keras LSTM train validation different length negative dimension error -

i have lstm task, input data inertial sensor , gyroscope measurements, 6d. output relative displacement 7d (3d pos, 4d angular).

now use 70% data training, [batchsize=1, time_steps=200, input_dimension=6], , set lstm stateful because want remember states last sequence. write customized loss function, owing fact last 4d of output should have norm 1 (quaternion angle representation, not important question)

after every epoch, want measure loss validation data set, single long sequence, can this? validation , training sequence have different lengths. got error this

• invalidargumenterror: shape [-1,-1,6] has negative dimensions

i attached detailed python notebook followed: https://github.com/matthewd1993/camerapose/blob/master/my_exp/sensor_fusioin_lstm-copy1.ipynb

• another question is: write metric function return vector 7d, in logs seems first value returned. idea?

thanks in advance!