Unified AI model training platform
Systems, methods, apparatuses and non-transitory computer executable media configured to unify preprocessing, configuration, training, monitoring, and evaluation of multiple neural network based object detection algorithms under a singular development environment/platform (i.e., a “unified training platform”). The unified training platform may include a neural network agnostic model training environment that addresses the deficiencies described above and may allow for unified data annotation formatting. In addition to incorporating a wide variety of state-of-the-art neural networks into the unified training platform, the unified training platform may also provide full accessibility to available network optimizations. The present disclosure may also include a universal model converter.
1 . A neural network agnostic method for unified data annotation, the method comprising:
retrieving a dataset from one or more databases based on a configuration file, the dataset comprising image files extracted from one or more video feeds along with corresponding label files comprising information about one or more annotations added to the image files;
determining that one or more of the image files and one or more of the label files are in a format that is not compatible with a required format of neural network architecture specified in the configuration file,
reformatting the one or more of the image files and the one or more of the label files, such that the dataset is formatted for the neural network architecture;
selecting the neural network architecture from a repository based on the configuration file;
training a machine learning (ML) model based on the formatted dataset and one or more hyperparameters;
evaluating a performance of the ML model over one or more object detection metrics;
adjusting the one or more hyperparameters and iterating the training until the performance of the ML model meets a determined threshold; and
once the performance of the ML model meets the determined threshold, converting the ML model to a file format that is compatible with a production platform.
2 . The method of claim 1 , further comprising:
performing an integrity check on the dataset; and
removing any of image file and label file that fails the integrity check from the dataset.
3 . The method of claim 2 , wherein the integrity check comprises comparing the image files and the label files to confirm they match.
4 . The method of claim 2 , wherein the integrity check comprises determining whether any of the image files and the label files are one or more of corrupted, missing, or incorrectly formatted.
5 . The method of claim 1 , wherein the ML model comprises an initial model obtained from pretrained weights.
6 . The method of claim 5 , wherein the training the ML model comprises fine tuning the pretrained weights using the formatted dataset.
7 . The method of claim 1 , wherein the evaluating a performance of the ML model over one or more object detection metrics comprises:
comparing the one or more object detection metrics against metrics of other ML models of to determine a relative performance and internal ranking of the ML model.
8 . The method of claim 1 , wherein the one or more object detection metrics comprise true positives and mean average precisions.
9 . The method of claim 1 , wherein the converting the ML model to a file format that is compatible with a production platform comprises:
generating one or more frozen model graphs from the training the ML model;
converting the one or more frozen model graphs to supported model graph files compatible for use in a streaming platform for inference;
generating one or more final weight files from the supported model graph files; and
running one or more inferences over testing images using the one or more final weight files to evaluate an accuracy of the ML model.
10 . The method of claim 9 , wherein the one or more weights comprise eight bit integer (INT8), floating point 16 (FP16), and floating point 32 (FP32).
11 . A system configured to provide a neural network agnostic method for unified data annotation, the system comprising:
a processor operatively coupled to a memory configured to store computer readable code that, when executed by the processor, causes the processor to:
retrieve a dataset from one or more databases based on a configuration file, the dataset comprising image files extracted from one or more video feeds along with corresponding label files comprising information about one or more annotations added to the image files;
determine that one or more of the image files and one or more of the label files are in a format that is not compatible with a required format of neural network architecture specified in the configuration file,
reformat the one or more of the image files and the one or more of the label files, such that the dataset is formatted for the neural network architecture;
select the neural network architecture from a repository based on the configuration file;
train a machine learning (ML) model based on the formatted dataset and one or more hyperparameters;
evaluate a performance of the ML model over one or more object detection metrics;
adjust the one or more hyperparameters and iterating the training until the performance of the ML model meets a determined threshold; and
once the performance of the ML model meets the determined threshold, convert the ML model to a file format that is compatible with a production platform.
12 . The system of claim 11 , wherein the computer readable code, when executed by the processor, further causes the processor to:
perform an integrity check on the dataset; and
removing any of image file and label file that fails the integrity check from the dataset.
13 . The system of claim 12 , wherein the integrity check comprises comparing the image files and the label files to confirm they match.
14 . The system of claim 12 , wherein the integrity check comprises determining whether any of the image files and the label files are one or more of corrupted, missing, or incorrectly formatted.
15 . The system of claim 11 , wherein the ML model comprises an initial model obtained from pretrained weights.
16 . The system of claim 15 , wherein the training the ML model comprises fine tuning the pretrained weights using the formatted dataset.
17 . The system of claim 11 , wherein the evaluating a performance of the ML model over one or more object detection metrics comprises:
comparing the one or more object detection metrics against metrics of other ML models of to determine a relative performance and internal ranking of the ML model.
18 . The system of claim 11 , wherein the one or more object detection metrics comprise true positives and mean average precisions.
19 . The system of claim 11 , wherein the converting the ML model to a file format that is compatible with a production platform comprises:
generating one or more frozen model graphs from the training the ML model;
converting the one or more frozen model graphs to supported model graph files compatible for use in a streaming platform for inference;
generating one or more final weight files from the supported model graph files; and
running one or more inferences over testing images using the one or more final weight files to evaluate an accuracy of the ML model.
20 . The system of claim 19 , wherein the one or more weights comprise eight bit integer (INT8), floating point 16 (FP16), and floating point 32 (FP32).