How to predict each class probability

[farhantandia]

how to predict the class probability? when I set the output to arcface output (softmax) w.r.t number of class, I got an error when run this model = Model(inputs=model.input[0], outputs=model.layers[-1].output) ,-1 instead of -3 `--------------------------------------------------------------------------- ValueError Traceback (most recent call last) in ----> 1 model = Model(inputs=model.input[0], outputs=model.layers[-1].output)

~\anaconda3\envs\kaggle\lib\site-packages\tensorflow\python\keras\engine\training.py in new(cls, *args, **kwargs) 240 # Functional model 241 from tensorflow.python.keras.engine import functional # pylint: disable=g-import-not-at-top --> 242 return functional.Functional(*args, **kwargs) 243 else: 244 return super(Model, cls).new(cls, *args, **kwargs)

~\anaconda3\envs\kaggle\lib\site-packages\tensorflow\python\training\tracking\base.py in _method_wrapper(self, *args, **kwargs) 455 self._self_setattr_tracking = False # pylint: disable=protected-access 456 try: --> 457 result = method(self, *args, **kwargs) 458 finally: 459 self._self_setattr_tracking = previous_value # pylint: disable=protected-access

~\anaconda3\envs\kaggle\lib\site-packages\tensorflow\python\keras\engine\functional.py in init(self, inputs, outputs, name, trainable) 113 # 'arguments during initialization. Got an unexpected argument:') 114 super(Functional, self).init(name=name, trainable=trainable) --> 115 self._init_graph_network(inputs, outputs) 116 117 @trackable.no_automatic_dependency_tracking

~\anaconda3\envs\kaggle\lib\site-packages\tensorflow\python\training\tracking\base.py in _method_wrapper(self, *args, **kwargs) 455 self._self_setattr_tracking = False # pylint: disable=protected-access 456 try: --> 457 result = method(self, *args, **kwargs) 458 finally: 459 self._self_setattr_tracking = previous_value # pylint: disable=protected-access

~\anaconda3\envs\kaggle\lib\site-packages\tensorflow\python\keras\engine\functional.py in _init_graph_network(self, inputs, outputs) 189 # Keep track of the network's nodes and layers. 190 nodes, nodes_by_depth, layers, _ = _map_graph_network( --> 191 self.inputs, self.outputs) 192 self._network_nodes = nodes 193 self._nodes_by_depth = nodes_by_depth

~\anaconda3\envs\kaggle\lib\site-packages\tensorflow\python\keras\engine\functional.py in _map_graph_network(inputs, outputs) 929 'The following previous layers ' 930 'were accessed without issue: ' + --> 931 str(layers_with_complete_input)) 932 for x in nest.flatten(node.outputs): 933 computable_tensors.add(id(x))

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_37:0", shape=(None, 51), dtype=float32) at layer "softmax_output". The following previous layers were accessed without issue:

[-1].output : <tf.Tensor 'class_output/Softmax_3:0' shape=(None, 10) dtype=float32> while [-3] return feature vectors which suitable for visualization. thanks.

[ozora-ogino]

Hi @farhantandia, Here is my DNN implementation with ASoftmax

class DNN(tf.keras.models.Model):
    def __init__(self, num_classes=10):
        super(DNN, self).__init__()
        weight_decay = 1e-4
        self.layer_1 = tf.keras.layers.Dense(32, activation='relu')
        self.layer_2 = tf.keras.layers.Dense(10)
        self.out = ArcFace(n_classes=num_classes, regularizer=regularizers.l2(weight_decay))

    def call(self, x, training=False):
        if training:
            x, y = x[0], x[1]
        x = self.layer_1(x)
        x = self.layer_2(x)
        if training:
            out = self.out([x, y])
        else:
            # Prediction 
            # Thanks to this, you don't need to pass labels to model when you predict
            out = tf.nn.softmax(x @ self.out.W)
        return out

Sample code is following;

odel = DNN()
optimizer = tf.keras.optimizers.Adam()
loss = tf.keras.losses.categorical_crossentropy

model.compile(loss=loss, optimizer=optimizer, metrics=['acc'])
model.fit([x_train, tf.keras.utils.to_categorical(y_train, 10)], tf.keras.utils.to_categorical(y_train, 10), batch_size=512, epochs=10)

pred = model.predict(x_test)

I hope it works in your situation.

[ozora-ogino]

I implemented my own ArcFace. https://github.com/ozora-ogino/asoftmax-tf

This works correctly and I believe this can solve your problem.