perf(imperative): add dim_expansion transform for conv/bn1d

GitOrigin-RevId: d14a69424d

imperative/python/megengine/functional/nn.py

@@ -41,7 +41,6 @@ from ..distributed import WORLD, is_distributed
 from ..jit import exclude_from_trace
 from ..tensor import Tensor
 from ..utils.deprecation import deprecated_func
-from ..utils.tuple_function import _pair, _pair_nonzero, _triple, _triple_nonzero
 from .debug_param import get_execution_strategy
 from .distributed import all_reduce_sum
 from .elemwise import _elwise, exp, log, log1p, maximum, minimum
@@ -94,14 +93,15 @@ __all__ = [
 
 
 def expand_hw(x):
-    # NOTE: >1d array is accepted, as long as 1 <= size <= 2
-    try:
-        x = int(x)
-        return [x, x]
-    except (TypeError, ValueError):
-        pass
-    h, w = x
-    return int(h), int(w)
+    if isinstance(x, Sequence):
+        return int(x[0]), int(x[1])
+    return int(x), int(x)
+
+
+def expand_dhw(x):
+    if isinstance(x, Sequence):
+        return int(x[0]), int(x[1]), int(x[2])
+    return int(x), int(x), int(x)
 
 
 def linear(
@@ -177,11 +177,8 @@ def conv1d(
         if weight.dtype != dtype:
             weight = weight.astype(dtype)
 
-    inp = expand_dims(inp, 3)
-    weight = expand_dims(weight, 3)
     if bias is not None:
         assert bias.ndim == 3, "the bias dimension of conv1d should be 3"
-        bias = expand_dims(bias, 3)
 
     stride_h = stride
     pad_h = padding
@@ -206,7 +203,6 @@ def conv1d(
     (output,) = apply(op, inp, weight)
     if bias is not None:
         output += bias
-    output = squeeze(output, 3)
     return output
 
 
@@ -314,9 +310,9 @@ def conv3d(
 
     D, H, W = 0, 1, 2
 
-    pad = _triple(padding)
-    stride = _triple_nonzero(stride)
-    dilate = _triple_nonzero(dilation)
+    pad = expand_dhw(padding)
+    stride = expand_dhw(stride)
+    dilate = expand_dhw(dilation)
 
     sparse_type = "dense" if groups == 1 else "group"
     op = builtin.Convolution3D(
@@ -572,9 +568,9 @@ def conv_transpose3d(
         output tensor.
     """
     D, H, W = 0, 1, 2
-    pad = _triple(padding)
-    stride = _triple_nonzero(stride)
-    dilate = _triple_nonzero(dilation)
+    pad = expand_dhw(padding)
+    stride = expand_dhw(stride)
+    dilate = expand_dhw(dilation)
 
     sparse_type = "dense" if groups == 1 else "group"
     op = builtin.Convolution3DBackwardData(
@@ -618,9 +614,9 @@ def max_pool2d(
     """
     if stride is None:
         stride = kernel_size
-    window_h, window_w = _pair_nonzero(kernel_size)
-    stride_h, stride_w = _pair_nonzero(stride)
-    padding_h, padding_w = _pair(padding)
+    window_h, window_w = expand_hw(kernel_size)
+    stride_h, stride_w = expand_hw(stride)
+    padding_h, padding_w = expand_hw(padding)
     conv_format = _config._get_actual_op_param("NCHW", _config.__conv_format)
 
     op = builtin.Pooling(
@@ -662,9 +658,9 @@ def avg_pool2d(
     """
     if stride is None:
         stride = kernel_size
-    window_h, window_w = _pair_nonzero(kernel_size)
-    stride_h, stride_w = _pair_nonzero(stride)
-    padding_h, padding_w = _pair(padding)
+    window_h, window_w = expand_hw(kernel_size)
+    stride_h, stride_w = expand_hw(stride)
+    padding_h, padding_w = expand_hw(padding)
     conv_format = _config._get_actual_op_param("NCHW", _config.__conv_format)
 
     op = builtin.Pooling(
@@ -1779,10 +1775,10 @@ def sliding_window(
         stride: stride of the window. Default: 1
         dilation: dilation of the window. Default: 1
     """
-    padding_h, padding_w = _pair(padding)
-    stride_h, stride_w = _pair_nonzero(stride)
-    dilation_h, dilation_w = _pair_nonzero(dilation)
-    window_h, window_w = _pair_nonzero(kernel_size)
+    padding_h, padding_w = expand_hw(padding)
+    stride_h, stride_w = expand_hw(stride)
+    dilation_h, dilation_w = expand_hw(dilation)
+    window_h, window_w = expand_hw(kernel_size)
 
     op = builtin.Images2Neibs(
         pad_h=padding_h,
@@ -1818,11 +1814,11 @@ def sliding_window_transpose(
         stride: stride of the window. Default: 1
         dilation: dilation of the window. Default: 1
     """
-    output_h, output_w = _pair_nonzero(output_size)
-    padding_h, padding_w = _pair(padding)
-    stride_h, stride_w = _pair_nonzero(stride)
-    dilation_h, dilation_w = _pair_nonzero(dilation)
-    window_h, window_w = _pair_nonzero(kernel_size)
+    output_h, output_w = expand_hw(output_size)
+    padding_h, padding_w = expand_hw(padding)
+    stride_h, stride_w = expand_hw(stride)
+    dilation_h, dilation_w = expand_hw(dilation)
+    window_h, window_w = expand_hw(kernel_size)
 
     expected_h = (
         output_h + 2 * padding_h - dilation_h * (window_h - 1) - 1

imperative/python/megengine/module/batchnorm.py

@@ -80,19 +80,6 @@ class _BatchNorm(Module):
                 self.track_running_stats == False
             ), "track_running_stats can not be initilized to False and changed to True later"
 
-        inp_shape = inp.shape
-        _ndims = len(inp_shape)
-        if _ndims != 4:
-            origin_shape = inp_shape
-            if _ndims == 2:
-                n, c = inp_shape[0], inp_shape[1]
-                new_shape = (n, c, 1, 1)
-            elif _ndims == 3:
-                n, c, h = inp_shape[0], inp_shape[1], inp_shape[2]
-                new_shape = (n, c, h, 1)
-
-            inp = inp.reshape(new_shape)
-
         _weight = self.weight
         _bias = self.bias
 
@@ -130,9 +117,6 @@ class _BatchNorm(Module):
             param_dim=self.param_dim,
         )
 
-        if _ndims != 4:
-            output = output.reshape(origin_shape)
-
         return output
 
     def _module_info_string(self) -> str:

imperative/python/src/tensor.cpp

@@ -15,6 +15,7 @@
 #include "megbrain/imperative/ops/backward_graph.h"
 #include "megbrain/imperative/ops/utility.h"
 #include "megbrain/imperative/profiler.h"
+#include "megbrain/imperative/transformations/dim_expansion.h"
 #include "megbrain/imperative/transformations/dtype_promote.h"
 #include "megbrain/imperative/transformations/eval.h"
 #include "megbrain/imperative/transformations/lazy.h"
@@ -61,11 +62,13 @@ struct SymbolVarContext {
     std::shared_ptr<SymbolTransformation> symbol_tsf;
     std::shared_ptr<ScalarTransformation> scalar_tsf;
     std::shared_ptr<DTypePromoteTransformation> dtype_promote_tsf;
+    std::shared_ptr<DimExpansionTransformation> dim_expansion_tsf;
 
     SymbolVarContext(cg::ComputingGraph* graph) {
         symbol_tsf = std::make_shared<SymbolTransformation>(graph);
         scalar_tsf = std::make_shared<ScalarTransformation>();
         dtype_promote_tsf = std::make_shared<DTypePromoteTransformation>();
+        dim_expansion_tsf = std::make_shared<DimExpansionTransformation>();
         Transformation::swap_context(context);
     }
 
@@ -73,6 +76,7 @@ struct SymbolVarContext {
         symbol_tsf->register_at(Transformation::top());
         scalar_tsf->register_at(Transformation::top());
         dtype_promote_tsf->register_at(Transformation::top());
+        dim_expansion_tsf->register_at(Transformation::top());
     }
 
     ValueRef symvar2val(py::handle py_symbol_var) {
@@ -452,6 +456,8 @@ void init_tensor(py::module m) {
             std::make_shared<ScalarTransformation>());
     transformations.register_at<Segment::DTypePromote>(
             std::make_shared<DTypePromoteTransformation>());
+    transformations.register_at<Segment::DimExpansion>(
+            std::make_shared<DimExpansionTransformation>());
 
     static py::exception<interpreter::AsyncError> py_async_error(
             m, "AsyncError", PyExc_RuntimeError);

imperative/python/src/transformation.h

@@ -26,13 +26,14 @@ struct TransformationManager {
     enum Segment {
         ModuleTrace,
         DTypePromote,
+        DimExpansion,
         Grad,
         Scalar,
         Trace,
         Eval,
     };
 
-    std::array<std::vector<std::shared_ptr<Transformation>>, 6> segments;
+    std::array<std::vector<std::shared_ptr<Transformation>>, 7> segments;
 
     template <Segment segment>
     void register_at(std::shared_ptr<Transformation> transformation) {

imperative/python/test/integration/test_dtr.py

@@ -91,7 +91,7 @@ class ResNet(M.Module):
 
 
 def run_dtr_resnet1202():
-    batch_size = 8
+    batch_size = 7
     resnet1202 = ResNet(BasicBlock, [200, 200, 200])
     opt = optim.SGD(resnet1202.parameters(), lr=0.05, momentum=0.9, weight_decay=1e-4)
     gm = GradManager().attach(resnet1202.parameters())

imperative/src/impl/transformations/dim_expansion.cpp

@@ -0,0 +1,95 @@
+#include "megbrain/imperative/transformations/dim_expansion.h"
+#include "megbrain/imperative/ops/autogen.h"
+
+namespace mgb::imperative {
+
+namespace {
+using DimExpansionRule = std::function<ValueRefList(const OpDef&, Span<ValueRef>)>;
+static std::unordered_map<Typeinfo*, DimExpansionRule> dim_expansion_rules;
+
+template <typename T>
+void register_dim_expansion_rules(const DimExpansionRule& rule) {
+    dim_expansion_rules[T::typeinfo()] = [rule](const OpDef& def,
+                                                Span<ValueRef> inputs) {
+        return rule(def.cast_final_safe<T>(), inputs);
+    };
+}
+
+ValueRefList conv1d_rule(const OpDef& op, Span<ValueRef> inputs) {
+    bool need_expand = inputs.at(0).shape()->ndim == 3;
+    if (!need_expand)
+        return imperative::apply(op, inputs);
+
+    ValueRefList converted(inputs.size());
+    std::vector<int32_t> axis = {(int32_t)3};
+    for (size_t i = 0; i < inputs.size(); ++i) {
+        converted[i] = imperative::apply(ApplyOp(*AddAxis::make(axis)), inputs[i])[0];
+    }
+
+    auto outputs = imperative::apply(op, converted);
+    outputs[0] = imperative::apply(ApplyOp(*RemoveAxis::make(axis)), outputs[0])[0];
+    return outputs;
+}
+
+ValueRefList bn1d_rule(const OpDef& op, Span<ValueRef> inputs) {
+    size_t ndim = inputs.at(0).shape()->ndim;
+    bool need_expand = (ndim == 2 || ndim == 3);
+    if (!need_expand)
+        return imperative::apply(op, inputs);
+
+    ValueRefList converted(inputs.size());
+    std::vector<int32_t> axis = {(int32_t)3};
+    if (ndim == 2) {
+        axis.insert(axis.begin(), (int32_t)2);
+    }
+    converted[0] = imperative::apply(ApplyOp(*AddAxis::make(axis)), inputs[0])[0];
+    for (size_t i = 1; i < inputs.size(); ++i) {
+        converted[i] = inputs[i];
+    }
+
+    std::reverse(std::begin(axis), std::end(axis));
+    auto outputs = imperative::apply(op, converted);
+    size_t idx = outputs.size() - 1;
+    outputs[idx] = imperative::apply(ApplyOp(*RemoveAxis::make(axis)), outputs[idx])[0];
+    return outputs;
+}
+
+struct DimExpansionRuleRegistry {
+    DimExpansionRuleRegistry() {
+        register_dim_expansion_rules<Convolution>(conv1d_rule);
+        register_dim_expansion_rules<BatchNorm>(bn1d_rule);
+    }
+} register_helper;
+
+}  // namespace
+
+ValueRefList DimExpansionTransformation::apply_transformation(
+        const Operator& op, Span<ValueRef> inputs) {
+    if (auto apply_op = op.as<ApplyOp>()) {
+        auto iter = dim_expansion_rules.find(apply_op->op().dyn_typeinfo());
+        if (iter != dim_expansion_rules.end()) {
+            return iter->second(apply_op->op(), inputs);
+        } else {
+            return imperative::apply(op, inputs);
+        }
+    }
+    return imperative::apply(op, inputs);
+}
+
+ValueRef DimExpansionTransformation::unwrap(ValueRef value) {
+    return value;
+}
+
+std::string DimExpansionTransformation::name() const {
+    return "DimExpansionTransformation";
+}
+
+void DimExpansionTransformation::on_register() {
+    // printf("DimExpansionTransformation has been registered\n");
+}
+
+void DimExpansionTransformation::on_unregister() noexcept {
+    // printf("DimExpansionTransformation has been unregistered\n");
+}
+
+}  // namespace mgb::imperative

imperative/src/include/megbrain/imperative/transformations/dim_expansion.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include "megbrain/imperative/dispatch.h"
+#include "megbrain/imperative/value.h"
+
+namespace mgb::imperative {
+
+class DimExpansionTransformation final : public Transformation {
+private:
+public:
+    ValueRefList apply_transformation(
+            const Operator& op, Span<ValueRef> inputs) override;
+    ValueRef unwrap(ValueRef value) override;
+    std::string name() const override;
+    void on_register() override;
+    void on_unregister() noexcept override;
+};
+
+}  // namespace mgb::imperative