Exploring Segformer but its giving out Value error for input size, and expects to be 128x128

🤗Transformers
1

I was kinda exploring segformer :
error I am
ValueError: Target size (torch.Size([12, 1, 512, 512])) must be the same as input size (torch.Size([12, 1, 128, 128]))

My Model is :

class My_model(nn.Module):
    def __init__(self):
        super(My_model, self).__init__()
#         configuration = SegformerConfig.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")
#         configuration.num_labels = 1 ## set output as 1 
#         self.model = SegformerForSemanticSegmentation(config=configuration)

        self.model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512",
                                                         num_labels=1,)
    
    def forward(self, image):
        img_segs = self.model(image)
        
        return img_segs.logits

Error :

ValueError: Target size (torch.Size([12, 1, 512, 512])) must be the same as input size (torch.Size([12, 1, 128, 128]))

my input size → 512 my batch_size —>12 Image shape —> (3,512,512)
@nielsr

2

what I think is model(input).logits is H/4 and W/4 but what if I wanted mask of (B,C,H,W) . . . How do I upsample from here

3

Upsampling can be done using torch.nn.functional.interpolate (assuming image is a Pillow image):

from torch import nn

# rescale logits to original image size
logits = nn.functional.interpolate(outputs.logits.detach().cpu(),
                size=image.size[::-1], # (height, width)
                mode='bilinear',
                align_corners=False)

I’ve taken the code snippet above from my SegFormer inference notebook which you can find here.

4 
class my_model(nn.Module):
    def __init__(self):
        super(my_model, self).__init__()
#         configuration = SegformerConfig.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")
#         configuration.num_labels = 1 ## set output as 1 
#         self.model = SegformerForSemanticSegmentation(config=configuration)

        self.model = SegformerForSemanticSegmentation.from_pretrained("../input/nvidia-segformer-b2-finetuned-ade-512-512/segformer-b2-finetuned-ade-512-512",
                                                         num_labels=1,ignore_mismatched_sizes=True)
    
    def forward(self, image):
        img_segs = self.model(image)
    
        upsampled_logits = nn.functional.interpolate(img_segs.logits,
#                 size=(image.shape[0],1,image.shape[2],image.shape[3]), # (height, width)
                scale_factor=4,
                mode='bilinear',
                align_corners=False)
#         print(upsampled_logits.shape)
        upsampled_logits = torch.sigmoid(upsampled_logits)
        return upsampled_logits

Actually i did this… It worked for me .