LAPiX DL - Utils for Computer Vision Deep Learning research

This package contains utilitary functions to support train and evaluation of Deep Learning models applied to images.

Three computer vision approaches are covered: Segmentation, Detection and Classification.

How to use

For Model Evaluation

This module exports the following functions for model evaluation:

from lapixdl.evaluation.evaluate import evaluate_segmentation
from lapixdl.evaluation.evaluate import evaluate_detection
from lapixdl.evaluation.evaluate import evaluate_classification

All model evaluation methods need two iterators: one for the ground truth itens and one for the predictions.

These iterators must be sorted equaly, assuring that the ground truth and the prediction of the same sample are at the same position.

Example of segmentation model evaluation using PyTorch:

from lapixdl.evaluation.evaluate import evaluate_segmentation

classes = ['background', 'object']

# Iterator for GT masks
# `dl` is a PyTorch DataLoader
def gt_mask_iterator_from_dl(dl):
  for imgs, masks in iter(dl):
    for mask in masks:
      yield mask

# Iterator for prediction masks
# `predict` a function that, given an image, predicts the mask.
def pred_mask_iterator_from_dl(dl, predict):
  for imgs, masks in iter(dl):
    for img in imgs:
      yield predict(img)

gt_masks = gt_mask_iterator_from_dl(validation_dl)
pred_masks = pred_mask_iterator_from_dl(validation_dl, prediction_function)

# Calculates and shows metrics
eval = evaluate_segmentation(gt_masks, pred_masks, classes)

# Shows confusion matrix and returns its Figure and Axes
fig, axes = eval.show_confusion_matrix()

Examples with third libraries

How to log the results of LAPiX DL evaluations in the Weights & Biases platform

About Weights & Biases.

from lapixdl.evaluation.evaluate import evaluate_segmentation
import wandb

# init wandb ...
...

eval_test = evaluate_segmentation(gt_masks, pred_masks, categories)

...

# If you want to log everything
wandb.log({'test_evaluation':  eval_test.to_dict()['By Class']})

# If you want to choose specific categories to log
selected_cats = ['A', 'B', 'C']
metrics_by_cat = {k: v for k, v in eval_test.to_dict()['By Class'].items() if k in selected_cats}
wandb.log({'test_evaluation': metrics_by_cat})

Computing using GPU with torchmetrics

About torchmetrics.

The lapixdl package calculates the confusion matrix first (on the CPU), which this will be slower than calculating using torchmetrics which uses pytorch tensors. So a trick here, to not calculate each metric separately in torchmetrics, is to calculate a confusion matrix using torchmetrics and then calculate all the metrics at once using lapixdl.

A simple example for a Segmentation case:

import torchmetrics
from lapixdl.evaluation.model import SegmentationMetrics

classes = ['background', 'object']

confMat = torchmetrics.ConfusionMatrix(
    reduce="macro", mdmc_reduce="global", num_classes=len(classes)
)

confusion_matrix = confMat(pred, target)
confusion_matrix = confusion_matrix.numpy()

metrics = SegmentationMetrics(
    classes=classes, confusion_matrix=confusion_matrix
)

For Results Visualization

This module exports the following functions for results visualization:

from lapixdl.evaluation.visualize import show_segmentations
from lapixdl.evaluation.visualize import show_classifications
from lapixdl.evaluation.visualize import show_detections

The available color maps are the ones from matplotlib.

For Data Conversion

This module exports the functions for data conversion.

from lapixdl.convert import labelbox_to_lapix
from lapixdl.convert import labelbox_to_coco

Example of conversion from Labelbox to COCO labels format:

import json

from lapixdl.formats import labelbox_to_coco

# A map categories between labelbox schematic id and category ID
map_categories = {
  '<schematic id from labelbox>': 1 # category id
}

# The categories section in the COCO format
categories_coco = [{
  'supercategory': None,
  'name': 'example_category',
  'id': 1
}]

# Convert it and create the COCO OD data
coco_dict = labelbox_to_coco(
  'labelbox_export_file.json',
  map_categories,
  categories_coco,
  target = 'object detection',
  image_shape = (1200, 1600)
)

# Saves converted json
with open('./coco.json', 'w') as out_file:
    json.dump(coco_dict, out_file)