This paper approaches the problem of texture classification from very challenging dataset, the describable texture dataset (DTD), using a combination of popular pre-trained convolutional neural networks architectures to improve the overall accuracy of the system. Different architectures include mixture of VGG, Resnet50, Inception, Xception models with different number of layers and parameters which are individually tweaked to attain maximum accuracy. The results obtained from these models are combined using different technique to obtain the best results. In order to better generalize our model we even tested for other well known datasets such as KTH-TIP-2b, FMD and CUReT. Using the ensemble techniques we were able to achieve comparable accuracy wrt to state of the art techniques.