PyTorch extensions for fast R&D prototyping and Kaggle farming
All encoders & decoders & heads are now inherit from HasOutputFeaturesSpecification interface to query number of output channels and strides this module outputs.
New loss class QualityFocalLoss from https://arxiv.org/abs/2006.04388
New function pad_tensor_to_size -A generic padding function for N-dimensional tensors [B,C, ...] shape.
Added DropPath layer (aka DropConnect)
Pretrained weights for SegFormer backbones
first_class_background_init for initializing last output convolution/linear block to have zeros in weights and bias layer set to [logit(bg_prob), logit(1-bg_prob), ...]
New function instantiate_normalization_block to create normalization layer by name. This is used in some decoder layers / heads.
Improve numeric accuracy of focal_loss_with_logits function by explicitly disabling AMP autocast for this function and casting preds & targets to float32.
MultiscaleTTA now allows setting interpolation mode and align_corners for resizing input and predictions.
BinaryFocalLoss now has __repr__
name_for_stride now accepts stride argument to be None. In this case the function is noop and returns input argument name.
RandomSubsetDataset now takes optional weights argument to select samples with given probability.
get_collate_fn for RandomSubsetDataset is not correct and returns collate function instead of calling itSignature of decoders changes to require first argument input_spec to be of type FeatureMapsSpecification.
Rewritten BiFPN decoder to support arbitrary number of input feature maps, user-defined normalization & activation & BiFPN block.
Rewritten UNetDecoder to allow setting upsample block as string type
WeightedLoss and JointLoss classes has been removed. If your code was using these classes here they are - copy paste them to your project and live happily, but I strongly suggest to use modern deep learning frameworks that support defining losses from configuration files.
class WeightedLoss(_Loss):
"""Wrapper class around loss function that applies weighted with fixed factor.
This class helps to balance multiple losses if they have different scales
"""
def __init__(self, loss, weight=1.0):
super().__init__()
self.loss = loss
self.weight = weight
def forward(self, *input):
return self.loss(*input) * self.weight
class JointLoss(_Loss):
"""
Wrap two loss functions into one. This class computes a weighted sum of two losses.
"""
def __init__(self, first: nn.Module, second: nn.Module, first_weight=1.0, second_weight=1.0):
super().__init__()
self.first = WeightedLoss(first, first_weight)
self.second = WeightedLoss(second, second_weight)
def forward(self, *input):
return self.first(*input) + self.second(*input)
All catalyst-related callbacks are moved to fork of the Catalyst library
get_collate_for_dataset function to get collate fn if a dataset instance (argument) exposes get_collate_fn method. Works also for ConcatDataset.DatasetMeanStdCalculator supports dtype to specify accumulator type (float64 by default)ApplySoftmaxTo and ApplySigmoidTo modules that could lead to activations not applied to input when it was a stringfs.find_images_in_dir_recursive
utils.describe_outputs to return a human-friendly representation of complex (dict, nested list, etc) outputs to see shape, mean/std of each tensor.More MyPy fixes & type annotations
fs.find_subdirectories_in_dir to retrieve list of subdirectories (non-recursive) in the given directory.logodd averaging of TTA predictions and counterpart logodd_mean function.plot_confusion_matrix one can disable plotting scores in each cell using show_scores argument (True by default).freeze_model method now returns input module argument.This is the major release update of Pytorch Toolbelt. It's been a long time since the last update and there are many improvements & updates since 0.4.4:
pytorch_toolbelt.datasets.DatasetMeanStdCalculator to compute mean & std of the dataset that does not fit entirely in memory.BiFPNDecoder
SwinTransformer, SwinB, SwinL, SwinT, SwinS
broadcast_from_master function to distributed utils. This method allows scattering a tensor from the master node to all nodes.reduce_dict_sum to gather & concatenate dictionary of lists from all nodes in DDP.master_print as a drop-in replacement to print that prints to stdout only on the zero-rank node.match_bboxes and match_bboxes_hungarian. The first method uses scores of predicted bboxes and matches most confident predictions first, while the match_bboxes_hungarian matches bboxes to maximize overall IoU.set_manual_seed now sets random seed for Numpy.to_numpy now correctly works for None and all iterables (Not only tuple & list)dim argument to ApplySoftmaxTo to specify channel for softmax operator (default value is 1, which was hardcoded previously)ApplySigmoidTo now applies in-place sigmoid (Purely performance optimization)TileMerger now supports specifying a device (Torch semantics) for storing intermediate tensors of accumulated tiles.MultiscaleTTA now supports a model that returns a single Tensor (Key-Value outputs still works as before)balanced_binary_cross_entropy_with_logits and BalancedBCEWithLogitsLoss now supports ignore_index argument.BiTemperedLogisticLoss & BinaryBiTemperedLogisticLoss also got support of ignore_index argument.focal_loss_with_logits now also supports ignore_index. Computation of ignored values has been moved from BinaryFocalLoss to this function.timm. Now GenericTimmEncoder queries output strides & feature maps directly from the timm's encoder instance.use_incre_features argument to specify whether output feature maps should have an increased number of features.change_extension, read_rgb_image, read_image_as_is functions now supports Path as input argument. Return type (str) remains unchanged.count_parameters now accepts human_friendly argument to print parameters count in human-friendly form 21.1M instead 21123123.plot_confusion_matrix now has format_string argument (None by default) to specify custom format string for values in confusion matrix.RocAucMetricCallback for Catalyst got fix_nans argument to fix NaN outputs, which caused roc_auc to raise an exception and break the training.BestWorstMinerCallbac now additionally logs batch with NaN value in monitored metricVolumeSlicer and VolumeMerger. Designed similarly to ImageSlicer. Not you can run 3D segmentation on huge volumes without risk of OOM.BalancedBCEWithLogitsLoss)SelectByIndex helper module to pick named output of the model (For use in nn.Sequential)MobileNetV3Large, MobileNetV3Small from torchvision.timm package (HRNets, ResNetD, EfficientNetV2 and others).PyTorch-based implementation for bbox matching (match_bboxes) that supports both CPU/GPU matching using hungarian algorithm.ignore renamed to ignore_index in BinaryLovaszLoss class.fpn_channels argument in constructor of FPNSumDecoder and FPNCatDecoder to channels.argument in constructor ofHRNetSegmentationDecoder to 'channels.conv1x1 not set bias to zero by defaultEnsembler class not correctly works with torch.jit.tracing
sigmoid activation support to get_activation_block
timm (They named with prefix Timm)rgb_image_from_tensor now clip valuesEnsembler now supports arithmetic, geometric & harmonic averaging via reduction parameter.read_binary_mask
SegmentationDataset to support strided masks for deep supervisionRandomSubsetDataset and RandomSubsetWithMaskDataset to sample dataset based on some condition (E.g. sample only samples of particular class)As usual, more tests, better type annotations & comments