lr_decay_epoch = 0 n_lr_decays = 0 train_ce, valid_ce = [], [] flag=True global insp_ insp_ = None res_dir = save_results(train_ce, valid_ce, res_dir, params=params) save_params(params, res_dir) lr.decay = 1 for epoch in xrange(tr.n_epochs): ce = [] print_params(params) #################################################################### #################################################################### print "\n%s\n\t epoch %d \n%s"%('-'*30, epoch, '-'*30) #################################################################### #################################################################### time_start = time() for i in range(loader.n_iter_train): #load data time_start_iter = time() loader.next_train_batch(x_, y_, x_skeleton_) tr.batch_size = y_.get_value(borrow=True).shape[0] ce.append(_batch(train_model, tr.batch_size, batch, True, apply_updates)) timing_report(i, time()-time_start_iter, tr.batch_size, res_dir) print "\t| "+ training_report(ce[-1]) + ", finish total of: 0." + str(i*1.0/loader.n_iter_train)
save_params(net_convnet3d_grbm_early_fusion.params, res_dir) # default learning rate lr.start = 0.0001 lr.stop = 0.00001 # Wudi makes thie to explicity control the learning rate learning_rate_map = linspace(lr.start, lr.stop, tr.n_epochs) for epoch in xrange(tr.n_epochs): learning_rate.set_value(float32(learning_rate_map[epoch])) ce = [] out_mean_train = [] out_std_train = [] print_params(net_convnet3d_grbm_early_fusion.params) #################################################################### print "\n%s\n\t epoch %d \n%s"%('-'*30, epoch, '-'*30) time_start = time() for i in range(loader.n_iter_train): #load data time_start_iter = time() loader.next_train_batch(x_, y_, x_skeleton_) ce_temp, out_mean_temp, out_std_temp = _batch(train_model, tr.batch_size, batch, True, apply_updates) ce.append(ce_temp) out_mean_train.append(out_mean_temp) out_std_train.append(out_std_temp) print "Training: No.%d iter of Total %d, %d s"% (i,loader.n_iter_train, time()-time_start_iter) \ + "\t| negative_log_likelihood "+ training_report(ce[-1]) # End of Epoch
lr_decay_epoch = 0 n_lr_decays = 0 train_ce, valid_ce = [], [] flag = True global insp_ insp_ = None res_dir = save_results(train_ce, valid_ce, res_dir, params=params) save_params(params, res_dir) lr.decay = 1 for epoch in xrange(tr.n_epochs): ce = [] print_params(params) #################################################################### #################################################################### print "\n%s\n\t epoch %d \n%s" % ('-' * 30, epoch, '-' * 30) #################################################################### #################################################################### time_start = time() for i in range(loader.n_iter_train): #load data time_start_iter = time() loader.next_train_batch(x_, y_, x_skeleton_) print('tr.batch_size_before=%d' % tr.batch_size) tr.batch_size = y_.get_value(borrow=True).shape[0] print('tr.batch_size_after=%d' % tr.batch_size) ce.append( _batch(train_model, tr.batch_size, batch, True, apply_updates)[0])
save_params(net_convnet3d_grbm_early_fusion.params, res_dir) # default learning rate lr.start = 0.0001 lr.stop = 0.00001 # Wudi makes thie to explicity control the learning rate learning_rate_map = linspace(lr.start, lr.stop, tr.n_epochs) for epoch in xrange(tr.n_epochs): learning_rate.set_value(float32(learning_rate_map[epoch])) ce = [] out_mean_train = [] out_std_train = [] print_params(net_convnet3d_grbm_early_fusion.params) #################################################################### print "\n%s\n\t epoch %d \n%s" % ('-' * 30, epoch, '-' * 30) time_start = time() for i in range(loader.n_iter_train): #load data time_start_iter = time() loader.next_train_batch(x_, y_, x_skeleton_) ce_temp, out_mean_temp, out_std_temp = _batch(train_model, tr.batch_size, batch, True, apply_updates) ce.append(ce_temp) out_mean_train.append(out_mean_temp) out_std_train.append(out_std_temp) print "Training: No.%d iter of Total %d, %d s"% (i,loader.n_iter_train, time()-time_start_iter) \