Method and apparatus for training neural network and computer server

1. A method for training a neural network, comprising the following process performed at a predetermined time period:

selecting input data automatically to obtain a set of data to be annotated, wherein a video containing a plurality of sequences of frames is inputted to the neural network to obtain a target detection result for each frame of image, wherein target detection results for all frames of images in the video are inputted to a target tracking model to obtain a tracking result for each frame of image, and wherein the frame of image is determined as the input data in response to that the target detection result and the tracking result for a frame of image are inconsistent with each other;

annotating the set of data to be annotated to obtain a new set of annotated data;

acquiring a set of newly added annotated data containing the new set of annotated data;

determining a union of the set of newly added annotated data and a set of training sample data for training the neural network in a previous period as a set of training sample data for a current period;

training the neural network iteratively, based on the set of training sample data for the current period, to obtain a neural network trained in the current period, wherein the neural network has a plurality of particular structures each provided with a corresponding sparse scaling operator for scaling an output from the particular structure, wherein sparsity constraints having different weights are applied to the particular structures based on different computational complexities of the particular structures; and

training the neural network iteratively using a sample data in the set of training sample data for the current period comprises training the neural network in a number of training iterations by:

optimizing a target function using a first optimization algorithm, with sparse scaling operators obtained from a previous training iteration being constants of the target function and the weights being variables of the target function, to obtain weights of a current training iteration;

optimizing the target function using a second optimization algorithm, with the weights of the current training iteration being constants of the target function and sparse scaling operators being variables of the target function, to obtain sparse scaling operators of the current training iteration; and

performing a next training iteration based on the weights and sparse scaling operators of the current training iteration.

2. The method of claim 1 , wherein said acquiring the set of newly added annotated data containing the new set of annotated data comprises:

receiving the set of newly added annotated data from an annotation platform, or

retrieving the set of newly added annotated data from an annotation database.

3. The method of claim 1 , further comprising, subsequent to obtaining the neural network trained in the current period:

pruning the neural network trained in the current period.

4. The method of claim 1 , wherein said training the neural network iteratively, based on the set of training sample data for the current period comprises:

training iteratively and pruning the neural network based on the set of training sample data for the current period.

5. The method of claim 4 , wherein said training iteratively and pruning the neural network based on the set of training sample data for the current period comprises:

training iteratively weights of the neural network and the respective sparse scaling operators for the particular structures based on the set of training sample data for the current period, and

removing any particular structure having a sparse scaling operator of zero from the trained neural network.

6. The method of claim 5 , wherein said training iteratively the weights of the neural network and the respective sparse scaling operators for the particular structures based on the set of training sample data for the current period comprises

determining to stop the training when a number of training iterations has reached a threshold or when a target function associated with the neural network satisfies a predetermined convergence condition, wherein the target function comprises a loss function or a sparse regular function.

7. The method of claim 1 , wherein the target function is:

min

W

,

λ

[

1

N

⁢

∑

i

=

1

N

ℒ

⁡

(

(

x

i

,

W

,

λ

)

)

+

(

W

)

+

s

(

λ

)

]

where s (λ)=Σ j=1 K γ j ∥λ j ∥ 1 , and

W denotes the weights of the neural network, λ denotes a vector of sparse scaling operators of the neural network, N denotes a number of pieces of sample data in the set of training sample data for the current period, ( (x i , W, λ)) denotes a loss of the neural network over sample data x i , (W) denotes a weight regular function, s (λ) denotes a sparse regular function, K denotes a number of particular structures in the neural network, λ j denotes a sparse scaling operator for the j-th particular structure, and γ j denotes a sparse penalty weight corresponding to the j-th particular structure and is calculated based on a computational complexity of the j-th particular structure.

8. An apparatus for training a neural network, comprising:

a control unit configured to trigger a selecting unit, an annotating unit, an acquiring unit and a training unit at a predetermined time period;

the selecting unit configured to select input data automatically to obtain a set of data to be annotated, wherein a video containing a plurality of sequences of frames is inputted to the neural network to obtain a target detection result for each frame of image, wherein target detection results for all frames of images in the video are inputted to a target tracking model to obtain a tracking result for each frame of image, and wherein the frame of image is determined as the input data in response to that the target detection result and the tracking result for a frame of image are inconsistent with each other;

the annotating unit configured to annotate the set of data to be annotated to obtain a new set of annotated data;

the acquiring unit configured to:

acquire a set of newly added annotated data containing the new set of annotated data; and

determine a union of the set of newly added annotated data and a set of training sample data for training the neural network in a previous period as a set of training sample data for a current period; and

the training unit configured to train the neural network iteratively, based on the set of training sample data for the current period, to obtain a neural network trained in the current period, wherein the neural network has a plurality of particular structures each provided with a corresponding sparse scaling operator for scaling an output from the particular structure, wherein sparsity constraints having different weights are applied to the particular structures based on different computational complexities of the particular structures,

wherein training unit comprises a training sub-unit that is configured to train the neural network iteratively using a sample data in the set of training sample data for the current period, the training sub-unit further configured to train the neural network in a number of training iterations by:

optimizing a target function using a first optimization algorithm, with sparse scaling operators obtained from a previous training iteration being constants of the target function and the weights being variables of the target function, to obtain weights of a current training iteration;

optimizing the target function using a second optimization algorithm, with the weights of the current training iteration being constants of the target function and sparse scaling operators being variables of the target function, to obtain sparse scaling operators of the current training iteration; and

performing a next training iteration based on the weights and sparse scaling operators of the current training iteration.

9. The apparatus of claim 8 , wherein the acquiring unit being configured to acquire the set of newly added annotated data containing the new set of annotated data comprises the acquiring unit being configured to:

receive the set of newly added annotated data from an annotation platform, or

retrieve the set of newly added annotated data from an annotation database.

10. The apparatus of claim 8 , further comprising:

a pruning unit configured to prune the neural network trained in the current period.

11. The apparatus of claim 8 , wherein the training unit is configured to:

train iteratively and prune the neural network based on the set of training sample data for the current period, to obtain the neural network trained in the current period.

12. The apparatus of claim 8 , wherein the training unit comprises:

the training sub-unit that is configured to train iteratively weights of the neural network and the respective sparse scaling operators for the particular structures based on the set of training sample data for the current period, and trigger a removing sub-unit when the iterative training has completed; and

the removing sub-unit configured to remove any particular structure having a sparse scaling operator of zero from the neural network trained by the training sub-unit, to obtain the neural network trained in the current period.

13. The apparatus of claim 12 , wherein the training sub-unit is configured to

determine to stop the training when a number of training iterations has reached a threshold or when a target function associated with the neural network satisfies a predetermined convergence condition, wherein the target function comprises a loss function or a sparse regular function.

14. The apparatus of claim 8 , wherein the target function is:

min

W

,

λ

[

1

N

⁢

∑

i

=

1

N

ℒ

⁡

(

(

x

i

,

W

,

λ

)

)

+

(

W

)

+

s

(

λ

)

]

where s (λ)=Σ j=1 K γ j ∥λ j ∥ 1 , and

W denotes the weights of the neural network, λ denotes a vector of sparse scaling operators of the neural network, N denotes a number of pieces of sample data in the set of training sample data for the current period, ( (x i , W, λ)) denotes a loss of the neural network over sample data x i , (W) denotes a weight regular function, s (λ) denotes a sparse regular function, K denotes a number of particular structures in the neural network, λ j denotes a sparse scaling operator for the j-th particular structure, and γ j denotes a sparse penalty weight corresponding to the j-th particular structure and is calculated based on a computational complexity of the j-th particular structure.

15. A computer server, comprising a memory and one or more processors communicatively connected to the memory, the memory storing instructions executable by the one or more processors, the instructions, when executed by the one or more processors, causing the one or more processors to perform a method comprising:

selecting input data automatically to obtain a set of data to be annotated, wherein a video containing a plurality of sequences of frames is inputted to a neural network to obtain a target detection result for each frame of image, wherein target detection results for all frames of images in the video are inputted to a target tracking model to obtain a tracking result for each frame of image, and wherein the frame of image is determined as the input data in response to that the target detection result and the tracking result for a frame of image are inconsistent with each other;

annotating the set of data to be annotated to obtain a new set of annotated data;

acquiring a set of newly added annotated data containing the new set of annotated data, and determining a union of the set of newly added annotated data and a set of training sample data for training the neural network in a previous period as a set of training sample data for a current period;

training the neural network iteratively, based on the set of training sample data for the current period, to obtain a neural network trained in the current period, wherein the neural network has a plurality of particular structures each provided with a corresponding sparse scaling operator for scaling an output from the particular structure, wherein sparsity constraints having different weights are applied to the particular structures based on different computational complexities of the particular structures; and

training the neural network iteratively using a sample data in the set of training sample data for the current period comprises training the neural network in a number of training iterations by:

optimizing a target function using a first optimization algorithm, with sparse scaling operators obtained from a previous training iteration being constants of the target function and the weights being variables of the target function, to obtain weights of a current training iteration;

optimizing the target function using a second optimization algorithm, with the weights of the current training iteration being constants of the target function and sparse scaling operators being variables of the target function, to obtain sparse scaling operators of the current training iteration; and

performing a next training iteration based on the weights and sparse scaling operators of the current training iteration.

16. The computer server of claim 15 , wherein the acquiring the set of newly added annotated data containing the new set of annotated data comprises:

receiving the set of newly added annotated data from an annotation platform, or

retrieving the set of newly added annotated data from an annotation database.

17. The computer server of claim 15 , wherein the method further comprises, subsequent to obtaining the neural network trained in the current period:

pruning the neural network trained in the current period.

18. The computer server of claim 15 , wherein said training the neural network iteratively, based on the set of training sample data for the current period comprises:

training iteratively and pruning the neural network based on the set of training sample data for the current period.