# # reset ADAM variables # sess.run(tf.initialize_variables(sigma_opt_vars)) # sigma_iter = 0 # that_change = sigma_opt_thresh * 2 # old_that = 0 # while that_change > sigma_opt_thresh and sigma_iter < sigma_opt_iter: # new_sigma, that_np, _ = sess.run([sigma, that, sigma_solver], # feed_dict={eval_real_PH: eval_eval_real, eval_sample_PH: eval_eval_sample}) # that_change = np.abs(that_np - old_that) # old_that = that_np # sigma_iter += 1 # opt_sigma = sess.run(sigma) # try: # mmd2, that_np = sess.run(mix_rbf_mmd2_and_ratio(eval_test_real, eval_test_sample, biased=False, sigmas=sigma)) # except ValueError: # mmd2 = 'NA' # that = 'NA' # # MMD[epoch, ] = mmd2 # -- save model parameters -- # model.dump_parameters(sub_id + '_' + str(seq_length) + '_' + str(epoch), sess) np.save('./experiments/plots/gs/' + identifier + '_' + 'MMD.npy', MMD) end = time() - begin print('Training terminated | Training time=%d s' % (end)) print("Training terminated | training time = %ds " % (time() - begin))
# print(new_sigma,that_np) that_change = np.abs(that_np - old_that) old_that = that_np sigma_iter += 1 opt_sigma = sess.run(sigma) mmd2, that_np = sess.run( mix_rbf_mmd2_and_ratio(eval_test_real, eval_test_sample, biased=False, sigmas=sigma)) ## save parameters if mmd2 < best_mmd2_so_far and epoch > 10: best_epoch = epoch best_mmd2_so_far = mmd2 model.dump_parameters(identifier + '_' + str(epoch), sess) pdf_sample = 'NA' pdf_real = 'NA' else: # report nothing this epoch mmd2 = 'NA' that = 'NA' pdf_sample = 'NA' pdf_real = 'NA' t = time() - t0 try: print('%d\t%.2f\t%.4f\t%.4f\t%.5f\t' % (epoch, t, D_loss_curr, G_loss_curr, mmd2)) except TypeError: # pdf are missing (format as strings)
# pdf_real = 'NA' # # MMD[epoch, ] = mmd2 # # ## print # # t = time() - t0 # print('epoch\ttime\tD_loss\tG_loss\tmmd2\tthat\tpdf_sample\tpdf_real') # try: # print('%d\t%.2f\t%.4f\t%.4f\t%.5f\t%.0f\t%.2f\t%.2f' % ( # epoch, t, D_loss_curr, G_loss_curr, mmd2, that_np, pdf_sample, pdf_real)) # except TypeError: # pdf are missing (format as strings) # print('%d\t%.2f\t%.4f\t%.4f\t%s\t%s\t %s\t %s' % ( # epoch, t, D_loss_curr, G_loss_curr, mmd2, that_np, pdf_sample, pdf_real)) # #-- save model parameters -- # model.dump_parameters( settings['identifier'] + '_' + str(settings['seq_length']) + '_' + str(epoch), sess) # model_parameters = dict() # for v in tf.trainable_variables(): # model_parameters[v.name] = sess.run(v) # print('Saved {} parameters'.format(len(model_parameters))) # np.save('./experiments/plots/gs/' + settings['identifier'] + '_' + settings['seq_length'] + '_' + 'MMD.npy', MMD) end = time() - begin # print('Training terminated | Training time=%ds' %(end) ) print("Training terminated | training time = %ds " % (time() - begin))