User Guide
-----------
Pre-requisites: Anaconda, CUDA 8.0

INSTALLATION:
--------------
1. Create a virtual environment called 'spg' and activate it
    conda create --name spg python=2.7 pip
    source activate spg

2. Install Pytorch
    conda install pytorch=0.3 cuda80 torchvision -c pytorch

3. Install OpenCV
    conda install -c menpo opencv

4. Install Matplotlib
    conda install -c conda-forge matplotlib

5. (Optional) Install Ipython-notebook
    conda install ipython-notebook

6. Install scikit-video and ffmpeg to write videos
    conda install -c menpo ffmpeg
    pip install scikit-video

7. Download the pretrained model and place it under snapshots/

DEMO (Single Image)
-------------------
demo_image.py
Input arguments:
    img_dir     (str): the directory where the test image is placed
    save_dir    (str): the directory where the result is saved
    image_name  (str): the name of the test image
    input_size  (int): the resolution where the test image is resized to [Default: 321]
    num_classes (int): number of classes of the pre-trained model. For model trained on ImageNet, this should be 1000.
    snapshots   (str): the snapshot of the pretrained model [Default: 'snapshots/imagenet_epoch_2_glo_step_128118.pth.tar']
    top_k       (int): the top-K predictions to be saved [Default: 1]
    save_spg_c  (bool): save the prediction from SPG-C branch? [Default: True]

Example 1:
    python demo_image.py --image_name 'ILSVRC2012_val_00000004.JPEG'

Example 2: save top-5 predictions
    python demo_image.py --image_name 'ILSVRC2012_val_00000004.JPEG' --top_k 5

Example 3: Save SPG-C output 
    python demo_image.py --image_name 'ILSVRC2012_val_00000004.JPEG' --save_spg_c True

Example 4: Run with different input size
    python demo_image.py --image_name 'ILSVRC2012_val_00000004.JPEG' --top_k 5 --input_size 224

**************REMARKS***************
-------------------------------------
1. Use '' instead of False to turn off the arguments. For example, to turn off the save_spg_c option, run the following:
    python demo_image.py --image_name 'ILSVRC2012_val_00000004.JPEG' --save_spg_c ''


DEMO (Video)
-----------
demo_video.py
Input arguments:
    video_dir   (str): the directory where the sequence of test images is placed
    save_dir    (str): the directory where the result is saved
    video_name  (str): the name of the folder of the video sequence
    input_size  (int): the resolution where each test frame is resized to [Default: 321]
    num_classes (int): number of classes of the pre-trained model. For model trained on ImageNet, this should be 1000.
    snapshots   (str): the snapshot of the pretrained model [Default: 'snapshots/imagenet_epoch_2_glo_step_128118.pth.tar']
    heatmap_type(str): there are 2 types of heatmaps to save [Default: 'loc']
                    (1) loc: attention localization
                    (2) spg_c: the prediction from SPG-C branch
    include_ori (bool): show original video and heatmaps side-by-side [Default: True]
    high_res    (bool): save as higher resolution videos [Default: False]

Example 1:
    python demo_video.py --video_name dog

Example 2: Save the spg-c output instead
    python demo_video.py --video_name dog --heatmap_type spg_c

Example 3: Do not include the original videos when saving the outputs
    python demo_video.py --video_name dog --include_ori ''

Example 4: Save as higher resolution output
    python demo_video.py --video_name dog --high_res True