Fast, fully automatic, data-centric out-of-the-box AI model building

1 . A computer-implemented method for fast training an artificial intelligence (AI) model, the method comprising:

accessing a training dataset comprising a plurality of labeled samples, each of the plurality of labeled samples labeled with a ground-truth label;

dividing the plurality of labeled samples into a training subset and a holdout test subset;

training the AI model using the training subset;

for each of a plurality of training epochs, checkpointing while training the AI model, wherein checkpointing comprises:

applying a partially trained AI model, which is partially trained from one or more prior training epochs, to the training subset to determine a train loss value; and

applying the partially trained AI model to the holdout test subset to determine a test loss value;

determining whether both train loss values and test loss values converge at a particular epoch;

responsive to determining that convergence occurred at the particular epoch, selecting the AI model trained at the particular epoch;

responsive to determining that convergence of train loss values and convergence of test loss values did not occur at a same epoch,

identifying an epoch corresponding to a lowest test loss value;

determining whether a difference between the train loss value and the test loss value at the epoch is below a margin;

responsive to determining that the difference between the train loss value and the test loss value is below the margin, selecting the AI model trained at the epoch for use; and

responsive to determining that the difference between the train loss value and the test loss value is greater than the margin, continuing on to a next training epoch.

2 . The computer-implemented method of claim 1 , wherein continuing on to the next training epoch comprises skipping one or more predetermined training epochs for checkpointing.

3 . The computer-implemented method of claim 1 , further comprising:

before the plurality of training epochs of checkpointing, training the AI model for a predetermined number of training epochs without checkpointing.

4 . The computer-implemented method of claim 1 , wherein the plurality of labeled samples comprise a plurality of images annotated with labels, and wherein the method further comprises:

analyzing the plurality of images to determine characteristics of the plurality of images; and

selecting an AI algorithm and a set of hyperparameters for training the AI model based in part on the characteristics of images.

5 . The computer-implemented method of claim 4 , wherein the characteristics of images comprise a characteristic indicative of a domain among a plurality of domains, the plurality of domains comprising an industrial domain, a medical domain, an agriculture domain, and a transportation domain.

6 . The computer-implemented method of claim 4 , wherein the set of hyperparameters comprises one or more of a batch size, a learning rate, and a training epoch number for initial training.

7 . The computer-implemented method of claim 4 , further comprising:

benchmarking a set of AI models against a plurality of benchmark training datasets, wherein benchmarking the set of AI models against the plurality of benchmark training datasets comprises logging a plurality of metrics of the set of AI models trained using the plurality of benchmark training datasets;

comparing the characteristics of the training dataset with characteristics of the plurality of benchmark training datasets to determine similarity; and

selecting an AI model from the set based in part on the similarity and the plurality of metrics of the set of AI models.

8 . The computer-implemented method of claim 7 , wherein selecting the AI model in the set comprises:

selecting a backbone network of the AI model.

9 . The computer-implemented method of claim 8 , wherein selecting the AI model in the set comprises freezing a portion of the selected backbone network.

10 . A computer system comprising:

one or more processors; and

a non-transitory computer-readable storage medium having instructions encoded thereon that, when executed by a processor, cause the one or more processors to:

access a training dataset comprising a plurality of labeled samples, each of the plurality of labeled samples labeled with a ground-truth label;

divide the plurality of labeled samples into a training subset and a holdout test subset;

train an AI model using the training subset;

for each of a plurality of training epochs, checkpoint while training the AI model, wherein checkpointing comprises:

apply a partially trained AI model, which is partially trained from one or more prior training epochs, to the training subset to determine a train loss value; and

apply the partially trained AI model to the holdout test subset to determine a test loss value;

determine whether both train loss values and test loss values converge at a particular epoch;

responsive to determining that convergence occurred at the particular epoch, select the AI model trained at the particular epoch;

responsive to determining that convergence of train loss values and convergence of test loss values did not occur at a same epoch,

identifying an epoch corresponding to a lowest test loss value;

determining whether a difference between the train loss value and the test loss value at the epoch is below a margin;

responsive to determining that the difference between the train loss value and the test loss value is below the margin, selecting the AI model trained at the epoch for use; and

responsive to determining that the difference between the train loss value and the test loss value is greater than the margin, continuing on to a next training epoch.

11 . The computer system of claim 10 , wherein continuing on to the next training epoch comprises skipping one or more predetermined training epochs for checkpointing.

12 . The computer system of claim 10 , the non-transitory computer-readable storage medium having additional instructions encoded thereon that, when executed by a processor, cause the one or more processors to:

before the plurality of training epochs of checkpointing, train the AI model for a predetermined number of training epochs without checkpointing.

13 . The computer system of claim 10 , wherein the plurality of labeled samples comprise a plurality of images annotated with labels, and wherein the non-transitory computer-readable storage medium having additional instructions encoded thereon that, when executed by a processor, cause the one or more processors to:

analyze the plurality of images to determine characteristics of the plurality of images; and

select an AI algorithm and a set of hyperparameters for training the AI model based in part on the characteristics of images.

14 . The computer system of claim 13 , wherein the characteristics of images comprise a characteristic indicative of a domain among a plurality of domains to which the plurality of images belong, the plurality of domains comprising an industrial domain, a medical domain, an agriculture domain, and a transportation domain.

15 . The computer system of claim 14 , the non-transitory computer-readable storage medium having additional instructions encoded thereon that, when executed by a processor, cause the one or more processors to:

benchmark a set of AI models against a plurality of benchmark training datasets, wherein benchmarking the set of AI models against the plurality of benchmark training datasets comprises logging a plurality of metrics of the set of AI models trained using the plurality of benchmark training datasets;

compare the characteristics of the training dataset with characteristics of the plurality of benchmark training datasets to determine similarity; and

select an AI model from the set based in part on the similarity and the plurality of metrics of the set of AI models.

16 . The computer system of claim 15 , wherein selecting the AI model in the set comprises:

selecting a backbone network of the AI model.

17 . The computer system of claim 15 , wherein the set of hyperparameters comprises one or more of a batch size, a learning rate, and a training epoch number for initial training.

18 . A non-transitory computer-readable storage medium having instructions encoded thereon that, when executed by a processor, cause one or more processors to:

access a training dataset comprising a plurality of labeled samples, each of the plurality of labeled samples labeled with a ground-truth label;

divide the plurality of labeled samples into a training subset and a holdout test subset;

train an AI model using the training subset;

for each of a plurality of training epochs, checkpoint while training the AI model, wherein checkpointing comprises:

apply a partially trained AI model, which is partially trained from one or more prior training epochs, to the training subset to determine a train loss value; and

apply the partially trained AI model to the holdout test subset to determine a test loss value;

determine whether both train loss values and test loss values converge at a particular epoch;

responsive to determining that convergence occurred at the particular epoch, select the AI model trained at the particular epoch;

responsive to determining that convergence of train loss values and convergence of test loss values did not occur at a same epoch,

identifying an epoch corresponding to a lowest test loss value;

determining whether a difference between the train loss value and the test loss value at the epoch is below a margin;

responsive to determining that the difference between the train loss value and the test loss value is below the margin, selecting the AI model trained at the epoch for use; and

responsive to determining that the difference between the train loss value and the test loss value is greater than the margin, continuing on to a next training epoch.