hossein1387 / u-net-fixed-point-quantization-for-medical-image-segmentation Goto Github PK

Hi,
Thanks for your good work. I have some questions when running the code:

The Parameter Size I obtained under all quantization settings is 18.48MB based on the EM dataset. According to your paper, The Parameter Size should be smaller than Full Precision. Could you please help me analyze the possible reasons? Thanks a lot!

For example, when I want to run UNET_EM_DATASET_a8_0_w0_8, the config.yaml is set to:
UNET:
dataset: 'emdataset'
lr: 0.001
num_epochs: 200
model_type: "unet"
init_type: glorot
quantization: "FIXED" # "INT", "BNN", "Normal", "FIXED"
activation_f_width: 0
activation_i_width: 6
weight_f_width: 4
weight_i_width: 0
gpu_core_num: 1
activation: "tanh"
trained_model: "./em_tanh_a6_0_w0_4/em_tanh_a6_0_w0_4.pkl"
experiment_name: "em_tanh_a6_0_w0_4"
log_output_dir: "./results/"
operation_mode: "normal" # normal, visualize, retrain, inference

Then, I run the code:
python em_unet.py -f config.yaml -t UNET

The environment I use:
Linux
Python 3.5.2
Nvidia 2080Ti GPU

Hi,
Thanks for the great work. I'm working on quantization for inference time speedup, but didn't seem to find any speed(up) reported in the paper. Is there any estimated value how much speedup (for inference) I can expect with an x memory reduction?

Thank you.

You have done a greate job.but I canont find you paper in the internet. Would you please share your paper? Thank you very much

hi,i appreciate the code you provided,but could you tell me what the method you did to show the model size after the quantification?i could't find the corresponding part in your code,thanks

Hi, I read your paper titled "U-Net Fixed-Point Quantization for Medical Image Segmentation" and find it very interesting. But I cannot figure out how you quantize the fractional part. Concretely, from equation (3), which is x_f = abs(x) − floor(abs(x)); x_i = floor(abs(x)), we know that x_f is always in [0, 1) and strictly smaller than 1.0. Thus, the quantization of x_f, i.e., quantize(x_f, n) = (round(clamp(x_f, n) << n)) >> n, will always be zero since you right-shift an integer smaller than 2^n by n bits. Hence, will the float part of Eq. (4) (the "to_float_point" function) always be zero? Am I misunderstanding your equation? Can you help me to figure it out?