Making min dist a little faster

bss_scale/algorithms.py

@@ -85,7 +85,7 @@ def minimum_distortion_l2(Y, ref):
 
 
 def minimum_distortion(
-    Y, ref, p=None, q=None, rtol=1e-3, max_iter=100,
+    Y, ref, p=None, q=None, rtol=1e-2, max_iter=100,
 ):
     """
     This function computes the frequency-domain filter that minimizes the sum
@@ -135,7 +135,7 @@ def minimum_distortion(
 
     eps = 1e-15
 
-    prev_res = None
+    prev_c = None
 
     epoch = 0
     while epoch < max_iter:
@@ -145,23 +145,23 @@ def minimum_distortion(
         # the current error
         error = ref[:, :, None] - c * Y
         if q is None or p == q:
-            res, weights = lp_norm(error, p=p)
+            weights = lp_norm(error, p=p)
         else:
-            res, weights = lpq_norm(error, p=p, q=q, axis=1)
+            weights = lpq_norm(error, p=p, q=q, axis=1)
 
         # minimize
         num = np.sum(ref[:, :, None] * np.conj(Y) * weights, axis=0)
         denom = np.sum(np.abs(Y) ** 2 * weights, axis=0)
         c = num / np.maximum(eps, denom)
 
         # condition for termination
-        if prev_res is None:
-            prev_res = res
+        if prev_c is None:
+            prev_c = c
             continue
 
         # relative step length
-        delta = (prev_res - res) / prev_res
-        prev_res = res
+        delta = np.linalg.norm(c - prev_c) / np.linalg.norm(prev_c)
+        prev_c = c
         if delta < rtol:
             break
 

bss_scale/surrogate.py

@@ -7,16 +7,14 @@
 
 def lp_norm(E, p=1):
     assert p > 0 and p < 2
-    cost = np.sum(np.abs(E) ** p)
     weights = p / np.maximum(eps, 2.0 * np.abs(E) ** (2 - p))
-    return cost, weights
+    return weights
 
 
 def lpq_norm(E, p=1, q=2, axis=1):
     assert p > 0 and q >= p and q <= 2.0
 
-    cost = np.sum(np.sum(np.abs(E) ** q, axis=axis, keepdims=True) ** (p / q))
     rn = np.sum(np.abs(E) ** q, axis=axis, keepdims=True) ** (1 - p / q)
     qfn = np.abs(E) ** (2 - q)
     weights = p / np.maximum(eps, 2.0 * rn * qfn)
-    return cost, weights
+    return weights

experiment2_config.json

@@ -0,0 +1,32 @@
+{
+  "metadata_fn": "./bss_speech_dataset/data/metadata.json",
+
+  "stft": {
+    "nfft": 512,
+    "hop": 256,
+    "window": "hamming"
+  },
+
+  "ref_mic": 0,
+  "si_metric": false,
+  "snr": 40,
+
+  "minimum_distortion": {
+    "p_list": [0.1, 2.0, 20],
+    "kwargs": {
+      "rtol": 1e-5,
+      "max_iter": 100
+    }
+  },
+
+  "bss_algorithms": {
+    "ilrma_t": {
+      "name": "ilrma_t",
+      "kwargs": {
+        "n_taps": 6,
+        "n_delays": 1
+      },
+      "n_iter_per_channel": 15
+    }
+  }
+}