def forward(self,
input_or_shape,
pool_shape,
pool_type="MAX",
strides=None):
self.init_input(input_or_shape)
self.pool_type = pool_type
self.pool_shape = (1, 1, pool_shape)
if strides is None:
self.strides = self.pool_shape
else:
self.strides = (1, 1, strides)
return T.poolNd(
input,
self.pool_shape,
strides=self.strides,
reducer=self.pool_type,
)
def mymean(x):
return (x[::2, ::2] + x[1::2, ::2] + x[::2, 1::2] + x[1::2, 1::2]) / 4
def mysum(x):
return mymean(x) * 4
def mymax(x):
return np.maximum(np.maximum(x[::2, ::2], x[1::2, ::2]),
np.maximum(x[::2, 1::2], x[1::2, 1::2]))
SHAPE = (4, 4)
w = T.Placeholder(SHAPE, "float32", name="w")
maxpool = T.poolNd(w, (2, 2), reducer="MAX")
meanpool = T.poolNd(w, (2, 2), reducer="AVG")
sumpool = T.poolNd(w, (2, 2), reducer="SUM")
f = symjax.function(w, outputs=[maxpool, meanpool, sumpool])
for i in range(10):
data = np.random.randn(*SHAPE)
output = f(data)
print(
np.allclose(mymean(data), output[1]),
np.allclose(mymax(data), output[0]),
np.allclose(mysum(data), output[2]),
)
import sys
sys.path.insert(0, "../")
import symjax
import symjax.tensor as T
def mymean(x):
return (x[::2,::2]+x[1::2,::2]+x[::2,1::2]+x[1::2,1::2])/4
def mysum(x):
return mymean(x)*4
def mymax(x):
return np.maximum(np.maximum(x[::2,::2],x[1::2,::2]),
np.maximum(x[::2,1::2],x[1::2,1::2]))
SHAPE = (4, 4)
w = T.Placeholder(SHAPE, 'float32', name='w')
maxpool = T.poolNd(w, (2, 2), reducer='MAX')
meanpool = T.poolNd(w, (2, 2), reducer='AVG')
sumpool = T.poolNd(w, (2, 2), reducer='SUM')
f = symjax.function(w, outputs=[maxpool, meanpool, sumpool])
for i in range(10):
data = np.random.randn(*SHAPE)
output = f(data)
print(np.allclose(mymean(data), output[1]),
np.allclose(mymax(data), output[0]),
np.allclose(mysum(data), output[2]))
def forward(self, input):
return T.poolNd(input,
self.pool_shape,
strides=self.strides,
reducer=self.pool_type)