Replaced for loop with vectorised mapping on ClusterOverlap regularization layer

deepof/model_utils.py

@@ -495,65 +495,67 @@ class ClusterOverlap(Layer):
         config.update({"samples": self.samples})
         return config
 
-    def call(self, inputs, **kwargs):
+    def call(self, inputs, training=None, **kwargs):
         """Updates Layer's call method"""
 
         encodings, categorical = inputs[0], inputs[1]
 
-        hard_groups = tf.math.argmax(categorical, axis=1)
-        max_groups = tf.reduce_max(categorical, axis=1)
+        if training:
 
-        get_local_neighbourhood_entropy = partial(
-            get_neighbourhood_entropy,
-            tensor=encodings,
-            clusters=hard_groups,
-            k=self.k,
-        )
+            hard_groups = tf.math.argmax(categorical, axis=1)
+            max_groups = tf.reduce_max(categorical, axis=1)
 
-        purity_vector = tf.map_fn(
-            get_local_neighbourhood_entropy,
-            tf.constant(list(range(self.batch_size))),
-            dtype=tf.dtypes.float32,
-        )
+            get_local_neighbourhood_entropy = partial(
+                get_neighbourhood_entropy,
+                tensor=encodings,
+                clusters=hard_groups,
+                k=self.k,
+            )
 
-        ### CANDIDATE FOR REMOVAL. EXPLORE HOW USEFUL THIS REALLY IS ###
-        neighbourhood_entropy = purity_vector * max_groups
-
-        number_of_clusters = tf.cast(
-            tf.shape(
-                tf.unique(
-                    tf.reshape(
-                        tf.gather(
-                            tf.cast(hard_groups, tf.dtypes.float32),
-                            tf.where(max_groups >= self.min_confidence),
-                            batch_dims=0,
+            purity_vector = tf.map_fn(
+                get_local_neighbourhood_entropy,
+                tf.constant(list(range(self.batch_size))),
+                dtype=tf.dtypes.float32,
+            )
+
+            ### CANDIDATE FOR REMOVAL. EXPLORE HOW USEFUL THIS REALLY IS ###
+            neighbourhood_entropy = purity_vector * max_groups
+
+            number_of_clusters = tf.cast(
+                tf.shape(
+                    tf.unique(
+                        tf.reshape(
+                            tf.gather(
+                                tf.cast(hard_groups, tf.dtypes.float32),
+                                tf.where(max_groups >= self.min_confidence),
+                                batch_dims=0,
+                            ),
+                            [-1],
                         ),
-                        [-1],
-                    ),
+                    )[0],
                 )[0],
-            )[0],
-            tf.dtypes.float32,
-        )
+                tf.dtypes.float32,
+            )
 
-        self.add_metric(
-            number_of_clusters,
-            name="number_of_populated_clusters",
-        )
+            self.add_metric(
+                number_of_clusters,
+                name="number_of_populated_clusters",
+            )
 
-        self.add_metric(
-            max_groups,
-            aggregation="mean",
-            name="average_confidence_in_selected_cluster",
-        )
+            self.add_metric(
+                max_groups,
+                aggregation="mean",
+                name="average_confidence_in_selected_cluster",
+            )
 
-        self.add_metric(
-            neighbourhood_entropy, aggregation="mean", name="neighbourhood_entropy"
-        )
+            self.add_metric(
+                neighbourhood_entropy, aggregation="mean", name="neighbourhood_entropy"
+            )
 
-        if self.loss_weight:
-            # minimize local entropy
-            self.add_loss(self.loss_weight * tf.reduce_mean(neighbourhood_entropy))
-            # maximize number of clusters
-            # self.add_loss(-self.loss_weight * tf.reduce_mean(number_of_clusters))
+            if self.loss_weight:
+                # minimize local entropy
+                self.add_loss(self.loss_weight * tf.reduce_mean(neighbourhood_entropy))
+                # maximize number of clusters
+                # self.add_loss(-self.loss_weight * tf.reduce_mean(number_of_clusters))
 
         return encodings

supplementary_notebooks/deepof_model_evaluation.ipynb

@@ -628,17 +628,9 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 193,
+   "execution_count": 195,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Tensor(\"cluster_overlap_44/Shape:0\", shape=(2,), dtype=int32)\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "(\n",
     "    encoder,\n",
@@ -664,101 +656,108 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 189,
+   "execution_count": 194,
    "metadata": {},
    "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Tensor(\"SEQ_2_SEQ_GMVAE/cluster_overlap_44/Shape:0\", shape=(2,), dtype=int32)\n"
+     ]
+    },
     {
      "data": {
       "text/plain": [
-       "array([[[-2.20274441e-02, -7.49340594e-01, -1.00606358e+00, ...,\n",
-       "          5.97163141e-01,  7.87078515e-02,  3.10142875e-01],\n",
-       "        [-1.86636951e-02, -7.35167682e-01, -1.04070246e+00, ...,\n",
-       "          5.36489308e-01,  6.65854439e-02,  2.50531852e-01],\n",
-       "        [-1.82860009e-02, -7.65824318e-01, -1.08549905e+00, ...,\n",
-       "          5.57325482e-01,  3.26592624e-02,  2.32314825e-01],\n",
+       "array([[[-2.33654119e-02, -7.18961537e-01, -1.04058743e+00, ...,\n",
+       "          6.04326367e-01, -8.63653272e-02,  3.18874061e-01],\n",
+       "        [-2.12193187e-02, -7.04667985e-01, -1.08008337e+00, ...,\n",
+       "          5.50029635e-01, -9.52750891e-02,  2.61153668e-01],\n",
+       "        [-2.03663222e-02, -7.36600041e-01, -1.12794685e+00, ...,\n",
+       "          5.78937709e-01, -1.28708437e-01,  2.42814869e-01],\n",
        "        ...,\n",
-       "        [-6.10359153e-03, -7.89098859e-01, -1.30699337e+00, ...,\n",
-       "          6.69562757e-01, -4.62342173e-01,  3.77381265e-01],\n",
-       "        [-4.65778168e-03, -8.11704040e-01, -1.28605306e+00, ...,\n",
-       "          7.13757634e-01, -4.67065245e-01,  4.34089184e-01],\n",
-       "        [-6.71512820e-03, -8.53176653e-01, -1.29082215e+00, ...,\n",
-       "          7.71026969e-01, -5.02031207e-01,  4.69406664e-01]],\n",
+       "        [-9.44020599e-03, -7.42647946e-01, -1.30498624e+00, ...,\n",
+       "          6.58422947e-01, -5.47523737e-01,  3.99996579e-01],\n",
+       "        [-7.29610212e-03, -7.65350342e-01, -1.27779627e+00, ...,\n",
+       "          6.95867777e-01, -5.41299820e-01,  4.53622282e-01],\n",
+       "        [-9.27564688e-03, -8.18071365e-01, -1.28233838e+00, ...,\n",
+       "          7.52779126e-01, -5.64222336e-01,  4.90554869e-01]],\n",
        "\n",
-       "       [[-2.22376399e-02, -7.70553827e-01, -9.43590641e-01, ...,\n",
-       "          5.78278065e-01,  2.24873498e-01,  3.14347565e-01],\n",
-       "        [-1.82316527e-02, -7.55171597e-01, -9.75605845e-01, ...,\n",
-       "          5.09953022e-01,  2.11454228e-01,  2.45110869e-01],\n",
-       "        [-1.79231763e-02, -7.78624117e-01, -1.02039325e+00, ...,\n",
-       "          5.19503474e-01,  1.74318790e-01,  2.36997604e-01],\n",
+       "       [[-2.22253725e-02, -8.41416478e-01, -1.12681007e+00, ...,\n",
+       "          6.94921017e-01, -3.09668630e-01,  4.00050104e-01],\n",
+       "        [-2.30817143e-02, -8.39012742e-01, -1.17822659e+00, ...,\n",
+       "          6.44349039e-01, -3.31067830e-01,  3.31252694e-01],\n",
+       "        [-2.34456696e-02, -8.76340389e-01, -1.22851527e+00, ...,\n",
+       "          6.79111958e-01, -3.67583066e-01,  3.47556919e-01],\n",
        "        ...,\n",
-       "        [-3.98765504e-03, -7.79918849e-01, -1.20278263e+00, ...,\n",
-       "          5.98293304e-01, -2.54329056e-01,  3.15500557e-01],\n",
-       "        [-3.29605211e-03, -8.03981423e-01, -1.18459630e+00, ...,\n",
-       "          6.46142304e-01, -2.72717953e-01,  3.54292005e-01],\n",
-       "        [-4.49736090e-03, -8.39645684e-01, -1.18976438e+00, ...,\n",
-       "          7.12202549e-01, -3.01250070e-01,  3.82732034e-01]],\n",
+       "        [-1.15935905e-02, -8.33023310e-01, -1.16872275e+00, ...,\n",
+       "          6.26452863e-01, -3.46652746e-01,  3.66459996e-01],\n",
+       "        [-7.99606554e-03, -8.50840628e-01, -1.14563727e+00, ...,\n",
+       "          6.63835168e-01, -3.26457262e-01,  3.89723182e-01],\n",
+       "        [-8.63283966e-03, -9.02453482e-01, -1.13931096e+00, ...,\n",
+       "          7.31720626e-01, -3.24028820e-01,  4.03006196e-01]],\n",
        "\n",
-       "       [[-2.36156955e-02, -7.91016340e-01, -1.11879790e+00, ...,\n",
-       "          7.13154435e-01, -3.65756422e-01,  4.06932175e-01],\n",
-       "        [-2.54947953e-02, -7.88331568e-01, -1.17535889e+00, ...,\n",
-       "          6.70483768e-01, -3.85737628e-01,  3.32547605e-01],\n",
-       "        [-2.56239660e-02, -8.24208558e-01, -1.23989546e+00, ...,\n",
-       "          7.14474797e-01, -4.28516060e-01,  3.39852810e-01],\n",
+       "       [[-2.67037526e-02, -7.14372754e-01, -1.01690137e+00, ...,\n",
+       "          5.67512035e-01, -7.96585530e-02,  2.87923396e-01],\n",
+       "        [-2.66271308e-02, -7.22530246e-01, -1.08599091e+00, ...,\n",
+       "          5.39661050e-01, -1.06031761e-01,  2.33209401e-01],\n",
+       "        [-2.72963755e-02, -7.53687024e-01, -1.14411652e+00, ...,\n",
+       "          5.67179024e-01, -1.51002675e-01,  2.36247778e-01],\n",
        "        ...,\n",
-       "        [-1.38379335e-02, -7.91984320e-01, -1.10805249e+00, ...,\n",
-       "          6.95663571e-01, -3.35209459e-01,  3.72855306e-01],\n",
-       "        [-8.09174031e-03, -8.20150554e-01, -1.07291603e+00, ...,\n",
-       "          7.31382430e-01, -3.07114869e-01,  3.89692128e-01],\n",
-       "        [-7.56193697e-03, -8.81913245e-01, -1.08008969e+00, ...,\n",
-       "          7.96405435e-01, -2.96295941e-01,  4.00389731e-01]],\n",
+       "        [-9.20080859e-03, -7.82252252e-01, -9.62321341e-01, ...,\n",
+       "          6.10508740e-01, -6.60327971e-02,  2.83200055e-01],\n",
+       "        [-3.12909577e-03, -7.93930829e-01, -9.27874446e-01, ...,\n",
+       "          6.45921230e-01, -4.59667295e-02,  2.94262946e-01],\n",
+       "        [-2.58489698e-03, -8.41321766e-01, -9.23104703e-01, ...,\n",
+       "          7.15546906e-01, -3.20333168e-02,  2.93513715e-01]],\n",
        "\n",
        "       ...,\n",
        "\n",
-       "       [[ 1.27419364e-03, -1.70991755e+00, -7.36668050e-01, ...,\n",
-       "          1.14400041e+00,  1.80199802e-01,  4.03101444e-01],\n",
-       "        [ 9.55599174e-03, -1.77840245e+00, -7.31509030e-01, ...,\n",
-       "          1.10214877e+00,  1.92172885e-01,  3.62863183e-01],\n",
-       "        [ 6.99162669e-03, -1.84455645e+00, -6.68665648e-01, ...,\n",
-       "          1.12744784e+00,  1.78093016e-01,  3.79255176e-01],\n",
+       "       [[ 1.11031905e-03, -1.75031507e+00, -8.52379441e-01, ...,\n",
+       "          1.21988916e+00,  1.03995144e-01,  4.59409773e-01],\n",
+       "        [ 8.80013034e-03, -1.83671558e+00, -8.69083345e-01, ...,\n",
+       "          1.19457781e+00,  1.23950966e-01,  4.14475203e-01],\n",
+       "        [ 6.59639947e-03, -1.91086435e+00, -8.22446942e-01, ...,\n",
+       "          1.22027707e+00,  1.20491169e-01,  4.16057706e-01],\n",
        "        ...,\n",
-       "        [ 1.46682467e-02, -1.67971981e+00, -5.29618204e-01, ...,\n",
-       "          1.07131672e+00,  2.70632327e-01,  3.96659851e-01],\n",
-       "        [ 1.90188847e-02, -1.64219010e+00, -5.86603642e-01, ...,\n",
-       "          1.06212115e+00,  2.97838628e-01,  4.20529246e-01],\n",
-       "        [ 1.77764259e-02, -1.60712445e+00, -5.87549746e-01, ...,\n",
-       "          1.08539033e+00,  3.32792282e-01,  4.31078553e-01]],\n",
+       "        [ 1.49972383e-02, -1.74408889e+00, -5.97882569e-01, ...,\n",
+       "          1.21293759e+00,  4.05578047e-01,  5.18048167e-01],\n",
+       "        [ 1.90434121e-02, -1.69634783e+00, -6.43523514e-01, ...,\n",
+       "          1.19492638e+00,  4.35988188e-01,  5.23940802e-01],\n",
+       "        [ 1.63217653e-02, -1.64555275e+00, -6.33343041e-01, ...,\n",
+       "          1.19980395e+00,  4.42245930e-01,  5.15006006e-01]],\n",
        "\n",
-       "       [[ 5.71209192e-03, -1.57354665e+00, -8.53239119e-01, ...,\n",
-       "          1.10862482e+00,  1.20560825e-01,  4.82771069e-01],\n",
-       "        [ 1.34701636e-02, -1.63361990e+00, -8.83263171e-01, ...,\n",
-       "          1.09039259e+00,  1.39529616e-01,  4.46821988e-01],\n",
-       "        [ 1.27676763e-02, -1.69738317e+00, -8.67947996e-01, ...,\n",
-       "          1.11015606e+00,  1.38534516e-01,  4.43767011e-01],\n",
+       "       [[ 2.67705321e-03, -1.70834327e+00, -7.96462834e-01, ...,\n",
+       "          1.17186177e+00,  1.32418081e-01,  4.49965090e-01],\n",
+       "        [ 1.02437194e-02, -1.78479028e+00, -8.03244650e-01, ...,\n",
+       "          1.13687229e+00,  1.46574080e-01,  4.06593144e-01],\n",
+       "        [ 7.09127262e-03, -1.85855865e+00, -7.63118446e-01, ...,\n",
+       "          1.15632105e+00,  1.33105025e-01,  4.12189573e-01],\n",
        "        ...,\n",
-       "        [ 2.44371835e-02, -1.57842243e+00, -5.92557728e-01, ...,\n",
-       "          1.22556579e+00,  5.59510589e-01,  6.62321389e-01],\n",
-       "        [ 2.63634715e-02, -1.54089200e+00, -6.52207255e-01, ...,\n",
-       "          1.19942415e+00,  5.94571948e-01,  6.66273296e-01],\n",
-       "        [ 2.19873693e-02, -1.52749789e+00, -6.52528405e-01, ...,\n",
-       "          1.23534441e+00,  5.85497737e-01,  6.72029912e-01]],\n",
+       "        [ 1.81343500e-02, -1.67693102e+00, -5.01757801e-01, ...,\n",
+       "          1.16755402e+00,  4.76390660e-01,  5.43144941e-01],\n",
+       "        [ 2.21438985e-02, -1.63842762e+00, -5.54517329e-01, ...,\n",
+       "          1.15032852e+00,  5.08659005e-01,  5.53249776e-01],\n",
+       "        [ 1.89741328e-02, -1.60885894e+00, -5.58988631e-01, ...,\n",
+       "          1.17562830e+00,  5.15346706e-01,  5.59121966e-01]],\n",
        "\n",
-       "       [[ 6.25598989e-03, -1.62235355e+00, -8.73309553e-01, ...,\n",
-       "          1.13996327e+00,  1.29154831e-01,  4.71493900e-01],\n",
-       "        [ 1.32705066e-02, -1.68581510e+00, -9.02929783e-01, ...,\n",
-       "          1.12689865e+00,  1.41210854e-01,  4.33707178e-01],\n",
-       "        [ 1.26603413e-02, -1.74876869e+00, -8.90606523e-01, ...,\n",
-       "          1.14830673e+00,  1.26072347e-01,  4.30647731e-01],\n",
+       "       [[ 1.44676864e-03, -1.54619968e+00, -9.54818726e-01, ...,\n",
+       "          1.10726678e+00, -8.50950629e-02,  4.39189941e-01],\n",
+       "        [ 6.98623247e-03, -1.59582186e+00, -9.81045246e-01, ...,\n",
+       "          1.09904671e+00, -8.24306607e-02,  3.94697845e-01],\n",
+       "        [ 7.70333409e-03, -1.66229522e+00, -9.75158274e-01, ...,\n",
+       "          1.12477064e+00, -9.11763161e-02,  3.99113178e-01],\n",
        "        ...,\n",
-       "        [ 2.57094558e-02, -1.64026582e+00, -6.49993658e-01, ...,\n",
-       "          1.23270714e+00,  4.76267278e-01,  6.13111377e-01],\n",
-       "        [ 2.90074851e-02, -1.60244536e+00, -6.78030908e-01, ...,\n",
-       "          1.21040225e+00,  5.14402866e-01,  6.22853637e-01],\n",
-       "        [ 2.53587011e-02, -1.59674680e+00, -6.75386965e-01, ...,\n",
-       "          1.23939729e+00,  5.17669916e-01,  6.38859391e-01]]],\n",
+       "        [ 2.38139462e-02, -1.57800007e+00, -6.76846921e-01, ...,\n",
+       "          1.23930001e+00,  4.26834762e-01,  5.95292032e-01],\n",
+       "        [ 2.62587424e-02, -1.54779446e+00, -7.09348798e-01, ...,\n",
+       "          1.21432519e+00,  4.73278880e-01,  5.99520683e-01],\n",
+       "        [ 2.12610681e-02, -1.53637493e+00, -6.99584544e-01, ...,\n",
+       "          1.24682486e+00,  4.80561942e-01,  6.04046106e-01]]],\n",
        "      dtype=float32)"
       ]
      },
-     "execution_count": 189,
+     "execution_count": 194,
      "metadata": {},
      "output_type": "execute_result"
     }