MMPose Model Quantization

[gusmcarreira]

Hi there,

I am trying to speed up inference speed of my mmpose model, for such im using PyTorch's quantisation. The code is as follows:

from mmpose.apis import init_model
import torch

config = "config_path"
checkpoint = "checkpoint_path"

pose_model = init_model(config, checkpoint, device="cpu")

# set quantization config for server (x86)
# 'fbgemm' for server, 'qnnpack' for mobile
pose_model.qconfig = torch.quantization.get_default_qconfig('fbgemm')

# Prepare the model for quantization
pose_model = torch.quantization.prepare(pose_model, inplace=True)

# Calibrate the model with some representative input data
# NOTE: replace this with your own calibration code
# for data in calibration_dataset:
#     pose_model(data)

# Convert to a quantized version
pose_model = torch.quantization.convert(pose_model, inplace=True)

# Save the quantized model
torch.save(pose_model.state_dict(), 'quantized_model.pth')

But when I try to then perform inference with that model, the results are way different. Any help?

Kind regards