nd.log(-1 * gate._limit_lo / gate._limit_hi))
print(cost)
# %%
[str(i) for i, item in enumerate(tree._embeddlayer._children.items())]
list(tree._embeddlayer._children.keys())
for i, key in enumerate(list(tree._embeddlayer._children.keys())):
item = tree._embeddlayer._children.pop(key)
tree._embeddlayer._children[str(i)] = item
# %%
tree._grow(nd.array([[0, 0], [3, 3], [7, 1]]))
tree._grow(nd.array([[10, 10]]))
tree._grow(nd.array([[6, 8]]))
tree._grow(nd.array([[9, 2], [1, 9]]))
tree(nd.array([[1, 1], [2, 2], [-1, -1]]))
# %%
[
node._decision._gate() for node in tree._embeddlayer._children.values()
if hasattr(node, "_decision")
]
# %%
net.collect_params().initialize(mx.init.Normal(sigma=0.1),
force_reinit=True,
ctx=model_ctx)
error = gluon.loss.SoftmaxCrossEntropyLoss()
# %%
X, y = shuffle(X, y)
for data, target in zip(np.split(X, 10), np.split(y, 10)):
data = nd.array(data).as_in_context(model_ctx)
target = nd.array(target).as_in_context(model_ctx)
tree._grow(prenet(data))
# tree._grow(data)
less = net.collect_params()
for key in list(less._params.keys()):
if less[key].shape is None:
less._params.pop(key)
trainer = gluon.Trainer(less, 'sgd', {'learning_rate': 2})
with mx.autograd.record():
loss = error(net(data), target)
# cost = []
# for decision in tree._routerlayer._children.values():
# gate = decision._gate
# cost.append(nd.sigmoid(
from Tree import Tree import mxnet as mx import numpy as np from mxnet import gluon, nd tree = Tree(units=2) tree.collect_params().initialize(force_reinit=True) tree.collect_params() tree._structure tree(nd.array([[1, 3], [5, 6], [1, 2], [3, 0], [8, 8], [3.5, 3.5]])) tree._grow(nd.array([[1, 1]])) # %% from Node import Node Node() node.collect_params() # %% a = nd.array([[1, 2], [3, 4], [-10, -10]]) a upper = nd.max(a, axis=0) lower = nd.min(a, axis=0)
net = gluon.nn.Sequential()
with net.name_scope():
net.add(tree)
net.add(gluon.nn.Dense(1))
error = gluon.loss.L2Loss()
net.collect_params().initialize(mx.init.Normal(sigma=0.1), force_reinit=True)
# %%
X, y = shuffle(X, y)
for data, target in zip(np.split(X, 10), np.split(y, 10)):
tree._grow(nd.array(data))
less = net.collect_params()
for key in list(less._params.keys()):
if less[key].shape is None:
less._params.pop(key)
trainer = gluon.Trainer(less, 'sgd', {'learning_rate': 0.5})
with mx.autograd.record():
loss = error(net(nd.array(data)), nd.array(target))
loss.backward()
trainer.step(data.shape[0], ignore_stale_grad=True)
after = [
