- Number of hidden layers = 1, number of neurons in hidden layer = 10.
- Tested for number of epochs = 100, batch size = 10, training sample size = 6000
- Here I limited to 100 epochs since the around 200 the process was killed (probably due to memory shortage).
- alpha used is :: 0.1
- Optimum Batch size came out to be 10
- The mininum epoch achieved was ::
0.0008333333
- The serial code is
neural_net.c - Time Taken :: 120.14 sec
- Accuracy :: 75 percent
- Grind rate :: 4994.17
-
The parallel code is
neural_net_omp.c -
Time Taken :: 30.8 sec
-
Accuracy :: 66 percent
-
Grind rate :: 19480.106
-
Since the major calculation portion was the matrix multiplications and other operations I parallelised all the matrix operations.
-
Apart from that I parallelised on places where I am looping through all the layers and not sharing resources for every input.
-
It was not possible according to my architecture to parallelize for
m for loopsince for that every input will have to have separate layer_a(activation layer array), layer_z(Z of layers) etc. -
The speed increased but the accuracy dropped I beleive its due to some race conditions while paralleslising which I could not debug.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent