python - Can I apply calculated gradient in tensorflow? -
what want simulate back-propagation process on different machines, 1 machine, gradient layer3 d(layer3_output)/d(layer2_output) numpy array, how able d(layer3_output)/d(layer1_output) efficiently given gradient received , passed previous layer?
when create network , attach loss call minimize on optimizer, (under hood) calls "apply_gradients". function adds gradient computing ops graph. have request op responsible partial derivative , pass precomputed partial derivative through feed_dict option. use tensorboard visualize graph , investigate names of gradients interested in. default in "gradient" namescope, , naming of each op analogous operations, among lines of gradient/output_op:0 etc.
Comments
Post a Comment