Contributed by Meenakshi Sarkar, email meenakshisar@iisc.ac.in for any queries.
The main.py generates the vgg16 cosine similarity between #n_future of ground truth and predicted frames given by the data_path. Image frames can be of two types color and grayscale.
The directory ../data/color has 60 color images (30 ground truth frames given by suffix: gt_ and 30 predicted frames given by suffix: pred_)
with 3 color channels (RGB in that order, [64,64,3]). The directory ../data/grayscale has 60 grayscale images (30 ground truth frames given by suffix: gt_ and 30 predicted frames given by suffix: pred_) with 3 color channels (All the 3 channels have the same value since its a grayscale image,[64,64,3])
data_root= root directory for data, default='../data'.
log_dir= path to the folder where you save your plot and .npz file, default='../log'.
seed= namually fixing the seed value for reproducibility of the code and values.
datatype= color/grayscale, default='color'.
n_past= number of past frames after which vgg16 cimilarity index to be calculated for the #n_future of image frames, type=int, default=5.
n_future= number of future frames on which we calculate vgg ccosine similarity, type=int, default=10.
image_h= Frame height,type=int, dest="image_h",default=64
image_w= Frame width,type=int, dest="image_w",default=64
gpu= gpu id on server, type=str, default="1". n.d: If you do not have gpu access you can modify this argument and write cpu compatible code.
You can add any additional input arguments in the main program, but I am going to run your code only using these PREDEFINED SET OF ARGUMENTS during testing.
I will also test the code with values other than the default values during testing excpet for seed, image_h, image_w, gpu and data_root.
generates a vgg16_plot.png, and a vgg16_similarity.npz file in the log_dir. vgg16_plot.png shows the plot of vgg16 cosine similarity between the ground truth frames and predicted frames for the n_future no of frames. for example if n_past= 5 and n_future= 10, then for datatype= 'color', it would evaluate the vgg16 cosine similarity between ground truth and predicted frames, starting with gt_0005 and pred_0005 till gt_0014 and pred_0014 (total 10 timesteps in future). These 10 values then need to be plotted in vgg16_plot.png also saved in vgg16_similarity.npz
python main.py --log_dir path/to/log_dir --datatype grayscale --n_past 10 --n_future 20