Persephone refactor

egs/voxceleb/ssl.v1/README.md

@@ -141,8 +141,9 @@ run_xxx_xxxx.sh --config-file global_conf/other_config.sh
 | | | | | | PLDA | 4.04 | 0.185 | 0.291 |
 | | | | | FT-1 | Cosine | 2.03 | 0.125 | 0.203 |
 | | | | | FT-1 | PLDA | 2.44 | 0.149 | 0.231 |
-| | | | | FT-2 | Cosine | 
-| | | | | FT-2 | PLDA | 
+| | | | | FT-2 | Cosine | 1.88 | 0.115 | 0.198 |
+| | | | | FT-2 | PLDA | 2.57 | 0.147 | 0.234 |
+
 
 
 ### VoxCeleb 1 Entire-Clean trial list
@@ -171,8 +172,9 @@ run_xxx_xxxx.sh --config-file global_conf/other_config.sh
 | | | | | | PLDA | 4.32 | 0.166 | 0.263 |
 | | | | | FT-1 | Cosine | 2.61 | 0.138 | 0.210 | 
 | | | | | FT-1 | PLDA | 2.72 | 0.1366 | 0.216 |
-| | | | | FT-2 | Cosine | 
-| | | | | FT-2 | PLDA | 
+| | | | | FT-2 | Cosine | 2.41 | 0.121 | 0.193 |
+| | | | | FT-2 | PLDA | 2.82 | 0.140 | 0.219 |
+
 
 
 ### VoxCeleb 1 Hard-Clean trial list
@@ -201,6 +203,5 @@ run_xxx_xxxx.sh --config-file global_conf/other_config.sh
 | | | | | | PLDA | 5.95 | 0.269 | 0.438 |
 | | | | | FT-1 | Cosine | 4.38 | 0.222 | 0.337 |
 | | | | | FT-1 | PLDA | 4.68 | 0.237 | 0.375 |
-| | | | | FT-2 | Cosine | 
-| | | | | FT-2 | PLDA | 
-
+| | | | | FT-2 | Cosine | 4.07 | 0.197 | 0.301 |
+| | | | | FT-2 | PLDA | 4.75 | 0.229 | 0.352 |

egs/voxceleb/ssl.v1/global_conf/config_fbank80_stmn_fwseresnet34.v1.2.2.sh

@@ -50,7 +50,7 @@ cluster_ft_s1_dir=exp/clustering/$nnet_ft_s1_2_name/$cluster_ft_s1_name
 nnet_ft_s2_1_base_cfg=conf/train_fwseresnet34_xvec_stage1.1_v1.2.2.yaml
 nnet_ft_s2_1_name=$nnet_name.s1.ft.s2.1
 nnet_ft_s2_1_dir=exp/xvector_nnets/$nnet_ft_s2_1_name
-nnet_ft_s2_1=$nnet_ft_s2_1_dir/model_ep0030.pth
+nnet_ft_s2_1=$nnet_ft_s2_1_dir/model_ep0025.pth
 
 # finetuning stage 2.2
 nnet_ft_s2_2_base_cfg=conf/train_fwseresnet34_xvec_stage1.2_v1.2.2.yaml

hyperion/np/pdfs/mixtures/gmm.py

@@ -2,6 +2,7 @@
  Copyright 2018 Johns Hopkins University  (Author: Jesus Villalba)
  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
 """
+
 import h5py
 import numpy as np
 import scipy.linalg as la
@@ -150,8 +151,8 @@ def _initialize_kmeans(self, num_comp, x):
             self.Lambda[0] = invert_pdmat(S, return_inv=True)[-1]
             return
 
-        kmeans = KMeans(num_clusters=num_comp)
-        loss, cluster_index = kmeans.fit(x, epochs=100)
+        kmeans = KMeans(num_clusters=num_comp, epochs=100)
+        loss, cluster_index = kmeans.fit(x)
 
         self.mu = kmeans.mu
         self.pi = np.zeros((self.num_comp,), dtype=float_cpu())
@@ -253,7 +254,7 @@ def split_comp(self, K=2):
         """
         num_comp = self.num_comp * K
         pi = np.repeat(self.pi, K) / K
-        Lambda = np.repeat(self.Lambda, K, axis=0) * (K ** 2)
+        Lambda = np.repeat(self.Lambda, K, axis=0) * (K**2)
         mu = np.repeat(self.mu, K, axis=0)
 
         for g in range(self.num_comp):
@@ -400,7 +401,7 @@ def load_from_kaldi(cls, file_path):
                             x_dim = len(fields)
                             eta1 = np.zeros((num_comp, x_dim), dtype=float_cpu())
                             eta2 = np.zeros(
-                                (num_comp, int((x_dim ** 2 + 3 * x_dim) / 2)),
+                                (num_comp, int((x_dim**2 + 3 * x_dim) / 2)),
                                 dtype=float_cpu(),
                             )
 
@@ -436,7 +437,7 @@ def _validate_Lambda(self):
 
     def _validate_eta(self):
         assert self.eta.shape[0] == self.num_comp
-        assert self.eta.shape[1] == (self.x_dim ** 2 + 3 * self.x_dim) / 2
+        assert self.eta.shape[1] == (self.x_dim**2 + 3 * self.x_dim) / 2
 
     def validate(self):
         """Validates the parameters of the distribution."""
@@ -454,7 +455,7 @@ def validate(self):
     def compute_eta(mu, Lambda):
         """Computes nat param. from mean and precision."""
         x_dim = mu.shape[-1]
-        eta_dim = int((x_dim ** 2 + 3 * x_dim) / 2)
+        eta_dim = int((x_dim**2 + 3 * x_dim) / 2)
         eta = np.zeros((mu.shape[0], eta_dim), dtype=float_cpu())
         for k in range(mu.shape[0]):
             eta[k] = Normal.compute_eta(mu[k], Lambda[k])

hyperion/np/pdfs/mixtures/gmm_diag_cov.py

@@ -121,8 +121,8 @@ def _initialize_kmeans(self, num_comp, x):
             self.Lambda = 1 / np.std(x, axis=0, keepdims=True) ** 2
             return
 
-        kmeans = KMeans(num_clusters=num_comp)
-        loss, cluster_index = kmeans.fit(x, epochs=100)
+        kmeans = KMeans(num_clusters=num_comp, epochs=100)
+        loss, cluster_index = kmeans.fit(x)
 
         self.mu = kmeans.mu
         self.pi = np.zeros((self.num_comp,), dtype=float_cpu())

hyperion/np/pdfs/mixtures/gmm_tied_diag_cov.py

@@ -2,6 +2,7 @@
  Copyright 2018 Johns Hopkins University  (Author: Jesus Villalba)
  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
 """
+
 import h5py
 import numpy as np
 from scipy.special import erf
@@ -83,8 +84,8 @@ def _initialize_kmeans(self, num_comp, x):
             self.Lambda = 1 / np.std(x, axis=0, keepdims=True) ** 2
             return
 
-        kmeans = KMeans(num_clusters=num_comp)
-        loss, cluster_index = kmeans.fit(x, epochs=100)
+        kmeans = KMeans(num_clusters=num_comp, epochs=100)
+        loss, cluster_index = kmeans.fit(x)
 
         self.mu = kmeans.mu
         self.pi = np.zeros((self.num_comp,), dtype=float_cpu())
@@ -93,7 +94,7 @@ def _initialize_kmeans(self, num_comp, x):
             r = cluster_index == k
             self.pi[k] = np.sum(r) / x.shape[0]
             delta = x[r] - self.mu[k]
-            C += np.sum(delta ** 2, axis=0)
+            C += np.sum(delta**2, axis=0)
 
         self.Lambda = x.shape[0] / C
 
@@ -111,7 +112,7 @@ def Mstep(self, N, u_x):
             self.mu = F / N[:, None]
 
         if self.update_Lambda:
-            S = S / N[:, None] - self.mu ** 2
+            S = S / N[:, None] - self.mu**2
             S_floor = self.var_floor * np.mean(S[N > self.min_N], axis=0)
             S = np.maximum(S, S_floor)
             Spool = np.sum(N[:, None] * S, axis=0) / np.sum(N)