Skip to content

P
pytensor
提交 8a23fb1c authored 作者: Brandon T. Willard's avatar Brandon T. Willard 提交者： Thomas Wiecki
浏览文件
操作

Rename theano.tensor.nnet.bn to theano.tensor.nnet.batchnorm

上级 e0e5b3b8
隐藏空白字符变更
内嵌 并排
正在显示 包含 84 行增加 和 72 行删除
tests/gpuarray/test_dnn.py
浏览文件 @ 8a23fb1c
......@@ -48,7 +48,7 @@ from theano.tensor.nnet import (
LogSoftmax,
Softmax,
SoftmaxGrad,
bn,
batchnorm,
conv2d,
softmax,
softmax_op,
......@@ -1869,7 +1869,7 @@ def test_dnn_batchnorm_train():
x_invstd_abstract,
out_running_mean_abstract,
out_running_var_abstract,
) = bn.batch_normalization_train(
) = batchnorm.batch_normalization_train(
x,
scale,
bias,
......@@ -1966,9 +1966,9 @@ def test_dnn_batchnorm_train():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f_abstract.maker.fgraph.toposort()
......@@ -2044,9 +2044,11 @@ def test_dnn_batchnorm_train_without_running_averages():
out_gpu, x_mean_gpu, x_invstd_gpu = dnn.dnn_batch_normalization_train(
x, scale, bias, "per-activation"
)
out_abstract, x_mean_abstract, x_invstd_abstract = bn.batch_normalization_train(
x, scale, bias, "per-activation"
)
(
out_abstract,
x_mean_abstract,
x_invstd_abstract,
) = batchnorm.batch_normalization_train(x, scale, bias, "per-activation")
# backward pass
grads_gpu = theano.grad(None, wrt=[x, scale, bias], known_grads={out_gpu: dy})
grads_abstract = theano.grad(
......@@ -2081,9 +2083,9 @@ def test_dnn_batchnorm_train_without_running_averages():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f_abstract.maker.fgraph.toposort()
......@@ -2113,9 +2115,11 @@ def test_without_dnn_batchnorm_train_without_running_averages():
param_shape = (1, 10, 30, 25)
# forward pass
out_abstract, x_mean_abstract, x_invstd_abstract = bn.batch_normalization_train(
x, scale, bias, "per-activation"
)
(
out_abstract,
x_mean_abstract,
x_invstd_abstract,
) = batchnorm.batch_normalization_train(x, scale, bias, "per-activation")
# backward pass
grads_abstract = theano.grad(
None, wrt=[x, scale, bias], known_grads={out_abstract: dy}
......@@ -2144,9 +2148,9 @@ def test_without_dnn_batchnorm_train_without_running_averages():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f_abstract.maker.fgraph.toposort()
......@@ -2243,7 +2247,7 @@ def test_batchnorm_inference():
x, scale, bias, mean, var, mode, eps
)
# forward pass, abstract interface
out_abstract = bn.batch_normalization_test(
out_abstract = batchnorm.batch_normalization_test(
x, scale, bias, mean, var, mode, eps
)
# reference forward pass
......@@ -2293,9 +2297,9 @@ def test_batchnorm_inference():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f_abstract.maker.fgraph.toposort()
......@@ -2389,10 +2393,12 @@ def test_dnn_batchnorm_valid_and_invalid_axes():
invalid_axes_lists = (tuple(range(1, ndim)),)
for axes in valid_axes_lists + invalid_axes_lists:
# forward pass, abstract interface
out_train, x_mean, x_invstd = bn.batch_normalization_train(
out_train, x_mean, x_invstd = batchnorm.batch_normalization_train(
x, scale, bias, axes
)
out_test = bn.batch_normalization_test(x, scale, bias, mean, var, axes)
out_test = batchnorm.batch_normalization_test(
x, scale, bias, mean, var, axes
)
# backward pass
dy = vartype("dy")
grads_train = theano.grad(
......@@ -2433,9 +2439,9 @@ def test_dnn_batchnorm_valid_and_invalid_axes():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f.maker.fgraph.toposort()
......@@ -2450,9 +2456,9 @@ def test_dnn_batchnorm_valid_and_invalid_axes():
(
dnn.GpuDnnBatchNorm,
dnn.GpuDnnBatchNormGrad,
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f.maker.fgraph.toposort()
......
tests/tensor/nnet/test_bn.py → tests/tensor/nnet/test_batchnorm.py
浏览文件 @ 8a23fb1c
......@@ -8,7 +8,7 @@ import theano.tensor as tt
from tests import unittest_tools as utt
from theano.configdefaults import config
from theano.tensor.math import sum as tt_sum
from theano.tensor.nnet import bn
from theano.tensor.nnet import batchnorm
from theano.tensor.type import (
TensorType,
matrix,
......@@ -51,7 +51,7 @@ def test_BNComposite():
f_ref = theano.function([x, b, g, m, v], [bn_ref_op])
res_ref = f_ref(X, G, B, M, V)
for mode in ["low_mem", "high_mem"]:
bn_op = bn.batch_normalization(x, g, b, m, v, mode=mode)
bn_op = batchnorm.batch_normalization(x, g, b, m, v, mode=mode)
f = theano.function([x, b, g, m, v], [bn_op])
res = f(X, G, B, M, V)
utt.assert_allclose(res_ref, res)
......@@ -79,13 +79,15 @@ def test_batch_normalization():
f_ref = theano.function([x, g, b, m, v], [bn_ref_op])
res_ref = f_ref(X, G, B, M, V)
for mode in ["low_mem", "high_mem"]:
bn_op = bn.batch_normalization(x, g, b, m, v, mode=mode)
bn_op = batchnorm.batch_normalization(x, g, b, m, v, mode=mode)
f = theano.function([x, g, b, m, v], [bn_op])
res = f(X, G, B, M, V)
utt.assert_allclose(res_ref, res)
def bn_f(inputs, gamma, beta, mean, std):
return bn.batch_normalization(inputs, gamma, beta, mean, std, mode=mode)
return batchnorm.batch_normalization(
inputs, gamma, beta, mean, std, mode=mode
)
utt.verify_grad(bn_f, [X, G, B, M, V])
......@@ -95,7 +97,7 @@ def test_batch_normalization():
f_ref = theano.function([x, b, g], [bn_ref_op])
res_ref = f_ref(X, G, B)
for mode in ["low_mem", "high_mem"]:
bn_op = bn.batch_normalization(
bn_op = batchnorm.batch_normalization(
x,
g,
b,
......@@ -108,7 +110,9 @@ def test_batch_normalization():
utt.assert_allclose(res_ref, res)
def bn_f(inputs, gamma, beta, mean, std):
return bn.batch_normalization(inputs, gamma, beta, mean, std, mode=mode)
return batchnorm.batch_normalization(
inputs, gamma, beta, mean, std, mode=mode
)
utt.verify_grad(
bn_f, [X, G, B, X.mean(axis=0)[np.newaxis], X.std(axis=0)[np.newaxis]]
......@@ -144,7 +148,7 @@ def test_bn_feature_maps():
res_ref = f_ref(X, G, B, M, V)
for mode in ["low_mem", "high_mem"]:
bn_op = bn.batch_normalization(
bn_op = batchnorm.batch_normalization(
x,
g.dimshuffle("x", 0, "x", "x"),
b.dimshuffle("x", 0, "x", "x"),
......@@ -157,7 +161,7 @@ def test_bn_feature_maps():
utt.assert_allclose(res_ref, res)
def conv_bn(inputs, gamma, beta, mean, std):
return bn.batch_normalization(
return batchnorm.batch_normalization(
inputs,
gamma.dimshuffle("x", 0, "x", "x"),
beta.dimshuffle("x", 0, "x", "x"),
......@@ -196,7 +200,7 @@ def test_batch_normalization_train():
x_invstd,
out_running_mean,
out_running_var,
) = bn.batch_normalization_train(
) = batchnorm.batch_normalization_train(
x,
scale,
bias,
......@@ -300,9 +304,9 @@ def test_batch_normalization_train():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f.maker.fgraph.toposort()
......@@ -378,19 +382,19 @@ def test_batch_normalization_train_grad_grad():
continue
def bn_grad_wrt_inputs_f(x, dy, scale, x_mean, x_invstd):
g_inputs, g_scale, g_bias = bn.AbstractBatchNormTrainGrad(axes)(
g_inputs, g_scale, g_bias = batchnorm.AbstractBatchNormTrainGrad(axes)(
x, dy, scale, x_mean, x_invstd
)
return g_inputs
def bn_grad_wrt_scale_f(x, dy, scale, x_mean, x_invstd):
g_inputs, g_scale, g_bias = bn.AbstractBatchNormTrainGrad(axes)(
g_inputs, g_scale, g_bias = batchnorm.AbstractBatchNormTrainGrad(axes)(
x, dy, scale, x_mean, x_invstd
)
return g_scale
def bn_grad_wrt_bias_f(x, dy, scale, x_mean, x_invstd):
g_inputs, g_scale, g_bias = bn.AbstractBatchNormTrainGrad(axes)(
g_inputs, g_scale, g_bias = batchnorm.AbstractBatchNormTrainGrad(axes)(
x, dy, scale, x_mean, x_invstd
)
return g_bias
......@@ -438,7 +442,7 @@ def test_batch_normalization_train_without_running_averages():
param_shape = (1, 10, 30, 25)
# forward pass
out, x_mean, x_invstd = bn.batch_normalization_train(
out, x_mean, x_invstd = batchnorm.batch_normalization_train(
x, scale, bias, "per-activation"
)
# backward pass
......@@ -451,9 +455,9 @@ def test_batch_normalization_train_without_running_averages():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f.maker.fgraph.toposort()
......@@ -508,7 +512,7 @@ def test_batch_normalization_train_broadcast():
running_var_bc = running_var.dimshuffle(params_dimshuffle)
# batch_normalization_train with original, non-broadcasted variables
train_non_bc = bn.batch_normalization_train(
train_non_bc = batchnorm.batch_normalization_train(
x,
scale,
bias,
......@@ -519,7 +523,7 @@ def test_batch_normalization_train_broadcast():
running_var,
)
# batch_normalization_train with broadcasted variables
train_bc = bn.batch_normalization_train(
train_bc = batchnorm.batch_normalization_train(
x,
scale_bc,
bias_bc,
......@@ -534,11 +538,11 @@ def test_batch_normalization_train_broadcast():
)
# batch_normalization_test with original, non-broadcasted variables
test_non_bc = bn.batch_normalization_test(
test_non_bc = batchnorm.batch_normalization_test(
x, scale, bias, running_mean, running_var, axes, eps
)
# batch_normalization_test with broadcasted variables
test_bc = bn.batch_normalization_test(
test_bc = batchnorm.batch_normalization_test(
x, scale_bc, bias_bc, running_mean_bc, running_var_bc, axes, eps
)
......@@ -588,7 +592,9 @@ def test_batch_normalization_test():
continue
# forward pass
out = bn.batch_normalization_test(x, scale, bias, mean, var, axes, eps)
out = batchnorm.batch_normalization_test(
x, scale, bias, mean, var, axes, eps
)
# reference forward pass
if axes == "per-activation":
axes2 = (0,)
......@@ -619,9 +625,9 @@ def test_batch_normalization_test():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f.maker.fgraph.toposort()
......@@ -660,10 +666,10 @@ def test_batch_normalization_broadcastable():
)
# forward pass
out_train, x_mean, x_invstd = bn.batch_normalization_train(
out_train, x_mean, x_invstd = batchnorm.batch_normalization_train(
x, scale, bias, "spatial"
)
out_test = bn.batch_normalization_test(x, scale, bias, mean, var, "spatial")
out_test = batchnorm.batch_normalization_test(x, scale, bias, mean, var, "spatial")
# backward pass
grads_train = tt.grad(None, wrt=[x, scale, bias], known_grads={out_train: dy})
grads_test = tt.grad(None, wrt=[x, scale, bias], known_grads={out_test: dy})
......@@ -677,9 +683,9 @@ def test_batch_normalization_broadcastable():
isinstance(
n.op,
(
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormInference,
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrainGrad,
),
)
for n in f.maker.fgraph.toposort()
......
theano/gpuarray/opt.py
浏览文件 @ 8a23fb1c
......@@ -179,7 +179,7 @@ from theano.tensor.basic import (
Tri,
)
from theano.tensor.math import MaxAndArgmax
from theano.tensor.nnet import bn, conv3d2d
from theano.tensor.nnet import batchnorm, conv3d2d
from theano.tensor.nnet.abstract_conv import (
AbstractConv2d,
AbstractConv2d_gradInputs,
......@@ -3122,9 +3122,9 @@ register_opt("fast_compile")(abstract_batch_norm_groupopt)
register_opt("fast_compile", name="abstract_batch_norm_db")(abstract_batch_norm_db)
register_opt2(
[
bn.AbstractBatchNormTrain,
bn.AbstractBatchNormTrainGrad,
bn.AbstractBatchNormInference,
batchnorm.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormInference,
],
"fast_compile",
name="abstract_batch_norm_db2",
......@@ -3132,19 +3132,19 @@ register_opt2(
for op, fct, cpu in [
(
bn.AbstractBatchNormTrain,
batchnorm.AbstractBatchNormTrain,
local_abstract_batch_norm_train_cudnn,
bn.local_abstract_batch_norm_train,
batchnorm.local_abstract_batch_norm_train,
),
(
bn.AbstractBatchNormTrainGrad,
batchnorm.AbstractBatchNormTrainGrad,
local_abstract_batch_norm_train_grad_cudnn,
bn.local_abstract_batch_norm_train_grad,
batchnorm.local_abstract_batch_norm_train_grad,
),
(
bn.AbstractBatchNormInference,
batchnorm.AbstractBatchNormInference,
local_abstract_batch_norm_inference_cudnn,
bn.local_abstract_batch_norm_inference,
batchnorm.local_abstract_batch_norm_inference,
),
]:
lifter = op_lifter([op])(fct)
......
theano/tensor/nnet/__init__.py
浏览文件 @ 8a23fb1c
......@@ -40,7 +40,7 @@ from theano.tensor.nnet.basic import (
softmax_with_bias,
softsign,
)
from theano.tensor.nnet.bn import batch_normalization
from theano.tensor.nnet.batchnorm import batch_normalization
from theano.tensor.nnet.sigm import (
hard_sigmoid,
scalar_sigmoid,
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论