Pooling Layers
SpatialMaxPooling
Scala:
val mp = SpatialMaxPooling(2, 2, dW=2, dH=2, padW=0, padH=0, format=DataFormat.NCHW)
Python:
mp = SpatialMaxPooling(2, 2, dw=2, dh=2, pad_w=0, pad_h=0, to_ceil=false, format="NCHW")
Applies 2D max-pooling operation in kWxkH regions by step size dWxdH steps. The number of output features is equal to the number of input planes. If the input image is a 3D tensor nInputPlane x height x width, the output image size will be nOutputPlane x oheight x owidth where
- owidth = op((width + 2*padW - kW) / dW + 1)
- oheight = op((height + 2*padH - kH) / dH + 1)
op is a rounding operator. By default, it is floor. It can be changed by calling ceil() or floor() methods.
As for padding, when padW and padH are both -1, we use a padding algorithm similar to the "SAME" padding of tensorflow. That is
outHeight = Math.ceil(inHeight.toFloat/strideH.toFloat)
outWidth = Math.ceil(inWidth.toFloat/strideW.toFloat)
padAlongHeight = Math.max(0, (outHeight - 1) * strideH + kernelH - inHeight)
padAlongWidth = Math.max(0, (outWidth - 1) * strideW + kernelW - inWidth)
padTop = padAlongHeight / 2
padLeft = padAlongWidth / 2
The format parameter is a string value (or DataFormat Object in Scala) of "NHWC" or "NCHW" to specify the input data format of this layer. In "NHWC" format data is stored in the order of [batch_size, height, width, channels], in "NCHW" format data is stored in the order of [batch_size, channels, height, width].
Scala example:
import com.intel.analytics.bigdl.nn.SpatialMaxPooling
import com.intel.analytics.bigdl.tensor.Tensor
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric.NumericFloat
val mp = SpatialMaxPooling(2, 2, 2, 2)
val input = Tensor(1, 3, 3)
input(Array(1, 1, 1)) = 0.5336726f
input(Array(1, 1, 2)) = 0.7963769f
input(Array(1, 1, 3)) = 0.5674766f
input(Array(1, 2, 1)) = 0.1803996f
input(Array(1, 2, 2)) = 0.2460861f
input(Array(1, 2, 3)) = 0.2295625f
input(Array(1, 3, 1)) = 0.3073633f
input(Array(1, 3, 2)) = 0.5973460f
input(Array(1, 3, 3)) = 0.4298954f
val gradOutput = Tensor(1, 1, 1)
gradOutput(Array(1, 1, 1)) = 0.023921491578221f
val output = mp.forward(input)
val gradInput = mp.backward(input, gradOutput)
println(output)
println(gradInput)
The output is,
(1,.,.) =
0.7963769
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x1x1]
The gradInput is,
(1,.,.) =
0.0 0.023921492 0.0
0.0 0.0 0.0
0.0 0.0 0.0
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x3x3]
Python example:
from bigdl.nn.layer import *
from bigdl.nn.criterion import *
from bigdl.optim.optimizer import *
from bigdl.util.common import *
mp = SpatialMaxPooling(2, 2, 2, 2)
input = np.array([0.5336726, 0.7963769, 0.5674766, 0.1803996, 0.2460861, 0.2295625, 0.3073633, 0.5973460, 0.4298954]).astype("float32")
input = input.reshape(1, 3, 3)
output = mp.forward(input)
print output
gradOutput = np.array([0.023921491578221]).astype("float32")
gradOutput = gradOutput.reshape(1, 1, 1)
gradInput = mp.backward(input, gradOutput)
print gradInput
The output is,
[array([[[ 0.79637688]]], dtype=float32)]
The gradInput is,
[array([[[ 0. , 0.02392149, 0. ],
[ 0. , 0. , 0. ],
[ 0. , 0. , 0. ]]], dtype=float32)]
SpatialAveragePooling
Scala:
val m = SpatialAveragePooling(kW, kH, dW=1, dH=1, padW=0, padH=0, globalPooling=false, ceilMode=false, countIncludePad=true, divide=true, format=DataFormat.NCHW)
Python:
m = SpatialAveragePooling(kw, kh, dw=1, dh=1, pad_w=0, pad_h=0, global_pooling=False, ceil_mode=False, count_include_pad=True, divide=True, format="NCHW")
SpatialAveragePooling is a module that applies 2D average-pooling operation in kWxkH regions by step size dWxdH.
The number of output features is equal to the number of input planes.
As for padding, when padW and padH are both -1, we use a padding algorithm similar to the "SAME" padding of tensorflow. That is
outHeight = Math.ceil(inHeight.toFloat/strideH.toFloat)
outWidth = Math.ceil(inWidth.toFloat/strideW.toFloat)
padAlongHeight = Math.max(0, (outHeight - 1) * strideH + kernelH - inHeight)
padAlongWidth = Math.max(0, (outWidth - 1) * strideW + kernelW - inWidth)
padTop = padAlongHeight / 2
padLeft = padAlongWidth / 2
The format parameter is a string value (or DataFormat Object in Scala) of "NHWC" or "NCHW" to specify the input data format of this layer. In "NHWC" format data is stored in the order of [batch_size, height, width, channels], in "NCHW" format data is stored in the order of [batch_size, channels, height, width].
Scala example:
scala>
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric.NumericFloat
import com.intel.analytics.bigdl.nn._
import com.intel.analytics.bigdl.tensor._
val input = Tensor(1, 3, 3).randn()
val m = SpatialAveragePooling(3, 2, 2, 1)
val output = m.forward(input)
val gradOut = Tensor(1, 2, 1).randn()
val gradIn = m.backward(input,gradOut)
scala> print(input)
(1,.,.) =
0.9916249 1.0299556 0.5608558
-0.1664829 1.5031902 0.48598626
0.37362042 -0.0966136 -1.4257964
[com.intel.analytics.bigdl.tensor.DenseTensor$mcF$sp of size 1x3x3]
scala> print(output)
(1,.,.) =
0.7341883
0.1123173
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x2x1]
scala> print(gradOut)
(1,.,.) =
-0.42837557
-1.5104272
[com.intel.analytics.bigdl.tensor.DenseTensor$mcF$sp of size 1x2x1]
scala> print(gradIn)
(1,.,.) =
-0.071395926 -0.071395926 -0.071395926
-0.3231338 -0.3231338 -0.3231338
-0.25173786 -0.25173786 -0.25173786
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x3x3]
Python example:
from bigdl.nn.layer import *
import numpy as np
input = np.random.randn(1,3,3)
print "input is :",input
m = SpatialAveragePooling(3,2,2,1)
out = m.forward(input)
print "output of m is :",out
grad_out = np.random.rand(1,3,1)
grad_in = m.backward(input,grad_out)
print "grad input of m is :",grad_in
produces output:
input is : [[[ 1.50602425 -0.92869054 -1.9393117 ]
[ 0.31447547 0.63450578 -0.92485516]
[-2.07858315 -0.05688643 0.73648798]]]
creating: createSpatialAveragePooling
output of m is : [array([[[-0.22297533],
[-0.22914261]]], dtype=float32)]
grad input of m is : [array([[[ 0.06282618, 0.06282618, 0.06282618],
[ 0.09333335, 0.09333335, 0.09333335],
[ 0.03050717, 0.03050717, 0.03050717]]], dtype=float32)]
VolumetricMaxPooling
Scala:
val layer = VolumetricMaxPooling(
kernelT, kernelW, kernelH,
strideT, strideW, strideH,
paddingT, paddingW, paddingH
)
Python:
layer = VolumetricMaxPooling(
kernelT, kernelW, kernelH,
strideT, strideW, strideH,
paddingT, paddingW, paddingH
)
Applies 3D max-pooling operation in kT x kW x kH regions by step size dT x dW x dH. The number of output features is equal to the number of input planes / dT. The input can optionally be padded with zeros. Padding should be smaller than half of kernel size. That is, padT < kT/2, padW < kW/2 and padH < kH/2
The input layout should be [batch, plane, time, height, width] or [plane, time, height, width]
Scala example:
import com.intel.analytics.bigdl.nn._
import com.intel.analytics.bigdl.utils.T
import com.intel.analytics.bigdl.tensor.Tensor
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric.NumericFloat
val layer = VolumetricMaxPooling(
2, 2, 2,
1, 1, 1,
0, 0, 0
)
val input = Tensor(T(T(
T(
T(1.0f, 2.0f, 3.0f),
T(4.0f, 5.0f, 6.0f),
T(7.0f, 8.0f, 9.0f)
),
T(
T(4.0f, 5.0f, 6.0f),
T(1.0f, 3.0f, 9.0f),
T(2.0f, 3.0f, 7.0f)
)
)))
layer.forward(input)
layer.backward(input, Tensor(T(T(T(
T(0.1f, 0.2f),
T(0.3f, 0.4f)
)))))
Its output should be
(1,1,.,.) =
5.0 9.0
8.0 9.0
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x1x2x2]
(1,1,.,.) =
0.0 0.0 0.0
0.0 0.1 0.0
0.0 0.3 0.4
(1,2,.,.) =
0.0 0.0 0.0
0.0 0.0 0.2
0.0 0.0 0.0
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x2x3x3]
Python example:
from bigdl.nn.layer import VolumetricMaxPooling
import numpy as np
layer = VolumetricMaxPooling(
2, 2, 2,
1, 1, 1,
0, 0, 0
)
input = np.array([[
[
[1.0, 2.0, 3.0],
[4.0, 5.0, 6.0],
[7.0, 8.0, 9.0]
],
[
[4.0, 5.0, 6.0],
[1.0, 3.0, 9.0],
[2.0, 3.0, 7.0]
]
]])
layer.forward(input)
layer.backward(input, np.array([[[
[0.1, 0.2],
[0.3, 0.4]
]]]))
Its output should be
array([[[[ 5., 9.],
[ 8., 9.]]]], dtype=float32)
array([[[[ 0. , 0. , 0. ],
[ 0. , 0.1 , 0. ],
[ 0. , 0.30000001, 0.40000001]],
[[ 0. , 0. , 0. ],
[ 0. , 0. , 0.2 ],
[ 0. , 0. , 0. ]]]], dtype=float32)
VolumetricAveragePooling
Scala:
val layer = VolumetricMaxPooling(
kT, kW, kH, dT, dW, dH,
padT=0, padW=0, padH=0,
countIncludePad=true, ceilMode=false
)
Python:
layer = VolumetricMaxPooling(
k_t, k_w, k_h, d_t, d_w, d_h
pad_t=0, pad_w=0, pad_h=0,
count_include_pad=True, ceil_mode=False
)
Applies 3D average-pooling operation in kernel kT x kW x kH regions by step size dT x dW x dH. The number of output features is equal to the number of input planes / dT. The input can optionally be padded with zeros. Padding should be smaller than half of kernel size. That is, padT < kT/2, padW < kW/2 and padH < kH/2
The input layout should be [batch, plane, time, height, width] or [plane, time, height, width]
By default, countIncludePad=true, which means to include padding when dividing the number of elements in pooling region. One can use ceilMode to control whether the output size is to be ceiled or floored.
Scala example:
import com.intel.analytics.bigdl.nn._
import com.intel.analytics.bigdl.utils.T
import com.intel.analytics.bigdl.tensor.Tensor
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric.NumericFloat
val layer = VolumetricAveragePooling(
2, 2, 2,
1, 1, 1,
0, 0, 0
)
val input = Tensor(T(T(
T(
T(1.0f, 2.0f, 3.0f),
T(4.0f, 5.0f, 6.0f),
T(7.0f, 8.0f, 9.0f)
),
T(
T(4.0f, 5.0f, 6.0f),
T(1.0f, 3.0f, 9.0f),
T(2.0f, 3.0f, 7.0f)
)
)))
layer.forward(input)
layer.backward(input, Tensor(T(T(T(
T(0.1f, 0.2f),
T(0.3f, 0.4f)
)))))
Its output should be
(1,1,.,.) =
3.125 4.875
4.125 6.25
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x1x2x2]
(1,1,.,.) =
0.0125 0.0375 0.025
0.05 0.125 0.075
0.0375 0.087500006 0.05
(1,2,.,.) =
0.0125 0.0375 0.025
0.05 0.125 0.075
0.0375 0.087500006 0.05
[com.intel.analytics.bigdl.tensor.DenseTensor of size 1x2x3x3]
Python example:
from bigdl.nn.layer import VolumetricAveragePooling
import numpy as np
layer = VolumetricAveragePooling(
2, 2, 2,
1, 1, 1,
0, 0, 0
)
input = np.array([[
[
[1.0, 2.0, 3.0],
[4.0, 5.0, 6.0],
[7.0, 8.0, 9.0]
],
[
[4.0, 5.0, 6.0],
[1.0, 3.0, 9.0],
[2.0, 3.0, 7.0]
]
]])
layer.forward(input)
layer.backward(input, np.array([[[
[0.1, 0.2],
[0.3, 0.4]
]]]))
Its output should be
array([[[[ 3.125 4.875]
[ 4.125 6.25 ]]]], dtype=float32)
array([[[[ 0.0125 0.0375 0.025 ]
[ 0.05 0.125 0.075 ]
[ 0.0375 0.08750001 0.05 ]]
[[ 0.0125 0.0375 0.025 ]
[ 0.05 0.125 0.075 ]
[ 0.0375 0.08750001 0.05 ]]]], dtype=float32)
RoiPooling
Scala:
val m = RoiPooling(pooled_w, pooled_h, spatial_scale)
Python:
m = RoiPooling(pooled_w, pooled_h, spatial_scale)
RoiPooling is a module that performs Region of Interest pooling.
It uses max pooling to convert the features inside any valid region of interest into a small feature map with a fixed spatial extent of pooledH × pooledW (e.g., 7 × 7).
An RoI is a rectangular window into a conv feature map. Each RoI is defined by a four-tuple (x1, y1, x2, y2) that specifies its top-left corner (x1, y1) and its bottom-right corner (x2, y2).
RoI max pooling works by dividing the h × w RoI window into an pooledH × pooledW grid of sub-windows of approximate size h/H × w/W and then max-pooling the values in each sub-window into the corresponding output grid cell. Pooling is applied independently to each feature map channel
forward accepts a table containing 2 tensors as input, the first tensor is the input image, the second tensor is the ROI regions. The dimension of the second tensor should be (*,5) (5 are batch_num, x1, y1, x2, y2).
Scala example:
scala>
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric.NumericFloat
import com.intel.analytics.bigdl.nn._
import com.intel.analytics.bigdl.tensor._
import com.intel.analytics.bigdl.tensor.Storage
val input_data = Tensor(2,2,6,8).randn()
val rois = Array(0, 0, 0, 7, 5, 1, 6, 2, 7, 5, 1, 3, 1, 6, 4, 0, 3, 3, 3, 3)
val input_rois = Tensor(Storage(rois.map(x => x.toFloat))).resize(4, 5)
val input = T(input_data,input_rois)
val m = RoiPooling(3, 2, 1)
val output = m.forward(input)
scala> print(input)
{
2: 0.0 0.0 0.0 7.0 5.0
1.0 6.0 2.0 7.0 5.0
1.0 3.0 1.0 6.0 4.0
0.0 3.0 3.0 3.0 3.0
[com.intel.analytics.bigdl.tensor.DenseTensor$mcF$sp of size 4x5]
1: (1,1,.,.) =
0.48066297 1.0994664 0.32474303 2.3391871 -0.79605865 0.836963950.36107457 1.2622415
0.657079 0.12720469 0.39894578 -0.41185552 -0.53111094 -0.36016005 -0.9726423 -2.5785272
0.3091435 -0.03613516 0.2375721 -1.1920663 -0.6757661 1.10612681.5409279 -0.17411499
0.23274016 -0.7149633 0.5473105 -0.40570387 -1.7966263 0.2071798-1.1530842 -0.010083453
-1.5769979 0.17043112 -0.28578365 -0.90779626 0.61457515 -0.1553582-0.3912479 -0.15326484
-0.24283029 1.3215472 1.3795123 -0.36933053 0.7077386 -0.56398267 0.6159163 0.5802894
(1,2,.,.) =
-1.1817129 -0.20470592 -1.3201113 0.36523122 -0.18260211 1.30210171.214403 1.1019816
0.7186407 0.78731173 1.5452348 0.0396181 0.5927014 1.17697431.0501136 -0.58295316
-0.96753055 0.6427254 -1.1396345 0.8701054 -0.22860864 -1.18719451.3372624 0.8616691
0.796831 -0.16609778 0.2950535 0.4595303 0.192339 0.6086106-0.76351887 -0.65964174
-0.12746814 -0.036058053 0.8858275 0.9677718 -1.1074747 -1.36859390.8783633 -0.11723315
-0.6947403 -0.23226547 -1.8510057 -1.3695518 -0.22317407 -0.36249024 -0.24097045 1.5691053
(2,1,.,.) =
0.84056973 1.144949 -1.0660682 0.4416162 -0.94440234 -0.24461010.91145027 -0.88650835
-0.81542057 0.14578317 -0.6531974 0.60776395 -0.32058007 -1.80771481.7660322 1.0680646
1.1328241 0.43677545 -0.9402618 -1.3002211 0.26012567 1.69481340.37849447 0.39286092
1.9443163 0.5415504 1.0793099 1.3312546 0.48346 1.2019655 0.3718734 0.21091922
0.5499047 1.6418253 0.8064177 0.37626198 0.8736181 -0.40816033 -0.5806787 1.286581
-0.5904657 -0.21188398 -0.040509004 1.2989452 1.6827602 1.3229258-0.68433124 0.87974
(2,2,.,.) =
-0.09759476 -0.32767114 0.16223079 2.3114302 -0.48496276 1.19290720.8572289 0.43429425
-1.0245247 0.19002944 1.5659521 -1.3689835 -1.4437296 -0.38216656 0.6333655 -0.57124794
-0.31111157 1.5184602 -1.3835855 -0.9295573 2.244521 -1.11849820.5451996 -0.4441631
-1.534093 -0.5599659 1.1980947 -1.0140935 1.3288999 0.19487387-0.1261734 -1.2222558
-0.070535585 0.9047848 -0.6719811 -1.6532638 -0.5290511 -0.18300447 0.69385433 0.018756092
0.24767837 0.620484 -0.5346291 1.0685066 -0.36903372 -0.26955062 1.1042496 0.5944603
[com.intel.analytics.bigdl.tensor.DenseTensor$mcF$sp of size 2x2x6x8]
}
scala> print(output)
(1,1,.,.) =
1.0994664 2.3391871 1.5409279
1.3795123 1.3795123 0.6159163
(1,2,.,.) =
1.5452348 1.5452348 1.3372624
0.8858275 0.9677718 1.5691053
(2,1,.,.) =
0.37849447 0.39286092 0.39286092
-0.5806787 1.286581 1.286581
(2,2,.,.) =
0.5451996 0.5451996 -0.4441631
1.1042496 1.1042496 0.5944603
(3,1,.,.) =
0.60776395 1.6948134 1.7660322
1.3312546 1.2019655 1.2019655
(3,2,.,.) =
2.244521 2.244521 0.6333655
1.3288999 1.3288999 0.69385433
(4,1,.,.) =
-0.40570387 -0.40570387 -0.40570387
-0.40570387 -0.40570387 -0.40570387
(4,2,.,.) =
0.4595303 0.4595303 0.4595303
0.4595303 0.4595303 0.4595303
[com.intel.analytics.bigdl.tensor.DenseTensor of size 4x2x2x3]
Python example:
from bigdl.nn.layer import *
import numpy as np
input_data = np.random.randn(2,2,6,8)
input_rois = np.array([0, 0, 0, 7, 5, 1, 6, 2, 7, 5, 1, 3, 1, 6, 4, 0, 3, 3, 3, 3],dtype='float64').reshape(4,5)
print "input is :",[input_data, input_rois]
m = RoiPooling(3,2,1.0)
out = m.forward([input_data,input_rois])
print "output of m is :",out
produces output:
input is : [array([[[[ 0.08500103, 0.33421796, 0.29084699, 1.60344635, -0.24289341,
-0.4793888 , 0.09452426, 0.16842477],
[-1.18575497, -0.53337542, 0.11661001, 0.9647904 , -0.25187936,
0.36516823, -0.16647209, -0.08095158],
[ 1.1982232 , -0.33549174, 0.11721347, -0.29319686, -0.01290122,
0.12344296, 0.30074829, -2.34951463],
[-0.60470899, -0.84657051, 0.1269276 , -0.06152321, -1.68838416,
-0.69808296, -2.06112892, -1.44790449],
[ 1.03944288, 0.13871728, 0.91478479, 0.47517105, 1.24638374,
0.98666841, 0.49403488, 1.26101127],
[-1.03949343, -0.39291108, 1.39107512, 1.73779253, 0.91656129,
0.103381 , 0.956243 , 0.44743548]],
[[ 0.79028054, 0.64244228, -0.37997334, -0.09130215, -2.3903429 ,
0.71919208, -0.14079786, 0.98304272],
[ 1.14678457, 1.58825227, 0.17137367, -0.62121819, -0.36103113,
-0.04452576, -0.0886136 , -1.32884721],
[ 0.06728957, -0.29701304, -0.52754207, -1.5785875 , 1.47354834,
-0.28545156, 0.49874194, 0.10277613],
[-0.10117571, -1.34902427, -1.40789327, 0.09853599, 0.60420022,
0.54869115, -0.49067696, 0.26696793],
[ 1.11780279, -0.77929016, 1.13772094, 0.14374057, 0.33199688,
-0.54057374, -0.45718861, 1.1577623 ],
[-1.4005645 , 1.15870496, 0.39292003, 0.88379515, 0.06440974,
0.65013063, 0.03759244, 0.18730126]]],
[[[-2.28272906, 0.06056305, 0.73632597, 0.10063274, -1.27497525,
-0.95597581, -0.22745785, 0.40146498],
[-1.37783475, 1.66000653, -1.80071745, -0.11805539, -0.27160583,
0.30691418, 2.62243232, -1.95274516],
[ 1.61364148, 1.91470546, -1.51984424, 2.13598224, -0.23156685,
-0.74203698, 0.65316888, 0.08018846],
[-1.8912854 , -0.50106158, 0.94937966, -0.10930541, 0.82136627,
-1.33209063, 1.43371302, -1.36370916],
[-0.52737928, -0.0681305 , -0.63472587, 0.41979229, -0.57093624,
-0.15968764, -1.005951 , -2.06873572],
[-2.34089346, 1.02593977, 0.90183415, 0.09504819, 0.53185448,
1.11305345, 1.290016 , -1.76216646]],
[[-0.10885459, -0.57089742, -0.55340708, -1.94445884, 1.30130049,
0.6333372 , -1.03100083, 0.0111167 ],
[ 0.59678149, -0.67601521, -1.25288718, -0.10922251, 3.06808996,
-1.46701513, -0.42140765, 1.12485412],
[ 1.21301567, -1.43304957, -0.56047239, 0.20716087, 1.40737646,
-0.08386437, -0.21916043, 0.85692906],
[ 1.59992399, -1.37044315, -0.71884386, 2.61830979, -0.74305496,
-0.32021174, 1.43275058, -0.3891857 ],
[-0.41355145, 0.22589689, 0.33154415, 0.86146815, -1.66326091,
0.37581697, -3.2250516 , -0.48807863],
[-2.52968957, 0.95801598, -1.20118154, 0.01141421, -0.11871498,
0.04555184, 1.3950473 , 0.37887998]]]]), array([[ 0., 0., 0., 7., 5.],
[ 1., 6., 2., 7., 5.],
[ 1., 3., 1., 6., 4.],
[ 0., 3., 3., 3., 3.]])]
creating: createRoiPooling
output of m is : [[[[ 1.19822323 1.60344636 0.36516821]
[ 1.39107513 1.73779249 1.26101124]]
[[ 1.58825231 1.47354829 0.98304272]
[ 1.158705 1.13772094 1.15776229]]]
[[[ 1.43371308 1.43371308 0.08018846]
[ 1.29001606 1.29001606 -1.7621665 ]]
[[ 1.43275058 1.43275058 0.85692906]
[ 1.39504731 1.39504731 0.37887999]]]
[[[ 2.13598228 0.30691418 2.62243223]
[ 0.82136625 0.82136625 1.43371308]]
[[ 3.06808996 3.06808996 -0.08386437]
[ 2.61830974 0.37581697 1.43275058]]]
[[[-0.06152321 -0.06152321 -0.06152321]
[-0.06152321 -0.06152321 -0.06152321]]
[[ 0.09853599 0.09853599 0.09853599]
[ 0.09853599 0.09853599 0.09853599]]]]