erprogs / genconvit Goto Github PK

I use the cuda 12.2, but it doesn't work. Then I try to use the torch 2.2.1(for cuda 12.2).

python train.py --d sample_train_data --m vae -e 5 -t y

I got this "RuntimeError: GET was unable to find an engine to execute this computation"

Hi, I am very interested in the 'Deepfake Video Detection Using Generative Convolutional Vision Transformer' research. Could you please provide the processed dataset? Thank you very much!!!

I used training scripts to get my VAE model: python train.py --d sample_train_data --m vae -e 5 -t y

Then I changed the path of the vae model in pred-func.py, but using the predicted script gave me an error：Missing key(s) in state_dict and Unexpected key(s).The structure of the two models is inconsistent.
I also found that the model size of the saved VAE was not consistent with the pre-trained model size you provided. The model size of my trained VAE became twice as large, about 5G in size.
I sincerely hope that you can help me, how to use their own trained model for inference.

do great work but why not publish paper of this work

Hello,
Thank you for the great end-to-end approach.

1. May I ask about the labels in this code (I am aware of the output but I'd like to clarify)

the first value of torch.sigmoid(model(df).squeeze()) is fake probability, right?
For example, after the sigmoid function we have tensor([[0.0468, 0.9539]], device='cuda:0'), then 0.0468 is the chance of the sample being fake, correct?

def pred_vid(df, model):
    with torch.no_grad():
        y_pred = model(df)
        return max_prediction_value(torch.sigmoid(model(df).squeeze()))

2. Also, the code is a bit counterintuitive... why we have to XOR?
   Since the prediction is 0, then the label then become 1, which is FAKE. Could you please explain.

def real_or_fake(prediction):
    return {0: "REAL", 1: "FAKE"}[prediction ^ 1]

Hi, Thank you for open sourcing your project. I have downloaded the provided checkpoints for both ed and vae and placed them inside the weight folder. However, I get the following errro:

raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for GenConViTED: Missing key(s) in state_dict: "backbone.patch_embed.backbone.layers.3.downsample.norm.weight", "backbone.patch_embed.backbone.layers.3.downsample.norm.bias", "backbone.patch_embed.backbone.layers.3.downsample.reduction.weight", "backbone.patch_embed.backbone.head.fc.weight", "backbone.patch_embed.backbone.head.fc.bias", "embedder.layers.3.downsample.norm.weight", "embedder.layers.3.downsample.norm.bias", "embedder.layers.3.downsample.reduction.weight", "embedder.head.fc.weight", "embedder.head.fc.bias". Unexpected key(s) in state_dict: "backbone.patch_embed.backbone.layers.0.downsample.norm.weight", "backbone.patch_embed.backbone.layers.0.downsample.norm.bias", "backbone.patch_embed.backbone.layers.0.downsample.reduction.weight", "backbone.patch_embed.backbone.layers.0.blocks.0.attn.relative_position_index", "backbone.patch_embed.backbone.layers.0.blocks.1.attn_mask", "backbone.patch_embed.backbone.layers.0.blocks.1.attn.relative_position_index", "backbone.patch_embed.backbone.layers.1.blocks.0.attn.relative_position_index", "backbone.patch_embed.backbone.layers.1.blocks.1.attn_mask", "backbone.patch_embed.backbone.layers.1.blocks.1.attn.relative_position_index", "backbone.patch_embed.backbone.layers.2.blocks.0.attn.relative_position_index", "backbone.patch_embed.backbone.layers.2.blocks.1.attn_mask", "backbone.patch_embed.backbone.layers.2.blocks.1.attn.relative_position_index", "backbone.patch_embed.backbone.layers.2.blocks.2.attn.relative_position_index", "backbone.patch_embed.backbone.layers.2.blocks.3.attn_mask", "backbone.patch_embed.backbone.layers.2.blocks.3.attn.relative_position_index", "backbone.patch_embed.backbone.layers.2.blocks.4.attn.relative_position_index", "backbone.patch_embed.backbone.layers.2.blocks.5.attn_mask", "backbone.patch_embed.backbone.layers.2.blocks.5.attn.relative_position_index", "backbone.patch_embed.backbone.layers.3.blocks.0.attn.relative_position_index", "backbone.patch_embed.backbone.layers.3.blocks.1.attn.relative_position_index", "backbone.patch_embed.backbone.head.weight", "backbone.patch_embed.backbone.head.bias", "embedder.layers.0.downsample.norm.weight", "embedder.layers.0.downsample.norm.bias", "embedder.layers.0.downsample.reduction.weight", "embedder.layers.0.blocks.0.attn.relative_position_index", "embedder.layers.0.blocks.1.attn_mask", "embedder.layers.0.blocks.1.attn.relative_position_index", "embedder.layers.1.blocks.0.attn.relative_position_index", "embedder.layers.1.blocks.1.attn_mask", "embedder.layers.1.blocks.1.attn.relative_position_index", "embedder.layers.2.blocks.0.attn.relative_position_index", "embedder.layers.2.blocks.1.attn_mask", "embedder.layers.2.blocks.1.attn.relative_position_index", "embedder.layers.2.blocks.2.attn.relative_position_index", "embedder.layers.2.blocks.3.attn_mask", "embedder.layers.2.blocks.3.attn.relative_position_index", "embedder.layers.2.blocks.4.attn.relative_position_index", "embedder.layers.2.blocks.5.attn_mask", "embedder.layers.2.blocks.5.attn.relative_position_index", "embedder.layers.3.blocks.0.attn.relative_position_index", "embedder.layers.3.blocks.1.attn.relative_position_index", "embedder.head.weight", "embedder.head.bias". size mismatch for backbone.patch_embed.backbone.layers.1.downsample.norm.weight: copying a param with shape torch.Size([768]) from checkpoint, the shape in current model is torch.Size([384]). size mismatch for backbone.patch_embed.backbone.layers.1.downsample.norm.bias: copying a param with shape torch.Size([768]) from checkpoint, the shape in current model is torch.Size([384]). size mismatch for backbone.patch_embed.backbone.layers.1.downsample.reduction.weight: copying a param with shape torch.Size([384, 768]) from checkpoint, the shape in current model is torch.Size([192, 384]). size mismatch for backbone.patch_embed.backbone.layers.2.downsample.norm.weight: copying a param with shape torch.Size([1536]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for backbone.patch_embed.backbone.layers.2.downsample.norm.bias: copying a param with shape torch.Size([1536]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for backbone.patch_embed.backbone.layers.2.downsample.reduction.weight: copying a param with shape torch.Size([768, 1536]) from checkpoint, the shape in current model is torch.Size([384, 768]). size mismatch for embedder.layers.1.downsample.norm.weight: copying a param with shape torch.Size([768]) from checkpoint, the shape in current model is torch.Size([384]). size mismatch for embedder.layers.1.downsample.norm.bias: copying a param with shape torch.Size([768]) from checkpoint, the shape in current model is torch.Size([384]). size mismatch for embedder.layers.1.downsample.reduction.weight: copying a param with shape torch.Size([384, 768]) from checkpoint, the shape in current model is torch.Size([192, 384]). size mismatch for embedder.layers.2.downsample.norm.weight: copying a param with shape torch.Size([1536]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for embedder.layers.2.downsample.norm.bias: copying a param with shape torch.Size([1536]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for embedder.layers.2.downsample.reduction.weight: copying a param with shape torch.Size([768, 1536]) from checkpoint, the shape in current model is torch.Size([384, 768]).