Aggregated zero grad (#16446)

* Trigger CI

* Aggregated zeroing of the gradients/arrays

* New files for aggregated zeroing of the gradients/arrays

* Adding possibility to reset the arrays of different types.

* Minor cleanup

Author: drivanov

python/mxnet/gluon/parameter.py

@@ -24,9 +24,10 @@
            'ParameterDict', 'tensor_types']
 
 
-from collections import OrderedDict
+from collections import OrderedDict, defaultdict
 import warnings
 import numpy as np
+import mxnet as mx
 
 from ..base import mx_real_t, MXNetError
 from .. import symbol, ndarray, initializer, context
@@ -887,8 +888,22 @@ def initialize(self, init=initializer.Uniform(), ctx=None, verbose=False,
 
     def zero_grad(self):
         """Sets all Parameters' gradient buffer to 0."""
-        for i in self.values():
-            i.zero_grad()
+        # collect gradient arrays for each ctx
+        arrays = defaultdict(list)
+        for p in self.values():
+            if p.grad_req == 'null' or p._grad is None:
+                continue
+            for g in p.list_grad():
+                if g.stype == 'row_sparse':
+                    mx.ndarray.zeros_like(g, out=g)
+                else:
+                    arrays[g.context].append(g)
+
+        if len(arrays) == 0:
+            return
+
+        for arr in arrays.values():
+            mx.nd.reset_arrays(*arr, num_arrays=len(arr))
 
     def reset_ctx(self, ctx):
         """Re-assign all Parameters to other contexts.

src/operator/contrib/reset_arrays-inl.h

@@ -0,0 +1,92 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software 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.
+ */
+
+/*!
+ *  Copyright (c) 2019 by Contributors
+ * \file reset_arrays-inl.h
+ * \brief setting all array element values to zeros
+ * \author Moises Hernandez-Fernandez, Andrei Ivanov
+ */
+
+#ifndef MXNET_OPERATOR_CONTRIB_RESET_ARRAYS_INL_H_
+#define MXNET_OPERATOR_CONTRIB_RESET_ARRAYS_INL_H_
+
+#include <vector>
+#include "../tensor/init_op.h"
+
+namespace mxnet {
+namespace op {
+
+struct ResetArraysParam : public dmlc::Parameter<ResetArraysParam> {
+  int num_arrays;
+  DMLC_DECLARE_PARAMETER(ResetArraysParam) {
+    DMLC_DECLARE_FIELD(num_arrays)
+    .describe("number of input arrays.");
+  }
+};
+
+inline bool ResetArraysShape(const NodeAttrs& attrs,
+                            std::vector<mxnet::TShape>* in_shape,
+                            std::vector<mxnet::TShape>* out_shape) {
+  const auto& param = dmlc::get<ResetArraysParam>(attrs.parsed);
+  CHECK_EQ(in_shape->size(), param.num_arrays);
+  for (auto s : *in_shape) {
+    if (s.ndim() == 0)
+      return false;
+  }
+
+  return true;
+}
+
+inline bool ResetArraysType(const NodeAttrs& attrs,
+                           std::vector<int>* in_type,
+                           std::vector<int>* out_type) {
+  const auto& param = dmlc::get<ResetArraysParam>(attrs.parsed);
+  CHECK_EQ(in_type->size(), param.num_arrays);
+  for (size_t i = 0; i < in_type->size(); ++i) {
+    if ((*in_type)[i] == -1)
+      return false;
+  }
+
+  return true;
+}
+
+template<typename xpu>
+void ResetMemory(void *pntr, size_t len, mshadow::Stream<xpu> *s);
+
+template<typename xpu>
+void ResetArrays(const nnvm::NodeAttrs& attrs,
+                 const OpContext &ctx,
+                 const std::vector<TBlob> &inputs,
+                 const std::vector<OpReqType> &req,
+                 const std::vector<TBlob> &outputs) {
+  auto s = ctx.get_stream<xpu>();
+  const auto& param = nnvm::get<ResetArraysParam>(attrs.parsed);
+  for (int i = 0; i < param.num_arrays; i++) {  // array index in inputs
+    const size_t size = inputs[i].shape_.Size();
+    MSHADOW_REAL_TYPE_SWITCH(inputs[i].type_flag_, DType,
+      ResetMemory(inputs[i].FlatTo2D<xpu, DType>(s).dptr_, size * sizeof(DType), s);
+    )
+  }
+}
+
+}  // namespace op
+}  // namespace mxnet
+
+#endif  // MXNET_OPERATOR_CONTRIB_RESET_ARRAYS_INL_H_

src/operator/contrib/reset_arrays.cc

@@ -0,0 +1,74 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software 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.
+ */
+
+/*!
+ *  Copyright (c) 2019 by Contributors
+ * \file reset_arrays.cc
+ * \brief setting all array element values to zeros
+ * \author Moises Hernandez-Fernandez, Andrei Ivanov
+ */
+
+#include "./reset_arrays-inl.h"
+
+namespace mxnet {
+namespace op {
+
+DMLC_REGISTER_PARAMETER(ResetArraysParam);
+
+NNVM_REGISTER_OP(reset_arrays)
+.describe(R"code(Set to zero multiple arrays
+)code" ADD_FILELINE)
+.set_num_inputs([](const nnvm::NodeAttrs& attrs) {
+    return static_cast<uint32_t>(dmlc::get<ResetArraysParam>(attrs.parsed).num_arrays);
+  })
+.set_attr<nnvm::FMutateInputs>("FMutateInputs",
+  [](const nnvm::NodeAttrs& attrs) {
+    const uint32_t num_args = dmlc::get<ResetArraysParam>(attrs.parsed).num_arrays;
+    std::vector<uint32_t> ret;
+    for (uint32_t i = 0; i < num_args; ++i) {
+      ret.push_back(i);
+    }
+    return ret;
+  })
+.set_num_outputs(0)
+.set_attr_parser(ParamParser<ResetArraysParam>)
+.set_attr<mxnet::FInferShape>("FInferShape", ResetArraysShape)
+.set_attr<nnvm::FInferType>("FInferType", ResetArraysType)
+.set_attr<nnvm::FListInputNames>("FListInputNames",
+  [](const NodeAttrs& attrs) {
+    const uint32_t num_args = dmlc::get<ResetArraysParam>(attrs.parsed).num_arrays;
+    std::vector<std::string> ret;
+    for (uint32_t i = 0; i < num_args; ++i) {
+      ret.push_back(std::string("array_") + std::to_string(i));
+    }
+    return ret;
+  })
+.add_argument("data", "NDArray-or-Symbol[]", "Arrays")
+.add_arguments(ResetArraysParam::__FIELDS__());
+
+NNVM_REGISTER_OP(reset_arrays)
+.set_attr<FCompute>("FCompute<cpu>", ResetArrays<cpu>);
+
+template<>
+void ResetMemory<cpu>(void *pntr, size_t len, mshadow::Stream<cpu> *s) {
+  memset(pntr, 0, len);
+}
+
+}  // namespace op
+}  // namespace mxnet

src/operator/contrib/reset_arrays.cu

@@ -0,0 +1,40 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software 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.
+ */
+
+/*!
+ *  Copyright (c) 2019 by Contributors
+ * \file reset_arrays.cu
+ * \brief setting all array element values to zeros
+ * \author Moises Hernandez-Fernandez, Andrei Ivanov
+ */
+#include "./reset_arrays-inl.h"
+
+namespace mxnet {
+namespace op {
+
+template<>
+void ResetMemory<gpu>(void *pntr, size_t len, mshadow::Stream<gpu> *s) {
+  CUDA_CALL(cudaMemsetAsync(pntr, 0, len, mshadow::Stream<gpu>::GetStream(s)));
+}
+
+NNVM_REGISTER_OP(reset_arrays)
+.set_attr<FCompute>("FCompute<gpu>", ResetArrays<gpu>);
+
+}  // namespace op
+}  // namespace mxnet

tests/python/unittest/test_gluon.py

@@ -33,6 +33,7 @@
 import warnings
 import json
 import unittest
+import random
 
 @with_seed()
 def test_parameter():
@@ -1504,15 +1505,62 @@ def test_hybrid_multi_context():
 
 @with_seed()
 def test_zero_grad():
-    data = mx.nd.random.uniform(shape=(3,3))
-    net = nn.Embedding(3, 4, sparse_grad=True, prefix='test_zero_grad_')
-    net.initialize()
-    with mx.autograd.record():
-        l = net(data)
-        l.backward()
-    net.collect_params().zero_grad()
-    grad = net.collect_params()['test_zero_grad_weight'].grad()
-    assert_almost_equal(grad.asnumpy(), grad.asnumpy() * 0)
+    def _test_grad_reset(ctx, dtype='float32', sparse=False, embeddingType=None):
+        data = mx.nd.random.uniform(shape=(3,3), dtype=dtype, ctx=ctx)
+        if embeddingType is None:
+            embeddingType = dtype
+        net = nn.Embedding(3, 4, sparse_grad=sparse, prefix='test_zero_grad_', dtype=embeddingType)
+        net.initialize(ctx=ctx)
+        with mx.autograd.record():
+            l = net(data)
+            l.backward()
+        net.collect_params().zero_grad()
+        grad = net.collect_params()['test_zero_grad_weight'].grad()
+        assert_almost_equal(grad.asnumpy(), grad.asnumpy() * 0)
+
+    def _test_multi_reset(nArrays, dtype, ctx):
+        # Construct the list of non-zeros arrays with random shapes
+        arr = []
+        for _ in range(nArrays):
+            arrType = random.choice(dtype) if isinstance(dtype, list) else dtype
+            shape = ()
+            for _ in range(np.random.randint(1, 5)):
+                shape = shape + (np.random.randint(1, 10),)
+            arr.append(mx.nd.random.uniform(shape=shape, dtype=arrType, ctx=ctx))
+
+        # Reset all arrays
+        mx.nd.reset_arrays(*arr, num_arrays=len(arr))
+
+        # Check results
+        for i in range(nArrays):
+            grad = arr[i].asnumpy()
+            assert_almost_equal(grad, grad * 0)
+
+
+    # Setting context for current test
+    ctx = mx.context.current_context()
+
+    # Launching _test_multi_reset 10 times with different types & randomly chosen nArrays
+    testedTypes = ['float16', 'float32', 'float64']
+    for _ in range(10):
+        for type in [testedTypes] + testedTypes:
+            _test_multi_reset(np.random.randint(1, 50), type, ctx)
+
+    # Saving value of environment variable, if it was defined
+    envVarKey = 'MXNET_STORAGE_FALLBACK_LOG_VERBOSE'
+    envVarValue = os.environ[envVarKey] if envVarKey in os.environ else None
+    # Changing value of environment variable
+    os.environ[envVarKey] = '0'
+    for type in ['float16', 'float32', 'float64']:
+        for embType in ['float32', 'float64']:
+            for sparse in [True, False]:
+                _test_grad_reset(ctx, dtype=type, sparse=sparse, embeddingType=embType)
+
+    # Remove or restore the value of environment variable
+    if envVarValue is None:
+        del os.environ[envVarKey]
+    else:
+        os.environ[envVarKey] = envVarValue
 
 def check_hybrid_static_memory(**kwargs):
     x = mx.nd.random.uniform(shape=(2, 3, 32, 32))