Fix reshape/transpose issues seen by pytorch

Pytorch does not accept numpy.ndarray shapes, we must use tuples (or
lists) of integers.

src/decomon/backward_layers/utils_conv.py

@@ -72,7 +72,7 @@ def get_toeplitz_channels_last(conv_layer: Conv2D, flatten: bool = True) -> Back
 
     diag_patches_ = K.reshape(diag_patches, (w_in, h_in, c_in, w_out, h_out, filter_size**2, c_in))
 
-    shape = np.arange(len(diag_patches_.shape))
+    shape = list(range(len(diag_patches_.shape)))
     shape[-1] -= 1
     shape[-2] += 1
     diag_patches_ = K.transpose(diag_patches_, shape)
@@ -135,7 +135,7 @@ def get_toeplitz_channels_first(conv_layer: Conv2D, flatten: bool = True) -> Bac
 
     diag_patches_ = K.reshape(diag_patches, (w_in, h_in, c_in, w_out, h_out, filter_size**2, c_in))
 
-    shape = np.arange(len(diag_patches_.shape))
+    shape = list(range(len(diag_patches_.shape)))
     shape[-1] -= 1
     shape[-2] += 1
     diag_patches_ = K.transpose(diag_patches_, shape)

src/decomon/layers/utils.py

@@ -560,9 +560,9 @@ def permute_dimensions(
         # not enough dim to permute
         return inputs
     index = np.arange(len(input_shape))
-    index = np.insert(np.delete(index, axis), axis_perm, axis)
+    index = tuple(np.insert(np.delete(index, axis), axis_perm, axis))
     index_w = np.arange(len(input_shape) + 1)
-    index_w = np.insert(np.delete(index_w, axis), axis_perm + 1, axis)
+    index_w = tuple(np.insert(np.delete(index_w, axis), axis_perm + 1, axis))
 
     if ibp:
         u_c_out = K.transpose(u_c, index)

src/decomon/layers/utils_pooling.py

@@ -65,7 +65,7 @@ def get_upper_linear_hull_max(
 
     # expand dim/broadcast
     mask = K.identity(n_dim, dtype=dtype)  # (n_dim, n_dim)
-    mask_shape = np.ones(len(u_c.shape) + 1)
+    mask_shape = [1] * (len(u_c.shape) + 1)
     mask_shape[-1] = n_dim
     if axis != -1:
         mask_shape[axis] = n_dim
@@ -101,11 +101,11 @@ def get_upper_linear_hull_max(
         bias_corner = o_value + K.sum(z_value * corners, -2, keepdims=True)
         corners_collapse = K.concatenate([corners_collapse, bias_corner], axis=-2)
 
-    dimensions = np.arange(len(corners.shape))
+    dimensions = list(range(len(corners.shape)))
     if axis != -1:
-        dim_permutation = np.concatenate([dimensions[:axis], dimensions[axis + 1 :], [dimensions[axis]]])
+        dim_permutation = dimensions[:axis] + dimensions[axis + 1 :] + [dimensions[axis]]
     else:
-        dim_permutation = np.concatenate([dimensions[:-2], dimensions[-1:], [dimensions[-2]]])
+        dim_permutation = dimensions[:-2] + dimensions[-1:] + [dimensions[-2]]
 
     corners_collapse = K.transpose(corners_collapse, dim_permutation)
     # tf.linalg.solve works only for float32