Add install instruction for new interface

README.md

@@ -1,12 +1,35 @@
 Code to accompany the paper <a href="https://arxiv.org/abs/1903.05895">Learning Fast Algorithms for Linear Transforms Using Butterfly Factorizations.
 
 ## Requirements
-Python 3.6+  
-PyTorch >=1.2
-Numpy
+python>=3.6  
+pytorch>=1.6  
+numpy  
+scipy
 
 ## Usage
 
+2020-08-03: The new interface to butterfly C++/CUDA code is in `csrc` and
+`torch_butterfly`.
+It is tested in `tests/test_torch_butterfly.py` (which also shows example
+usage).
+
+To install:
+```
+python setup.py install
+```
+That is, use the `setup.py` file in this root directory.
+
+The file `torch_butterfly/special.py` shows how to construct butterfly matrices
+that performs FFT, inverse FFT, circulant matrix multiplication,
+Hadamard transform, and torch.nn.Conv1d with circular padding. The tests in
+`tests/test_special.py` show that these butterfly matrices exactly perform
+those operations.
+
+## Old interface
+
+Note: this interface is being rewritten. Only use this if you need some feature
+that's not supported in the new interface.
+
 * The module `Butterfly` in `butterfly/butterfly.py` can be used as a drop-in
 replacement for a `nn.Linear` layer. The files in `butterfly` directory are all
 that are needed for this use.

torch_butterfly/butterfly.py

@@ -17,10 +17,10 @@ class Butterfly(nn.Module):
         bias: If set to False, the layer will not learn an additive bias.
                 Default: ``True``
         complex: whether complex or real
-        increasing_stride: whether to multiply with increasing stride (e.g. 1, 2, ..., n/2) or
-            decreasing stride (e.g., n/2, n/4, ..., 1).
+        increasing_stride: whether the first butterfly block will multiply with increasing stride
+            (e.g. 1, 2, ..., n/2) or decreasing stride (e.g., n/2, n/4, ..., 1).
         ortho_init: whether the weight matrix should be initialized to be orthogonal/unitary.
-        nblocks: number of (BB^T) blocks.
+        nblocks: number of B or B^T blocks. The B and B^T will alternate.
     """
 
     def __init__(self, in_size, out_size, bias=True, complex=False,