accuracy drops a lot in fp16 mode

[itachi1232gg]

my model's accuracy drops a lot when I convert it into fp16 mode, even a pretrained resnet34 experienced accuracy drop in fp16 mode.

import os  
os.environ['CUDA_MODULE_LOADING'] = 'LAZY'    
from torchvision.models.resnet import resnet34, resnet18    
data = torch.ones((1,3,224,224)).cuda()    
model = resnet18(pretrained=True)    
model.eval()     
model.cuda()     
model_trt = torch2trt(model, [data], fp16_mode=True)     
with torch.no_grad():     
    print((model(data)-model_trt(data)).abs().sum())      

tensor(5.3151, device='cuda:0')

if I set fp16_mode=False
then the output is

tensor(0.0017, device='cuda:0')

[Lu-tju]

Hi, the result in my computer is 0.6877 with your code when set fp16_mode=False (and 5.3 when fp16_mode=True). I don't know is this error normal?

[itachi1232gg]

Hi, the result in my computer is 0.6877 with your code when set fp16_mode=False (and 5.3 when fp16_mode=True). I don't know is this error normal?

0.6877 is likely to give you the wrong outputs.

[JWLee89]

Instead of testing the absolute sum of differences between the two models, I believe an element-wise difference check to see if all the element-wise differences do not exceed a certain threshold might be a more accurate measure.

For example (did not test the code), we can check whether all elements inside of source and target tensor are within a certain absolute threshold.

output_pt = model(data)
output_trt = model_trt(data)

# Set this value to something that seems appropriate to you
# 1e-5 is generally reasonable
import numpy as np
absolute_tolerance = 1e-5
np.allclose(output_pt, output_trt, atol=absolute_tolerance)

[itachi1232gg]

Instead of testing the absolute sum of differences between the two models, I believe an element-wise difference check to see if all the element-wise differences do not exceed a certain threshold might be a more accurate measure.

For example (did not test the code), we can check whether all elements inside of source and target tensor are within a certain absolute threshold.

output_pt = model(data)
output_trt = model_trt(data)

# Set this value to something that seems appropriate to you
# 1e-5 is generally reasonable
import numpy as np
absolute_tolerance = 1e-5
np.allclose(output_pt, output_trt, atol=absolute_tolerance)

np.allclose(output_pt, output_trt, atol=absolute_tolerance) returns False