deepcell.model_zoo.fpn.__create_semantic_head applies ReLU for n_channels=1

[JLrumberger]

Describe the bug

deepcell.model_zoo.fpn.__create_semantic_head which is used by default in your segmentation models applies ReLU activation function on the output if n_classes == 1. I would expect that segmentation models with single channel output use sigmoid activation in their last layer. Furthermore, applying ReLU makes the logits unusable for binary_cross_entropy_with_logits since sigmoid(relu(x)) has a lower bound of 0.5.

To Reproduce

Something like the below code would return values bound between [0.5, 1]

x = tf.constant(np.random.rand(1,256,256,2), tf.float32)
model = PanopticNet(
    backbone="resnet50", input_shape=[256,256,2],
    norm_method="std", num_semantic_classes=[1],
)
out = model(x)
print(tf.reduce_min(tf.sigmoid(out)))

Expected behavior
Expected would be that the output is run through sigmoid, not ReLU.

Desktop (please complete the following information):

• OS: Win 10
• deepcell 0.12.3
• Python 3.8.0

Thanks for maintaining this library :)