fgnt / pb_bss Goto Github PK

Hi,

What's the motivation behind those lines?
Why not float32 for scale invariant SDR?

dear friend:
Could you give one instruction to show how to run the example please?
Thanks.

Steps to reproduce

On a fresh python install

git clone https://github.com/fgnt/pb_bss
cd pb_bss
pip install numpy Cython
pip install -e .

Solution so far

• Comment these lines
• pip install -e .
• Uncomment the same lines
• pip install -e .

Ends up working

Is there any plan to support install pb_bss from pypi?

Hi, I'm a little confused about the SDR in the output metrics. In the paper https://arxiv.org/pdf/1811.02508.pdf, it referred to the original SNR. Does the SDR also refer to SNR in this toolkit? Or it refers to the SD-SDR.

Hi,

Both InputMetrics and OutputMetrics have the as_dict function which returns all the metrics as a dict. I think this is a nice UI but it would be even nicer to be able to control which metric is going to be computed without calling them one by one, what do you think?

Something like this might do, with more

Index: pb_bss/evaluation/wrapper.py
===================================================================
--- pb_bss/evaluation/wrapper.py	(revision 8c31ef0b1e32d355f468170f07e41f989d8cf4c6)
+++ pb_bss/evaluation/wrapper.py	(date 1579701195206)
@@ -318,6 +318,18 @@
             return_dict=True,
         )
 
+    def sdr(self):
+        return self.mir_eval['sdr']
+
+    def sir(self):
+        return self.mir_eval['sir']
+
+    def sar(self):
+        return self.mir_eval['sar']
+
+    def selection(self):
+        return self.mir_eval['selection']
+
     @cached_property.cached_property
     def pesq(self):
         return pb_bss.evaluation.pesq(
@@ -342,6 +354,15 @@
         )
         return invasive_sxr
 
+    def invasive_sdr(self):
+        return self.invasive_sxr['sdr']
+
+    def invasive_sir(self):
+        return self.invasive_sxr['sir']
+
+    def invasive_sar(self):
+        return self.invasive_sxr['sar']
+
     @cached_property.cached_property
     def stoi(self):
         return pb_bss.evaluation.stoi(
@@ -383,3 +404,8 @@
             metrics['invasive_sxr_snr'] = self.invasive_sxr['snr']
 
         return metrics
+
+    def get_as_dict(self, *metric_names):
+        metrics = dict()
+        for m in metric_names:
+            metrics[m] = self.__getattribute__(m)

I'm willing to put more efforts into it if you'd be interested of course.

I'm annoying, I know but this is not really user friendly ;-)

Is paderbox open-source? I cannot find it

After some discussions in #25, it turned out, that it would be good to update the example.

• use paderbox stft
• use beamforming instead of masking

Would you consider making pb_bss.evaluation a standalone python package?

Hello,
When i run activity_alignment.py(Generate finetuned time annotations from kaldi worn alignments), there is something wrong ，The following information is an error message：
ERROR - Chime5 Array Enhancement - Failed after 1 day, 10:27:41!
Traceback (most recent calls WITHOUT Sacred internals):
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/scripts/run.py", line 91, in main
run(_run)
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/scripts/run.py", line 128, in run
audio_dir_exist_ok=True
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/core_chime6.py", line 381, in enhance_session
x_hat = self.enhance_example(ex)
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/core_chime6.py", line 495, in enhance_example
debug=debug,
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/core_chime6.py", line 533, in enhance_observation
masks = self.gss_block(Obs, acitivity_freq, debug=debug)
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/core_chime6.py", line 183, in call
source_activity_mask=source_active_mask[f, ..., :T],
File "/kaldi/egs/chime6/pb_chime5/pb_chime5/pb_bss/pb_bss/distribution/cacgmm.py", line 223, in fit
initialization.shape, affiliation_shape
AssertionError: ((5, 1140), (5, 1165))

Do you know what I should do to solve this problem

hi dear author:
Could you give an instruction to show how to run your project please?
Thanks.