Tensorflow1.15 pointnet part_seg

それは、さて置き、早速修正した、コードを掲載します。
part_seg/train_py update 2020.1.15 by nishi
import argparse import subprocess import tensorflow as tf import numpy as np from datetime import datetime import json import os import sys BASE_DIR = os.path.dirname(os.path.abspath(__file__)) sys.path.append(BASE_DIR) sys.path.append(os.path.dirname(BASE_DIR)) import provider import pointnet_part_seg as model # DEFAULT SETTINGS parser = argparse.ArgumentParser() parser.add_argument('--gpu', type=int, default=0, help='GPU to use [default: GPU 0]') #parser.add_argument('--batch', type=int, default=32, help='Batch Size during training [default: 32]') parser.add_argument('--batch', type=int, default=16, help='Batch Size during training [default: 32]') #parser.add_argument('--epoch', type=int, default=200, help='Epoch to run [default: 50]') parser.add_argument('--epoch', type=int, default=8, help='Epoch to run [default: 50]') parser.add_argument('--point_num', type=int, default=2048, help='Point Number [256/512/1024/2048]') parser.add_argument('--output_dir', type=str, default='train_results', help='Directory that stores all training logs and trained models') parser.add_argument('--wd', type=float, default=0, help='Weight Decay [Default: 0.0]') FLAGS = parser.parse_args() #hdf5_data_dir = os.path.join(BASE_DIR, './hdf5_data') hdf5_data_dir = '/tmp/pointnet/part_seg/hdf5_data' # MAIN SCRIPT point_num = FLAGS.point_num batch_size = FLAGS.batch output_dir = FLAGS.output_dir if not os.path.exists(output_dir): os.mkdir(output_dir) color_map_file = os.path.join(hdf5_data_dir, 'part_color_mapping.json') color_map = json.load(open(color_map_file, 'r')) all_obj_cats_file = os.path.join(hdf5_data_dir, 'all_object_categories.txt') fin = open(all_obj_cats_file, 'r') lines = [line.rstrip() for line in fin.readlines()] all_obj_cats = [(line.split()[0], line.split()[1]) for line in lines] fin.close() all_cats = json.load(open(os.path.join(hdf5_data_dir, 'overallid_to_catid_partid.json'), 'r')) NUM_CATEGORIES = 16 NUM_PART_CATS = len(all_cats) print('#### Batch Size: {0}'.format(batch_size)) print('#### Point Number: {0}'.format(point_num)) print('#### Training using GPU: {0}'.format(FLAGS.gpu)) DECAY_STEP = 16881 * 20 DECAY_RATE = 0.5 LEARNING_RATE_CLIP = 1e-5 BN_INIT_DECAY = 0.5 BN_DECAY_DECAY_RATE = 0.5 BN_DECAY_DECAY_STEP = float(DECAY_STEP * 2) BN_DECAY_CLIP = 0.99 BASE_LEARNING_RATE = 0.001 MOMENTUM = 0.9 TRAINING_EPOCHES = FLAGS.epoch print('### Training epoch: {0}'.format(TRAINING_EPOCHES)) TRAINING_FILE_LIST = os.path.join(hdf5_data_dir, 'train_hdf5_file_list.txt') TESTING_FILE_LIST = os.path.join(hdf5_data_dir, 'val_hdf5_file_list.txt') #MODEL_STORAGE_PATH = os.path.join(output_dir, 'trained_models') MODEL_STORAGE_PATH = os.path.join('.', 'trained_models') if not os.path.exists(MODEL_STORAGE_PATH): os.mkdir(MODEL_STORAGE_PATH) LOG_STORAGE_PATH = os.path.join(output_dir, 'logs') if not os.path.exists(LOG_STORAGE_PATH): os.mkdir(LOG_STORAGE_PATH) SUMMARIES_FOLDER = os.path.join(output_dir, 'summaries') if not os.path.exists(SUMMARIES_FOLDER): os.mkdir(SUMMARIES_FOLDER) # add by nishi class Acc_Saver(): def __init__(self,path,thres=0.6): self.path=path self.acc_m=None self.thres=thres if os.path.exists(path): f = open(path) data = f.read() f.close() self.acc_m=float(data) def check(self,acc): ok=False if self.acc_m == None: self.acc_m=acc ok=True elif acc > self.acc_m: self.acc_m = acc ok=True if ok == True and self.acc_m > self.thres: return True else: return False def save(self): f = open(self.path,'w') data = f.write(str(self.acc_m)) f.close() # add by nishi class Next_Epoch(): def __init__(self,path,start=0): self.path=path self.next_epoch_m=start self.steps=0 if os.path.exists(path): f = open(path) data = f.read() f.close() j = data.split(',') self.next_epoch_m=int(j[0]) self.steps=int(j[1]) def set(self,epoch): self.next_epoch_m=epoch def inc(self): self.next_epoch_m +=1 def get(self): return self.next_epoch_m def add_steps(self,step): self.steps += step def get_steps(self): return self.steps def save(self): f = open(self.path,'w') s=str(self.next_epoch_m)+','+str(self.steps) data = f.write(s) f.close() def printout(flog, data): print(data) if flog != None: flog.write(data + '\n') def placeholder_inputs(): pointclouds_ph = tf.compat.v1.placeholder(tf.float32, shape=(batch_size, point_num, 3)) input_label_ph = tf.compat.v1.placeholder(tf.float32, shape=(batch_size, NUM_CATEGORIES)) labels_ph = tf.compat.v1.placeholder(tf.int32, shape=(batch_size)) seg_ph = tf.compat.v1.placeholder(tf.int32, shape=(batch_size, point_num)) return pointclouds_ph, input_label_ph, labels_ph, seg_ph def convert_label_to_one_hot(labels): label_one_hot = np.zeros((labels.shape[0], NUM_CATEGORIES)) for idx in range(labels.shape[0]): label_one_hot[idx, labels[idx]] = 1 return label_one_hot # add by nishi next_epoch=Next_Epoch('trained_models/next_epoch.txt') START_EPOCH = next_epoch.get() print('### Start epoch: {0}'.format(START_EPOCH)) print('### Start steps: {0}'.format(next_epoch.get_steps())) # make Acc_Saver add by nishi acc_saver=Acc_Saver('trained_models/eval_acc.txt',thres=0.3) def train(): with tf.Graph().as_default(): with tf.device('/gpu:'+str(FLAGS.gpu)): pointclouds_ph, input_label_ph, labels_ph, seg_ph = placeholder_inputs() is_training_ph = tf.compat.v1.placeholder(tf.bool, shape=()) # batch is equal with test steps. by nishi # changed by nishi if START_EPOCH > 0: batch = tf.Variable(next_epoch.get_steps(), trainable=False) else: batch = tf.Variable(0, trainable=False) learning_rate = tf.compat.v1.train.exponential_decay( BASE_LEARNING_RATE, # base learning rate batch * batch_size, # global_var indicating the number of steps DECAY_STEP, # step size DECAY_RATE, # decay rate staircase=True # Stair-case or continuous decreasing ) learning_rate = tf.maximum(learning_rate, LEARNING_RATE_CLIP) bn_momentum = tf.compat.v1.train.exponential_decay( BN_INIT_DECAY, batch*batch_size, BN_DECAY_DECAY_STEP, BN_DECAY_DECAY_RATE, staircase=True) bn_decay = tf.minimum(BN_DECAY_CLIP, 1 - bn_momentum) lr_op = tf.compat.v1.summary.scalar('learning_rate', learning_rate) batch_op = tf.compat.v1.summary.scalar('batch_number', batch) bn_decay_op = tf.compat.v1.summary.scalar('bn_decay', bn_decay) # model out put # labels_pred / net -> label predict # seg_pred /net2 -> segemt predict # end_points -> end_points['transform'] labels_pred, seg_pred, end_points = model.get_model(pointclouds_ph, input_label_ph, \ is_training=is_training_ph, bn_decay=bn_decay, cat_num=NUM_CATEGORIES, \ part_num=NUM_PART_CATS, batch_size=batch_size, num_point=point_num, weight_decay=FLAGS.wd) # model.py defines both classification net and segmentation net, which share the common global feature extractor network. # In model.get_loss, we define the total loss to be weighted sum of the classification and segmentation losses. # Here, we only train for segmentation network. Thus, we set weight to be 1.0. # loss function() #wx= 1.0 -> orginal # set by nishi wx=0.99 loss, label_loss, per_instance_label_loss, seg_loss, per_instance_seg_loss, per_instance_seg_pred_res \ = model.get_loss(labels_pred, seg_pred, labels_ph, seg_ph, wx, end_points) total_training_loss_ph = tf.compat.v1.placeholder(tf.float32, shape=()) total_testing_loss_ph = tf.compat.v1.placeholder(tf.float32, shape=()) label_training_loss_ph = tf.compat.v1.placeholder(tf.float32, shape=()) label_testing_loss_ph = tf.compat.v1.placeholder(tf.float32, shape=()) seg_training_loss_ph = tf.compat.v1.placeholder(tf.float32, shape=()) seg_testing_loss_ph = tf.compat.v1.placeholder(tf.float32, shape=()) label_training_acc_ph = tf.compat.v1.placeholder(tf.float32, shape=()) label_testing_acc_ph = tf.compat.v1.placeholder(tf.float32, shape=()) label_testing_acc_avg_cat_ph = tf.compat.v1.placeholder(tf.float32, shape=()) seg_training_acc_ph = tf.compat.v1.placeholder(tf.float32, shape=()) seg_testing_acc_ph = tf.compat.v1.placeholder(tf.float32, shape=()) seg_testing_acc_avg_cat_ph = tf.compat.v1.placeholder(tf.float32, shape=()) total_train_loss_sum_op = tf.compat.v1.summary.scalar('total_training_loss', total_training_loss_ph) total_test_loss_sum_op = tf.compat.v1.summary.scalar('total_testing_loss', total_testing_loss_ph) label_train_loss_sum_op = tf.compat.v1.summary.scalar('label_training_loss', label_training_loss_ph) label_test_loss_sum_op = tf.compat.v1.summary.scalar('label_testing_loss', label_testing_loss_ph) seg_train_loss_sum_op = tf.compat.v1.summary.scalar('seg_training_loss', seg_training_loss_ph) seg_test_loss_sum_op = tf.compat.v1.summary.scalar('seg_testing_loss', seg_testing_loss_ph) label_train_acc_sum_op = tf.compat.v1.summary.scalar('label_training_acc', label_training_acc_ph) label_test_acc_sum_op = tf.compat.v1.summary.scalar('label_testing_acc', label_testing_acc_ph) label_test_acc_avg_cat_op = tf.compat.v1.summary.scalar('label_testing_acc_avg_cat', label_testing_acc_avg_cat_ph) seg_train_acc_sum_op = tf.compat.v1.summary.scalar('seg_training_acc', seg_training_acc_ph) seg_test_acc_sum_op = tf.compat.v1.summary.scalar('seg_testing_acc', seg_testing_acc_ph) seg_test_acc_avg_cat_op = tf.compat.v1.summary.scalar('seg_testing_acc_avg_cat', seg_testing_acc_avg_cat_ph) train_variables = tf.compat.v1.trainable_variables() # trainnr function() trainer = tf.compat.v1.train.AdamOptimizer(learning_rate) train_op = trainer.minimize(loss, var_list=train_variables, global_step=batch) saver = tf.compat.v1.train.Saver() config = tf.compat.v1.ConfigProto() config.gpu_options.allow_growth = True config.allow_soft_placement = True config.log_device_placement = False # add by nishi #allow_soft_placement=True, log_device_placement=True sess = tf.compat.v1.Session(config=config) # write logs to the disk flog = open(os.path.join(LOG_STORAGE_PATH, 'log.txt'), 'w') # changed by nishi if START_EPOCH > 0: MODEL_PATH=os.path.join(MODEL_STORAGE_PATH, 'epoch_' + str(START_EPOCH-1)+'.ckpt') #MODEL_PATH="model_sav/model-%d.ckpt" % (START_EPOCH-1) saver.restore(sess, MODEL_PATH) #s="Model restored. %s" %(MODEL_PATH) printout(flog, "Model restored. %s" %(MODEL_PATH)) else: init = tf.compat.v1.global_variables_initializer() sess.run(init) train_writer = tf.compat.v1.summary.FileWriter(SUMMARIES_FOLDER + '/train', sess.graph) test_writer = tf.compat.v1.summary.FileWriter(SUMMARIES_FOLDER + '/test') train_file_list = provider.getDataFiles(TRAINING_FILE_LIST) num_train_file = len(train_file_list) test_file_list = provider.getDataFiles(TESTING_FILE_LIST) num_test_file = len(test_file_list) fcmd = open(os.path.join(LOG_STORAGE_PATH, 'cmd.txt'), 'w') fcmd.write(str(FLAGS)) fcmd.close() # write logs to the disk #flog = open(os.path.join(LOG_STORAGE_PATH, 'log.txt'), 'w') flog=None def train_one_epoch(train_file_idx, epoch_num): is_training = True for i in range(num_train_file): cur_train_filename = os.path.join(hdf5_data_dir, train_file_list[train_file_idx[i]]) printout(flog, 'Loading train file ' + cur_train_filename) # test data # cur_data -> point cloud data # cur_labels -> class # cur_seg -> segement data cur_data, cur_labels, cur_seg = provider.loadDataFile_with_seg(cur_train_filename) if False: cur_data, cur_labels, order = provider.shuffle_data(cur_data, np.squeeze(cur_labels)) cur_seg = cur_seg[order, ...] else: cur_data, cur_labels, order = provider.shuffle_data(cur_data, cur_labels) cur_seg = cur_seg[order, ...] cur_labels=np.squeeze(cur_labels) cur_labels_one_hot = convert_label_to_one_hot(cur_labels) num_data = len(cur_labels) num_batch = num_data // batch_size next_epoch.add_steps(num_batch) total_loss = 0.0 total_label_loss = 0.0 total_seg_loss = 0.0 total_label_acc = 0.0 total_seg_acc = 0.0 for j in range(num_batch): begidx = j * batch_size endidx = (j + 1) * batch_size feed_dict = { pointclouds_ph: cur_data[begidx: endidx, ...], # input pointclouds data labels_ph: cur_labels[begidx: endidx, ...], # class label input_label_ph: cur_labels_one_hot[begidx: endidx, ...], # class label one hot seg_ph: cur_seg[begidx: endidx, ...], # segument data is_training_ph: is_training, } _, loss_val, label_loss_val, seg_loss_val, per_instance_label_loss_val, \ per_instance_seg_loss_val, label_pred_val, seg_pred_val, pred_seg_res \ = sess.run([train_op, loss, label_loss, seg_loss, per_instance_label_loss, \ per_instance_seg_loss, labels_pred, seg_pred, per_instance_seg_pred_res], \ feed_dict=feed_dict) per_instance_part_acc = np.mean(pred_seg_res == cur_seg[begidx: endidx, ...], axis=1) average_part_acc = np.mean(per_instance_part_acc) total_loss += loss_val total_label_loss += label_loss_val total_seg_loss += seg_loss_val per_instance_label_pred = np.argmax(label_pred_val, axis=1) # changed by nishi --> get label_acc label_acc = np.mean(np.float32(per_instance_label_pred == cur_labels[begidx: endidx, ...])) total_label_acc += label_acc total_seg_acc += average_part_acc # add by nishi if j % 10 == 0: print('\rsteps ' + repr(j) +'/' + repr(num_batch) + ' || Loss val: %.4f ||' % (loss_val) + ' label_acc: %.4f' % (label_acc), end=' ') total_loss = total_loss * 1.0 / num_batch total_label_loss = total_label_loss * 1.0 / num_batch total_seg_loss = total_seg_loss * 1.0 / num_batch total_label_acc = total_label_acc * 1.0 / num_batch total_seg_acc = total_seg_acc * 1.0 / num_batch lr_sum, bn_decay_sum, batch_sum, train_loss_sum, train_label_acc_sum, \ train_label_loss_sum, train_seg_loss_sum, train_seg_acc_sum = sess.run(\ [lr_op, bn_decay_op, batch_op, total_train_loss_sum_op, label_train_acc_sum_op, \ label_train_loss_sum_op, seg_train_loss_sum_op, seg_train_acc_sum_op], \ feed_dict={total_training_loss_ph: total_loss, label_training_loss_ph: total_label_loss, \ seg_training_loss_ph: total_seg_loss, label_training_acc_ph: total_label_acc, \ seg_training_acc_ph: total_seg_acc}) pt = i + epoch_num * num_train_file train_writer.add_summary(train_loss_sum, pt) train_writer.add_summary(train_label_loss_sum, pt) train_writer.add_summary(train_seg_loss_sum, pt) train_writer.add_summary(lr_sum, pt) train_writer.add_summary(bn_decay_sum, pt) train_writer.add_summary(train_label_acc_sum, pt) train_writer.add_summary(train_seg_acc_sum, pt) train_writer.add_summary(batch_sum, pt) print('') printout(flog, 'Training Total Mean_loss: %f' % total_loss) printout(flog, '\tTraining Label Mean_loss: %f' % total_label_loss) printout(flog, '\tTraining Label Accuracy: %f' % total_label_acc) printout(flog, '\tTraining Seg Mean_loss: %f' % total_seg_loss) printout(flog, '\tTraining Seg Accuracy: %f' % total_seg_acc) # test by nishi #batch_x = sess.run(batch) #print('batch_x=',batch_x) #print('next_epoch.get_steps()=',next_epoch.get_steps()) def eval_one_epoch(epoch_num): is_training = False total_loss = 0.0 total_label_loss = 0.0 total_seg_loss = 0.0 total_label_acc = 0.0 total_seg_acc = 0.0 total_seen = 0 total_label_acc_per_cat = np.zeros((NUM_CATEGORIES)).astype(np.float32) total_seg_acc_per_cat = np.zeros((NUM_CATEGORIES)).astype(np.float32) total_seen_per_cat = np.zeros((NUM_CATEGORIES)).astype(np.int32) for i in range(num_test_file): cur_test_filename = os.path.join(hdf5_data_dir, test_file_list[i]) printout(flog, 'Loading test file ' + cur_test_filename) cur_data, cur_labels, cur_seg = provider.loadDataFile_with_seg(cur_test_filename) cur_labels = np.squeeze(cur_labels) cur_labels_one_hot = convert_label_to_one_hot(cur_labels) num_data = len(cur_labels) num_batch = num_data // batch_size for j in range(num_batch): begidx = j * batch_size endidx = (j + 1) * batch_size feed_dict = { pointclouds_ph: cur_data[begidx: endidx, ...], labels_ph: cur_labels[begidx: endidx, ...], input_label_ph: cur_labels_one_hot[begidx: endidx, ...], seg_ph: cur_seg[begidx: endidx, ...], is_training_ph: is_training, } loss_val, label_loss_val, seg_loss_val, per_instance_label_loss_val, \ per_instance_seg_loss_val, label_pred_val, seg_pred_val, pred_seg_res \ = sess.run([loss, label_loss, seg_loss, per_instance_label_loss, \ per_instance_seg_loss, labels_pred, seg_pred, per_instance_seg_pred_res], \ feed_dict=feed_dict) per_instance_part_acc = np.mean(pred_seg_res == cur_seg[begidx: endidx, ...], axis=1) average_part_acc = np.mean(per_instance_part_acc) total_seen += 1 total_loss += loss_val total_label_loss += label_loss_val total_seg_loss += seg_loss_val per_instance_label_pred = np.argmax(label_pred_val, axis=1) total_label_acc += np.mean(np.float32(per_instance_label_pred == cur_labels[begidx: endidx, ...])) total_seg_acc += average_part_acc for shape_idx in range(begidx, endidx): total_seen_per_cat[cur_labels[shape_idx]] += 1 total_label_acc_per_cat[cur_labels[shape_idx]] += np.int32(per_instance_label_pred[shape_idx-begidx] == cur_labels[shape_idx]) total_seg_acc_per_cat[cur_labels[shape_idx]] += per_instance_part_acc[shape_idx - begidx] total_loss = total_loss * 1.0 / total_seen total_label_loss = total_label_loss * 1.0 / total_seen total_seg_loss = total_seg_loss * 1.0 / total_seen total_label_acc = total_label_acc * 1.0 / total_seen total_seg_acc = total_seg_acc * 1.0 / total_seen test_loss_sum, test_label_acc_sum, test_label_loss_sum, test_seg_loss_sum, test_seg_acc_sum = sess.run(\ [total_test_loss_sum_op, label_test_acc_sum_op, label_test_loss_sum_op, seg_test_loss_sum_op, seg_test_acc_sum_op], \ feed_dict={total_testing_loss_ph: total_loss, label_testing_loss_ph: total_label_loss, \ seg_testing_loss_ph: total_seg_loss, label_testing_acc_ph: total_label_acc, seg_testing_acc_ph: total_seg_acc}) pt= (epoch_num+1) * num_train_file-1 test_writer.add_summary(test_loss_sum, pt) test_writer.add_summary(test_label_loss_sum, pt) test_writer.add_summary(test_seg_loss_sum, pt) test_writer.add_summary(test_label_acc_sum, pt) test_writer.add_summary(test_seg_acc_sum, pt) printout(flog, 'Testing Total Mean_loss: %f' % total_loss) printout(flog, '\tTesting Label Mean_loss: %f' % total_label_loss) printout(flog, '\tTesting Label Accuracy: %f' % total_label_acc) printout(flog, '\tTesting Seg Mean_loss: %f' % total_seg_loss) printout(flog, '\tTesting Seg Accuracy: %f' % total_seg_acc) for cat_idx in range(NUM_CATEGORIES): if total_seen_per_cat[cat_idx] > 0: printout(flog, '\n\tCategory %s Object Number: %d' % (all_obj_cats[cat_idx][0], total_seen_per_cat[cat_idx])) printout(flog, '\tCategory %s Label Accuracy: %f' % (all_obj_cats[cat_idx][0], total_label_acc_per_cat[cat_idx]/total_seen_per_cat[cat_idx])) printout(flog, '\tCategory %s Seg Accuracy: %f' % (all_obj_cats[cat_idx][0], total_seg_acc_per_cat[cat_idx]/total_seen_per_cat[cat_idx])) # add by nishi return total_label_acc for epoch in range(START_EPOCH,START_EPOCH+TRAINING_EPOCHES): printout(flog, '\n<<< Testing on the test dataset ...') eval_acc = eval_one_epoch(epoch) # add by nishi if acc_saver.check(eval_acc) == True: model_sav_path = "eval_acc_model-%d.ckpt" % epoch if not os.path.exists(os.path.join(MODEL_STORAGE_PATH, model_sav_path)): #save_path = saver.save(sess, os.path.join(MODEL_SAV_DIR, "model.ckpt")) save_path = saver.save(sess, os.path.join(MODEL_STORAGE_PATH, model_sav_path)) s="eval_acc Model saved in file: %s" % save_path printout(flog,s) printout(flog, '\n>>> Training for the epoch %d/%d ...' % (epoch, START_EPOCH+TRAINING_EPOCHES)) train_file_idx = np.arange(0, len(train_file_list)) np.random.shuffle(train_file_idx) train_one_epoch(train_file_idx, epoch) #if epoch % 10 == 0 and epoch != START_EPOCH: if epoch % 2 == 0 and epoch != START_EPOCH: #cp_filename = saver.save(sess, os.path.join(MODEL_STORAGE_PATH, 'epoch_' + str(epoch+1)+'.ckpt')) cp_filename = saver.save(sess, os.path.join(MODEL_STORAGE_PATH, 'epoch_' + str(epoch)+'.ckpt')) printout(flog, 'Successfully store the checkpoint model into ' + cp_filename) # add by nishi next_epoch.inc() if flog != None: flog.flush() #cp_filename = saver.save(sess, os.path.join(MODEL_STORAGE_PATH, 'epoch_' + str(epoch+1)+'.ckpt')) cp_filename = saver.save(sess, os.path.join(MODEL_STORAGE_PATH, 'epoch_' + str(epoch)+'.ckpt')) printout(flog, 'Successfully store the checkpoint model into ' + cp_filename) if flog != None: flog.flush() flog.close() if __name__=='__main__': train() # add by nishi acc_saver.save() next_epoch.save()

一度に、実行したいエポック数を、
parser.add_argument('--epoch', type=int, default=5, help='Epoch to run [default: 50]')
で指定して於けば、

後は、
python train.py
を実行するだけで、学習が継続できます。
後は、別ターミナルから、Tensorboard を起動すれば、経過観察できます。
> cd part_seg
> tensorboard --logdir=train_results\summaries

改めて言いますが、なかなか、 label_acc が上がってきません。
これで、本当によいのか?

batch の再設定が、間違っていたようなので、下記に改めてました。
それでも、 label_acc は、上がってこないみたいです。
# batch is equal with test steps. by nishi # changed by nishi if START_EPOCH > 0: batch = tf.Variable(next_epoch.get_steps(), trainable=False) else: batch = tf.Variable(0, trainable=False)

余りに、label_acc が上がって来ないので、pointnet_part_seg.py の get_loss() をいじってみました。
part_seg\pointnet_part_seg.py
... def get_loss(l_pred, seg_pred, label, seg, weight, end_points): ... #total_loss = weight * seg_loss + (1 - weight) * label_loss + mat_diff_loss * 1e-3 # test by nishi total_loss = weight * seg_loss + weight * label_loss + mat_diff_loss * 1e-3 return total_loss, label_loss, per_instance_label_loss, seg_loss, per_instance_seg_loss, per_instance_seg_pred_res

これで、 label_acc が、いきなり上がるようになりました。
これでは、いけないのだろうか?
seg_acc が、上がってこないのだろうか?
過学習になるのだろうか?

5 epoch で、 label_testing_acc が一番よくなるみたいだが、
これは、何かの誤りだろうか?

もしかしたら、先に、 seg_acc を改善させて、十分に、seg_acc が良くなった後で、
label_acc を上げてくる狙いあるのか?
ここは、我慢して、 epoch=250 まで一度試してみないといけないのかも?
気の短いおんちゃんには、大変ぞね。

test.py を見てみると、 epoch_190.ckpt をロードしているので、
190 epoch まで必要か?

オリジナルのままで、学習を続けてみます。
今日も、必死こ必死こ、学習させちょります。
只今、167 epoch ぞね。label_acc は上がる気配なし。 2021.1.15 by nishi

只今、195 epoch が終わった所ぞね。label_acc 上がる気配なし。
このプログラムは、いかさまか?
わざと、不具合を入れてあるのか?

model.get_los() のコメントに、記載がありました。
# model.py defines both classification net and segmentation net, which share the common global feature extractor network.
# In model.get_loss, we define the total loss to be weighted sum of the classification and segmentation losses.
# Here, we only train for segmentation network. Thus, we set weight to be 1.0. # loss function()

どうやら、ここでは、segmentation network の学習のみ行うので、weight=1.0 を、loss function() に渡す。との事でした。

オーマイガー! 何と言う事じゃ!! 全然コメントなんぞ、見てなかったぞね!!
と、言う訳で、weight < 1.0 を渡せば、label loss が計算されて、label acc が上がると言う事でした。

part_seg\train.py
... def train(): with tf.Graph().as_default(): with tf.device('/gpu:'+str(FLAGS.gpu)): ... # model.py defines both classification net and segmentation net, which share the common global feature extractor network. # In model.get_loss, we define the total loss to be weighted sum of the classification and segmentation losses. # Here, we only train for segmentation network. Thus, we set weight to be 1.0. # loss function() #wx= 1.0 -> orginal # set by nishi wx=0.99 loss, label_loss, per_instance_label_loss, seg_loss, per_instance_seg_loss, per_instance_seg_pred_res \ = model.get_loss(labels_pred, seg_pred, labels_ph, seg_ph, wx, end_points) .....

結論。
最初は、wx= 1.0 で、 seg_acc が上がるまで、80 epochs 位学習させて、
その後、wx= 0.9 にして、 label_acc を上げれば良い。
でした。
でも、最初から、 wx=0.99 で、OK ぞね。

やっと、解決しました。
この後は、part_seg\test.py を試してみるぞね!!