How did you convert the model to pytorch mobile?

[ramtiin]

Hi,

May I know how you converted the original python model to the mobile (lite) version?

[juanjaho]

Hi @ramtiin ,

Using the codes attached below: both are done by tracing the data.
- For ArcaneGAN model, I have converted .jit to .ptl
- As for the AnimeGAN model, .pt model is converted to .ptl
- https://github.com/bryandlee/animegan2-pytorch also has the detailed documentation on how TensorFlow weights were converted to .pt

convert_model.zip

.jit to .ptl (ArcaneGAN model)

import torch
from torch.utils.mobile_optimizer import optimize_for_mobile
from torchvision import transforms
from PIL import Image

model = torch.jit.load("model.jit", map_location="cpu").to('cpu').float().eval().cpu()
image = Image.open("test.jpg").convert("RGB")


means = [0.485, 0.456, 0.406]
stds = [0.229, 0.224, 0.225]

img_transforms = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(means, stds)])

with torch.no_grad():
    image = img_transforms(image)[None, ...].cpu()
    traced_script_module = torch.jit.trace(model, image)
    optimized_traced_model = optimize_for_mobile(traced_script_module)
    optimized_traced_model._save_for_lite_interpreter("model.ptl")

.pt to .ptl (AnimeGAN model)

import torch
from torch.utils.mobile_optimizer import optimize_for_mobile
from PIL import Image
from torchvision.transforms.functional import to_tensor, to_pil_image

from model import Generator

model = Generator()
model.load_state_dict(torch.load("model.pt", map_location="cpu"))
model.to('cpu').eval()

image = Image.open("test.jpg").convert("RGB")


with torch.no_grad():
    image = to_tensor(image).unsqueeze(0) * 2 - 1
    traced_script_module = torch.jit.trace(model, image)
    optimized_traced_model = optimize_for_mobile(traced_script_module)
    optimized_traced_model._save_for_lite_interpreter("model.ptl")

[ramtiin]

Thank you so much!