conv_split_awa.py [443:488]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                                model.keep_prob: 1.0, model.train_phase: True}
                        logit_mask_dict = {m_t: i_t for (m_t, i_t) in zip(model.output_mask, a_gem_logit_mask)}
                        ref_feed_dict.update(logit_mask_dict)
                        ref_feed_dict[model.mem_batch_size] = float(len(mem_sample_mask))
                        sess.run(model.store_ref_grads, feed_dict=ref_feed_dict)
                        # Compute the gradient for current task and project if need be
                        a_gem_logit_mask[:] = 0
                        logit_mask_offset = task * TOTAL_CLASSES
                        classes_adjusted_for_head = [cls + logit_mask_offset for cls in task_labels[task]]
                        a_gem_logit_mask[task][classes_adjusted_for_head] = 1.0
                        logit_mask_dict = {m_t: i_t for (m_t, i_t) in zip(model.output_mask, a_gem_logit_mask)}
                        feed_dict.update(logit_mask_dict)
                        feed_dict[model.mem_batch_size] = batch_size
                        _, loss = sess.run([model.train_subseq_tasks, model.agem_loss], feed_dict=feed_dict)

                elif model.imp_method == 'RWALK':
                    feed_dict[model.output_mask] = logit_mask
                    # If first iteration of the first task then set the initial value of the running fisher
                    if task == 0 and iters == 0:
                        sess.run([model.set_initial_running_fisher], feed_dict=feed_dict)
                        # Store the current value of the weights
                        sess.run(model.weights_delta_old_grouped)
                    # Update fisher and importance score after every few iterations
                    if (iters + 1) % model.fisher_update_after == 0:
                        # Update the importance score using distance in riemannian manifold   
                        sess.run(model.update_big_omega_riemann)
                        # Now that the score is updated, compute the new value for running Fisher
                        sess.run(model.set_running_fisher)
                        # Store the current value of the weights
                        sess.run(model.weights_delta_old_grouped)
                        # Reset the delta_L
                        sess.run([model.reset_small_omega])

                    _, _, _, _, loss = sess.run([model.set_tmp_fisher, model.weights_old_ops_grouped, 
                        model.train, model.update_small_omega, model.reg_loss], feed_dict=feed_dict)


                if (iters % 100 == 0):
                    print('Step {:d} {:.3f}'.format(iters, loss))

                if (math.isnan(loss)):
                    print('ERROR: NaNs NaNs NaNs!!!')
                    break_training = 1
                    break

            print('\t\t\t\tTraining for Task%d done!'%(task))
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



conv_split_awa_hybrid.py [452:496]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                                model.keep_prob: 1.0, model.train_phase: True}
                        logit_mask_dict = {m_t: i_t for (m_t, i_t) in zip(model.output_mask, a_gem_logit_mask)}
                        ref_feed_dict.update(logit_mask_dict)
                        ref_feed_dict[model.mem_batch_size] = float(len(mem_sample_mask))
                        sess.run(model.store_ref_grads, feed_dict=ref_feed_dict)
                        # Compute the gradient for current task and project if need be
                        a_gem_logit_mask[:] = 0
                        logit_mask_offset = task * TOTAL_CLASSES
                        classes_adjusted_for_head = [cls + logit_mask_offset for cls in task_labels[task]]
                        a_gem_logit_mask[task][classes_adjusted_for_head] = 1.0
                        logit_mask_dict = {m_t: i_t for (m_t, i_t) in zip(model.output_mask, a_gem_logit_mask)}
                        feed_dict.update(logit_mask_dict)
                        feed_dict[model.mem_batch_size] = batch_size
                        _, loss = sess.run([model.train_subseq_tasks, model.agem_loss], feed_dict=feed_dict)

                elif model.imp_method == 'RWALK':
                    feed_dict[model.output_mask] = logit_mask
                    # If first iteration of the first task then set the initial value of the running fisher
                    if task == 0 and iters == 0:
                        sess.run([model.set_initial_running_fisher], feed_dict=feed_dict)
                        # Store the current value of the weights
                        sess.run(model.weights_delta_old_grouped)
                    # Update fisher and importance score after every few iterations
                    if (iters + 1) % model.fisher_update_after == 0:
                        # Update the importance score using distance in riemannian manifold   
                        sess.run(model.update_big_omega_riemann)
                        # Now that the score is updated, compute the new value for running Fisher
                        sess.run(model.set_running_fisher)
                        # Store the current value of the weights
                        sess.run(model.weights_delta_old_grouped)
                        # Reset the delta_L
                        sess.run([model.reset_small_omega])

                    _, _, _, _, loss = sess.run([model.set_tmp_fisher, model.weights_old_ops_grouped, 
                        model.train, model.update_small_omega, model.reg_loss], feed_dict=feed_dict)

                if (iters % 100 == 0):
                    print('Step {:d} {:.3f}'.format(iters, loss))

                if (math.isnan(loss)):
                    print('ERROR: NaNs NaNs NaNs!!!')
                    break_training = 1
                    break

            print('\t\t\t\tTraining for Task%d done!'%(task))
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -