DeepLabV3+ with various trunks supported Always stride8

DeepLabV3Plus-based mscale segmentation model

Resnet Global Module for Initialization

DeepLabV3+ with Attention-to-scale style attention Attn head: conv 3x3 512 ch relu conv 1x1 3 ch -> 1.0, 0.75, 0.5 train

Panoptic DeepLab-style semantic segmentation network stride8 only

Panoptic DeepLab-style semantic segmentation network stride8 only

Some of the code in this class is borrowed from: https://github.com/HRNet/HRNet-Semantic-Segmentation/tree/HRNet-OCR

This creates an HTML page of embedded images, useful for showing evaluation results. Usage: ip = ImagePage(html_fn) # Add a t

Sequeeze Excitation Module

Cross Entroply NLL Loss

Sampler that restricts data loading to a subset of the dataset. It is especially useful in conjunction with :class:`torch.nn.parallel.Distrib

Global average pooling over the input's spatial dimensions

Identity Residual Block for WideResnet

Take a random crop from the image. First the image or crop size may need to be adjusted if the incoming image is too small... If th

Main Squeeze Excitation Network Module

Modified Alighed Xception

DeepLabV3+ with Attention-to-scale style attention Attn head: conv 3x3 512 ch relu conv 1x1 3 ch -> 1.0, 0.75, 0.5 train

ASPP-based Mscale

ASPP-based Segmentation network

operations performed: 1x1 x depth 3x3 x depth dilation 6 3x3 x depth dilation 12 3x3 x depth dilation 18 image pool

Basic segmentation network, no ASPP, no Mscale

From: Searching for Efficient Multi-scale architectures for dense prediction

DeepLabV3 with various trunks supported

Image Weighted Cross Entropy Loss

Image dumping class You pass images/tensors from training pipeline into this object and it first converts them to images (doing tran

Relax Loss

OCR net

DeepLabV3Plus-based mscale segmentation model

OCR net

The basic implementation for object context block Input: N X C X H X W Parameters: in_channels : the dimension of t

RAdam optimizer

region mutual information I(A, B) = H(A) + H(B) - H(A, B) This version need a lot of memory if do not dwonsample.

Scale image such that longer side is == size

Aggregate the context features according to the initial predicted probability distribution. Employ the soft-weighted method t

Implementation of the OCR module: We aggregate the global object representation to update the representation for each pixel.

This is wider resnet 38, output_stride=8

Basic Block for Resnet

Base class for bottlenecks that implements `forward()` method.

Bottleneck Layer for Resnet

Randomly change the brightness, contrast and saturation of an image. Args: brightness (float): How much to jitter brightness. brightness_

Reference : https://github.com/quark0/darts/blob/master/cnn/utils.py

Flip around the x-axis

Lighting noise(AlexNet - style PCA - based noise)

Null Dataset for Performance

Multi-scale attention segmentation model base class

Multi-scale attention segmentation model base class

Implements data parallelism at the module level. This container parallelizes the application of the given module by splitting the input a

OCR net

Point Class For X and Y Location

Apply Bilateral Filtering

Apply Gaussian Blur

Boundary Relaxation

Resize image to exact size of crop

Bottleneck for SENet154.

ResNeXt bottleneck type C with a Squeeze-and-Excitation module.

ResNet bottleneck with a Squeeze-and-Excitation module. It follows Caffe implementation and uses `stride=stride` in `conv1` and not in `conv2

Scale image such that shorter side is == size

WideResnet Global Module for Initialization

Wider ResNet with pre-activation (identity mapping) blocks This variant uses down-sampling by max-pooling in the first two blocks and b

Scale image such that longer side is == size