The transform trellis code is an optimal source code as a block size and the constraint length of a shift register go to infinite for stationary Gaussian sources with the squared-error distortion measure. However to implement this code, we have to choose the finite block size and constraint length. Moreover real-world sources are inherently non stationary. To overcome these difficulties, we developed a training algorithm for the transform trellis code. The trained transform trellis code which uses the same rates to each block led to a variation in the resulting distortion from one block to another. To alleviate this non-uniformity in the encoded image, we constructed clusters from the variance of the training data and assigned different rates for each cluster.