segmentation_models.pytorch

segmentation_models.pytorch copied to clipboard

Mix ViT in_channels restriction

2

What's the reasoning behind limiting the Mix Visual Transformer encoder #632 to 3 input channels? https://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/encoders/mix_transformer.py#L468 I couldn't spot anything in the paper or the original SegFormer implementation.

chrislfathom

## Issue Description: **Code Location:** `metrics/functional.py` ### Problem: In the file `functional.py`, specifically in the the function get_stats, the line: ```python if torch.is_floating_point(output) and threshold is None: raise ValueError(f"Output should...

hxfdanger

Hi, is there any reference to modify this code for instance segmentation task? how hard it would be?

ramdhan1989

tracing mit-b1 on cpu and then running on gpu fails

1

Hi I traced the Unet model with mit-b1 backbone on cpu. I can load and convert it to gpu but once I run inference with it I get the following...

ghazalehtrb96

ModuleNotFoundError: No module named 'segmentation_models_pytorch.unetplusplus'

6

I was using an older version from segmentation models and everything was working fine, when i've upgraded to the latest version i got this error message ``` barcode_det_model_path = os.path.join("passport_barcode_det_UnetPlusPlus_resnext50_32x4d_e39_lv0.026_ltr0.014_ltest0.038_iv0.95_itr0.973_itest0.932.pth")...

ShroukMansour

preprocess_input = get_preprocessing_fn('resnet18', pretrained='imagenet') ... img = preprocess_input(img) What range of pixels should i put there? [0, 1] or [0,255]. Does it depends on specific preprocessing function? Or it's always...

isayoften

Bumps [timm](https://github.com/huggingface/pytorch-image-models) from 0.9.7 to 0.9.16. Release notes Sourced from timm's releases. Release v0.9.16 Feb 19, 2024 Next-ViT models added. Adapted from https://github.com/bytedance/Next-ViT HGNet and PP-HGNetV2 models added. Adapted from...

dependabot[bot]

Adding FeatUp

1

@qubvel happen to see [FeatUp](https://github.com/mhamilton723/FeatUp)? Might make for some interesting gains in model performance, where applicable.

Jordan-Pierce

Get score per class

1

When trying with predictions and targets of shape (batch_size, num_classes, image_height, image_width): ``` tp, fp, fn, tn = smp.metrics.get_stats(predictions, targets, mode='multiclass', num_classes=9) iou_score = smp.metrics.iou_score(tp, fp, fn, tn, reduction=None) recall...

tabikhm

Extracting features from the model

1

Hi, I trained a model for semantic segmentation with good results. Could I extract the feature vector from the encoder of my trained network in any way?

PierpaoloV