remove the last softmax activation

graphgallery/nn/layers/__init__.py

@@ -11,4 +11,4 @@
 from graphgallery.nn.layers.edgeconv import GraphEdgeConvolution
 from graphgallery.nn.layers.mediansage import MedianAggregator, MedianGCNAggregator
 from graphgallery.nn.layers.gcnf import GraphConvFeature
-from graphgallery.nn.layers.misc import SparseConversion, Scale, Sample
+from graphgallery.nn.layers.misc import SparseConversion, Scale, Sample, Gather

graphgallery/nn/layers/misc.py

@@ -57,3 +57,24 @@ def get_config(self):
 
     def compute_output_shape(self, input_shapes):
         return tf.TensorShape(input_shapes[0])
+
+
+class Gather(Layer):
+    def __init__(self, axis=0, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.axis = axis
+
+    def call(self, inputs):
+        params, indices = inputs
+        output = tf.gather(params, indices, axis=self.axis)
+        return output
+
+    def get_config(self):
+        base_config = super().get_config()
+        return base_config
+
+    def compute_output_shape(self, input_shapes):
+        axis = self.axis
+        params_shape, indices_shape = input_shapes
+        output_shape = params_shape[:axis] + indices_shape + params_shape[axis + 1:]
+        return tf.TensorShape(output_shape)

graphgallery/nn/models/__init__.py

@@ -4,7 +4,6 @@
 # (semi-)supervised model
 from graphgallery.nn.models.semisupervised.semi_supervised_model import SemiSupervisedModel
 from graphgallery.nn.models.semisupervised.gcn import GCN
-from graphgallery.nn.models.semisupervised.gcn_mix import GCN_MIX
 from graphgallery.nn.models.semisupervised.sgc import SGC
 from graphgallery.nn.models.semisupervised.gat import GAT
 from graphgallery.nn.models.semisupervised.clustergcn import ClusterGCN
@@ -17,11 +16,14 @@
 from graphgallery.nn.models.semisupervised.lgcn import LGCN
 from graphgallery.nn.models.semisupervised.obvat import OBVAT
 from graphgallery.nn.models.semisupervised.sbvat import SBVAT
-from graphgallery.nn.models.semisupervised.s_obvat import SimplifiedOBVAT
 from graphgallery.nn.models.semisupervised.gmnn import GMNN
-from graphgallery.nn.models.semisupervised.edgeconv import EdgeGCN
-from graphgallery.nn.models.semisupervised.mediansage import MedianSAGE
-from graphgallery.nn.models.semisupervised.gcnf import GCNF
+
+# experimental
+from graphgallery.nn.models.semisupervised.experimental.gcnf import GCNF
+from graphgallery.nn.models.semisupervised.experimental.gcn_mix import GCN_MIX
+from graphgallery.nn.models.semisupervised.experimental.s_obvat import SimplifiedOBVAT
+from graphgallery.nn.models.semisupervised.experimental.edgeconv import EdgeGCN
+from graphgallery.nn.models.semisupervised.experimental.mediansage import MedianSAGE
 
 
 # unsupervised model

graphgallery/nn/models/semisupervised/chebynet.py

@@ -1,10 +1,11 @@
 import tensorflow as tf
 from tensorflow.keras import Model, Input
-from tensorflow.keras.layers import Dropout, Softmax
+from tensorflow.keras.layers import Dropout
 from tensorflow.keras.optimizers import Adam
 from tensorflow.keras import regularizers
+from tensorflow.keras.losses import SparseCategoricalCrossentropy
 
-from graphgallery.nn.layers import ChebyConvolution
+from graphgallery.nn.layers import ChebyConvolution, Gather
 from graphgallery.sequence import FullBatchNodeSequence
 from graphgallery.nn.models import SemiSupervisedModel
 from graphgallery.utils.misc import chebyshev_polynomials
@@ -37,7 +38,7 @@ class ChebyNet(SemiSupervisedModel):
                 i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}})
             norm_x (String, optional):
                 How to normalize the node feature matrix. See `graphgallery.normalize_x`
-                (default :str: `l1`)
+                (default :obj: `None`)
             device (String, optional):
                 The device where the model is running on. You can specified `CPU` or `GPU`
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
@@ -51,7 +52,7 @@ class ChebyNet(SemiSupervisedModel):
     """
 
     def __init__(self, adj, x, labels, order=2, norm_adj=-0.5,
-                 norm_x='l1', device='CPU:0', seed=None, name=None, **kwargs):
+                 norm_x=None, device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels,
                          device=device, seed=seed, name=name, **kwargs)
@@ -103,11 +104,11 @@ def build(self, hiddens=[16], activations=['relu'], dropouts=[0.5], l2_norms=[5e
                 h = Dropout(rate=dropout)(h)
 
             h = ChebyConvolution(self.n_classes, order=self.order, use_bias=use_bias)([h, adj])
-            h = tf.gather(h, index)
-            output = Softmax()(h)
+            h = Gather()([h, index])
 
-            model = Model(inputs=[x, *adj, index], outputs=output)
-            model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=lr), metrics=['accuracy'])
+            model = Model(inputs=[x, *adj, index], outputs=h)
+            model.compile(loss=SparseCategoricalCrossentropy(from_logits=True),
+                          optimizer=Adam(lr=lr), metrics=['accuracy'])
             self.model = model
 
     def train_sequence(self, index):

graphgallery/nn/models/semisupervised/clustergcn.py

@@ -2,9 +2,10 @@
 import networkx as nx
 import tensorflow as tf
 from tensorflow.keras import Model, Input
-from tensorflow.keras.layers import Dropout, Softmax
+from tensorflow.keras.layers import Dropout
 from tensorflow.keras.optimizers import Adam
 from tensorflow.keras import regularizers
+from tensorflow.keras.losses import SparseCategoricalCrossentropy
 
 from graphgallery.nn.layers import GraphConvolution, SparseConversion
 from graphgallery.nn.models import SemiSupervisedModel
@@ -46,7 +47,7 @@ class ClusterGCN(SemiSupervisedModel):
                 i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}})
             norm_x (String, optional):
                 How to normalize the node feature matrix. See `graphgallery.normalize_x`
-                (default :str: `l1`)
+                (default :obj: `None`)
             device (String, optional):
                 The device where the model is running on. You can specified `CPU` or `GPU`
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
@@ -60,7 +61,7 @@ class ClusterGCN(SemiSupervisedModel):
     """
 
     def __init__(self, adj, x, labels, graph=None, n_clusters=None,
-                 norm_adj=-0.5, norm_x='l1', device='CPU:0', seed=None, name=None, **kwargs):
+                 norm_adj=-0.5, norm_x=None, device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels=labels, device=device, seed=seed, name=name, **kwargs)
 
@@ -118,7 +119,7 @@ def build(self, hiddens=[32], activations=['relu'], dropouts=[0.5], l2_norms=[1e
             mask = Input(batch_shape=[None],  dtype=tf.bool, name='mask')
 
             adj = SparseConversion()([edge_index, edge_weight])
-            
+
             h = x
             for hid, activation, dropout, l2_norm in zip(hiddens, activations, dropouts, l2_norms):
                 h = Dropout(rate=dropout)(h)
@@ -128,12 +129,10 @@ def build(self, hiddens=[32], activations=['relu'], dropouts=[0.5], l2_norms=[1e
             h = Dropout(rate=dropout)(h)
             h = GraphConvolution(self.n_classes, use_bias=use_bias)([h, adj])
             h = tf.boolean_mask(h, mask)
-            output = Softmax()(h)
-
-            model = Model(inputs=[x, edge_index, edge_weight, mask], outputs=output)
 
-            model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=lr),
-                          metrics=['accuracy'])
+            model = Model(inputs=[x, edge_index, edge_weight, mask], outputs=h)
+            model.compile(loss=SparseCategoricalCrossentropy(from_logits=True),
+                          optimizer=Adam(lr=lr), metrics=['accuracy'])
 
             self.model = model
 

graphgallery/nn/models/semisupervised/densegcn.py

@@ -5,8 +5,9 @@
 from tensorflow.keras.layers import Dropout, Softmax
 from tensorflow.keras.optimizers import Adam
 from tensorflow.keras import regularizers
+from tensorflow.keras.losses import SparseCategoricalCrossentropy
 
-from graphgallery.nn.layers import DenseGraphConv
+from graphgallery.nn.layers import DenseGraphConv, Gather
 from graphgallery.nn.models import SemiSupervisedModel
 from graphgallery.sequence import FullBatchNodeSequence
 from graphgallery.utils.shape_utils import set_equal_in_length
@@ -42,7 +43,7 @@ class DenseGCN(SemiSupervisedModel):
                 i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}})
             norm_x (String, optional):
                 How to normalize the node feature matrix. See `graphgallery.normalize_x`
-                (default :str: `l1`)
+                (default :obj: `None`)
             device (String, optional):
                 The device where the model is running on. You can specified `CPU` or `GPU`
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
@@ -56,7 +57,7 @@ class DenseGCN(SemiSupervisedModel):
 
     """
 
-    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x='l1',
+    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x=None,
                  device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels, device=device, seed=seed, name=name, **kwargs)
@@ -112,11 +113,11 @@ def build(self, hiddens=[16], activations=['relu'], dropouts=[0.5], l2_norms=[5e
                 h = Dropout(rate=dropout)(h)
 
             h = DenseGraphConv(self.n_classes, use_bias=use_bias)([h, adj])
-            h = tf.gather(h, index)
-            output = Softmax()(h)
+            h = Gather()([h, index])
 
-            model = Model(inputs=[x, adj, index], outputs=output)
-            model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=lr), metrics=['accuracy'])
+            model = Model(inputs=[x, adj, index], outputs=h)
+            model.compile(loss=SparseCategoricalCrossentropy(from_logits=True),
+                          optimizer=Adam(lr=lr), metrics=['accuracy'])
             self.model = model
 
     def train_sequence(self, index):

graphgallery/nn/models/semisupervised/edgeconv.py → graphgallery/nn/models/semisupervised/experimental/edgeconv.py

@@ -1,11 +1,12 @@
 import numpy as np
 import tensorflow as tf
 from tensorflow.keras import Model, Input
-from tensorflow.keras.layers import Dropout, Softmax
+from tensorflow.keras.layers import Dropout
 from tensorflow.keras.optimizers import Adam
 from tensorflow.keras import regularizers
+from tensorflow.keras.losses import SparseCategoricalCrossentropy
 
-from graphgallery.nn.layers import GraphEdgeConvolution
+from graphgallery.nn.layers import GraphEdgeConvolution, Gather
 from graphgallery.nn.models import SemiSupervisedModel
 from graphgallery.sequence import FullBatchNodeSequence
 from graphgallery.utils.shape_utils import set_equal_in_length
@@ -44,7 +45,7 @@ class EdgeGCN(SemiSupervisedModel):
                 i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}})
             norm_x (String, optional):
                 How to normalize the node feature matrix. See `graphgallery.normalize_x`
-                (default :str: `l1`)
+                (default :obj: `None`)
             device (String, optional):
                 The device where the model is running on. You can specified `CPU` or `GPU`
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
@@ -57,7 +58,7 @@ class EdgeGCN(SemiSupervisedModel):
 
     """
 
-    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x='l1',
+    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x=None,
                  device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels, device=device, seed=seed, name=name, **kwargs)
@@ -109,11 +110,11 @@ def build(self, hiddens=[16], activations=['relu'], dropouts=[0.5], l2_norms=[5e
                 h = Dropout(rate=dropout)(h)
 
             h = GraphEdgeConvolution(self.n_classes, use_bias=use_bias)([h, edge_index, edge_weight])
-            h = tf.gather(h, index)
-            output = Softmax()(h)
-
-            model = Model(inputs=[x, edge_index, edge_weight, index], outputs=output)
-            model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=lr), metrics=['accuracy'])
+            h = Gather()([h, index])
+
+            model = Model(inputs=[x, edge_index, edge_weight, index], outputs=h)
+            model.compile(loss=SparseCategoricalCrossentropy(from_logits=True),
+                          optimizer=Adam(lr=lr), metrics=['accuracy'])
             self.model = model
 
     def train_sequence(self, index):

graphgallery/nn/models/semisupervised/gcn_mix.py → graphgallery/nn/models/semisupervised/experimental/gcn_mix.py

@@ -36,7 +36,7 @@ class GCN_MIX(SemiSupervisedModel):
                 i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}}) 
             norm_x (String, optional): 
                 How to normalize the node feature matrix. See `graphgallery.normalize_x`
-                (default :str: `l1`)
+                (default :obj: `None`)
             device (String, optional): 
                 The device where the model is running on. You can specified `CPU` or `GPU` 
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
@@ -49,7 +49,7 @@ class GCN_MIX(SemiSupervisedModel):
 
     """
 
-    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x='l1',
+    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x=None,
                  device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels, device=device, seed=seed, name=name, **kwargs)

graphgallery/nn/models/semisupervised/gcnf.py → graphgallery/nn/models/semisupervised/experimental/gcnf.py

@@ -1,10 +1,11 @@
 import tensorflow as tf
 from tensorflow.keras import Model, Input
-from tensorflow.keras.layers import Dropout, Softmax
+from tensorflow.keras.layers import Dropout
 from tensorflow.keras.optimizers import Adam
 from tensorflow.keras import regularizers
+from tensorflow.keras.losses import SparseCategoricalCrossentropy
 
-from graphgallery.nn.layers import GraphConvFeature
+from graphgallery.nn.layers import GraphConvFeature, Gather
 from graphgallery.nn.models import SemiSupervisedModel
 from graphgallery.sequence import FullBatchNodeSequence
 from graphgallery.utils.shape_utils import set_equal_in_length
@@ -17,34 +18,34 @@ class GCNF(SemiSupervisedModel):
 
         Arguments:
         ----------
-            adj: shape (N, N), Scipy sparse matrix if  `is_adj_sparse=True`, 
+            adj: shape (N, N), Scipy sparse matrix if  `is_adj_sparse=True`,
                 Numpy array-like (or matrix) if `is_adj_sparse=False`.
-                The input `symmetric` adjacency matrix, where `N` is the number 
+                The input `symmetric` adjacency matrix, where `N` is the number
                 of nodes in graph.
-            x: shape (N, F), Scipy sparse matrix if `is_x_sparse=True`, 
+            x: shape (N, F), Scipy sparse matrix if `is_x_sparse=True`,
                 Numpy array-like (or matrix) if `is_x_sparse=False`.
                 The input node feature matrix, where `F` is the dimension of features.
             labels: Numpy array-like with shape (N,)
                 The ground-truth labels for all nodes in graph.
-            norm_adj (Float scalar, optional): 
-                The normalize rate for adjacency matrix `adj`. (default: :obj:`-0.5`, 
-                i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}}) 
-            norm_x (Boolean, optional): 
-                Whether to use row-wise normalization for node feature matrix. 
-                (default :bool: `True`)
-            device (String, optional): 
-                The device where the model is running on. You can specified `CPU` or `GPU` 
+            norm_adj (Float scalar, optional):
+                The normalize rate for adjacency matrix `adj`. (default: :obj:`-0.5`,
+                i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}})
+            norm_x (String, optional):
+                How to normalize the node feature matrix. See `graphgallery.normalize_x`
+                (default :obj: `None`)
+            device (String, optional):
+                The device where the model is running on. You can specified `CPU` or `GPU`
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
-            seed (Positive integer, optional): 
-                Used in combination with `tf.random.set_seed` & `np.random.seed` & `random.seed`  
-                to create a reproducible sequence of tensors across multiple calls. 
+            seed (Positive integer, optional):
+                Used in combination with `tf.random.set_seed` & `np.random.seed` & `random.seed`
+                to create a reproducible sequence of tensors across multiple calls.
                 (default :obj: `None`, i.e., using random seed)
-            name (String, optional): 
+            name (String, optional):
                 Specified name for the model. (default: :str: `class.__name__`)
 
     """
 
-    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x='l1',
+    def __init__(self, adj, x, labels, norm_adj=-0.5, norm_x=None,
                  device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels, device=device, seed=seed, name=name, **kwargs)
@@ -100,11 +101,11 @@ def build(self, hiddens=[16], activations=['relu'], dropouts=[0.5], l2_norms=[5e
             # of the input data to remain the same
             # if ensure_shape:
             #     h = tf.ensure_shape(h, [self.n_nodes, self.n_classes])
-            h = tf.gather(h, index)
-            output = Softmax()(h)
+            h = Gather()([h, index])
 
-            model = Model(inputs=[x, adj, index], outputs=output)
-            model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=lr), metrics=['accuracy'])
+            model = Model(inputs=[x, adj, index], outputs=h)
+            model.compile(loss=SparseCategoricalCrossentropy(from_logits=True),
+                          optimizer=Adam(lr=lr), metrics=['accuracy'])
 
             self.model = model
 

graphgallery/nn/models/semisupervised/mediansage.py → graphgallery/nn/models/semisupervised/experimental/mediansage.py

@@ -1,9 +1,10 @@
 import numpy as np
 import tensorflow as tf
 from tensorflow.keras import Model, Input
-from tensorflow.keras.layers import Dropout, Softmax
+from tensorflow.keras.layers import Dropout
 from tensorflow.keras.optimizers import Adam
 from tensorflow.keras import regularizers
+from tensorflow.keras.losses import SparseCategoricalCrossentropy
 
 from graphgallery.nn.layers import MedianAggregator, MedianGCNAggregator
 from graphgallery.nn.models import SemiSupervisedModel
@@ -34,7 +35,7 @@ class MedianSAGE(SemiSupervisedModel):
                 `5` sencond-order neighbors, and the radius for `GraphSAGE` is `2`)
             norm_x (String, optional): 
                 How to normalize the node feature matrix. See `graphgallery.normalize_x`
-                (default :str: `l1`)
+                (default :obj: `None`)
             device (String, optional): 
                 The device where the model is running on. You can specified `CPU` or `GPU` 
                 for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`)
@@ -48,7 +49,7 @@ class MedianSAGE(SemiSupervisedModel):
 
     """
 
-    def __init__(self, adj, x, labels, n_samples=[15, 3], norm_x='l1',
+    def __init__(self, adj, x, labels, n_samples=[15, 3], norm_x=None,
                  device='CPU:0', seed=None, name=None, **kwargs):
 
         super().__init__(adj, x, labels, device=device, seed=seed, name=name, **kwargs)
@@ -124,10 +125,10 @@ def build(self, hiddens=[32], activations=['relu'], dropouts=[0.5], l2_norms=[5e
             h = h[0]
             if output_normalize:
                 h = tf.nn.l2_normalize(h, axis=1)
-            output = Softmax()(h)
 
-            model = Model(inputs=[x, nodes, *neighbors], outputs=output)
-            model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=lr), metrics=['accuracy'])
+            model = Model(inputs=[x, nodes, *neighbors], outputs=h)
+            model.compile(loss=SparseCategoricalCrossentropy(from_logits=True),
+                          optimizer=Adam(lr=lr), metrics=['accuracy'])
 
             self.model = model