NoiseInjector¶
Description¶
This module implements the function of introducing noise into the data tensor.
This operation can be used both for data augmentation, as well as to increase the noise immunity of the model when training on prepared intentionally noisy data.
Additional sources¶
Initializing¶
def __init__(self, mode="add", noisetype="uniform", params=(0.0, 1.0), rng=globalRng, inplace=False, slicing=None,
name=None):
Parameters
| Parameter | Allowed types | Description | Default |
|---|---|---|---|
| mode | str | Noise adding mode | "add" |
| noisetype | str | Noise type | "uniform" |
| params | tuple | Noise model parameters | (0.0, 1.0) |
| rng | str | Random number generator object (backend-dependent) | globalRng |
| inplace | bool | If True, then the output tensor will be written in the place of the input tensor in memory | False |
| slicing | slice | To which slice the operation is applied | None |
| name | str | Layer name | None |
Explanations
mode - noise adding mode. It can take values "add" and "mul";
noisetype - type of added noise. It can take values "gaussian" and "uniform";
rng - random number generator, which depends on the platform where the calculations are performed, i.e. the generator is backend-dependent: for example, when using Nvidia-GPU and CUDA, the generator is formed using the curand library;
params - params - tuple of noise parameters: (a, b) for "uniform" and (mean, sigma) for "gaussian";
inplace - flag showing whether additional memory resources should be allocated for the result. If True, then the output tensor will be written in the place of the input tensor in memory, which can negatively affect the network, if the input tensor takes part in calculations on other branches of the graph;
slicing - slice is set through the Python built-in slice object. Please have in mind that the slice indexes are set for the flattened data tensor;
Examples¶
Necessary imports.
import numpy as np
from PuzzleLib.Backend import gpuarray
from PuzzleLib.Modules import NoiseInjector
Info
gpuarray is required to properly place the tensor in the GPU
Take the normally distributed noise and the summing noise values mode with data values:
np.random.seed(123)
data = gpuarray.to_gpu(np.random.randint(0, 9, (1, 3, 4, 4)).astype(np.float32))
print(data)
[[[[2. 2. 6. 1.]
[3. 6. 1. 0.]
[1. 0. 0. 3.]
[4. 0. 0. 4.]]
[[1. 7. 3. 2.]
[4. 7. 2. 4.]
[8. 0. 7. 3.]
[4. 6. 1. 5.]]
[[6. 2. 1. 8.]
[3. 5. 0. 2.]
[6. 2. 4. 4.]
[6. 3. 0. 6.]]]]
injector = NoiseInjector(mode="add", noisetype="gaussian", params=(0.0, 10.0))
print(injector(data))
[[[[ -4.3309712 2.08201 16.573734 8.57326 ]
[ 21.06646 3.758148 14.312004 -6.5693135 ]
[ 9.087677 -9.566303 -6.01458 0.16788411]
[ -1.3159542 0.38866973 -11.592874 23.892897 ]]
[[ 3.4345775 8.709841 -4.4099264 11.415678 ]
[ 2.3037925 5.180489 5.9057817 6.890024 ]
[ -9.175585 16.199835 22.050224 9.048243 ]
[ 0.18314695 9.058003 0.78339946 1.8816617 ]]
[[ 13.431959 8.329308 -7.429866 20.178535 ]
[ 5.6269503 2.4493244 4.348013 11.406551 ]
[ 10.589058 21.149744 15.779724 2.0710073 ]
[ 8.479838 12.597048 11.60874 -19.87299 ]]]]
Let us change the distribution to uniform and choose the mode of multiplying noise values by data values; this time the noise will be introduced through slicing:
injector = NoiseInjector(mode="mul", noisetype="uniform", params=(0.0, 10.0), slicing = slice(1*2*3*2, 1*3*3*2))
print(injector(data))
[[[[ 2. 2. 6. 1. ]
[ 3. 6. 1. 0. ]
[ 1. 0. 0. 3. ]
[ 6.6883607 0. 0. 36.555717 ]]
[[ 9.848382 65.91567 3. 2. ]
[ 4. 7. 2. 4. ]
[ 8. 0. 7. 3. ]
[ 4. 6. 1. 5. ]]
[[ 6. 2. 1. 8. ]
[ 3. 5. 0. 2. ]
[ 6. 2. 4. 4. ]
[ 6. 3. 0. 6. ]]]]