-  dec_nf = [dim, dim, dim, dim, dim, int(dim/2)]
-  nb_upsample_skips = 0
-  nb_dec_convs = len(enc_nf)
-  final_convs = dec_nf[nb_dec_convs:]
-  dec_nf = dec_nf[:nb_dec_convs]
-  nb_levels = int(nb_dec_convs / nb_conv_per_level) + 1
-  source = tf.keras.Input(shape=volshape, name='source_input')
-  target = tf.keras.Input(shape=volshape, name='target_input')
-  inputs = [source, target]
-  unet_input = concatenate(inputs, name='input_concat')
-  #Define lyers
-  ndims = len(unet_input.get_shape()) - 2
-  MaxPooling = getattr(tf.keras.layers, 'MaxPooling%dD' % ndims)
-  Conv = getattr(tf.keras.layers, 'Conv%dD' % ndims)
-  UpSampling = getattr(tf.keras.layers, 'UpSampling%dD' % ndims)
-  # Encoder
-  enc_layers = []
-  lyr = unet_input
-  for level in range(nb_levels - 1):
-      for conv in range(nb_conv_per_level):
-          nfeat = enc_nf[level * nb_conv_per_level + conv]
-          lyr = Conv(nfeat, kernel_size=3, padding='same', strides=1,activation = LeakyReLU(0.2), kernel_initializer = 'he_normal')(lyr)
-      enc_layers.append(lyr)
-      lyr = MaxPooling(2)(lyr)
-
-  # Decoder
-  for level in range(nb_levels - 1):
-      real_level = nb_levels - level - 2
-      for conv in range(nb_conv_per_level):
-          nfeat = dec_nf[level * nb_conv_per_level + conv]
-          lyr = Conv(nfeat, kernel_size=3, padding='same', strides=1,activation = LeakyReLU(0.2), kernel_initializer = 'he_normal')(lyr)
-      # upsample
-      if level < (nb_levels - 1 - nb_upsample_skips):
-          upsampled = UpSampling(size=(2,) * ndims)(lyr)
-          lyr = concatenate([upsampled, enc_layers.pop()])
-
-  # Final convolution
-  for num, nfeat in enumerate(final_convs):
-      lyr = Conv(nfeat, kernel_size=3, padding='same', strides=1,activation = LeakyReLU(0.2), kernel_initializer = 'he_normal')(lyr)
-
-  unet =  tf.keras.models.Model(inputs=inputs, outputs=lyr)
-  # transform the results into a flow field.
-  disp_tensor = Conv(ndims, kernel_size=3, padding='same', name='disp')(unet.output)
-  # using keras, we can easily form new models via tensor pointers
-  def_model = tf.keras.models.Model(inputs, disp_tensor)
-  # build transformer layer
-  spatial_transformer = SpatialTransformer()
-  # warp the moving image with the transformer
-  moved_image_tensor = spatial_transformer([source, disp_tensor])
-  outputs = [moved_image_tensor, disp_tensor]
-  vxm_model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
-  return vxm_model
-'''