This paper presents an evaluation of the efficacy and efficiency of various transfer learning methods in wood knot classification. We compared the wood knot classification results from four different convolutional neural networks (Xception, Inception V3, ResNet50 and VGG16) in order to determine whether they are suitable for this task or not. We conducted an experiment in which we prototyped a series of classifiers using two main approaches - using a neural network with pre- trained weights and training weights from scratch. We used four network architectures, two approaches and eight optimizers to build a total of 64 classifiers. Comparison of the classifiers performance lead as towards the direction of future work.