Big Big Bug find in tensorflow workspace!!



  • i use my own datesheet(5classification) ,i used the commmend as follow:

    python train_softmax.py
    --model_def models.mobilenet_v1
    --data_dir data/pic/
    --gpu_memory_fraction 0.85
    --gpus 0
    --image_size 224
    --logs_base_dir backup_classifier
    --models_base_dir backup_classifier
    --batch_size 64
    --epoch_size 100
    --learning_rate -1
    --max_nrof_epochs 50
    --class_num 5
    --use_fixed_image_standardization
    --optimizer MOM
    --learning_rate_schedule_file data/learning_rate.txt
    --keep_probability 1.0

    after a few epoch ,luckily the accuraury reach 97.9%

    but to my big surprise , when i use the command as the follow to test the model 's performance:
    python predict_one_pic.py backup_classifier\22 1_jpeg
    python predict_one_pic.py backup_classifier\1_0_jpeg
    thise two picture belongs to 2 different classifictions

    i print the sess's out

    they are the same ,maybe it's a bug

    ![0_1548599143578_J4VG}]3ZJP@L%D%UUU3~S)X.png](http://s3.cn-north-1.amazonaws.com.cn/kendryte-forum/uploads/489a9c5a-f481-4da3-8de7-32ae95a653cd.png)

    ![0_1548599149471_TF0QUYTTIGL)1GW_34ZH$D.png



  • 0_1548639576601_e2458157-fd56-4f2a-880c-eed728b39c53-image.png
    this is in the predict——one——pic.py



  • sorry this is the 4 classification model

    python train_softmax.py
    --model_def models.mobilenet_v1
    --data_dir data/pic/
    --gpu_memory_fraction 0.85
    --gpus 0
    --image_size 224
    --logs_base_dir backup_classifier
    --models_base_dir backup_classifier
    --batch_size 64
    --epoch_size 100
    --learning_rate -1
    --max_nrof_epochs 50
    --class_num 4
    --use_fixed_image_standardization
    --optimizer MOM
    --learning_rate_schedule_file data/learning_rate.txt
    --keep_probability 1.0



  • more detail :
    i full the image array with the “0” or“0.5” or“1”
    the predict out and logits find the are the same !!

    i add some code in train_softmax.py to dirctly to test the picture ,in order to avoid the bug maybe accured when saving or loading ckpt
    this is my add code ![0_1548599620313_78~H3S)[55C~RD~E1ODQXB.png

    the result ,with the training continue the accuracy become higer ,but two different picture's logit and predict are the same
    0_1548599782348_5HKZ[R(G]AYR3_M%FY`%9T4.png

    the two pictures, i choose from the trainset

    i tried to choose them from val set ,but the result are the same