Add nce op #5480

wanghaoshuang · 2017-11-08T07:03:33Z

Add forward and backward kernel
Add unitest
Add comment and doc

1. Add nce forward and backward kernel for CPU

… nce_op

zhouxiao-coder · 2017-11-16T05:04:31Z

I think TensorFlow has a nice design about candidate sampling methods and we can use it as a reference. It supports not only NCE, but also importance sampling and negative sampling, which are also popular methods. It also naturally adapts to GPU version.

lcy-seso · 2017-11-16T06:10:00Z

@zhouxiao-coder We have already discussed with @wanghaoshuang about systematically support the sampling methods. It will be refined later. In this PR, we only want to quickly port the old NCE layer into the new framework.

guoshengCS · 2017-11-20T07:34:09Z

paddle/operators/nce_op.cc

+    // set dims of output(Out)
+    std::vector<int64_t> out_dims;
+    out_dims.push_back(x_dims[0]);
+    ctx->SetOutputDim("Cost", framework::make_ddim(out_dims));


The output shape is [batch_size, 1] in smooth_l1, cos_sim, sequared_l2_distance, softmax_with_cross_entropy, and this can be discussed and unified later.

guoshengCS · 2017-11-20T07:35:15Z

paddle/operators/nce_op.cc

+    AddOutput("SampleLogits", "An intermediate tensor.").AsIntermediate();
+    AddOutput("SampleLabels", "An intermediate tensor.").AsIntermediate();
+    AddAttr<int>("num_classes", "Total number of classes.");
+    AddAttr<int>("num_sampled_classes", "The number of negative classes.")


Might num_neg_samples be better if it indicates the number of negative samples for each positive sample. I just compare with the NCELayer, but I am not sure what num_sampled_classes represents for and which is better. Maybe I just haven't got the point.

Agree with @guoshengCS . But maybe can use num_neg_classes to also incorporate @wanghaoshuang ‘s original consideration.

Renamed num_sampled_classes to num_neg_samples

guoshengCS · 2017-11-20T07:36:20Z

paddle/operators/nce_op.cc

+    std::vector<int> sampled_labels =
+        ctx->Attrs().Get<std::vector<int>>("sampled_labels");
+    PADDLE_ENFORCE_EQ(num_classes, ctx->GetInputDim("Weight")[0]);
+    PADDLE_ENFORCE_LT(num_sampled_classes, num_classes);


If num_sampled_classes indicates the number of negative samples, the above PADDLE_ENFORCE_LT may not be needed.

num_classes means the total number of classes in all samples. So i think this checking is necessary.
如果sampled_classes中有重复的class, 是有可能num_sampled_classes > num_classes的, 也不能说噪声样本数量大于num_classes就是错的，但是从计算性能上考虑，噪声样本数量大于num_classes是没有必要的，所以有了这里的checking.
另外paddle v2和tesoflow中都没有这个限制。

我还是先去掉这个checking吧

guoshengCS · 2017-11-20T09:44:48Z

paddle/operators/nce_op.cc

+    AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim].");
+    AddInput("Label",
+             "(Tensor) A tensor of shape [batch_size, num_true_class]. "
+             "'num_true_class' is the number of target class in each sample.");


Is this used for multi-label and must all samples have the same label number.

It would be useful to allow a variable number of target classes per example. For now, if you have a variable number of target classes, you can pad them out to a constant number by either repeating them or by padding with an otherwise unused class.

guoshengCS · 2017-11-20T12:56:53Z

paddle/operators/nce_op.h

+      }
+    } else {
+      for (; j < sample_labels_dims[1]; ++j) {
+        sample_labels_data[index++] = rand(rng);


It seems that uniform distribution is the only supported temporarily. TODO can be added for future work.

guoshengCS · 2017-11-20T15:35:15Z

paddle/operators/nce_op.h

+    auto input_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
+    auto weight_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
+    for (size_t i = 0; i < sample_labels->numel(); ++i) {
+      Eigen::Tensor<float, 0, Eigen::RowMajor, Eigen::DenseIndex> result =


Can EigenScalar in the framework be used here. I am not sure.

EigenScalar will cause compiling error:

错误：请求从‘const Eigen::TensorReductionOp<Eigen::internal::SumReducer<float>, const Eigen::DimensionList<long int, 1ul>, const Eigen::TensorCwiseBinaryOp<Eigen::internal::scalar_product_op<const float, const float>, const Eigen::TensorChippingOp<-1l, const Eigen::TensorMap<Eigen::Tensor<const float, 2, 1, long int>, 0, Eigen::MakePointer> >, const Eigen::TensorChippingOp<-1l, const Eigen::TensorMap<Eigen::Tensor<const float, 2, 1, long int>, 0, Eigen::MakePointer> > >, Eigen::MakePointer>’转换到非标量类型‘paddle::operators::EigenScalar<float, 1, long int> {aka paddle::framework::EigenScalar<float, 1, long int>}’ .sum();

guoshengCS · 2017-11-21T04:27:16Z

paddle/operators/nce_op.h

+              .sum();
+      sample_out_data[i] += result(0);
+      // activation_->forward
+      sample_out_data[i] = (1. / (1. + exp(-sample_out_data[i])));


Maybe only output values of positive samples are needed to compute, and some computation can be reduced in future.

guoshengCS · 2017-11-21T04:51:19Z

paddle/operators/nce_op.h

+    if (d_bias != nullptr) {
+      T* d_bias_data = d_bias->mutable_data<T>(context.GetPlace());
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_bias_data[sample_labels_data[i]] += sample_grad_data[i];


I am not sure if the grad_data should be clear to zero first.

guoshengCS · 2017-11-21T05:03:12Z

python/paddle/v2/fluid/tests/test_nce.py

+            samples.append((i, num, False, w))
+            sample_labels.append(num)
+    # forward bias
+    sampleOut = np.zeros(len(samples)).astype(np.float32)


sample_out might be better to unify name style.

lcy-seso

Some simple comments first.

lcy-seso · 2017-11-21T06:31:00Z

paddle/operators/nce_op.cc

+             "total number of class.");
+    AddInput("Bias",
+             "(Tensor) A tensor of shape [num_class]. 'num_class' is the total "
+             "number of class. It is a dispensable input.")


I remember we have discussed before and decide to use the 2-dimensional tensor to represent a vector to distinguish it is a row vector or a column vector explicitly.

lcy-seso · 2017-11-21T06:37:36Z

paddle/operators/nce_op.cc

+    AddOutput("SampleLogits", "An intermediate tensor.").AsIntermediate();
+    AddOutput("SampleLabels", "An intermediate tensor.").AsIntermediate();
+    AddAttr<int>("num_classes", "Total number of classes.");
+    AddAttr<int>("num_sampled_classes", "The number of negative classes.")


Agree with @guoshengCS . But maybe can use num_neg_classes to also incorporate @wanghaoshuang ‘s original consideration.

lcy-seso · 2017-11-21T06:41:57Z

paddle/operators/nce_op.cc

+    AddInput("SampleWeight",
+             "(Tensor) A tensor of shape [batch_size] storing a weight for "
+             "each sample. And it is a dispensable input. The default value of "
+             "sample is 1.")


The same problem as Input(Bias) about tensor's shape.

lcy-seso · 2017-11-21T06:43:04Z

paddle/operators/nce_op.cc

+             "sample is 1.")
+        .AsDispensable();
+    AddOutput("Cost",
+              "(Tensor) A tensor of shape [batch_size]. Cost of samples.");


The same problem as Input(Bias) about tensor's shape.

lcy-seso · 2017-11-21T06:46:28Z

paddle/operators/nce_op.cc

+        .SetDefault(10);
+    AddAttr<std::vector<int>>("sampled_labels", "");
+    AddComment(R"DOC(
+Computes and returns the noise-contrastive estimation training loss.


Compute and return

lcy-seso · 2017-11-21T07:02:10Z

paddle/operators/nce_op.cc

+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */


Please make the indention of license follows that in accuracy_op.h.

lcy-seso · 2017-11-21T07:14:07Z

paddle/operators/nce_op.cc

+    AddOutput("Cost",
+              "(Tensor) A tensor of shape [batch_size]. Cost of samples.");
+    AddOutput("SampleLogits", "An intermediate tensor.").AsIntermediate();
+    AddOutput("SampleLabels", "An intermediate tensor.").AsIntermediate();


Comments on Output(SampleLogits) and Output(SampleLabels). What are these intermediate outputs used for?

Fixed by adding more comments.

lcy-seso · 2017-11-21T07:15:43Z

paddle/operators/nce_op.cc

+    PADDLE_ENFORCE(ctx->HasInput("SampleLogits"));
+    PADDLE_ENFORCE(ctx->HasInput("SampleLabels"));
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Cost")),
+                   "The input(Out@GRAD) should not be null");


If the comment is a complete sentence, please add a period at the end of the sentence.

lcy-seso · 2017-11-21T07:18:37Z

paddle/operators/nce_op.h

+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;


using framework::Tensor;

lcy-seso · 2017-11-21T07:40:05Z

paddle/operators/nce_op.cc

+ public:
+  NCEOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim].");


I have a question here. From the current implementation, this operator only supports uniform distribution which is hardcoded. How can we extend the codes to support more distribution? Should we make distribution method an attribute, or leave the sampling process outside this operator? What is your suggestion?

Yes. You are right. More distribution sampler is necessary. The terms in NLP application always meet heavy-tailed distribution. So we need a long uniform distribution sampler. And the distribution`s PDF is 1/(x+1)lnR in which R is the range of sampling. I will implement the log uniform sampler and other common distribution samplers as independent math function in another pr.

1. Remove checking for num_neg_samples. 2. Fix dims of Output(Cost) and Input(Bias). 3. Renamed num_sampled_classes to num_neg_samples. 4. Add TODO for add more distribution sampler. 5. Init grad_data of bias by zero. 6. Refine comments. 7. Register a kernel for type double.

… nce_op

lcy-seso

LGTM

lcy-seso · 2017-12-01T06:58:55Z

paddle/operators/nce_op.h

+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */


The indentation should like that in nce_op.cc

wanghaoshuang added 2 commits November 8, 2017 14:17

Add nce op

3d27627

1. Add nce forward and backward kernel for CPU

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

c2dd75b

… nce_op

qingqing01 added the OpPorting label Nov 8, 2017

wanghaoshuang added 3 commits November 16, 2017 00:01

Add unitest and comments.

09d32b0

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

668dc53

… nce_op

fix unitest

e60eb1e

qingqing01 requested review from lcy-seso, qingqing01 and guoshengCS November 16, 2017 04:38

qingqing01 requested a review from chengduoZH November 20, 2017 11:31

guoshengCS reviewed Nov 21, 2017

View reviewed changes

lcy-seso reviewed Nov 21, 2017

View reviewed changes

wanghaoshuang added 5 commits November 28, 2017 12:47

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

d0246e2

… nce_op

Fix double type error while using eigen api

ab9d59c

Fix comments style

76a65a8

Fix unitest.

29262ab

lcy-seso approved these changes Dec 1, 2017

View reviewed changes

wanghaoshuang merged commit 10b2534 into PaddlePaddle:develop Dec 1, 2017

wanghaoshuang deleted the nce_op branch May 20, 2022 03:56

Add nce op #5480

Add nce op #5480

Conversation

wanghaoshuang commented Nov 8, 2017 • edited Loading

zhouxiao-coder commented Nov 16, 2017

lcy-seso commented Nov 16, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wanghaoshuang Nov 24, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wanghaoshuang Nov 24, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lcy-seso left a comment

Choose a reason for hiding this comment

lcy-seso Nov 21, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wanghaoshuang Nov 27, 2017 • edited Loading

Choose a reason for hiding this comment

lcy-seso left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wanghaoshuang commented Nov 8, 2017 •

edited

Loading

lcy-seso commented Nov 16, 2017 •

edited

Loading

wanghaoshuang Nov 24, 2017 •

edited

Loading

wanghaoshuang Nov 24, 2017 •

edited

Loading

lcy-seso Nov 21, 2017 •

edited

Loading

wanghaoshuang Nov 27, 2017 •

edited

Loading