An utility that visualizes your convolutional neural network's "attention" on the region of input image, all operations can be done in a console command, supports frozen model(pb and saved_model format). This is a TensorFlow and Python implementation of Grad-CAM method(https://arxiv.org/abs/1610.02391).
Python 3.x with following packages:
- tensorflow 1.x (>= 1.6)
- pillow
- matplotlib
python grad_cam.py -h
usage: grad_cam.py [-h] [--model_path MODEL_PATH] [--print_tensors]
[--input_path INPUT_PATH] [--output_path OUTPUT_PATH]
[--input_width INPUT_WIDTH] [--input_height INPUT_HEIGHT]
[--preprocess PREPROCESS] [--class_index CLASS_INDEX]
[--input_tensor_name INPUT_TENSOR_NAME]
[--featuremap_tensor_name FEATUREMAP_TENSOR_NAME]
[--logits_tensor_name LOGITS_TENSOR_NAME]
[--prediction_tensor_name PREDICTION_TENSOR_NAME]
[--per_process_gpu_memory_fraction PER_PROCESS_GPU_MEMORY_FRACTION]
Produce Class Activation Maps (CAM) by Grad-CAM.
optional arguments:
-h, --help show this help message and exit
--model_path MODEL_PATH
a frozen model path, either pb or saved_model is
supported. NOTE that you MUST assign a directory path
which contains model files for save_model format
--print_tensors load model then print all tensors' name, it will do
nothing else
--input_path INPUT_PATH
an image file path or a directory path that contains
images (jpg/png/bmp are supported)
--output_path OUTPUT_PATH
a directory path for CAMs output, the default is the
same as input_path
--input_width INPUT_WIDTH
image width of model input
--input_height INPUT_HEIGHT
image height of model input
--preprocess PREPROCESS
vgg or inception, the default is do nothing for
preprocess
--class_index CLASS_INDEX
a class index you interested, the default is the most
probable class
--input_tensor_name INPUT_TENSOR_NAME
input tensor's name (it doesn't matter whether suffix
':0' is added)
--featuremap_tensor_name FEATUREMAP_TENSOR_NAME
feature map tensor's name (it doesn't matter whether
suffix ':0' is added)
--logits_tensor_name LOGITS_TENSOR_NAME
logits tensor's name (it doesn't matter whether suffix
':0' is added)
--prediction_tensor_name PREDICTION_TENSOR_NAME
(optional) prediction tensor's name (it doesn't matter
whether suffix ':0' is added)
--per_process_gpu_memory_fraction PER_PROCESS_GPU_MEMORY_FRACTION
(optional) fraction of the available GPU memory to
allocate for each process
There is a classification model zoo called TensorFlow-Slim, which is officially supported by TensorFlow team. I am going to test one of models to demonstrate the usage of the utility.
First of all, download and unzip mobilenet_v2_1.0_224.tgz (check out this page if the link is unavailable) into the the path TensorFlow_CAM/mobilenet_v2_1.0_224. Run with the following command:
python grad_cam.py \
--input_path=images/cat_dog.jpg \
--input_width=224 \
--input_height=224 \
--preprocess=inception \
--model_path=mobilenet_v2_1.0_224/mobilenet_v2_1.0_224_frozen.pb \
--input_tensor_name=input \
--featuremap_tensor_name=MobilenetV2/Conv_1/Relu6 \
--logits_tensor_name=MobilenetV2/Logits/Squeeze
The input is at the left, the ouput image file result1.png is at the right:
The region that is close to red represents high response to the class 'bull mastiff'(244).
This utility visualizes the most probable class by default, which was 'bull mastiff'(244) in the previous example.
For this case, just assigns your targeting class index to argument --class_index, so the complete command would be:
python grad_cam.py \
--input_path=images/cat_dog.jpg \
--input_width=224 \
--input_height=224 \
--preprocess=inception \
--class_index=283 \
--model_path=mobilenet_v2_1.0_224/mobilenet_v2_1.0_224_frozen.pb \
--input_tensor_name=input \
--featuremap_tensor_name=MobilenetV2/Conv_1/Relu6 \
--logits_tensor_name=MobilenetV2/Logits/Squeeze
The targeting class is 'tiger cat'(283), the output becomes:
If there are a bunch of images to be processed, you don't have to run the command many times.
Put them all under the same directory, then assign the directory path to argument --input_path. For instance, there are two images under TensorFlow_CAM/images, modify the command to:
python grad_cam.py \
--input_path=images \
--input_width=224 \
--input_height=224 \
--preprocess=inception \
--model_path=mobilenet_v2_1.0_224/mobilenet_v2_1.0_224_frozen.pb \
--input_tensor_name=input \
--featuremap_tensor_name=MobilenetV2/Conv_1/Relu6 \
--logits_tensor_name=MobilenetV2/Logits/Squeeze
The input images(at left) and the corresponding output images(at right) as follows:
'bull mastiff'(244)
'Border collie'(233)