habana_frameworks.mediapipe.fn.Where
habana_frameworks.mediapipe.fn.Where¶
- Class:
habana_frameworks.mediapipe.fn.Where(**kwargs)
- Define graph call:
__call__(perdicate, input0, input1)
- Parameter:
predicate - Predicate tensor to operator, size=[batch_size]. Supported dimensions: minimum = 1, maximum = 1. Supported data types: UINT8.
input0 - Input0 tensor, Supported dimensions: minimum = 1, maximum = 5. Supported data types: FLOAT16, FLOAT32, BFLOAT16.
input1 - Input1 tensor, Supported dimensions: minimum = 1, maximum = 5. Supported data types: FLOAT16, FLOAT32, BFLOAT16.
Description:
Outputs a tensor of elements selected from either input0 or input1, depending on predicate. If predicate is true then input0 is selected, else input1 is selected.
- Supported backend:
HPU, CPU
Keyword Arguments
kwargs |
Description |
|---|---|
dtype |
Output data type.
|
Note
All input0, input1, output tensors must be of the same data type and must have the same dimensionality.
Example: Where Operator
The following code snippet shows usage of Where operator:
from habana_frameworks.mediapipe import fn
from habana_frameworks.mediapipe.mediapipe import MediaPipe
from habana_frameworks.mediapipe.media_types import dtype as dt
import numpy as np
import os
g_batch_size = 3
g_predicate = np.array(np.arange(g_batch_size).reshape(
[g_batch_size, 1, 1, 1]), dtype=dt.INT8)
g_predicate = np.mod(g_predicate, 2)
print(g_predicate)
print()
g_scale1 = np.ones([g_batch_size, 1, 1, 1], dtype=dt.FLOAT32) * 0.3
g_scale2 = np.ones([g_batch_size, 1, 1, 1], dtype=dt.FLOAT32) * 0.7
# Create MediaPipe derived class
class myMediaPipe(MediaPipe):
def __init__(self, device, queue_depth, batch_size, num_threads,
op_device, dir):
super(myMediaPipe, self).__init__(
device,
queue_depth,
batch_size,
num_threads,
self.__class__.__name__)
self.inp = fn.ReadNumpyDatasetFromDir(num_outputs=1,
shuffle=False,
dir=dir,
pattern="inp_x_*.npy",
dense=True,
dtype=dt.FLOAT32,
device="cpu")
self.predicate = fn.MediaConst(data=g_predicate,
shape=[1, 1, 1, batch_size],
dtype=dt.INT8,
device="cpu",
batch_broadcast=False)
self.scale1 = fn.MediaConst(data=g_scale1,
shape=[1, 1, 1, batch_size],
dtype=dt.FLOAT32,
device="cpu",
batch_broadcast=False)
self.scale2 = fn.MediaConst(data=g_scale2,
shape=[1, 1, 1, batch_size],
dtype=dt.FLOAT32,
device="cpu",
batch_broadcast=False)
self.where = fn.Where(dtype=dt.FLOAT32,
device=op_device)
self.mul = fn.Mult(device=op_device)
self.op_device = op_device
def definegraph(self):
inp = self.inp()
scale1 = self.scale1()
scale2 = self.scale2()
predicate = self.predicate()
scale = self.where(predicate, scale1, scale2)
out = self.mul(inp, scale)
return inp, scale, out
def run(device, op_device):
queue_depth = 1
num_threads = 1
base_dir = os.environ['DATASET_DIR']
dir = base_dir+"/npy_data/fp32/"
# Create MediaPipe object
pipe = myMediaPipe(device, queue_depth, g_batch_size, num_threads,
op_device, dir)
# Build MediaPipe
pipe.build()
# Initialize MediaPipe iterator
pipe.iter_init()
# Run MediaPipe
inp, scale, out = pipe.run()
def as_cpu(tensor):
if (callable(getattr(tensor, "as_cpu", None))):
tensor = tensor.as_cpu()
return tensor
inp = as_cpu(inp).as_nparray()
scale = as_cpu(scale).as_nparray()
out = as_cpu(out).as_nparray()
del pipe
print("\ninp tensor shape:", inp.shape)
print("inp tensor dtype:", inp.dtype)
print("inp tensor data:\n", inp)
print("\nscale tensor shape:", scale.shape)
print("scale tensor dtype:", scale.dtype)
print("scale tensor data:\n", scale)
print("\nout tensor shape:", out.shape)
print("out tensor dtype:", out.dtype)
print("out tensor data:\n", out)
return inp, scale, out
def compare_ref(inp, scale, out):
ref_scale = np.where(g_predicate == 1, g_scale1, g_scale2)
if np.array_equal(ref_scale, scale) == False:
raise ValueError(f"Mismatch w.r.t scale ref")
ref_out = inp * ref_scale
if np.array_equal(ref_out, out) == False:
raise ValueError(f"Mismatch w.r.t out ref")
if __name__ == "__main__":
dev_opdev = {'cpu': ['cpu'],
'mixed': ['hpu'],
'legacy': ['hpu']}
for dev in dev_opdev.keys():
for op_dev in dev_opdev[dev]:
inp, scale, out = run(dev, op_dev)
compare_ref(inp, scale, out)
The following is the output of Where operator:
inp tensor shape: (3, 3, 2, 3)
inp tensor dtype: float32
inp tensor data:
[[[[182. 227. 113.]
[175. 128. 253.]]
[[ 58. 140. 136.]
[ 86. 80. 111.]]
[[175. 196. 178.]
[ 20. 163. 108.]]]
[[[186. 254. 96.]
[180. 64. 132.]]
[[149. 50. 117.]
[213. 6. 111.]]
[[ 77. 11. 160.]
[129. 102. 154.]]]
[[[100. 95. 164.]
[134. 131. 112.]]
[[ 77. 38. 127.]
[123. 87. 71.]]
[[227. 186. 223.]
[ 35. 56. 166.]]]]
scale tensor shape: (3, 1, 1, 1)
scale tensor dtype: float32
scale tensor data:
[[[[0.7]]]
[[[0.3]]]
[[[0.7]]]]
out tensor shape: (3, 3, 2, 3)
out tensor dtype: float32
out tensor data:
[[[[127.4 158.9 79.1 ]
[122.5 89.6 177.09999 ]]
[[ 40.6 98. 95.2 ]
[ 60.2 56. 77.7 ]]
[[122.5 137.2 124.6 ]
[ 14. 114.1 75.6 ]]]
[[[ 55.800003 76.200005 28.800001 ]
[ 54.000004 19.2 39.600002 ]]
[[ 44.7 15.000001 35.100002 ]
[ 63.9 1.8000001 33.300003 ]]
[[ 23.1 3.3000002 48. ]
[ 38.7 30.6 46.2 ]]]
[[[ 70. 66.5 114.799995 ]
[ 93.799995 91.7 78.4 ]]
[[ 53.899998 26.6 88.9 ]
[ 86.1 60.899998 49.7 ]]
[[158.9 130.2 156.09999 ]
[ 24.5 39.2 116.2 ]]]]