Fix bugs in inference; update DistributedDataParallel; release a pre-…

…trained FullSubNet using cumulative norm

README.md

@@ -9,7 +9,7 @@ This Git repository for the official PyTorch implementation
 of ["FullSubNet: A Full-Band and Sub-Band Fusion Model for Real-Time Single-Channel Speech Enhancement"](https://arxiv.org/abs/2010.15508), accepted
 to ICASSP 2021. 
 
-:bulb:[[Demo\]](https://www.haoxiangsnr.com/demo/fullsubnet/) | :page_facing_up:[[PDF\]](https://arxiv.org/abs/2010.15508) | :floppy_disk:[[Model Checkpoint\]](https://github.com/haoxiangsnr/FullSubNet/releases) | :satellite:[[Loss Curve\]](https://tensorboard.dev/experiment/63WgyAXOSbiBzHg4AdVfYw/#scalars)
+:bulb:[[Demo\]](https://www.haoxiangsnr.com/demo/fullsubnet/) | :page_facing_up:[[PDF\]](https://arxiv.org/abs/2010.15508) | :floppy_disk:[[Model Checkpoint\]](https://github.com/haoxiangsnr/FullSubNet/releases)
 
 ## Introduction
 

audio_zen/acoustics/beamforming.py

@@ -122,8 +122,7 @@ def mvdr_beamformer(
 
     psd_noise_inverse = noise_psd.inverse()
     w1 = FC.matmul(psd_noise_inverse, steering_vector)  # [B * F, C, 1]
-    w2 = FC.matmul(steering_vector.conj().transpose(2, 1),
-                   psd_noise_inverse)  # [B * F, 1, C] * [B * F, C, C] * [B * F, C， 1] = [B * F, 1, 1]
+    w2 = FC.matmul(steering_vector.conj().transpose(2, 1), psd_noise_inverse)  # [B * F, 1, C] * [B * F, C, C] * [B * F, C， 1] = [B * F, 1, 1]
     w2 = FC.matmul(w2, steering_vector)  # [B * F, 1, 1]
 
     w = w1 / w2  # [B * F, C, 1]