Face recognition method facing face recognition training method of big data processing
A big data processing and face recognition technology, applied in the field of face recognition, can solve problems such as affecting model accuracy and losing face parameters.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0029] This embodiment 1 tests that using large data sets on the LFRTrainer system can effectively improve the model accuracy. Note the number of sample pairs in each node as (pos:neg), where pos is the number of positive sample pairs and neg is the number of negative sample pairs. According to the actual tuning experience: when the ratio of positive and negative sample pairs is 1:8 under the same conditions, the model accuracy is optimal. Therefore, this section selects three test sets for each node, namely (1Ok:80k), (20k:160k) and (30k:240k). Test results such as figure 1 shown.
[0030] from figure 1 It can be seen from the figure that: 1) All models trained on multi-nodes have higher accuracy than those on single nodes, because the total number of sample pairs increases with the number of nodes; 2) When using 16 nodes, each node The number of sample pairs in is (1Ok:80k), that is, when the total number of sample pairs is 1440k, the optimal accuracy of the trained mode...
Embodiment 2
[0032] This embodiment 2 tests the vertical acceleration performance of the distributed SVM trainer of LFRTrainer using multi-core technology. This test performs thread-level and instruction-level parallel optimization on the training process of the optimal model (that is, using 16 nodes, and the number of sample pairs in each node is (1Ok:8Ok)).
[0033] number of cores 1 2 4 8 12 24 overall speedup 1× 1.8× 3.0× 4.8× 5.4× 6.1× Computational speedup 1× 1.9× 3.4× 6.0× 7.1× 8.4×
[0034] Table 1 shows the multi-core speedup of the distributed SVM trainer. It can be seen from Table 1 that: 1) the use of multi-core technology can reduce the model training time, and the total training time on 24 cores is only one-fifth of that of single core; There is still a long way to go for linear acceleration, but since not all calculation processes can be parallelized, and both thread-level parallelism and instruction-level parallelism are fine-grain...
Embodiment 3
[0036]This embodiment 3 tests the horizontal scalability of the distributed SVM trainer of LFRTrainer on hundreds of nodes. Fix the number of sample pairs on each node, and increase the total number of sample pairs by increasing the number of nodes, such as figure 2 , image 3 shown.
[0037] from figure 2 with image 3 It can be seen from the figure that: 1) All the time spent on training on 8 nodes is nearly half less than that on a single node, indicating that the use of large data sets can speed up the speed of model convergence; 2) All training on 8-128 nodes The time spent is kept in a range of about 10% up and down, indicating that the distributed SVM trainer exhibits good linear scalability with the increase of the data set, and the up and down fluctuations are due to the interference of other tasks on the Endeavor cluster leading to communication overhead. There is an error in the measurement; 3) The percentage of all communication overheads in the total trainin...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More