Skip to content

Commit

Permalink
[TOP] Level1 complete (#3)
Browse files Browse the repository at this point in the history
  • Loading branch information
@tqchen
tqchen committed May 29, 2018
1 parent be1660b commit 986caf7
Show file tree
Hide file tree
Showing 9 changed files with 355 additions and 170 deletions.
58 changes: 58 additions & 0 deletions nnvm/include/nnvm/top/nn.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ namespace top {
struct DenseParam : public dmlc::Parameter<DenseParam> {
int units;
bool use_bias;

DMLC_DECLARE_PARAMETER(DenseParam) {
DMLC_DECLARE_FIELD(units).set_lower_bound(1)
.describe("Number of hidden units of the dense transformation.");
Expand All @@ -27,6 +28,63 @@ struct DenseParam : public dmlc::Parameter<DenseParam> {
static const constexpr int kBias = 2;
};

struct DropoutParam : public dmlc::Parameter<DropoutParam> {
float rate;

DMLC_DECLARE_PARAMETER(DropoutParam) {
DMLC_DECLARE_FIELD(rate).set_default(0.5)
.set_range(0, 1)
.describe("Fraction of the input that gets dropped out during training time.");
}
};

struct BatchNormParam : public dmlc::Parameter<BatchNormParam> {
int axis;
float epsilon;
float momentum;
bool center;
bool scale;

DMLC_DECLARE_PARAMETER(BatchNormParam) {
DMLC_DECLARE_FIELD(axis).set_default(1)
.describe("Specify which shape axis the channel is specified.");
DMLC_DECLARE_FIELD(epsilon).set_default(1e-5f)
.describe("Small float added to variance to avoid dividing by zero.");
DMLC_DECLARE_FIELD(center).set_default(true)
.describe("If True, add offset of `beta` to normalized tensor."
"If False, `beta` is ignored.");
DMLC_DECLARE_FIELD(scale).set_default(true)
.describe("If True, multiply by `gamma`. If False, `gamma` is not used."
"When the next layer is piecewise linear (also e.g. `nn.relu`),"
"this can be disabled since the scaling"
"will be done by the next layer.");
}
// constants
static const constexpr int kData = 0;
static const constexpr int kGamma = 1;
static const constexpr int kBeta = 2;
static const constexpr int kMovingMean = 3;
static const constexpr int kMovingVariance = 4;
};

struct SoftmaxParam : public dmlc::Parameter<SoftmaxParam> {
int axis;

DMLC_DECLARE_PARAMETER(SoftmaxParam) {
DMLC_DECLARE_FIELD(axis).set_default(-1)
.describe("The axis to sum over when computing softmax.");
}
};

struct LogSoftmaxParam : public dmlc::Parameter<LogSoftmaxParam> {
int axis;

DMLC_DECLARE_PARAMETER(LogSoftmaxParam) {
DMLC_DECLARE_FIELD(axis).set_default(-1)
.describe("The axis to sum over when computing softmax.");
}
};

} // namespace top
} // namespace nnvm

Expand Down
31 changes: 27 additions & 4 deletions nnvm/include/nnvm/top/tensor.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9,14 +9,37 @@
namespace nnvm {
namespace top {

struct ConcatParam : public dmlc::Parameter<ConcatParam> {
int dim;
DMLC_DECLARE_PARAMETER(ConcatParam) {
DMLC_DECLARE_FIELD(dim).set_range(0, 4).set_default(1)
struct ConcatenateParam : public dmlc::Parameter<ConcatenateParam> {
int axis;
DMLC_DECLARE_PARAMETER(ConcatenateParam) {
DMLC_DECLARE_FIELD(axis).set_lower_bound(0).set_default(1)
.describe("the axis to be concated.");
}
};

enum TypeFlag {
kFloat32 = 0,
kFloat64 = 1,
kFloat16 = 2,
kUint8 = 3,
kInt32 = 4,
kInt8 = 5,
kInt64 = 6,
};

struct CastParam : public dmlc::Parameter<CastParam> {
int dtype;
DMLC_DECLARE_PARAMETER(CastParam) {
DMLC_DECLARE_FIELD(dtype)
.add_enum("float32", kFloat32)
.add_enum("float64", kFloat64)
.add_enum("float16", kFloat16)
.add_enum("uint8", kUint8)
.add_enum("int32", kInt32)
.describe("Output data type.");
}
};

} // namespace top
} // namespace nnvm

Expand Down
25 changes: 20 additions & 5 deletions nnvm/src/top/elemwise_op_common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ inline bool ElemwiseAttr(const nnvm::NodeAttrs& attrs,
}

template<int n_in, int n_out>
inline bool ElemwiseShape(const nnvm::NodeAttrs& attrs,
inline bool ElemwiseShape(const NodeAttrs& attrs,
std::vector<TShape> *in_attrs,
std::vector<TShape> *out_attrs) {
if (n_in != -1) {
Expand All @@ -71,7 +71,7 @@ inline bool ElemwiseShape(const nnvm::NodeAttrs& attrs,
}

template<int n_in, int n_out>
inline bool ElemwiseType(const nnvm::NodeAttrs& attrs,
inline bool ElemwiseType(const NodeAttrs& attrs,
std::vector<int> *in_attrs,
std::vector<int> *out_attrs) {
if (n_in != -1) {
Expand All @@ -88,13 +88,28 @@ inline bool ElemwiseType(const nnvm::NodeAttrs& attrs,
NNVM_REGISTER_OP(name) \
.set_num_inputs(1) \
.set_num_outputs(1) \
.set_attr<nnvm::FInferShape>("FInferShape", ElemwiseShape<1, 1>) \
.set_attr<nnvm::FInferType>("FInferType", ElemwiseType<1, 1>) \
.set_attr<nnvm::FInplaceOption>("FInplaceOption", \
.set_attr<FInferShape>("FInferShape", ElemwiseShape<1, 1>) \
.set_attr<FInferType>("FInferType", ElemwiseType<1, 1>) \
.set_attr<FInplaceOption>("FInplaceOption", \
[](const NodeAttrs& attrs){ \
return std::vector<std::pair<int, int> >{{0, 0}}; \
}) \
.add_argument("data", "Tensor", "The input tensor.")


#define NNVM_REGISTER_ELEMWISE_BINARY_OP(name) \
NNVM_REGISTER_OP(name) \
.set_num_inputs(2) \
.set_num_outputs(1) \
.set_attr<FInferShape>("FInferShape", ElemwiseShape<2, 1>) \
.set_attr<FInferType>("FInferType", ElemwiseType<2, 1>) \
.set_attr<FInplaceOption>("FInplaceOption", \
[](const NodeAttrs& attrs) { \
return std::vector<std::pair<int, int> >{{0, 0}, {1, 0}}; \
}) \
.add_argument("lhs", "NDArray-or-Symbol", "first input") \
.add_argument("rhs", "NDArray-or-Symbol", "second input")

} // namespace top
} // namespace nnvm
#endif // NNVM_TOP_ELEMWISE_OP_COMMON_H_
118 changes: 117 additions & 1 deletion nnvm/src/top/nn.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -73,7 +73,7 @@ If ``use_bias`` is set to be false, then the ``bias`` term is ignored.
.set_attr_parser(ParamParser<DenseParam>)
.set_num_outputs(1)
.set_num_inputs([](const NodeAttrs& attrs) {
const DenseParam& param = nnvm::get<DenseParam>(attrs.parsed);
const DenseParam& param = get<DenseParam>(attrs.parsed);
return param.use_bias ? 3 : 2;
})
.set_attr<FListInputNames>("FListInputNames", DenseListInputNames)
Expand All @@ -90,5 +90,121 @@ NNVM_REGISTER_ELEMWISE_UNARY_OP(relu)
)code" NNVM_ADD_FILELINE)
.set_support_level(1);

// dropout
DMLC_REGISTER_PARAMETER(DropoutParam);

NNVM_REGISTER_OP(dropout)
.describe(R"(Applies dropout operation to input array.
- During training, each element of the input is set to zero with probability p.
The whole array is rescaled by :math:`1/(1-p)` to keep the expected
sum of the input unchanged.
)" NNVM_ADD_FILELINE)
.add_argument("data", "Tensor", "Input to which dropout will be applied")
.set_num_inputs(1)
.set_num_outputs(2)
.set_attr_parser(ParamParser<DropoutParam>)
.set_attr<FInferShape>("FInferShape", ElemwiseShape<1, 2>)
.set_attr<FInferType>("FInferType", ElemwiseType<1, 2>)
.set_attr<FNumVisibleOutputs>("FNumVisibleOutputs", [](const NodeAttrs& attrs) {
return 1;
})
.set_attr<FListOutputNames>("FListOutputNames", [](const NodeAttrs& attrs) {
return std::vector<std::string>{"output", "mask"};
})
.set_support_level(1);

// batchnorm
DMLC_REGISTER_PARAMETER(BatchNormParam);

NNVM_REGISTER_OP(batch_norm)
.describe(R"(Batch normalization layer (Ioffe and Szegedy, 2014).
Normalizes the input at each batch, i.e. applies a transformation
that maintains the mean activation close to 0 and the activation
standard deviation close to 1.
.. math::
data\_mean[i] = mean(data[:,i,:,...]) \\
data\_var[i] = var(data[:,i,:,...])
Then compute the normalized output, which has the same shape as input, as following:
.. math::
out[:,i,:,...] = \frac{data[:,i,:,...] - data\_mean[i]}{\sqrt{data\_var[i]+\epsilon}} * gamma[i] + beta[i]
Both *mean* and *var* returns a scalar by treating the input as a vector.
Assume the input has size *k* on axis 1, then both ``gamma`` and ``beta`` have shape *(k,)*.
Besides the inputs and the outputs, this operator accepts two auxiliary
states, ``moving_mean`` and ``moving_var``, which are *k*-length
vectors. They are global statistics for the whole dataset, which are updated
by::
moving_mean = moving_mean * momentum + data_mean * (1 - momentum)
moving_var = moving_var * momentum + data_var * (1 - momentum)
The parameter ``axis`` specifies which axis of the input shape denotes
the 'channel' (separately normalized groups). The default is 1. Specifying -1 sets the channel
axis to be the last item in the input shape.
)" NNVM_ADD_FILELINE)
.add_argument("data", "Tensor", "Input to which dropout will be applied")
.add_argument("gamma", "Tensor", "The gamma scale factor")
.add_argument("beta", "Tensor", "The beta offset factor")
.add_argument("moving_mean", "Tensor", "running mean of input")
.add_argument("moving_var", "Tensor", "running variance of input")
.set_num_inputs(5)
.set_num_outputs(3)
.set_attr_parser(ParamParser<BatchNormParam>)
.set_attr<FListInputNames>("FListInputNames", [](const NodeAttrs& attrs) {
return std::vector<std::string>{"data", "gamma", "beta", "moving_mean", "moving_var"};
})
.set_attr<FListOutputNames>("FListOutputNames", [](const NodeAttrs& attrs) {
return std::vector<std::string>{"output", "mean", "var"};
})
.set_attr<FNumVisibleOutputs>("FNumVisibleOutputs", [](const NodeAttrs& attrs) {
return 1;
})
.set_attr<FMutateInputs>("FListMutateInputs", [](const NodeAttrs& attrs) {
return std::vector<uint32_t>{3, 4};
})
.set_support_level(1);

// softmax
DMLC_REGISTER_PARAMETER(SoftmaxParam);

NNVM_REGISTER_OP(softmax)
.describe(R"code(Computes softmax.
.. math:: \text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}
)code" NNVM_ADD_FILELINE)
.set_num_inputs(1)
.set_num_outputs(1)
.set_attr_parser(ParamParser<SoftmaxParam>)
.set_attr<FInferShape>("FInferShape", ElemwiseShape<1, 1>)
.set_attr<FInferType>("FInferType", ElemwiseType<1, 1>)
.set_support_level(1);

// log_softmax
DMLC_REGISTER_PARAMETER(LogSoftmaxParam);

NNVM_REGISTER_OP(log_softmax)
.describe(R"code(Computes softmax.
.. math:: \text{log_softmax}(x)_i = \log \frac{exp(x_i)}{\sum_j exp(x_j)}
)code" NNVM_ADD_FILELINE)
.set_num_inputs(1)
.set_num_outputs(1)
.set_attr_parser(ParamParser<LogSoftmaxParam>)
.set_attr<FInferShape>("FInferShape", ElemwiseShape<1, 1>)
.set_attr<FInferType>("FInferType", ElemwiseType<1, 1>)
.set_support_level(1);

} // namespace top
} // namespace nnvm
Loading

0 comments on commit 986caf7

Please sign in to comment.