Skip to content

P
Paddle

wmsofts / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 0
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 c5360a3f 编写于 作者: X xuezhong
浏览文件
操作

refine code

上级 44240216
隐藏空白更改
内联 并排
Showing with 163 addition and 158 deletion
paddle/fluid/operators/sample_logits_op.cc
浏览文件 @ c5360a3f
......@@ -25,63 +25,64 @@ class SampleLogitsOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor, default: Tensor<float>), The unscaled log probabilities "
"which is a 2-D tensor with shape [N x K]. N is the batch_size, "
"and K is the class number.");
AddInput("Label",
"(Tensor) The ground truth which is a 2-D tensor. Label is a "
AddInput("Labels",
"(Tensor) The ground truth which is a 2-D tensor. Labels is a "
"Tensor<int64> with shape [N x NT], where NT is the number of"
"true labels for each example.");
AddInput(
"CustomSamples",
"(Tensor, default: Tensor<int64_t>), A 2-D tensor with shaoe [N x "
"S+NT]."
"The customized sample labels with true labels at first. This tensor"
"is only use_custom_samples is true.")
AddInput("CustomizedSamples",
"(Tensor, default: Tensor<int64_t>), A 2-D tensor with shape [N, "
"NT + S],"
" where N is the batch size, NT is the number of true labels "
"and S is the number of negtive sample for each example."
"The first NT elements of each row should be the same with true "
"labels, "
"followed by S custom negtive samples. This tensor"
"is only used when use_customized_samples is true.")
.AsDispensable();
AddInput(
"CustomProbabilities",
"(Tensor, default: Tensor<float>), A 2-D tensor with shaoe [N x S+NT]."
"The customized sample probabilities with true labels at first. This "
"tensor is only use_custom_samples is true.")
"CustomizedProbabilities",
"(Tensor, default: Tensor<float>), A 2-D tensor with shape [N, NT + S]."
"The tensor has the same shape with CustomSamples,"
"and each element represents probability of element in CustomSamples. "
"This "
"tensor is only used when use_customized_samples is true.")
.AsDispensable();
AddOutput(
"Samples",
"(Tensor, default: Tensor<int64_t>), A 2-D tensor with shape [N x "
"S+NT]."
"The outputs value of sampler by given the true label, where S is the "
"number of negative sample for each example. So Samples includes NT "
"true"
"labels and S negative labels for each example. This will be used in"
"backward calculation.")
AddOutput("Samples",
"(Tensor, default: Tensor<int64_t>), A 2-D tensor with shape [N, "
"NT + S]."
"The outputs value of sampler, including NT true lables and S "
"negetive samples "
"for each example. This will be used in"
"backward calculation.")
.AsIntermediate();
AddOutput(
"Probabilities",
"(Tensor, default: Tensor<float>), A 2-D tensor with shape [N x "
"S+NT]."
"The outputs value of progabilites of samples by given the true label, "
"where S is the "
"number of negative sample for each example. So Samples includes NT "
"true"
"labels and S negative labels for each example.")
"(Tensor, default: Tensor<float>), A 2-D tensor with shape [N, NT + S]."
"The probabilites of sampled positive and negtive labels.")
.AsIntermediate();
AddOutput("SampledLogits",
"(Tensor, default: Tensor<float>), A 2-D tensor with shape"
"[N x S+NT]. The outputs value of sample logits, which will be"
"used in backward calculation.")
"[N, NT + S]. The outputs value of sampled logits, which will be"
"used in backward propagation.")
.AsIntermediate();
AddOutput(
"SampledLabel",
"(Tensor, default: Tensor<int64>), A 2-D tensor. The sampled label"
"with shape [N x S + NT].");
"SampledLabels",
"(Tensor, default: Tensor<int64>), A 2-D tensor. The sampled labels"
"with shape [N, NT]. The tonsor contains hard labels as input to "
" softmax op, that is 0, 1, …, NT-1 because of the first NT elements"
" of Sampels are positive lables.");
AddAttr<bool>(
"use_custom_samples",
"An indicator whether to use custom samples with probabilities, if True"
"the operator will use custom samples and custom probabilities"
"use_customized_samples",
"An indicator whether to use customized samples with probabilities, if "
"True"
"the operator will use customized samples and customized probabilities"
"otherwise, the operator will generate them by itself.")
.SetDefault(false);
AddAttr<bool>(
"uniq",
"An indicator whether to sample non-repetitive negtive labels, if True"
"the operator will sample negtive labels without replacement."
"otherwise, the operator will sample negtive labels with replacement.")
"Otherwise, the operator will sample negtive labels with replacement.")
.SetDefault(true);
AddAttr<bool>(
"remove_accidental_hits",
......@@ -95,8 +96,7 @@ class SampleLogitsOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
"""
Computes sampled output training logits and labels suitable for implementing
sampled softmax.
sampled softmax.
"""
)DOC");
......@@ -110,7 +110,8 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Logits"),
"Input(Logits) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Labels"),
"Input(Labels) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("Samples"),
"Output(Samples) should be not null.");
......@@ -118,11 +119,11 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
"Output(Probabilities) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("SampledLogits"),
"Output(SampledLogits) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("SampledLabel"),
"Output(SampledLabel) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("SampledLabels"),
"Output(SampledLabels) should be not null.");
auto logits_dims = ctx->GetInputDim("Logits");
auto labels_dims = ctx->GetInputDim("Label");
auto labels_dims = ctx->GetInputDim("Labels");
PADDLE_ENFORCE_EQ(
logits_dims.size(), 2UL,
......@@ -135,7 +136,7 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
ctx->SetOutputDim("Samples", {logits_dims[0], num_sampled_classes});
ctx->SetOutputDim("Probabilities", {logits_dims[0], num_sampled_classes});
ctx->SetOutputDim("SampledLogits", {logits_dims[0], num_sampled_classes});
ctx->SetOutputDim("SampledLabel", {logits_dims[0], labels_dims[1]});
ctx->SetOutputDim("SampledLabels", {logits_dims[0], labels_dims[1]});
}
protected:
......@@ -144,7 +145,6 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Logits"));
framework::OpKernelType kt =
framework::OpKernelType(data_type, ctx.device_context());
// kt.place_ = platform::CPUPlace();
return kt;
}
};
......@@ -157,7 +157,8 @@ class SampleLogitsOpGrad : public framework::OperatorWithKernel {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Logits"),
"Input(Logits) should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Labels"),
"Input(Labels) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Samples"),
"Input(Samples) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("SampledLogits"),
......@@ -168,7 +169,7 @@ class SampleLogitsOpGrad : public framework::OperatorWithKernel {
"Output(Logits@Grad) should be not null.");
auto logit_dims = ctx->GetInputDim("Logits");
auto label_dims = ctx->GetInputDim("Label");
auto label_dims = ctx->GetInputDim("Labels");
PADDLE_ENFORCE_EQ(label_dims.size(), 2UL,
"The label should be a 2-D tensor.");
PADDLE_ENFORCE_EQ(logit_dims.size(), 2UL,
......@@ -185,7 +186,6 @@ class SampleLogitsOpGrad : public framework::OperatorWithKernel {
ctx.InputVar(framework::GradVarName("SampledLogits")));
framework::OpKernelType kt =
framework::OpKernelType(data_type, ctx.device_context());
// kt.place_ = platform::CPUPlace();
return kt;
}
};
......@@ -200,7 +200,7 @@ class SampleLogitsGradMaker : public framework::SingleGradOpDescMaker {
auto* grad_op = new framework::OpDesc();
grad_op->SetType("sample_logits_grad");
grad_op->SetInput("Logits", Input("Logits"));
grad_op->SetInput("Label", Input("Label"));
grad_op->SetInput("Labels", Input("Labels"));
grad_op->SetInput("Samples", Output("Samples"));
grad_op->SetInput("SampledLogits", Output("SampledLogits"));
grad_op->SetInput(framework::GradVarName("SampledLogits"),
......
paddle/fluid/operators/sample_logits_op.cu
浏览文件 @ c5360a3f
......@@ -109,25 +109,26 @@ class SampleLogitsCUDAKernel : public framework::OpKernel<T> {
void Compute(const framework::ExecutionContext& context) const override {
// get necessary inputs
const Tensor* logits = context.Input<Tensor>("Logits");
const Tensor* label = context.Input<Tensor>("Label");
const Tensor* labels = context.Input<Tensor>("Labels");
VLOG(3) << "Enter SampleLogitsCUDAKernel";
// get necessary outputs
Tensor* samples = context.Output<Tensor>("Samples");
Tensor* probabilities = context.Output<Tensor>("Probabilities");
Tensor* sampled_logits = context.Output<Tensor>("SampledLogits");
Tensor* sampled_label = context.Output<Tensor>("SampledLabel");
Tensor* sampled_labels = context.Output<Tensor>("SampledLabels");
// shapes
const auto batch_size = logits->dims()[0];
const auto num_classes = logits->dims()[1];
const auto label_dim = label->dims();
const auto num_true = label_dim[1];
const auto labels_dim = labels->dims();
const auto num_true = labels_dim[1];
const auto samples_dim = samples->dims();
// attrs
const auto num_samples = context.Attr<int>("num_samples");
const bool use_custom_samples = context.Attr<bool>("use_custom_samples");
const bool use_customized_samples =
context.Attr<bool>("use_customized_samples");
const bool uniq = context.Attr<bool>("uniq");
const bool remove_accidental_hits =
context.Attr<bool>("remove_accidental_hits");
......@@ -140,21 +141,22 @@ class SampleLogitsCUDAKernel : public framework::OpKernel<T> {
math::SetConstant<platform::CUDADeviceContext, T> set_zero;
set_zero(dev_ctx, sampled_logits, static_cast<T>(0));
auto sampled_label_data =
sampled_label->mutable_data<int64_t>(label_dim, context.GetPlace());
auto sampled_labels_data =
sampled_labels->mutable_data<int64_t>(labels_dim, context.GetPlace());
int threads = 512;
size_t size = batch_size * num_true;
int grid = (size + threads - 1) / threads;
GPUSetLabel<
T><<<grid, threads, 0, context.cuda_device_context().stream()>>>(
size, num_true, sampled_label_data);
if (use_custom_samples) {
const Tensor* custom_samples = context.Input<Tensor>("CustomSamples");
const Tensor* custom_probabilities =
context.Input<Tensor>("CustomProbabilities");
samples->ShareDataWith(*custom_samples);
probabilities->ShareDataWith(*custom_probabilities);
size, num_true, sampled_labels_data);
if (use_customized_samples) {
const Tensor* customized_samples =
context.Input<Tensor>("CustomizedSamples");
const Tensor* customized_probabilities =
context.Input<Tensor>("CustomizedProbabilities");
samples->ShareDataWith(*customized_samples);
probabilities->ShareDataWith(*customized_probabilities);
} else {
samples->mutable_data<int64_t>(context.GetPlace());
probabilities->mutable_data<T>(samples_dim, context.GetPlace());
......@@ -162,7 +164,7 @@ class SampleLogitsCUDAKernel : public framework::OpKernel<T> {
const auto seed = context.Attr<int>("seed");
auto sampler_with_prob = math::GPUSampleWithProb<T>();
sampler_with_prob(context.cuda_device_context(), seed, num_classes, uniq,
num_samples, label, samples, probabilities);
num_samples, labels, samples, probabilities);
}
// UNDERSTAND: gather sampled logits and remove accidental hits if needed
......
paddle/fluid/operators/sample_logits_op.h
浏览文件 @ c5360a3f
......@@ -150,24 +150,25 @@ class SampleLogitsKernel : public framework::OpKernel<T> {
VLOG(3) << "Enter SampleLogitsKernel";
// get necessary inputs
const Tensor* logits = context.Input<Tensor>("Logits");
const Tensor* label = context.Input<Tensor>("Label");
const Tensor* labels = context.Input<Tensor>("Labels");
// get necessary outputs
Tensor* samples = context.Output<Tensor>("Samples");
Tensor* probabilities = context.Output<Tensor>("Probabilities");
Tensor* sampled_logits = context.Output<Tensor>("SampledLogits");
Tensor* sampled_label = context.Output<Tensor>("SampledLabel");
Tensor* sampled_labels = context.Output<Tensor>("SampledLabels");
// shapes
const auto batch_size = logits->dims()[0];
const auto num_classes = logits->dims()[1];
const auto label_dim = label->dims();
const auto num_true = label_dim[1];
const auto labels_dim = labels->dims();
const auto num_true = labels_dim[1];
const auto samples_dim = samples->dims();
// attrs
const auto num_samples = context.Attr<int>("num_samples");
const bool use_custom_samples = context.Attr<bool>("use_custom_samples");
const bool use_customized_samples =
context.Attr<bool>("use_customized_samples");
const bool remove_accidental_hits =
context.Attr<bool>("remove_accidental_hits");
......@@ -177,18 +178,21 @@ class SampleLogitsKernel : public framework::OpKernel<T> {
// UNDERSTAND: allocate memories for temporaries
sampled_logits->mutable_data<T>(samples_dim, context.GetPlace());
auto sampled_label_data =
sampled_label->mutable_data<int64_t>(label_dim, context.GetPlace());
for (int i = 0; i < batch_size; ++i)
for (int j = 0; j < num_true; ++j)
sampled_label_data[i * num_true + j] = j;
if (use_custom_samples) {
const Tensor* custom_samples = context.Input<Tensor>("CustomSamples");
const Tensor* custom_probabilities =
context.Input<Tensor>("CustomProbabilities");
samples->ShareDataWith(*custom_samples);
probabilities->ShareDataWith(*custom_probabilities);
auto sampled_labels_data =
sampled_labels->mutable_data<int64_t>(labels_dim, context.GetPlace());
for (int i = 0; i < batch_size; ++i) {
for (int j = 0; j < num_true; ++j) {
sampled_labels_data[i * num_true + j] = j;
}
}
if (use_customized_samples) {
const Tensor* customized_samples =
context.Input<Tensor>("CustomizedSamples");
const Tensor* customized_probabilities =
context.Input<Tensor>("CustomizedProbabilities");
samples->ShareDataWith(*customized_samples);
probabilities->ShareDataWith(*customized_probabilities);
} else {
samples->mutable_data<int64_t>(context.GetPlace());
probabilities->mutable_data<T>(samples_dim, context.GetPlace());
......@@ -197,7 +201,7 @@ class SampleLogitsKernel : public framework::OpKernel<T> {
auto sampler_with_prob =
math::SampleWithProb<platform::CPUDeviceContext, T>();
sampler_with_prob(dev_ctx, math::LogUniformSampler(num_classes, seed),
num_samples, label, samples, probabilities);
num_samples, labels, samples, probabilities);
}
// UNDERSTAND: gather sampled logits and remove accidental hits if needed
......
python/paddle/fluid/layers/nn.py
浏览文件 @ c5360a3f
......@@ -5771,9 +5771,9 @@ def sampled_softmax_with_cross_entropy(logits,
num_samples,
num_true=1,
remove_accidental_hits=True,
use_custom_samples=False,
custom_samples=None,
custom_probabilities=None,
use_customized_samples=False,
customized_samples=None,
customized_probabilities=None,
seed=0):
"""
**Sampled Softmax With Cross Entropy Operator.**
......@@ -5789,7 +5789,7 @@ def sampled_softmax_with_cross_entropy(logits,
For examples with T true labels (T >= 1), we assume that each true label has
a probability of 1/T. For each sample, S samples are generated using a
log uniform distribution. True labels are concatenated with hese samples to
log uniform distribution. True labels are concatenated with these samples to
form T + S samples for each example. So, assume the shape of logits is
[N x K], the shape for samples is [N x (T+S)]. For each sampled label, a
probability is calculated, which corresponds to the Q(y|x) in
......@@ -5798,7 +5798,7 @@ def sampled_softmax_with_cross_entropy(logits,
Logits are sampled according to the sampled labels. Then if
remove_accidental_hits is True, if a sample[i, j] accidentally hits true
labels, then the corresponding sampled_logits[i, j] is minus by 1e20 to
make its softmax result close to zero. Then samled logits are subtracted by
make its softmax result close to zero. Then sampled logits are subtracted by
logQ(y|x), these sampled logits and re-indexed labels are used to compute
a softmax with cross entropy.
......@@ -5816,14 +5816,16 @@ def sampled_softmax_with_cross_entropy(logits,
accidentally hits true labels, then the corresponding
sampled_logits[i, j] is minus by 1e20 to make its softmax result
close to zero. Default is True.
use_custom_samples (bool): Whether to use custom samples and probabities to sample
use_customized_samples (bool): Whether to use custom samples and probabities to sample
logits.
custom_samples (Variable): User defined samples, which is a 1-D tensor with shape [S]. S is the num_samples.
custom_probabilities (Variable): User defined probabilities of samples, a 1-D tensor which has the same shape with custom_samples.
customized_samples (Variable): User defined samples, which is a 2-D tensor
with shape [N, T + S]. S is the num_samples, and T is the number of true
labels per example.
customized_probabilities (Variable): User defined probabilities of samples,
a 2-D tensor which has the same shape with customized_samples.
seed (int): The random seed for generating random number, which is used
in the process of sampling. Default is 0.
Returns:
Variable: Return the cross entropy loss which is a 2-D tensor with shape
[N x 1].
......@@ -5849,18 +5851,18 @@ def sampled_softmax_with_cross_entropy(logits,
type='sample_logits',
inputs={
'Logits': logits,
'Label': label,
'Labels': label,
'CustomSamples': custom_samples,
'CustomProbabilities': custom_probabilities
},
outputs={
'Samples': samples,
'Probabilities': probabilities,
'SampledLabel': sampled_label,
'SampledLabels': sampled_label,
'SampledLogits': sampled_logits
},
attrs={
'use_custom_samples': use_custom_samples,
'use_customized_samples': use_customized_samples,
'uniq': True,
'remove_accidental_hits': remove_accidental_hits,
'num_samples': num_samples,
......
python/paddle/fluid/tests/unittests/test_sample_logits.py
浏览文件 @ c5360a3f
......@@ -61,8 +61,8 @@ def take_along_axis1(array, index):
return out
def sample_prob(sampler, num_samples, label):
batch_size, num_true = label.shape
def sample_prob(sampler, num_samples, labels):
batch_size, num_true = labels.shape
num_sampled_classes = num_samples + num_true
samples = np.zeros((batch_size, num_sampled_classes), dtype=np.int64)
......@@ -74,8 +74,8 @@ def sample_prob(sampler, num_samples, label):
j = 0
while j < num_true:
for i in range(batch_size):
samples[i, j] = label[i, j]
probabilities[i, j] = sampler.probability(label[i, j])
samples[i, j] = labels[i, j]
probabilities[i, j] = sampler.probability(labels[i, j])
j += 1
while j < num_sampled_classes:
v = sampler.sample()
......@@ -103,33 +103,30 @@ def compute_remove_accidental_hits(sampled_logits, samples, num_true):
def sample_logits(logits,
label,
labels,
num_samples,
seed,
remove_accidental_hits,
use_custom_samples,
custom_samples=None,
custom_probabilities=None):
use_customized_samples,
customized_samples=None,
customized_probabilities=None):
batch_size, num_classes = logits.shape
num_true = label.shape[1]
num_true = labels.shape[1]
num_sampled_classes = num_true + num_samples
if use_custom_samples:
samples = custom_samples
probabilities = custom_probabilities
if use_customized_samples:
samples = customized_samples
probabilities = customized_probabilities
else:
sampler = LogUniformSampler(num_classes, seed)
samples, probabilities = sample_prob(sampler, num_samples, label)
samples, probabilities = sample_prob(sampler, num_samples, labels)
sampled_logits = take_along_axis1(logits, samples)
#print(samples)
#print(probabilities)
#print(sampled_logits)
if remove_accidental_hits:
compute_remove_accidental_hits(sampled_logits, samples, num_true)
sampled_logits -= np.log(probabilities)
sampled_label = np.tile(np.arange(num_true), (batch_size, 1))
return (sampled_logits, samples, sampled_label, probabilities)
sampled_labels = np.tile(np.arange(num_true), (batch_size, 1))
return (sampled_logits, samples, sampled_labels, probabilities)
class TestSampleLogitsOp(OpTest):
......@@ -138,51 +135,51 @@ class TestSampleLogitsOp(OpTest):
in python and just test the non-random part.
'''
def generate_data(self, logits, label, num_samples, seed,
remove_accidental_hits, use_custom_samples,
custom_samples, custom_probabilities):
def generate_data(self, logits, labels, num_samples, seed,
remove_accidental_hits, use_customized_samples,
customized_samples, customized_probabilities):
self.attrs = {
'num_samples': num_samples,
'use_custom_samples': use_custom_samples,
'use_customized_samples': use_customized_samples,
'remove_accidental_hits': remove_accidental_hits,
'seed': seed
}
self.inputs = {
'Logits': logits,
'Label': label,
'CustomSamples': custom_samples,
'CustomProbabilities': custom_probabilities
'Labels': labels,
'CustomizedSamples': customized_samples,
'CustomizedProbabilities': customized_probabilities
}
def set_data(self, batch_size, num_classes, num_true, num_samples, seed,
remove_accidental_hits):
logits = np.random.randn(batch_size, num_classes)
label = np.stack([
labels = np.stack([
np.random.choice(
range(0, num_classes), num_true, replace=False)
for _ in range(batch_size)
])
sampler = LogUniformSampler(num_classes, seed)
custom_samples, custom_probabilities = \
sample_prob(sampler, num_samples, label)
use_custom_samples = True
customized_samples, customized_probabilities = \
sample_prob(sampler, num_samples, labels)
use_customized_samples = True
remove_accidental_hits = remove_accidental_hits
self.generate_data(logits, label, num_samples, seed,
remove_accidental_hits, use_custom_samples,
custom_samples, custom_probabilities)
self.generate_data(logits, labels, num_samples, seed,
remove_accidental_hits, use_customized_samples,
customized_samples, customized_probabilities)
def compute(self):
out = sample_logits(self.inputs["Logits"], self.inputs["Label"],
out = sample_logits(self.inputs["Logits"], self.inputs["Labels"],
self.attrs["num_samples"], self.attrs["seed"],
self.attrs["remove_accidental_hits"],
self.attrs["use_custom_samples"],
self.inputs["CustomSamples"],
self.inputs["CustomProbabilities"])
self.attrs["use_customized_samples"],
self.inputs["CustomizedSamples"],
self.inputs["CustomizedProbabilities"])
self.outputs = {
'SampledLogits': out[0],
'Samples': out[1],
'SampledLabel': out[2],
'SampledLabels': out[2],
'Probabilities': out[3]
}
......@@ -255,29 +252,29 @@ class TestSampleLogitsOpV2(OpTest):
in C++ and copied to python and just test the non-random part.
'''
def generate_data(self, logits, label, num_samples, seed,
remove_accidental_hits, use_custom_samples):
def generate_data(self, logits, labels, num_samples, seed,
remove_accidental_hits, use_customized_samples):
self.attrs = {
'num_samples': num_samples,
'use_custom_samples': use_custom_samples,
'use_customized_samples': use_customized_samples,
'remove_accidental_hits': remove_accidental_hits,
'seed': seed
}
self.inputs = {'Logits': logits, 'Label': label.astype(np.int64)}
self.inputs = {'Logits': logits, 'Labels': labels.astype(np.int64)}
def set_data(self, num_classes, num_samples, seed, remove_accidental_hits):
label = np.array([[6, 12, 15, 5, 1], [0, 9, 4, 1, 10],
[0, 2, 10, 16, 13], [14, 4, 7, 2, 1],
[3, 18, 11, 8, 14]])
batch_size, num_true = label.shape
use_custom_samples = False
labels = np.array([[6, 12, 15, 5, 1], [0, 9, 4, 1, 10],
[0, 2, 10, 16, 13], [14, 4, 7, 2, 1],
[3, 18, 11, 8, 14]])
batch_size, num_true = labels.shape
use_customized_samples = False
num_sampled_classes = num_samples + num_true
logits = np.random.randn(batch_size, num_classes)
remove_accidental_hits = remove_accidental_hits
self.generate_data(logits, label, num_samples, seed,
remove_accidental_hits, use_custom_samples)
self.generate_data(logits, labels, num_samples, seed,
remove_accidental_hits, use_customized_samples)
# python and c++ use different random generator
# use fetched samples from c++ for python code
......@@ -302,7 +299,7 @@ class TestSampleLogitsOpV2(OpTest):
self.probabilities = probabilities
def compute(self):
out = sample_logits(self.inputs["Logits"], self.inputs["Label"],
out = sample_logits(self.inputs["Logits"], self.inputs["Labels"],
self.attrs["num_samples"], self.attrs["seed"],
self.attrs["remove_accidental_hits"], True,
self.fetched_samples.astype(np.int64),
......@@ -310,7 +307,7 @@ class TestSampleLogitsOpV2(OpTest):
self.outputs = {
'SampledLogits': out[0],
'Samples': out[1],
'SampledLabel': out[2],
'SampledLabels': out[2],
'Probabilities': out[3]
}
......@@ -339,18 +336,18 @@ class TestSampleLogitsOpV3(OpTest):
in C++ and copied to python and just test the non-random part.
'''
def generate_data(self, logits, label, num_samples, seed,
remove_accidental_hits, use_custom_samples):
def generate_data(self, logits, labels, num_samples, seed,
remove_accidental_hits, use_customized_samples):
self.attrs = {
'num_samples': num_samples,
'use_custom_samples': use_custom_samples,
'use_customized_samples': use_customized_samples,
'remove_accidental_hits': remove_accidental_hits,
'seed': seed
}
self.inputs = {'Logits': logits, 'Label': label.astype(np.int64)}
self.inputs = {'Logits': logits, 'Labels': labels.astype(np.int64)}
def set_data(self, num_classes, num_samples, seed, remove_accidental_hits):
label = [52, 2, 2, 17, 96, 2, 17, 96, 37, 2]
labels = [52, 2, 2, 17, 96, 2, 17, 96, 37, 2]
samples = [
3, 12, 74, 28, 1, 79, 2, 42, 8, 13, 0, 18, 88, 49, 14, 46, 39, 57,
26, 75, 9, 50, 16, 66, 6, 23, 5, 11, 17, 54, 35, 20, 53, 10, 47, 80,
......@@ -359,19 +356,19 @@ class TestSampleLogitsOpV3(OpTest):
63, 81, 59, 48, 91, 68, 72, 61, 52, 86
]
self.fetched_samples = np.array([[x] + samples for x in label])
self.fetched_samples = np.array([[x] + samples for x in labels])
fectched_num_tries = 323
label = self.fetched_samples[:, 0:1]
batch_size, num_true = label.shape
use_custom_samples = False
labels = self.fetched_samples[:, 0:1]
batch_size, num_true = labels.shape
use_customized_samples = False
num_sampled_classes = num_samples + num_true
logits = np.random.randn(batch_size, num_classes)
remove_accidental_hits = remove_accidental_hits
self.generate_data(logits, label, num_samples, seed,
remove_accidental_hits, use_custom_samples)
self.generate_data(logits, labels, num_samples, seed,
remove_accidental_hits, use_customized_samples)
# python and c++ use different random generator
# use fetched samples from c++ for python code
......@@ -388,7 +385,7 @@ class TestSampleLogitsOpV3(OpTest):
self.probabilities = probabilities
def compute(self):
out = sample_logits(self.inputs["Logits"], self.inputs["Label"],
out = sample_logits(self.inputs["Logits"], self.inputs["Labels"],
self.attrs["num_samples"], self.attrs["seed"],
self.attrs["remove_accidental_hits"], True,
self.fetched_samples.astype(np.int64),
......@@ -396,7 +393,7 @@ class TestSampleLogitsOpV3(OpTest):
self.outputs = {
'SampledLogits': out[0],
'Samples': out[1],
'SampledLabel': out[2],
'SampledLabels': out[2],
'Probabilities': out[3]
}
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或