Very low accuracy results with FLOWERS17

Thanks for putting up this code. I'm just a bit confused as to why I'm getting vey low accuracy results from running your code. Any idea where it might be going wrong?

rank-1: 4.66%
rank-5: 28.19%

             precision    recall  f1-score   support

          0       0.09      0.13      0.11        23
          1       0.00      0.00      0.00        23
          2       0.00      0.00      0.00        27
          3       0.04      0.03      0.04        31
          4       0.04      0.03      0.04        29
          5       0.10      0.10      0.10        21
          6       0.06      0.04      0.05        25
          7       0.05      0.04      0.04        26
          8       0.00      0.00      0.00        23
          9       0.00      0.00      0.00        24
         10       0.07      0.05      0.06        21
         11       0.00      0.00      0.00        22
         12       0.08      0.10      0.09        20
         13       0.04      0.03      0.04        30
         14       0.09      0.09      0.09        22
         15       0.04      0.06      0.05        18
         16       0.12      0.13      0.12        23

My conf/config.json

{
        "model"                 : "vgg19",
        "weights"               : "imagenet",
        "include_top"           : false,

        "train_path"            : "/home/ubuntu/flower/flowers",
        "features_path"         : "output/flowers_17/vgg19/features.h5",
        "labels_path"           : "output/flowers_17/vgg19/labels.h5",
        "results"               : "output/flowers_17/vgg19/results.txt",
        "classifier_path"       : "output/flowers_17/vgg19/classifier.cpickle",

        "test_size"             : 0.30,
        "seed"                  : 9,
        "num_classes"           : 17
}

I've downloaded the FLOWERS17 dataset, and ran organize_flowers17.py. It all looks good:

ubuntu@ubuntu:~/flower/flowers$ ls
bluebell   coltsfoot  crocus    daisy      fritillary  lilyvalley  snowdrop   tigerlily  windflower
buttercup  cowslip    daffodil  dandelion  iris        pansy       sunflower  tulip

ubuntu@ubuntu:~/flower/flowers$ ls -l windflower/
total 3568
-rwxr-xr-x 1 ubuntu ubuntu  28427 Jul 23 23:32 image_0006.jpg
-rwxr-xr-x 1 ubuntu ubuntu  41580 Jul 23 23:32 image_0029.jpg
-rwxr-xr-x 1 ubuntu ubuntu  37322 Jul 23 23:32 image_0034.jpg
...

they all seem to have 80 images:

ubuntu@ubuntu:~/flower/flowers$ ls -1 windflower/ | wc -l
80
ubuntu@ubuntu:~/flower/flowers$ ls -1 daisy/ | wc -l
80

the hd5 files seem to be filled:

ubuntu@ubuntu:~/flower/output/flowers_17/vgg19$ ls -l
total 23448
-rw-rw-r-- 1 ubuntu ubuntu  1700188 Jul 24 02:23 classifier.cpickle
-rw-rw-r-- 1 ubuntu ubuntu 22284384 Jul 24 02:07 features.h5
-rw-rw-r-- 1 ubuntu ubuntu    13024 Jul 24 02:07 labels.h5
-rw-rw-r-- 1 ubuntu ubuntu     1040 Jul 24 02:23 results.txt

Interestingly, if I download your output/flowers_17/vgg19/features.h5 and, using that, run train.py I get your results. So it seems like the problem has something to do with extract_features.py

Since I'm using Linux, I had to change a few parts of the code to let me use unix-style paths (e.g., '/' vs '\'). But I don't really see why it should matter since it still seems to be picking up all the files.

Gogul09 / flower-recognition

Very low accuracy results with FLOWERS17 #2