scale of one of the image crops in multi_crop_batch seems to be off
nitred opened this issue · 1 comments
nitred commented
I'm using the gender classification script:
python guess.py --class_type gender --model_type inception --model_dir ./21936 --requested_step 14999 --filename ./test.jpg
In the make_multi_crop_batch function, a batch of cropped images is being created from the image file. Every image crop seems to be standardized except for the one that is horizontally flipped (i.e. flip_left_right) on line 152.
I printed out one pixel (3 channels) for each of the evaluated image crops. Here's how they look like:
standardized_original: [1.5681903 1.5681903 1.5681903]
flip_left_right: [236. 236. 236.] # scale is around 256, i.e. not standardized
standardized_crop1: [1.4424354 1.4424354 1.4424354]
standardized_crop1_flip: [-1.2171873 -1.2171873 -1.2171873]
standardized_crop2: [0.92570674 0.92570674 0.92570674]
standardized_crop2_flip: [1.5456676 1.5456676 1.5456676]
standardized_crop3: [1.455786 1.455786 1.455786]
standardized_crop3_flip: [-0.8199551 -0.8199551 -0.8199551]
standardized_crop4: [-0.29447666 -0.29447666 -0.29447666]
standardized_crop4_flip: [-0.48242104 -0.48242104 -0.48242104]
standardized_crop5: [1.4569467 1.4569467 1.4569467]
standardized_crop5_flip: [0.35454485 0.35454485 0.35454485]
I'm currently using a pre-trained model (checkpoint) that was provided in the readme. I have a couple of doubts on what the effects of this are?
- Is this expected behavior?
- Has the pre-trained model been trained using the
make_multi_crop_batch
? If so, how do you think this affects the performance of the pre-trained model? - How would you suggest I proceed if this isn't expected behavior? Should I now start standardizing the the
flip_left_right
image or should I drop it from the list of crops?