Training

Training

class alonet.deformable_detr.train.LitDeformableDetr(args=None, model=None, **kwargs)

Bases: alonet.detr.train.LitDetr

static add_argparse_args(parent_parser, parser=None)

Add argument for Deformable DETR

Parameters

parent_parserArgumentParser

parser to be added new argument

parser[type], optional

parser to overwrite parent_parser if not None, by default None

Returns

ArgumentParser

Parser with arguments for Deformable DETR added

build_criterion(matcher, loss_label_weight=1, loss_boxes_weight=5, loss_giou_weight=2, losses=['labels', 'boxes'], aux_loss_stage=6, eos_coef=0.1)

Build criterion module to calculate losses

Parameters

matcherDeformableDetrHungarianMatcher

Hungarian matcher to match between predictions and targets

loss_label_weightint, optional

Weight of the classification loss, by default 1

loss_boxes_weightint, optional

Weight of the L1 loss of the bounding box coordinates, by default 5

loss_giou_weightint, optional

Weight of the giou loss of the bounding box, by default 2

losseslist, optional

Type of loss use in training, by default [‘labels’, ‘boxes’]

aux_loss_stageint, optional

Number of auxiliary decoder stages, by default 6

eos_coeffloat, optional

Relative classification weight applied to the no-object category, by default 0.1. This factor is applied only when softmax activation is used in model.

Returns

DeformableCriterion

rtype

DeformableCriterion ..

build_matcher(cost_class=1, cost_boxes=5, cost_giou=2)

Build matcher to match between predictions and targets

Parameters

cost_classint, optional

Weight of the classification error in the matching cost, by default 1

cost_boxesint, optional

Weight of the L1 error of the bounding box coordinates in the matching cost, by default 5

cost_giouint, optional

Weight of the giou loss of the bounding box in the matching cost, by default 2

Returns

DeformableDetrHungarianMatcher

rtype

DeformableDetrHungarianMatcher ..

build_model(num_classes=91, aux_loss=True, weights=None, activation_fn='sigmoid')

Build model for training

Parameters

num_classesint, optional

Number of classes to detect, by default 91

aux_lossbool, optional

aux_loss : bool, optional If True, the model will returns auxilary outputs at each decoder layer to calculate auxiliary decoding losses. By default True.

weightsstr, optional

Pretrained weights, by default None

activation_fnstr, optional

Activation function for classification head. Either “sigmoid” or “softmax”. By default “sigmoid”.

Returns

Union[DeformableDetrR50Refinement, DeformableDetrR50]

Deformable DETR for training

rtype

Union[DeformableDetrR50Refinement, DeformableDetrR50] ..

configure_optimizers()

Configure optimzier using AdamW

training: bool

Callbacks

Callbacks for Deformable DETR training is the same as DETR training

Criterion

class alonet.deformable_detr.criterion.DeformableCriterion(loss_label_weight, focal_alpha=0.25, **kwargs)

Bases: alonet.detr.criterion.DetrCriterion

This class computes the loss for Deformable DETR. The process happens in two steps

1. we compute hungarian assignment between ground truth boxes and the outputs of the model

2. we supervise each pair of matched ground-truth / prediction (supervise class and box)

This Criterion is rouhgly smilar to DetrCriterion except for the labels loss and the metrics.

get_metrics(outputs, frames, indices, num_boxes)

Compute some usefull metrics related to the model performance

Parameters

outputsdict

model forward outputs

framesaloscene.Frame

Target frame with ground truth boxes2d and labels

indiceslist

List of tuple with matching predicted indices and target indices. len(indices) is equal to batch size.

num_boxestorch.Tensor

Number of total target boxes

Returns

dict

• “recall” : Percentage of detected class among all the GT class

• “precision”: Among the positive prediction of the model. How much are well classify ?

Notes

Metrics are calculated based on threshold = 0.3 to filter out positive prediction. Positive prediction = prediction with score >= threshold, otherwise negative prediction This threshold value can be change in inference. That being said, it is still a usefull information to monitor the training progress.

Return type

dict

loss_labels(outputs, frames, indices, num_boxes, **kwargs)

Compute the clasification loss

Parameters

outputsdict

model forward outputs

framesaloscene.Frame

Target frame with ground truth boxes2d and labels

indiceslist

List of tuple with matching predicted indices and target indices. len(indices) is equal to batch size.

num_boxestorch.Tensor

Number of total target boxes

Returns

torch.Tensor

Classification loss

rtype

Tensor ..

training: bool

alonet.deformable_detr.criterion.sigmoid_focal_loss(inputs, targets, num_boxes, alpha=0.25, gamma=2)

Sigmoid focal loss for classification

Parameters

inputstorch.Tensor

The predictions for each example.

targetstorch.Tensor

A float tensor with the same shape as inputs. Stores the binary classification label for each element in inputs (0 for the negative class and 1 for the positive class).

num_boxestorch.Tensor

alphafloat, optional

(optional) Weighting factor in range (0,1) to balance positive vs negative examples. -1 for no weighting. By default 0.25.

gammafloat, optional

Exponent of the modulating factor (1 - p_t) to balance easy vs hard examples. By default 2

Returns

torch.Tensor

Scalar

rtype

Tensor ..

Matcher

class alonet.deformable_detr.matcher.DeformableDetrHungarianMatcher(cost_class=1, cost_boxes=1, cost_giou=1)

Bases: alonet.detr.matcher.DetrHungarianMatcher

hungarian_cost_class(tgt_boxes, m_outputs, **kwargs)

Compute the cost class for the Hungarian matcher

Parameters

m_outputs: dict

Dict output of the alonet.detr.Detr model. This is a dict that contains at least these entries:

• “pred_logits”: Tensor of dim [batch_size, num_queries, num_classes] with the classification logits

• “pred_boxes”: Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates

tgt_boxes: aloscene.BoundingBoxes2D

Target boxes2d across the batch

training: bool