Dependency: numpy, matplotlib, Tensorflow, cv2(For visualization)

A Stand-alone version for ImageRecognition project here

• Inception-v3 model for this project can be downloaded here
• Extract the zipped file and put Inception-v3 model (which should be renamed from 'tensorflow_inception_graph.pb' to 'Model.pb') to 'Models/Extractor/v3' folder
• Put your training set FOLDERS into '_Data' folder, please use English names for your folders to ensure that cv2 works correctly
  • Each folder name should be treated as the 'label' of the pictures contained in the folder
• Put your test set PICTURES into 'Test' folder
  • If possible, provide a ONE-HOT answer naming '_answer.npy' into 'Test' folder as well for better visualization
  • If you don't want to struggling for these, just leave 'Test' folder empty (Reference the Notice below)
• Run 'Main.py'!

• If 'Test' folder remains empty when the program is running, min(196, 0.2 * n_data)pictures will be MOVED from '_Data' folder to 'Test' folder if 'gen_test' FLAG is True
  • An '_answer.npy' ndarray will also be generated automatically!
• After processing all images in '_Data' folder, a '_Cache' folder which contains 'features.npy' and 'labels.npy' (shuffled) will be generated
  • If you want to train on new dataset, '_Cache' folder should be deleted
  • You can train your own classifier using 'features.npy' and 'labels.npy'
• After the program is done, a Predictor will be stored in 'Models/Predictors/v3' folder. If you want to train on new dataset, this folder should be deleted

--args:

parser.add_argument(

"--gen_test",
type=bool,
default=True,
help="Whether generate test images"

)

parser.add_argument(

"--images_dir",
type=str,
default="Test",
help="Path to test set"

)

parser.add_argument(

"--extract_only",
type=bool,
default=False,
help="Whether extract features only"

)

parser.add_argument(

"--visualize_only",
type=bool,
default=False,
help="Whether visualize only"

)

parser.add_argument(

"--overview",
type=bool,
default=True,
help="Whether use cv2 to overview"

)

(Not so elegant, but (maybe) better than nothing...)