Adjust color saturation of an image. Args: img (PIL Image): PIL Image to be adjusted. saturation_factor (float): How much to adj

Call this function in your script after you have finished setting all cfg values that are necessary (e.g., merging a config from a file, merging

Args: matrix, 4-D tensor, [N, C, M, M]. matrix must be a symmetric positive define matrix.

test for function batch_cholesky_inverse()

Batched inverse of lower triangular matrices Args: L : a lower triangular matrix Ref: https://www.pugetsystems.com/labs/hpc/PyTorch-for-Scient

calculate class-specific error masks

calculate class-agnostic error masks

Divide an image into tiles to help us cover classes that are spread out. tile_size: size of tile to distribute image_size: original image

Find cities that correspond to a given split of the data. We split the data such that a given city belongs to either train or val, but never

intermediate function to call pooled_class_centroid

For one image, calculate centroids for all classes present in image. item: image, image_name tile_size: num_classes: id2trainid:

Find coarse cities

Colorize the segmentation mask

NLL Relaxed Loss Implementation

Compute mean,recall and decay from per-frame evaluation. Calculates precision/recall for boundaries between foreground_mask and gt_mask using morp

Takes an image and detect the peaks usingthe local maximum filter. Returns a boolean mask of the peaks (i.e. 1 when the pixel

Turn off using coarse images in training

Do transformations to image and mask :returns: image, mask

Evaluate a single minibatch of images. * calculate metrics * dump images There are two primary multi-scale inference types:

Forward Pass through the each layer of SE network

Mapillary code for color map and class names Outputs ------- self.trainid_to_name self.color_mapping

Find image and segmentation mask files and return a list of tuples of them.

Get the criterion based on the loss function args: commandline arguments return: criterion, criterion_val

Fetch Network Function Pointer

Decide Optimizer (Adam or SGD)

Get a randomized transform to be applied on image. Arguments are same as that of __init__. Returns: Transform which rand

Args: matrix: matrix must be a positive define matrix. shape [N, C, D, D]. Ref: https://github.com/tensorflow/tensorflow/blob/r1.13/tensorf

AvgPool and Linear Layer

Main Function

Calculate tp, fp, fn for one image

Hierarchical attention, primarily used for getting best inference results. We use attention at multiple scales, giving prior

Hierarchical attention, primarily used for getting best inference results. We use attention at multiple scales, giving prior

Hierarchical attention, primarily used for getting best inference results. We use attention at multiple scales, giving prior

multi-scale evaluation for model with fused_aspp feature Evaluation must happen in two directions: from low to high to feed

Turn on using coarse images in training

Calculate class centroids for all classes for all images for all tiles. items: list of (image_fn, label_fn) tile size: size of tile r

Make output directories, setup logging, Tensorboard, snapshot code.

If single scale: just print results for default scale else print all scale results Inputs: hist = histogram for defaul

recursive eval for n-scales target resolution is fixed at 1.0 [0.5, 1.0]: p_0.5, aspp_0.5, attn_0.5 = fwd(attn,

Constructs a ResNet-101 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet

Constructs a ResNet-50 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet

Restore weights and optimizer (if needed ) for resuming job.

calculate the lower bound of the region mutual information. Args: labels_4D : [N, C, H, W], dtype=float32

Defination For SE Resnext101

Defination For SE Resnext50

Runs the training loop per epoch train_loader: Data loader for train net: thet network optimizer: optimizer curr_epoch: current e

Do we supervised both aux outputs, lo and high scale? Should attention be used to combine the aux output? Normally we only su

Find worse case failures ... Only single GPU for now First pass = calculate TP, FP, FN pixels per image per class Take these stat

write out tensorboard write out html webpage summary only update tensorboard if was a best epoch always update webpa