From 8ee64f30bdc229e15b5d03072e8c520b2c45c4bc Mon Sep 17 00:00:00 2001 From: smthemex <138738845+smthemex@users.noreply.github.com> Date: Fri, 6 Sep 2024 21:45:02 +0800 Subject: [PATCH] Update encoder_plus.py --- detail_encoder/encoder_plus.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/detail_encoder/encoder_plus.py b/detail_encoder/encoder_plus.py index 250fcdc..b922972 100644 --- a/detail_encoder/encoder_plus.py +++ b/detail_encoder/encoder_plus.py @@ -25,7 +25,7 @@ class detail_encoder(torch.nn.Module): self.dtype = dtype # load image encoder - clip_encoder = OriginalCLIPVisionModel.from_pretrained(image_encoder_path) + clip_encoder = OriginalCLIPVisionModel.from_pretrained(image_encoder_path,attn_implementation="eager") self.image_encoder = CLIPVisionModel(clip_encoder.config) state_dict = clip_encoder.state_dict() self.image_encoder.load_state_dict(state_dict, strict=False) @@ -73,7 +73,7 @@ class detail_encoder(torch.nn.Module): for pil in pil_image: tensor_image = self.clip_image_processor(images=pil, return_tensors="pt").pixel_values.to(self.device, dtype=self.dtype) clip_image.append(tensor_image) - clip_image = torch.cat(clip_image, dim=0) + clip_image = torch.cat(clip_image, dim=0) #torch.Size([1, 3, 224, 224]) # cond clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True)['hidden_states'][2::2] # 1 257*12 1024 @@ -81,7 +81,7 @@ class detail_encoder(torch.nn.Module): uncond_clip_image_embeds = self.image_encoder(torch.zeros_like(clip_image), output_hidden_states=True)['hidden_states'][2::2] uncond_clip_image_embeds = torch.cat(uncond_clip_image_embeds, dim=1) clip_image_embeds = self.resampler(clip_image_embeds) - uncond_clip_image_embeds = self.resampler(uncond_clip_image_embeds) + uncond_clip_image_embeds = self.resampler(uncond_clip_image_embeds) #torch.Size([1, 3084, 1024]) return clip_image_embeds, uncond_clip_image_embeds def generate( @@ -110,4 +110,4 @@ class detail_encoder(torch.nn.Module): **kwargs, ).images[0] - return image \ No newline at end of file + return image