7fd49e352027b5e24d0f8ede917877e5d5ef945f,dataset/models/tf/vgg.py,VGG,_build,#VGG#,45

Before Change

        if batch_norm:
            kwargs["batch_norm"] = batch_norm

        net = self.body(dim, inputs["images"], arch, **kwargs)
        net = self.head(dim, net, style="dense", layout="fff", num_classes=num_classes, units=[100, 100], **kwargs)

        logits = tf.identity(net, name="predictions")
        pred_proba = tf.nn.softmax(logits, name="predicted_proba")
        pred_labels = tf.argmax(pred_proba, axis=-1, name="predicted_labels")
        true_labels = tf.argmax(inputs["labels"], axis=-1, name="true_labels")
        equality = tf.equal(pred_labels, true_labels)
        equality = tf.cast(equality, dtype=tf.float32)

After Change

        - number of filters in each layer
    

    def _build(self):
        names = ["images", "labels"]
        _, inputs = self._make_inputs(names)

        num_classes = self.num_classes("labels")
        data_format = self.data_format("images")
        dim = self.spatial_dim("images")
        batch_norm = self.get_from_config("batch_norm", {"momentum": 0.1})
        arch = self.get_from_config("arch", "VGG16")

        input_block_config = self.get_from_config("input_block", {})
        body_config = self.get_from_config("body", {})
        head_config = self.get_from_config("head", {"units": [100, 100]})
        head_config["num_classes"] = num_classes

        kwargs = {"training": self.is_training}
        if batch_norm:
            kwargs["batch_norm"] = batch_norm

        with tf.variable_scope("VGG"):
            x = self.input_block(dim, inputs["images"], **{**kwargs, **input_block_config})
            x = self.body(dim, x, arch, **{**kwargs, **body_config})
            output = self.head(dim, x, **{**kwargs, **head_config})

        logits = tf.identity(output, name="predictions")
        tf.nn.softmax(logits, name="predicted_proba")

    @classmethod
    def body(cls, dim, inputs, arch, **kwargs):
         Create base VGG layers

Instances