Hi, I didnt find any code about adaface loss in your training source code for pretrained adaface AdaFace + CFSM model. Can you provide adaface loss source code for that model?
The readme for CFSM model is hard to understand in the training with custom dataset part, can you also provide the custom dataset config for it too?

Hi, your work is very interesting and enlightens me a lot.
The get_model (from backbones import get_model) function is not available in your released source code.

Thanks for the great work and comprehensive repository.
I tried to reproduce the results (FR part). Still, after training (FR backbone is being trained from scratch, and for the generator, I used the pre-trained weights of WiderFace from the repository), the results on Tinyface are lower than what is reported on the paper and also lower than the performance of the pre-trained R50 from the repository?
Am I missing something?

I appreciate any help you can provide.

When i use the following testing-generation script i am facing that issue.

import torch
from PIL import Image
import os
import torchvision.transforms as transforms
from synthesis_network import Generator # Import the Generator model from synthesis_network.py

Define paths

input_image_path = '/workspace/saransh/tinyface/Testing_Set/Gallery_Match/1_64.jpg'
model_path = 'model_4.pt'
output_dir = 'output'

Ensure the output directory exists

os.makedirs(output_dir, exist_ok=True)

Set device

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

Load the model

model = Generator()
model.load_state_dict(torch.load(model_path, map_location=device))
model.to(device)
model.eval()

Define image transformation

transform = transforms.Compose([
transforms.Resize((256, 256)), # Resize image to model input size
transforms.ToTensor(),
])

Load and preprocess input image

input_image = Image.open(input_image_path).convert('RGB')
input_tensor = transform(input_image).unsqueeze(0).to(device)

Generate output

with torch.no_grad():
output_tuple = model(input_tensor)
output_image = output_tuple[1].squeeze(0).cpu() # Access the synthesized image (out)

Save output image

output_image_path = os.path.join(output_dir, 'synthesized_image.jpg')
transforms.ToPILImage()(output_image).save(output_image_path)

print(f"Synthesized image saved at: {output_image_path}")

Missing key(s) in state_dict: "encoder.model.1.weight", "encoder.model.1.bias", "encoder.model.4.weight", "encoder.model.4.bias", "encoder.model.7.weight", "encoder.model.7.bias", "encoder.model.10.block.1.weight", "encoder.model.10.block.1.bias", "encoder.model.10.block.5.weight", "encoder.model.10.block.5.bias", "encoder.model.11.block.1.weight", "encoder.model.11.block.1.bias", "encoder.model.11.block.5.weight", "encoder.model.11.block.5.bias", "encoder.model.12.block.1.weight", "encoder.model.12.block.1.bias", "encoder.model.12.block.5.weight", "encoder.model.12.block.5.bias", "decoder.model.0.block.1.weight", "decoder.model.0.block.1.bias", "decoder.model.0.block.2.running_mean", "decoder.model.0.block.2.running_var", "decoder.model.0.block.5.weight", "decoder.model.0.block.5.bias", "decoder.model.0.block.6.running_mean", "decoder.model.0.block.6.running_var", "decoder.model.1.block.1.weight", "decoder.model.1.block.1.bias", "decoder.model.1.block.2.running_mean", "decoder.model.1.block.2.running_var", "decoder.model.1.block.5.weight", "decoder.model.1.block.5.bias", "decoder.model.1.block.6.running_mean", "decoder.model.1.block.6.running_var", "decoder.model.2.block.1.weight", "decoder.model.2.block.1.bias", "decoder.model.2.block.2.running_mean", "decoder.model.2.block.2.running_var", "decoder.model.2.block.5.weight", "decoder.model.2.block.5.bias", "decoder.model.2.block.6.running_mean", "decoder.model.2.block.6.running_var", "decoder.model.4.weight", "decoder.model.4.bias", "decoder.model.5.gamma", "decoder.model.5.beta", "decoder.model.8.weight", "decoder.model.8.bias", "decoder.model.9.gamma", "decoder.model.9.beta", "decoder.model.12.weight", "decoder.model.12.bias", "decoder.mlp.model.0.weight", "decoder.mlp.model.0.bias", "decoder.mlp.model.2.weight", "decoder.mlp.model.2.bias", "decoder.mlp.model.4.weight", "decoder.mlp.model.4.bias", "pca.U", "pca.mu".
Unexpected key(s) in state_dict: "conv1.weight", "bn1.weight", "bn1.bias", "bn1.running_mean", "bn1.running_var", "bn1.num_batches_tracked", "prelu.weight", "layer1.0.bn1.weight", "layer1.0.bn1.bias", "layer1.0.bn1.running_mean", "layer1.0.bn1.running_var", "layer1.0.bn1.num_batches_tracked", "layer1.0.conv1.weight", "layer1.0.bn2.weight", "layer1.0.bn2.bias", "layer1.0.bn2.running_mean", "layer1.0.bn2.running_var", "layer1.0.bn2.num_batches_tracked", "layer1.0.prel

Something wrong in your code?

`
File "/u02/thanhtm/tmp/CFSM/GuidedFaceRecognition/losses.py", line 91, in forward
m_arc.scatter_(1, label[index, None], 1.0)
RuntimeError: Index tensor must have the same number of dimensions as self tensor

`

Thank you for the great work.
I noticed that the training code of CFSM is failing with the input size 112 with the following error:
ValueError: Expected more than 1 spatial element when training, got input size torch.Size([32, 512, 1, 1])
Apparently, it's coming from the InstanceNorm2d https://github.com/liuf1990/CFSM/blob/main/CFSM/synthesis_network.py#L200

Any advice on how to solve it?