Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 4b269baa 编写于 作者: W Weilong Wu 提交者： GitHub
浏览文件
操作

Revert "[Eager Grad] Support eager grad interface (#40170)" 

This reverts commit 4db8cf24.
上级 06fee998
显示空白变更内容
内联 并排
Showing with 163 addition and 1217 deletion
paddle/fluid/eager/accumulation/accumulation_node.cc
浏览文件 @ 4b269baa
......@@ -24,7 +24,7 @@
#include "paddle/fluid/platform/errors.h"
#include "glog/logging.h"
DECLARE_bool(retain_grad_for_all_tensor);
namespace egr {
static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
......@@ -39,8 +39,8 @@ static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
}
std::vector<std::vector<paddle::experimental::Tensor>> GradNodeAccumulation::
operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph) {
operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads) {
VLOG(3) << "Running Eager Backward Node: GradNodeAccumulation";
PADDLE_ENFORCE(grads.size() == 1,
paddle::platform::errors::Fatal(
......@@ -62,7 +62,7 @@ operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
grad_out = grads[0][0];
}
if (!weak_grad_.expired() && FLAGS_retain_grad_for_all_tensor) {
if (!weak_grad_.expired()) {
auto grad = weak_grad_.lock();
CopyOrAddTensor(grad.get(), grad_out);
}
......
paddle/fluid/eager/accumulation/accumulation_node.h
浏览文件 @ 4b269baa
......@@ -35,15 +35,8 @@ class GradNodeAccumulation : public GradNodeBase {
// Functor: perform backward computations
virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph = false) override;
void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
bool IsTensorWrappersCleared() override {
VLOG(6) << "Do nothing here now";
return false;
}
const std::vector<std::vector<paddle::experimental::Tensor>>& grads)
override;
std::string name() { return "GradNodeAccumulation"; }
......
paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc
浏览文件 @ 4b269baa
......@@ -145,8 +145,8 @@ void GradNodeScale::SetTensorWrappers_X(
void GradNodeScale::SetAttributes_scale(float scale) { scale_ = scale; }
std::vector<std::vector<paddle::experimental::Tensor>> GradNodeScale::
operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph) {
operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads) {
// 1. Check Output Size
PADDLE_ENFORCE(
((grads.size() == 1) && (grads[0].size() == 1)),
......
paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h
浏览文件 @ 4b269baa
......@@ -39,15 +39,8 @@ class GradNodeScale : public GradNodeBase {
// Functor: perform backward computations
virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph = false) override;
void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
bool IsTensorWrappersCleared() override {
VLOG(6) << "Do nothing here now";
return false;
}
const std::vector<std::vector<paddle::experimental::Tensor>>& grads)
override;
void SetTensorWrappers_X(
const std::vector<paddle::experimental::Tensor>& tensors);
......
paddle/fluid/eager/auto_code_generator/eager_generator.cc
浏览文件 @ 4b269baa
......@@ -2074,8 +2074,7 @@ static std::string GenerateGradNodeCCContents(
const char* GRAD_FUNCTION_TEMPLATE =
"std::vector<std::vector<paddle::experimental::Tensor>> "
"GradNode%s::operator()(const "
"std::vector<std::vector<paddle::experimental::Tensor>>& grads, "
"bool create_graph) {\n%s\n}";
"std::vector<std::vector<paddle::experimental::Tensor>>& grads) {\n%s\n}";
std::string grad_function_str = paddle::string::Sprintf(
GRAD_FUNCTION_TEMPLATE, fwd_op_type, generated_grad_function_body);
......@@ -2110,28 +2109,18 @@ static std::string GenerateGradNodeHeaderContents(
"\n"
" virtual std::vector<std::vector<paddle::experimental::Tensor>> "
"operator()(const "
"std::vector<std::vector<paddle::experimental::Tensor>>& grads, const "
"bool create_graph = false) "
"std::vector<std::vector<paddle::experimental::Tensor>>& grads) "
"override;\n"
"\n"
" void ClearTensorWrappers() override { \n"
"%s\n"
" is_tensor_wrappers_cleared = true;\n"
" }\n"
" std::string name() override { return \" GradNode%s \"; } \n "
"\n"
" // SetX, SetY, ...\n"
"%s\n"
" // SetAttrMap\n"
"%s\n"
" bool IsTensorWrappersCleared() override { \n"
" return is_tensor_wrappers_cleared;\n"
" }\n"
" private:\n"
" // TensorWrappers\n"
"%s\n"
" bool is_tensor_wrappers_cleared = false;\n"
"\n"
" // Attribute Map\n"
"%s\n"
"};";
......@@ -2165,7 +2154,6 @@ static std::string GenerateGradNodeHeaderContents(
std::string set_tensor_wrappers_str = "";
std::string tensor_wrapper_members_str = "";
std::string clear_tensor_wrappers_str = "";
for (const auto& iter : op_base_infos) {
const std::map<std::string, std::string>& grad_ins_fwd_slotname_map =
iter.GetGradInsFwdSlotnameMap();
......@@ -2197,13 +2185,6 @@ static std::string GenerateGradNodeHeaderContents(
SET_TENSOR_WRAPPER_BODY_TEMPLATE, tensor_wrapper_name,
struct_tensor_wrapper_name);
const char* CLEAR_TENSOR_WRAPPER_TEMPLATE =
"for (auto tw: %s) {\n"
" tw.clear();\n"
" }\n";
clear_tensor_wrappers_str += paddle::string::Sprintf(
CLEAR_TENSOR_WRAPPER_TEMPLATE, struct_tensor_wrapper_name);
} else {
const char* ATTR_TENSOR_WRAPPER_ARG_TEMPLATE =
"const paddle::experimental::Tensor& %s";
......@@ -2216,14 +2197,10 @@ static std::string GenerateGradNodeHeaderContents(
TENSOR_WRAPPER_MEMBER_TEMPLATE, struct_tensor_wrapper_name);
const char* SET_TENSOR_WRAPPER_BODY_TEMPLATE =
"%s = egr::TensorWrapper(%s, %s /*full_reserved*/);\n";
"%s = egr::TensorWrapper(%s, %s /*full_reserved*/);";
tensor_wrapper_body_str = paddle::string::Sprintf(
SET_TENSOR_WRAPPER_BODY_TEMPLATE, struct_tensor_wrapper_name,
tensor_wrapper_name, full_reserved_str);
const char* CLEAR_TENSOR_WRAPPER_TEMPLATE = " %s.clear();\n";
clear_tensor_wrappers_str += paddle::string::Sprintf(
CLEAR_TENSOR_WRAPPER_TEMPLATE, struct_tensor_wrapper_name);
}
std::string full_reserved_signature_str = "bool full_reserved";
const char* SET_TENSOR_WRAPPER_TEMPLATE =
......@@ -2238,8 +2215,8 @@ static std::string GenerateGradNodeHeaderContents(
std::string grad_node_str = paddle::string::Sprintf(
GRAD_NODE_TEMPLATE, op_type, op_type, op_type, op_type, op_type, op_type,
op_type, clear_tensor_wrappers_str, op_type, set_tensor_wrappers_str,
set_attr_map_str, tensor_wrapper_members_str, attr_members_str);
op_type, op_type, set_tensor_wrappers_str, set_attr_map_str,
tensor_wrapper_members_str, attr_members_str);
return grad_node_str;
}
......
paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
浏览文件 @ 4b269baa
......@@ -478,7 +478,6 @@ def GenerateNodeDeclaration(fwd_api_name, backward_fwd_input_map,
# SetTensorWrapper Methods & TensorWrapper Members
set_tensor_wrapper_methods_str = ""
tensor_wrapper_members_str = ""
clear_tensor_wrapper_str = ""
for tname, (ttype, is_fwd_input, _) in backward_fwd_input_map.items():
if tname in no_need_buffer_set:
no_need_buffer = "true"
......@@ -500,13 +499,6 @@ def GenerateNodeDeclaration(fwd_api_name, backward_fwd_input_map,
"""
tensor_wrapper_members_str += PLAIN_TENSOR_MEMBER_TEMPLATE.format(
tensor_wrapper_name)
CLEAR_TENSOR_WRAPPERS_TEMPLATE = """
{}.clear();
"""
clear_tensor_wrapper_str += CLEAR_TENSOR_WRAPPERS_TEMPLATE.format(
tensor_wrapper_name)
else:
assert IsVectorTensorType(ttype)
SET_VECTOR_TENSOR_WRAPPER_TEMPLATE = """
......@@ -524,15 +516,6 @@ def GenerateNodeDeclaration(fwd_api_name, backward_fwd_input_map,
"""
tensor_wrapper_members_str += VECTOR_TENSOR_MEMBER_TEMPLATE.format(
tensor_wrapper_name)
CLEAR_TENSOR_WRAPPERS_TEMPLATE = """
for (auto tw: {}) {
tw.clear();
};
"""
clear_tensor_wrapper_str += CLEAR_TENSOR_WRAPPERS_TEMPLATE.format(
tensor_wrapper_name)
# End: SetTensorWrapper Methods & TensorWrapper Members
# SetAttributes & Attribute Members
......@@ -572,37 +555,25 @@ class {} : public egr::GradNodeBase {{
~{}() override = default;
virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph = false) override;
const std::vector<std::vector<paddle::experimental::Tensor>>& grads) override;
std::string name() override {{ return \" {} \"; }}
void ClearTensorWrappers() override {{
{}
is_tensor_wrappers_cleared = true;
}}
// SetTensorWrapperX, SetTensorWrapperY, ...
{}
// SetAttributes
{}
bool IsTensorWrappersCleared() override {{
return is_tensor_wrappers_cleared;
}}
private:
// TensorWrappers
{}
bool is_tensor_wrappers_cleared = false;
// Attributes
{}
}};
"""
node_declaration_str = NODE_DECLARATION_TEMPLATE.format(
grad_node_name, grad_node_name, grad_node_name, grad_node_name,
grad_node_name, clear_tensor_wrapper_str,
set_tensor_wrapper_methods_str, set_attribute_methods_str,
tensor_wrapper_members_str, attribute_members_str)
grad_node_name, set_tensor_wrapper_methods_str,
set_attribute_methods_str, tensor_wrapper_members_str,
attribute_members_str)
return node_declaration_str
......@@ -666,7 +637,7 @@ def GenerateNodeDefinition(fwd_api_name, bwd_api_name, backward_fwd_input_map,
grad_api_namespace = f"paddle::experimental"
FUNCTION_TEMPLATE = """
std::vector<std::vector<paddle::experimental::Tensor>> {}::operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph) {{
std::vector<std::vector<paddle::experimental::Tensor>> {}::operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads) {{
// Call grad_api function
auto grad_api_returns = {}::{}({});
{}
......
paddle/fluid/eager/backward.cc
浏览文件 @ 4b269baa
......@@ -39,21 +39,12 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
// Copy nodes
std::queue<GradNodeBase*> queue = init_queue;
std::unordered_set<GradNodeBase*> visited;
size_t potential_startup_ops_cnt = queue.size();
size_t cnt = 0;
// Visit each node exactly once in any order
while (!queue.empty()) {
GradNodeBase* node = queue.front();
queue.pop();
if (cnt < potential_startup_ops_cnt) {
if (!node_in_degree_map.count(node)) {
node_in_degree_map[node] = 0;
}
cnt += 1;
}
if (visited.count(node)) {
continue;
}
......@@ -85,248 +76,23 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
return node_in_degree_map;
}
// Remove some nodes those doesn't need to be
// stored in potential_stop_nodes、potential_startup_nodes
void UpdateGraphInfo(
std::unordered_map<GradNodeBase*, AutogradMeta*>*
target_nodes_inputmeta_map,
std::unordered_map<GradNodeBase*, std::unordered_set<GradNodeBase*>>*
depending_nodes,
std::unordered_set<GradNodeBase*>* potential_stop_nodes,
std::unordered_set<GradNodeBase*>* potential_startup_nodes) {
// Updated potential_sotp_nodes by depending_nodes,
// make sure the path from root to target_node is ok
std::unordered_set<GradNodeBase*> _startup_ops;
VLOG(6) << "Running in UpdateGraphInfo";
std::queue<GradNodeBase*> queue;
for (auto& target_nodes_inputmeta_pair : *target_nodes_inputmeta_map) {
queue.emplace(target_nodes_inputmeta_pair.first);
}
while (!queue.empty()) {
auto* target_node = queue.front();
queue.pop();
if (!(*depending_nodes)[target_node].empty()) {
auto precedding_nodes = (*depending_nodes)[target_node];
for (auto pre_nodes : precedding_nodes) {
queue.emplace(pre_nodes);
if (potential_stop_nodes->find(pre_nodes) !=
potential_stop_nodes->end()) {
potential_stop_nodes->erase(pre_nodes);
}
}
} else { // startup_ops have no precedding nodes
VLOG(6) << "Emplace _startup_ops";
_startup_ops.emplace(target_node);
}
}
// Purify potential_startup_nodes again, remove some
// potential startup_nodes that unreach to input target nodes
if (!_startup_ops.empty()) {
std::unordered_set<GradNodeBase*> potential_startup_nodes_to_be_erased;
for (auto node : *potential_startup_nodes) {
if (_startup_ops.count(node) == 0) {
VLOG(6) << "Set up potential_startup_nodes_to_be_erased";
potential_startup_nodes_to_be_erased.emplace(node);
}
}
if (!potential_startup_nodes_to_be_erased.empty()) {
for (auto node : potential_startup_nodes_to_be_erased) {
VLOG(6) << "Erase nodes in potential_startup_nodes_to_be_erased";
potential_startup_nodes->erase(node);
}
}
}
}
// Get Graph Info Betweent input target gradnode and outputs，
// record depending_nodes、 potential_stop_nodes、potential_startup_nodes
void GetGraphInfoBetweenTargets(
const std::queue<GradNodeBase*>& init_queue,
std::unordered_map<GradNodeBase*, AutogradMeta*>*
input_target_nodes_inputmeta_map,
std::unordered_map</*child node*/ GradNodeBase*,
/*father nodes*/ std::unordered_set<GradNodeBase*>>*
depending_nodes,
std::unordered_set<GradNodeBase*>* potential_stop_nodes,
std::unordered_set<GradNodeBase*>* potential_startup_nodes) {
if (input_target_nodes_inputmeta_map->empty()) return;
VLOG(6) << "Runing In GetGraphInfoBetweenTargets";
// Calculate in_degree for each node
std::unordered_map<GradNodeBase*, int> node_in_degree_map;
// Copy nodes
std::queue<GradNodeBase*> queue = init_queue;
std::unordered_set<GradNodeBase*> visited;
// Visit each node exactly once in any order
while (!queue.empty()) {
GradNodeBase* node = queue.front();
queue.pop();
if (visited.count(node)) {
continue;
}
visited.insert(node);
// Check node is target_nodes or not, if node is not target_node,
// all the next_node will be marked in potential_stop_nodes
bool is_potential_stop_nodes =
input_target_nodes_inputmeta_map->count(node);
// Find and append next nodes
const std::vector<std::vector<Edge>>& edges = node->GetEdges();
for (const auto& edge_list : edges) {
for (const Edge& edge : edge_list) {
GradNodeBase* next_node = edge.GetMutableGradNode().get();
// Next node could be nullptr if it is leaf tensor with no
// AccumulationNode attached
// Or it could also originated from dispensable inputs
if (!next_node) continue;
// if node not in input_target_nodes,
// all the next_nodes of current node will be inserted to
// potential_stop_node
if (is_potential_stop_nodes) {
potential_stop_nodes->emplace(next_node);
}
// Update in_degree
if (!node_in_degree_map.count(next_node))
node_in_degree_map[next_node] = 0;
node_in_degree_map[next_node]++;
// Record depending relationship
(*depending_nodes)[next_node].emplace(node);
queue.push(next_node);
}
}
}
// Update Graph Info, remove some stop_node in potential_stop_nodes
UpdateGraphInfo(input_target_nodes_inputmeta_map, depending_nodes,
potential_stop_nodes, potential_startup_nodes);
}
void GetTargetNodesInfo(const std::vector<paddle::experimental::Tensor>& inputs,
std::unordered_map<GradNodeBase*, AutogradMeta*>*
target_nodes_inputmeta_map) {
VLOG(6) << "Running in GetTargetNodesInfo";
if (!inputs.empty()) {
VLOG(6) << "Inputs are not empty";
size_t num_inputs = inputs.size();
for (size_t i = 0; i < num_inputs; i++) {
AutogradMeta* auto_grad_meta =
EagerUtils::unsafe_autograd_meta(inputs[i]);
auto target_node = auto_grad_meta->GetMutableGradNode().get();
PADDLE_ENFORCE_NOT_NULL(target_node,
paddle::platform::errors::Fatal(
"There is no grad op for input:%d or it's"
"stop_gradient=True",
i));
(*target_nodes_inputmeta_map)[target_node] = auto_grad_meta;
}
}
}
std::vector<paddle::experimental::Tensor> GetResults(
const std::vector<paddle::experimental::Tensor>& inputs,
std::unordered_map<GradNodeBase*, paddle::experimental::Tensor>*
results_map,
bool allow_unused, bool create_graph) {
VLOG(6) << "Running in GetResults";
if (inputs.empty()) return {};
std::vector<paddle::experimental::Tensor> results;
results.reserve(inputs.size());
for (size_t i = 0; i < inputs.size(); ++i) {
auto& input = inputs[i];
AutogradMeta* auto_grad_meta = EagerUtils::unsafe_autograd_meta(input);
auto target_node = auto_grad_meta->GetMutableGradNode().get();
auto iter = results_map->find(target_node);
if (iter != results_map->end()) {
// set StopGradient = !create_graph
AutogradMeta* tensor_auto_grad_meta =
EagerUtils::autograd_meta(&(iter->second));
tensor_auto_grad_meta->SetStopGradient(!create_graph);
results.emplace_back(iter->second);
} else {
PADDLE_ENFORCE_EQ(allow_unused, true,
paddle::platform::errors::InvalidArgument(
"The %d-th input does not appear in the backward "
"graph. Please check the input variable or set "
"allow_unused=True to get None result.",
i));
results.emplace_back();
}
}
return results;
}
// Enforce GradNode has TensorWrappers as Input
void EnforceGradNodeHasInput(GradNodeBase* node) {
VLOG(6) << "Running in EnforceGradNodeHasInput";
PADDLE_ENFORCE_NE(
node->IsTensorWrappersCleared(), true,
paddle::platform::errors::Fatal(
"The TensorWrappers of %s do not exist. This may be because:\n"
"You calculate backward twice for the same subgraph without "
"setting retain_graph=True. Please set retain_graph=True in the "
"first backward/grad call.\n",
node->name()));
}
// Purify potential_startup_nodes, remove nodes those are the same as
// input_target_nodes
void PurifyPotentialStartUpNodes(
std::unordered_set<GradNodeBase*>* potential_startup_nodes,
std::unordered_map<GradNodeBase*, AutogradMeta* /* InputMeta */>*
input_target_nodes_inputmeta_map) {
VLOG(6) << "Running in PurifyPotentialStartUpNodes";
if (input_target_nodes_inputmeta_map->empty()) return;
std::unordered_set<GradNodeBase*> potential_startup_nodes_to_be_erased;
for (auto startup_op : *potential_startup_nodes) {
auto iter = input_target_nodes_inputmeta_map->find(startup_op);
if (iter != input_target_nodes_inputmeta_map->end()) {
potential_startup_nodes_to_be_erased.emplace(iter->first);
}
}
if (!potential_startup_nodes_to_be_erased.empty()) {
for (auto nodes : potential_startup_nodes_to_be_erased) {
potential_startup_nodes->erase(nodes);
}
}
}
std::vector<paddle::experimental::Tensor> RunBackward(
const std::vector<paddle::experimental::Tensor>& tensors, // output
void RunBackward(const std::vector<paddle::experimental::Tensor>& tensors,
const std::vector<paddle::experimental::Tensor>& grad_tensors,
bool retain_graph, bool create_graph = false,
const std::vector<paddle::experimental::Tensor>& inputs = {},
bool allow_unused = false,
const std::vector<paddle::experimental::Tensor>& no_grad_vars = {}) {
bool retain_graph) {
paddle::platform::RecordEvent backward_record_event(
"backward", paddle::platform::TracerEventType::Operator, 1);
VLOG(6) << "Start Backward";
// *Gradient Hook should happen at node-level
// *Inplace version check should perform at node-level
// *Cross-batch accumulation happens at forward pass
std::unordered_map<GradNodeBase*, AutogradMeta*>
no_grad_var_nodes_inputmeta_map;
// Get no_grad_vars's GradNodes and InputMeta Info
GetTargetNodesInfo(no_grad_vars, &no_grad_var_nodes_inputmeta_map);
/* --- Initialization --- */
// 1. Init queue with starting nodes
// 2. Prepare initial input buffers
std::queue<GradNodeBase*> queue;
std::unordered_map<GradNodeBase*, std::unique_ptr<GradTensorHolder>>
node_input_buffers_dict;
std::unordered_set<GradNodeBase*> potential_startup_nodes;
for (size_t i = 0; i < tensors.size(); i++) {
const paddle::experimental::Tensor& tensor = tensors[i];
......@@ -366,17 +132,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
"size = 0 or same size as tensors"));
// Feed given tensor if it's provided
VLOG(6) << "Fill grad input tensor " << i << "with give grad tensor";
if (grad_tensors[i].is_initialized()) {
// Deep copy
paddle::experimental::Tensor tmp_tensor;
tmp_tensor.copy_(grad_tensors[i], true);
node_input_buffers_dict[grad_node]->add(input_info.first,
input_info.second, tmp_tensor);
} else {
node_input_buffers_dict[grad_node]->add(
input_info.first, input_info.second, grad_tensors[i]);
}
} else {
VLOG(6) << "Fill grad input tensor " << i << " with 1.0";
......@@ -389,9 +146,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
input_info.first, input_info.second, tensor, true /*fill_one=true*/);
}
// Prepare queue, potential startup_nodes
// Prepare queue
queue.push(grad_node);
potential_startup_nodes.emplace(grad_node);
}
VLOG(6) << "Update In degree Map for backward";
......@@ -399,74 +155,25 @@ std::vector<paddle::experimental::Tensor> RunBackward(
std::unordered_map<GradNodeBase*, int> node_in_degree_map =
getInDegreeMap(queue);
// Get input's GradNodes and InputMeta Info
std::unordered_map<GradNodeBase*, AutogradMeta* /* InputMeta */>
input_target_nodes_inputmeta_map;
GetTargetNodesInfo(inputs, &input_target_nodes_inputmeta_map);
// Purify potential_startup_ops, remove those nodes that are the same as
// input_target_nodes
PurifyPotentialStartUpNodes(&potential_startup_nodes,
&input_target_nodes_inputmeta_map);
// Get Graph Info Betweent input target gradnode and outputs
// Record the depending_nodes and potential_stop_nodes
std::unordered_map<GradNodeBase* /* child node */,
std::unordered_set<GradNodeBase*> /* father node */>
depending_nodes;
std::unordered_set<GradNodeBase*> potential_stop_nodes;
// std::unordered_set<GradNodeBase*> startup_ops;
GetGraphInfoBetweenTargets(queue, &input_target_nodes_inputmeta_map,
&depending_nodes, &potential_stop_nodes,
&potential_startup_nodes);
// ready_queue store all startup nodes
std::queue<GradNodeBase*> ready_queue;
// startup op's indegree should be 0
for (auto node : potential_startup_nodes) {
if (node_in_degree_map[node] == 0) {
ready_queue.emplace(node);
}
}
VLOG(1) << " startup_ops' size is :" << ready_queue.size();
std::unordered_map<GradNodeBase*, paddle::experimental::Tensor> results_map;
// read_queue is empty only when 1.input equals to output. 2.input can not
// reach to output.
if (ready_queue.size() == 0) {
for (auto input_target_node : input_target_nodes_inputmeta_map) {
// out rank_info of forward op
auto rank_info = input_target_node.second->OutRankInfo();
if (node_input_buffers_dict[input_target_node.first]) {
auto& target_result =
node_input_buffers_dict[input_target_node.first]
->Buffers()[rank_info.first][rank_info.second];
// save the target result
results_map[input_target_node.first] = target_result;
}
}
}
/* --- Topological Visit --- */
// 1. Pop queue
// 2. Run node
// |- Check and capture target result
// |- node(grads)
// |- Prepare for next node
// 3. Update queue
VLOG(6) << "Run Backward";
while (!ready_queue.empty()) {
GradNodeBase* node = ready_queue.front();
VLOG(6) << "Running GradNode:" << node->name();
ready_queue.pop();
while (!queue.empty()) {
GradNodeBase* node = queue.front();
paddle::platform::RecordEvent node_record_event(
std::string(typeid(*node).name()) + " grad_node",
paddle::platform::TracerEventType::Operator, 1);
if (queue.size() > 1 && node_in_degree_map[node] != 0) {
queue.pop();
continue;
}
queue.pop();
// Run node: This is where Hook happens
PADDLE_ENFORCE(
node_input_buffers_dict.count(node),
......@@ -477,45 +184,10 @@ std::vector<paddle::experimental::Tensor> RunBackward(
std::unique_ptr<GradTensorHolder> node_input_buffer =
std::move(node_input_buffers_dict[node]);
// get target grad_var from node_input_buffer by inputmeta
if (input_target_nodes_inputmeta_map.find(node) !=
input_target_nodes_inputmeta_map.end()) {
VLOG(6) << "Get target result by by inputmeta";
// out rank_info of forward op
auto rank_info = input_target_nodes_inputmeta_map[node]->OutRankInfo();
// rank_info is a pair, first means slot_id, second means rank.
auto& target_result =
node_input_buffer->Buffers()[rank_info.first][rank_info.second];
// save the target result
results_map[node] = target_result;
}
// no_grad_vars
if (no_grad_var_nodes_inputmeta_map.find(node) !=
no_grad_var_nodes_inputmeta_map.end()) {
VLOG(6) << "Change the input buffer[slot][rank] by Zeros";
auto rank_info = no_grad_var_nodes_inputmeta_map[node]->OutRankInfo();
node_input_buffer->SetBufferSlotRankZeros(rank_info.first,
rank_info.second);
}
VLOG(6) << "Running GradNode:" << node->name();
// check input
EnforceGradNodeHasInput(node);
VLOG(6) << "Run Backward Kernel with GradTensorHolder";
// Run Pre Backward Node and get outputs
std::vector<std::vector<paddle::experimental::Tensor>> grad_output_tensors =
(*node)(node_input_buffer->Buffers(), create_graph);
// retain_grad or not
if (!retain_graph) {
VLOG(6)
<< "retain_graph is false, need to clear the TensorWrapper of nodes.";
node->ClearTensorWrappers();
}
(*node)(node_input_buffer->Buffers());
// TODO(jiabin): Should we erase it or find a more efficient way.
node_input_buffers_dict.erase(node);
......@@ -580,44 +252,18 @@ std::vector<paddle::experimental::Tensor> RunBackward(
// Update queue
node_in_degree_map[next_node]--;
PADDLE_ENFORCE(
node_in_degree_map[next_node] >= 0,
paddle::platform::errors::Fatal(
"Detected in-degree value smaller than zero. For Node: %s"
"Node's in-degree cannot be negative",
next_node->name()));
bool is_potential_stop_node = potential_stop_nodes.count(next_node);
if (node_in_degree_map[next_node] == 0 && !is_potential_stop_node) {
ready_queue.emplace(std::move(next_node));
if (node_in_degree_map[next_node] == 0) {
queue.emplace(std::move(next_node));
}
}
}
}
return GetResults(inputs, &results_map, allow_unused, create_graph);
}
void Backward(
const std::vector<paddle::experimental::Tensor>& tensors, // output
const std::vector<paddle::experimental::Tensor>& grad_tensors,
bool retain_graph) {
VLOG(6) << "Run in Backward";
paddle::platform::RecordEvent backward_record_event(
"backward", paddle::platform::TracerEventType::Operator, 1);
RunBackward(tensors, grad_tensors, retain_graph);
}
std::vector<paddle::experimental::Tensor> Grad(
const std::vector<paddle::experimental::Tensor>& tensors, // output
const std::vector<paddle::experimental::Tensor>& inputs,
const std::vector<paddle::experimental::Tensor>& grad_tensors,
bool retain_graph, bool create_graph, bool only_inputs, bool allow_unused,
const std::vector<paddle::experimental::Tensor>& no_grad_vars) {
VLOG(6) << "Run in Grad";
return RunBackward(tensors, grad_tensors, retain_graph, create_graph, inputs,
allow_unused, no_grad_vars);
}
} // namespace egr
paddle/fluid/eager/backward.h
浏览文件 @ 4b269baa
......@@ -19,21 +19,13 @@
namespace egr {
// Backward():
// run_backward():
// tensors corresponds to those lived in the backward graph
// each grad_tensors[i] keeps the value for its corresponding tensors[i]
void Backward(const std::vector<paddle::experimental::Tensor>& tensors,
const std::vector<paddle::experimental::Tensor>& grad_tensors,
void RunBackward(const std::vector<paddle::experimental::Tensor> &tensors,
const std::vector<paddle::experimental::Tensor> &grad_tensors,
bool retain_graph = false);
std::vector<paddle::experimental::Tensor> Grad(
const std::vector<paddle::experimental::Tensor>& tensors,
const std::vector<paddle::experimental::Tensor>& inputs,
const std::vector<paddle::experimental::Tensor>& grad_tensors = {},
bool retain_graph = false, bool create_graph = false,
bool only_inputs = false, bool allow_unused = false,
const std::vector<paddle::experimental::Tensor>& no_grad_vars = {});
// Reserved for gradient()
} // namespace egr
paddle/fluid/eager/custom_operator/custom_operator_node.cc
浏览文件 @ 4b269baa
......@@ -20,8 +20,8 @@
namespace egr {
std::vector<std::vector<paddle::experimental::Tensor>> RunCustomOpNode::
operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph) {
operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads) {
paddle::CustomOpKernelContext ctx;
auto grad_inputs_name = paddle::framework::OpMetaInfoHelper::GetInputs(
egr::Controller::Instance().GetOpMetaInfoMap().at(op_type_)[1]);
......
paddle/fluid/eager/custom_operator/custom_operator_node.h
浏览文件 @ 4b269baa
......@@ -37,8 +37,8 @@ class RunCustomOpNode : public GradNodeBase {
// Functor: perform backward computations
virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph) override;
const std::vector<std::vector<paddle::experimental::Tensor>>& grads)
override;
std::string name() {
return paddle::string::Sprintf("RunCustomOpNode: %s_grad", op_type_);
......@@ -62,12 +62,6 @@ class RunCustomOpNode : public GradNodeBase {
return res;
}
void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
bool IsTensorWrappersCleared() override {
VLOG(6) << "Do nothing here now";
return false;
}
void SetAttrs(const std::vector<paddle::any>& attr) { attrs_ = attr; }
public:
......
paddle/fluid/eager/grad_node_info.h
浏览文件 @ 4b269baa
......@@ -95,12 +95,8 @@ class GradNodeBase {
* is better choice to fit this format.
* **/
virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph = false) = 0;
const std::vector<std::vector<paddle::experimental::Tensor>>& grads) = 0;
virtual void ClearTensorWrappers() = 0;
virtual bool IsTensorWrappersCleared() = 0;
/**
* AddEdges is designed to set input tensors' backward Node as current
* node's Edges.
......
paddle/fluid/eager/grad_tensor_holder.cc
浏览文件 @ 4b269baa
......@@ -21,11 +21,6 @@
namespace egr {
void GradTensorHolder::SetBufferSlotRankZeros(size_t slot_id, size_t rank) {
buffer_[slot_id][rank] =
paddle::experimental::zeros_like(buffer_[slot_id][rank]);
}
void GradTensorHolder::add(size_t slot_id, size_t rank,
const paddle::experimental::Tensor& t,
bool fill_one) {
......
paddle/fluid/eager/grad_tensor_holder.h
浏览文件 @ 4b269baa
......@@ -56,8 +56,6 @@ class GradTensorHolder {
return buffer_;
}
void SetBufferSlotRankZeros(size_t slot_id, size_t rank);
private:
std::vector<std::vector<paddle::experimental::Tensor>> buffer_;
};
......
paddle/fluid/eager/tensor_wrapper.h
浏览文件 @ 4b269baa
......@@ -98,8 +98,6 @@ class TensorWrapper {
}
}
void clear() { intermidiate_tensor_.reset(); }
private:
bool full_reserved_ = false;
std::pair<size_t, size_t> out_rank_info_;
......
paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h
浏览文件 @ 4b269baa
......@@ -32,8 +32,8 @@ class GradTestNode : public egr::GradNodeBase {
GradTestNode() : GradNodeBase() { val_ = 1.0; }
std::string name() override { return "GradTestNode"; }
std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
bool create_graph = false) override {
const std::vector<std::vector<paddle::experimental::Tensor>>& grads)
override {
val_ = std::dynamic_pointer_cast<phi::DenseTensor>(grads[0][0].impl())
->data<float>()[0];
phi::DenseTensorMeta meta =
......@@ -49,11 +49,6 @@ class GradTestNode : public egr::GradNodeBase {
std::vector<std::vector<paddle::experimental::Tensor>> res = {{et1}};
return res;
}
void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
bool IsTensorWrappersCleared() override {
VLOG(6) << "Do nothing here now";
return false;
}
float val_;
};
} // namespace eager_test
paddle/fluid/eager/tests/performance_tests/benchmark_utils.cc
浏览文件 @ 4b269baa
......@@ -58,7 +58,7 @@ void benchmark_eager_scale(const paddle::experimental::Tensor& tensor,
}
std::vector<paddle::experimental::Tensor> target_tensors = {input_tensor};
Backward(target_tensors, {});
RunBackward(target_tensors, {});
if (accuracy_check) {
// Examine Forward Grad (w.r.t max_num_runs = 10)
......@@ -80,7 +80,7 @@ void benchmark_eager_matmul(const paddle::experimental::Tensor& X,
}
std::vector<paddle::experimental::Tensor> target_tensors = {input_tensor0};
Backward(target_tensors, {});
RunBackward(target_tensors, {});
if (accuracy_check) {
// Examine Forward Grad (w.r.t max_num_runs = 2)
......@@ -106,7 +106,7 @@ void benchmark_eager_intermediate_matmul(const paddle::experimental::Tensor& X,
}
std::vector<paddle::experimental::Tensor> target_tensors = {input_tensor0};
Backward(target_tensors, {});
RunBackward(target_tensors, {});
if (accuracy_check) {
// Examine Forward Grad (w.r.t max_num_runs = 2)
......@@ -137,7 +137,7 @@ void benchmark_eager_intermediate_mlp(
reduce_sum_dygraph_function(input0, {{"reduce_all", true}});
std::vector<paddle::experimental::Tensor> target_tensors = {Out};
Backward(target_tensors, {});
RunBackward(target_tensors, {});
if (accuracy_check) {
std::unordered_map<std::string, float> result =
......
paddle/fluid/eager/tests/task_tests/CMakeLists.txt
浏览文件 @ 4b269baa
......@@ -5,7 +5,6 @@ cc_test(test_egr_task_backward SRCS backward_test.cc DEPS ${eager_deps} ${fluid_
cc_test(test_egr_task_hook SRCS hook_test.cc DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
cc_test(test_egr_task_cross_batch SRCS cross_batch_accumulation_test.cc DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
cc_test(test_egr_task_fwd_bwd_joint SRCS fwd_bwd_joint_test.cc DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
cc_test(test_egr_task_grad SRCS grad_test.cc DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
if(NOT ((NOT WITH_PYTHON) AND ON_INFER))
cc_test(test_egr_task_hook_intermidiate SRCS hook_test_intermidiate.cc DEPS ${eager_deps} ${fluid_deps} ${generated_deps} dygraph_node)
......
paddle/fluid/eager/tests/task_tests/backward_test.cc
浏览文件 @ 4b269baa
......@@ -33,7 +33,6 @@
#include "paddle/phi/core/kernel_registry.h"
PD_DECLARE_KERNEL(full, CPU, ALL_LAYOUT);
PD_DECLARE_KERNEL(copy, CPU, ALL_LAYOUT);
namespace egr {
......@@ -80,7 +79,7 @@ TEST(Backward, SingleNodeEmptyGrad) {
}
std::vector<paddle::experimental::Tensor> outs = {target_tensor};
// Run Backward
Backward(outs, {});
RunBackward(outs, {});
// Check Output Value
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 5.0);
......@@ -139,7 +138,7 @@ TEST(Backward, SingleNodeCustomGrad) {
}
// Run Backward
Backward(target_tensors, grad_tensors);
RunBackward(target_tensors, grad_tensors);
// Check Output Value
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 50.0);
......@@ -212,7 +211,7 @@ TEST(Backward, LinearNodes) {
}
// Use Empty Grad Tensor
Backward(target_tensors, {});
RunBackward(target_tensors, {});
// Check Output Value
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 50.0);
......@@ -316,7 +315,7 @@ TEST(Backward, WithAccumulation) {
node2_ptr->AddEdges(&res2, 0);
}
Backward(target_tensors, grad_tensors);
RunBackward(target_tensors, grad_tensors);
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 2500.0);
}
......
paddle/fluid/eager/tests/task_tests/cross_batch_accumulation_test.cc
浏览文件 @ 4b269baa
......@@ -71,12 +71,12 @@ TEST(CrossBatchAccumulation, SingleScaleNode) {
std::vector<egr::AutogradMeta*> res = {meta};
scale_node_ptr->AddEdges(&res, 0);
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(target_tensor, 1.0);
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 5.0);
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(target_tensor, 1.0);
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 10.0);
......
paddle/fluid/eager/tests/task_tests/fwd_bwd_joint_test.cc
浏览文件 @ 4b269baa
......@@ -86,7 +86,7 @@ TEST(FwdBwdJoint, SingleNode) {
std::vector<paddle::experimental::Tensor> outs = {out};
// 4. Run Backward
Backward(outs, {});
RunBackward(outs, {});
VLOG(7) << "Target Grad is: "
<< std::static_pointer_cast<phi::DenseTensor>(
......@@ -137,7 +137,7 @@ TEST(FwdBwdJoint, LinearNodes) {
std::vector<paddle::experimental::Tensor> outs = {out1};
// 4. Run Backward
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
eager_test::CompareGradTensorWithValue<float>(tensor, 10.0);
......@@ -203,7 +203,7 @@ TEST(FwdBwdJoint, BranchedNodes) {
// 4. Run Backward
std::vector<paddle::experimental::Tensor> outs = {out1, out2};
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
eager_test::CompareGradTensorWithValue<float>(tensor, 30.0);
......@@ -260,7 +260,7 @@ TEST(FwdBwdJoint, GradientHook) {
// 4. Run Backward
std::vector<paddle::experimental::Tensor> outs = {out1, out2};
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
// leaf grad
......@@ -318,13 +318,13 @@ TEST(FwdBwdJoint, CrossBatchAccumulation) {
// 4. Run Backward
std::vector<paddle::experimental::Tensor> outs = {out1, out2};
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
eager_test::CompareGradTensorWithValue<float>(tensor, 30.0);
// Cross Batch Accumulation
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
eager_test::CompareGradTensorWithValue<float>(tensor, 60.0);
......@@ -356,7 +356,7 @@ TEST(FwdBwdJoint, SingleNodeCUDA) {
std::vector<paddle::experimental::Tensor> outs = {out};
// 4. Run Backward
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
eager_test::CompareGradTensorWithValue<float>(tensor, 2.0);
......@@ -412,7 +412,7 @@ TEST(FwdBwdJoint, BranchedNodesCUDA) {
// TODO(jiabin): fix this with add functor
// 4. Run Backward
std::vector<paddle::experimental::Tensor> outs = {out1, out2};
Backward(outs, {});
RunBackward(outs, {});
// Examine Backward Grad
eager_test::CompareGradTensorWithValue<float>(tensor, 30.0);
......
paddle/fluid/eager/tests/task_tests/generated_test.cc
浏览文件 @ 4b269baa
......@@ -57,7 +57,7 @@ TEST(Generated, Sigmoid) {
std::vector<paddle::experimental::Tensor> target_tensors = {output_tensor};
VLOG(6) << "Runing Backward";
Backward(target_tensors, {});
RunBackward(target_tensors, {});
VLOG(6) << "Finish Backward";
eager_test::CompareGradTensorWithValue<float>(tensor, 0.25);
......@@ -89,7 +89,7 @@ TEST(Generated, Matmul_v2) {
eager_test::CompareTensorWithValue<float>(output_tensor, 96);
std::vector<paddle::experimental::Tensor> target_tensors = {output_tensor};
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(X, 2.0 * 20);
eager_test::CompareGradTensorWithValue<float>(Y, 3.0 * 4);
......@@ -120,7 +120,7 @@ TEST(Generated, ElementwiseAdd) {
eager_test::CompareTensorWithValue<float>(output_tensor, 5);
std::vector<paddle::experimental::Tensor> target_tensors = {output_tensor};
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(X, 1.0);
eager_test::CompareGradTensorWithValue<float>(Y, 1.0);
......
paddle/fluid/eager/tests/task_tests/grad_test.cc 已删除 100644 → 0
浏览文件 @ 06fee998
// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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.
#include <sstream>
#include "glog/logging.h"
#include "gtest/gtest.h"
#include "paddle/fluid/eager/accumulation/accumulation_node.h"
#include "paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h"
#include "paddle/fluid/eager/api/utils/tensor_utils.h"
#include "paddle/fluid/eager/autograd_meta.h"
#include "paddle/fluid/eager/backward.h"
#include "paddle/fluid/eager/grad_node_info.h"
#include "paddle/fluid/eager/tests/test_utils.h"
#include "paddle/fluid/eager/api/all.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/tensor_meta.h"
PD_DECLARE_KERNEL(full, CPU, ALL_LAYOUT);
PD_DECLARE_KERNEL(copy, CPU, ALL_LAYOUT);
namespace egr {
TEST(Grad, SingleNodeEmptyGrad) {
// Prepare Device Contexts
eager_test::InitEnv(paddle::platform::CPUPlace());
// Prepare Inputs
paddle::framework::DDim ddim = phi::make_ddim({4, 16, 16, 32});
// Create Target Tensor (output)
paddle::experimental::Tensor output_tensor =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, false /*is_leaf*/);
// Create input tensor
const paddle::experimental::Tensor leaf_tensor =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, true /*is_leaf*/);
{
// Create Scale Node
auto node0_ptr = std::make_shared<GradNodeScale>(1, 1);
node0_ptr->SetAttributes_scale(5.0 /*scale*/);
// Set grad in/out meta
node0_ptr->SetDefaultGradInOutMeta();
// Output_tensor set GradNode、OutRank、StopGradient propertis
AutogradMeta* auto_grad_meta = EagerUtils::autograd_meta(&output_tensor);
auto_grad_meta->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(node0_ptr));
auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta->SetStopGradient(false);
// Get autograd_meta from input tensor
AutogradMeta* auto_grad_meta1 =
EagerUtils::unsafe_autograd_meta(leaf_tensor);
// Connect Tensor and AccumulationNode via AutoGradMeta
auto acc_node_ptr =
std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta1);
// input tensor set GradNode、OutRank、StopGradient propertis
auto_grad_meta1->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta1->SetStopGradient(false);
// grad_node Add Edges
std::vector<egr::AutogradMeta*> res = {auto_grad_meta1};
node0_ptr->AddEdges(&res, 0);
}
std::vector<paddle::experimental::Tensor> outs = {output_tensor};
// Run Grad
auto result = Grad(outs, {leaf_tensor}, {});
// Check Output Value
eager_test::CompareTensorWithValue<float>(result[0], 5.0);
}
TEST(Grad, SingleNodeCustomGrad) {
// Prepare Device Contexts
eager_test::InitEnv(paddle::platform::CPUPlace());
// Prepare Inputs
std::vector<paddle::experimental::Tensor> target_tensors;
paddle::framework::DDim ddim = phi::make_ddim({4, 16, 16, 32});
// Create Target Tensor
paddle::experimental::Tensor tensor = egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, false /*is_leaf*/);
target_tensors.emplace_back(std::move(tensor));
std::vector<paddle::experimental::Tensor> grad_tensors;
// Create Grad Tensor
paddle::experimental::Tensor grad_tensor =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 10.0 /*value*/, false /*is_leaf*/);
grad_tensors.emplace_back(std::move(grad_tensor));
paddle::experimental::Tensor leaf_tensor =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, true /*is_leaf*/);
{
// Create Scale Node
auto node0_ptr = std::make_shared<GradNodeScale>(1, 1);
node0_ptr->SetAttributes_scale(5.0 /*scale*/);
// Set grad in/out meta
node0_ptr->SetDefaultGradInOutMeta();
// Connect Tensor and Node via AutoGradMeta
AutogradMeta* auto_grad_meta =
EagerUtils::autograd_meta(&(target_tensors[0]));
auto_grad_meta->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(node0_ptr));
auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta->SetStopGradient(false);
AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
// Connect Tensor and AccumulationNode via AutoGradMeta
auto acc_node_ptr =
std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta1);
auto_grad_meta1->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta1->SetStopGradient(false);
std::vector<egr::AutogradMeta*> res = {auto_grad_meta1};
node0_ptr->AddEdges(&res, 0);
}
auto result = Grad(target_tensors, {leaf_tensor}, grad_tensors);
// Check Output Value
eager_test::CompareTensorWithValue<float>(result[0], 50.0);
}
/*
Node1
|
Node0
|
{ } // empty grad tensor
*/
TEST(Grad, LinearNodes) {
// Prepare Device Contexts
eager_test::InitEnv(paddle::platform::CPUPlace());
// Prepare Target Tensor
std::vector<paddle::experimental::Tensor> target_tensors;
paddle::framework::DDim ddim = phi::make_ddim({4, 16, 16, 32});
// Create Target Tensor
paddle::experimental::Tensor tensor = egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, false /*is_leaf*/);
target_tensors.emplace_back(std::move(tensor));
paddle::experimental::Tensor leaf_tensor =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, true /*is_leaf*/);
{
// Create Node0
auto node0_ptr = std::make_shared<GradNodeScale>(1, 1);
node0_ptr->SetAttributes_scale(5.0 /*scale*/);
// Set grad in/out meta for node0
node0_ptr->SetDefaultGradInOutMeta();
// Create Node1
auto node1_ptr = std::make_shared<GradNodeScale>(1, 1);
node1_ptr->SetAttributes_scale(10.0 /*scale*/);
// Set grad in/out meta for node1
node1_ptr->SetDefaultGradInOutMeta();
// Connect Input Tensor and Node0 via AutoGradMeta
AutogradMeta* auto_grad_meta =
EagerUtils::autograd_meta(&(target_tensors[0]));
auto_grad_meta->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(node0_ptr));
auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta->SetStopGradient(false);
// Connect Node0 -> Node1 via Edge
auto meta0 = egr::AutogradMeta();
meta0.SetStopGradient(false);
meta0.SetSingleOutRankWithSlot(0, 0);
meta0.SetGradNode(node1_ptr);
std::vector<egr::AutogradMeta*> res0 = {&meta0};
node0_ptr->AddEdges(&res0, 0);
AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
// Connect Tensor and AccumulationNode via AutoGradMeta
auto acc_node_ptr =
std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta1);
auto_grad_meta1->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta1->SetStopGradient(false);
std::vector<egr::AutogradMeta*> res1 = {auto_grad_meta1};
node1_ptr->AddEdges(&res1, 0);
}
// Use Empty Grad Tensor
auto result = Grad(target_tensors, {leaf_tensor}, {});
// Check Output Value
eager_test::CompareTensorWithValue<float>(result[0], 50.0);
}
/*
Node2
| |
Node0 Node1
| |
in0 in1
*/
TEST(Grad, WithAccumulation) {
// Prepare Device Contexts
eager_test::InitEnv(paddle::platform::CPUPlace());
// Prepare Inputs
paddle::framework::DDim ddim = phi::make_ddim({4, 16, 16, 32});
// Create Target Tensor
std::vector<paddle::experimental::Tensor> target_tensors;
paddle::experimental::Tensor tensor0 = egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, false /*is_leaf*/);
paddle::experimental::Tensor tensor1 = egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 1.0 /*value*/, false /*is_leaf*/);
target_tensors.emplace_back(std::move(tensor0));
target_tensors.emplace_back(std::move(tensor1));
// Create Grad Tensor
std::vector<paddle::experimental::Tensor> grad_tensors;
paddle::experimental::Tensor grad_tensor0 =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 5.0 /*value*/, false /*is_leaf*/);
paddle::experimental::Tensor grad_tensor1 =
egr_utils_api::CreateTensorWithValue(
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 10.0 /*value*/, false /*is_leaf*/);
grad_tensors.emplace_back(std::move(grad_tensor0));
grad_tensors.emplace_back(std::move(grad_tensor1));
paddle::experimental::Tensor leaf_tensor;
{
// Create Node0
auto node0_ptr = std::make_shared<GradNodeScale>(1, 1);
node0_ptr->SetAttributes_scale(5.0 /*scale*/);
node0_ptr->SetDefaultGradInOutMeta();
// Create Node1
auto node1_ptr = std::make_shared<GradNodeScale>(1, 1);
node1_ptr->SetAttributes_scale(10.0 /*scale*/);
node1_ptr->SetDefaultGradInOutMeta();
// Create Node2
auto node2_ptr = std::make_shared<GradNodeScale>(1, 1);
node2_ptr->SetAttributes_scale(20.0 /*scale*/);
node2_ptr->SetDefaultGradInOutMeta();
// Connect Inp0 and Node0 via AutoGradMeta
AutogradMeta* auto_grad_meta0 =
EagerUtils::autograd_meta(&(target_tensors[0]));
auto_grad_meta0->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(node0_ptr));
auto_grad_meta0->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta0->SetStopGradient(false);
// Connect Inp1 and Node1 via AutoGradMeta
AutogradMeta* auto_grad_meta1 =
EagerUtils::autograd_meta(&(target_tensors[1]));
auto_grad_meta1->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(node1_ptr));
auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta1->SetStopGradient(false);
// Connect Node0 -> Node2 via Edge
auto meta0 = egr::AutogradMeta();
meta0.SetStopGradient(false);
meta0.SetSingleOutRankWithSlot(0, 0);
meta0.SetGradNode(node2_ptr);
std::vector<egr::AutogradMeta*> res0 = {&meta0};
node0_ptr->AddEdges(&res0, 0);
// Connect Node1 -> Node2 via Edge
auto meta1 = egr::AutogradMeta();
meta1.SetStopGradient(false);
meta1.SetSingleOutRankWithSlot(0, 0);
meta1.SetGradNode(node2_ptr);
std::vector<egr::AutogradMeta*> res1 = {&meta1};
node1_ptr->AddEdges(&res1, 0);
AutogradMeta* auto_grad_meta2 = EagerUtils::autograd_meta(&leaf_tensor);
// Connect Tensor and AccumulationNode via AutoGradMeta
auto acc_node_ptr =
std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta2);
auto_grad_meta2->SetGradNode(
std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
auto_grad_meta2->SetSingleOutRankWithSlot(0, 0);
auto_grad_meta2->SetStopGradient(false);
std::vector<egr::AutogradMeta*> res2 = {auto_grad_meta2};
node2_ptr->AddEdges(&res2, 0);
}
auto result = Grad(target_tensors, {leaf_tensor}, grad_tensors);
eager_test::CompareTensorWithValue<float>(result[0], 2500.0);
}
} // namespace egr
paddle/fluid/eager/tests/task_tests/hook_test.cc
浏览文件 @ 4b269baa
......@@ -132,7 +132,7 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
leaf_tensor); // result: 4.0*5.0 + 3.0 = 23.0
}
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(target_tensor, 4.0);
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 23.0);
......@@ -199,7 +199,7 @@ TEST(RetainGrad, HookAfterRetainGrad) {
leaf_tensor, std::make_shared<egr::CppTensorHook>(hook_function));
}
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(target_tensor, 1.0);
eager_test::CompareGradTensorWithValue<float>(leaf_tensor, 23.0);
}
......
paddle/fluid/eager/tests/task_tests/hook_test_intermidiate.cc
浏览文件 @ 4b269baa
......@@ -108,7 +108,7 @@ void test_sigmoid(bool is_remove_gradient_hook) {
}
VLOG(6) << "Runing Backward";
Backward(target_tensors, {});
RunBackward(target_tensors, {});
VLOG(6) << "Finish Backward";
eager_test::CompareGradTensorWithValue<float>(
......@@ -166,7 +166,7 @@ void test_elementwiseAdd(bool is_remove_gradient_hook) {
grad_node_tmp->RemoveGradientHook(hook_id);
}
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(X, 1.0);
eager_test::CompareGradTensorWithValue<float>(
......@@ -224,7 +224,7 @@ void test_matmul(bool is_remove_gradient_hook) {
grad_node_tmp->RemoveGradientHook(hook_id);
}
Backward(target_tensors, {});
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(X, 2.0 * 20);
eager_test::CompareGradTensorWithValue<float>(
......
paddle/fluid/eager/to_static/run_program_op_node.h
浏览文件 @ 4b269baa
......@@ -370,8 +370,8 @@ class GradNodeRunProgram : public egr::GradNodeBase {
~GradNodeRunProgram() override = default;
// Functor: perform backward computations
virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
const std::vector<std::vector<paddle::experimental::Tensor>> &grads,
bool create_graph) override {
const std::vector<std::vector<paddle::experimental::Tensor>> &grads)
override {
VLOG(3) << "Running Eager Backward Node: GradNodeRunProgram";
PADDLE_ENFORCE_EQ(
grads.size(), 1,
......@@ -415,12 +415,6 @@ class GradNodeRunProgram : public egr::GradNodeBase {
// return {x_grad, details::DereferenceTensors(params_grad_ptr)};
}
void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
bool IsTensorWrappersCleared() override {
VLOG(6) << "Do nothing here now";
return false;
}
// SetAttrMap
void SetAttrMap(const paddle::framework::AttributeMap &attrs) {
attrs_ = attrs;
......
paddle/fluid/pybind/eager_functions.cc
浏览文件 @ 4b269baa
......@@ -122,33 +122,13 @@ static PyObject* eager_api_run_backward(PyObject* self, PyObject* args,
EAGER_TRY
auto tensors = CastPyArg2VectorOfTensor(PyTuple_GET_ITEM(args, 0), 0);
auto grad_tensors = CastPyArg2VectorOfTensor(PyTuple_GET_ITEM(args, 1), 1);
egr::Backward(tensors, grad_tensors,
egr::RunBackward(tensors, grad_tensors,
CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 2), 2));
Py_INCREF(Py_None);
return Py_None;
EAGER_CATCH_AND_THROW_RETURN_NULL
}
static PyObject* eager_api_run_partial_grad(PyObject* self, PyObject* args,
PyObject* kwargs) {
EAGER_TRY
auto tensors = CastPyArg2VectorOfTensor(PyTuple_GET_ITEM(args, 0), 0);
auto inputs = CastPyArg2VectorOfTensor(PyTuple_GET_ITEM(args, 1), 1);
auto grad_tensors = CastPyArg2VectorOfTensor(PyTuple_GET_ITEM(args, 2), 2);
auto retain_graph = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 3), 3);
auto create_graph = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 4), 4);
auto only_inputs = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 5), 5);
auto allow_unused = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 6), 6);
auto no_grad_vars = CastPyArg2VectorOfTensor(PyTuple_GET_ITEM(args, 7), 7);
std::vector<paddle::experimental::Tensor> result =
egr::Grad(tensors, inputs, grad_tensors, retain_graph, create_graph,
only_inputs, allow_unused, no_grad_vars);
VLOG(1) << " in eager_api_run_partial_grad, after runing egr::Grad";
return ToPyObject(result, true /* return_py_none_if_not_initialize */);
EAGER_CATCH_AND_THROW_RETURN_NULL
}
static PyObject* eager_api_tensor_copy(PyObject* self, PyObject* args,
PyObject* kwargs) {
EAGER_TRY
......@@ -472,9 +452,6 @@ PyMethodDef variable_functions[] = {
METH_VARARGS | METH_KEYWORDS, NULL},
{"run_backward", (PyCFunction)(void (*)(void))eager_api_run_backward,
METH_VARARGS | METH_KEYWORDS, NULL},
{"run_partial_grad",
(PyCFunction)(void (*)(void))eager_api_run_partial_grad,
METH_VARARGS | METH_KEYWORDS, NULL},
{"_run_custom_op", (PyCFunction)(void (*)(void))eager_api_run_costum_op,
METH_VARARGS | METH_KEYWORDS, NULL},
{"tensor_copy", (PyCFunction)(void (*)(void))eager_api_tensor_copy,
......
paddle/fluid/pybind/eager_utils.cc
浏览文件 @ 4b269baa
......@@ -492,15 +492,10 @@ PyObject* ToPyObject(const std::vector<double>& value) {
return result;
}
PyObject* ToPyObject(const std::vector<paddle::experimental::Tensor>& value,
bool return_py_none_if_not_initialize) {
PyObject* ToPyObject(const std::vector<paddle::experimental::Tensor>& value) {
PyObject* result = PyList_New((Py_ssize_t)value.size());
for (size_t i = 0; i < value.size(); i++) {
if (!value[i].initialized() && return_py_none_if_not_initialize) {
Py_INCREF(Py_None);
PyList_SET_ITEM(result, static_cast<Py_ssize_t>(i), Py_None);
} else {
PyObject* obj = p_tensor_type->tp_alloc(p_tensor_type, 0);
if (obj) {
auto v = reinterpret_cast<TensorObject*>(obj);
......@@ -512,7 +507,6 @@ PyObject* ToPyObject(const std::vector<paddle::experimental::Tensor>& value,
}
PyList_SET_ITEM(result, static_cast<Py_ssize_t>(i), obj);
}
}
return result;
}
......
paddle/fluid/pybind/eager_utils.h
浏览文件 @ 4b269baa
......@@ -68,8 +68,7 @@ PyObject* ToPyObject(const std::vector<int>& value);
PyObject* ToPyObject(const std::vector<int64_t>& value);
PyObject* ToPyObject(const std::vector<float>& value);
PyObject* ToPyObject(const std::vector<double>& value);
PyObject* ToPyObject(const std::vector<paddle::experimental::Tensor>& value,
bool return_py_none_if_not_initialize = false);
PyObject* ToPyObject(const std::vector<paddle::experimental::Tensor>& value);
PyObject* ToPyObject(const platform::Place& value);
PyObject* ToPyObject(const framework::LoDTensor* value);
PyObject* ToPyObject(const paddle::framework::proto::VarType::Type& dtype);
......
python/paddle/fluid/dygraph/base.py
浏览文件 @ 4b269baa
......@@ -565,25 +565,16 @@ def grad(outputs,
if isinstance(in_out_list, (list, tuple)):
assert len(in_out_list) > 0, "{} cannot be empty".format(name)
for each_var in in_out_list:
if core._in_eager_mode():
assert isinstance(
each_var, core.eager.
Tensor), "Elements of {} must be Tensor".format(name)
else:
assert isinstance(
each_var,
core.VarBase), "Elements of {} must be Variable".format(
name)
return in_out_list
else:
if core._in_eager_mode():
assert isinstance(
in_out_list, core.eager.
Tensor), "{} must be Tensor or list of Tensor".format(name)
else:
assert isinstance(
in_out_list, core.VarBase
), "{} must be Variable or list of Variable".format(name)
in_out_list,
core.VarBase), "{} must be Variable or list of Variable".format(
name)
return [in_out_list]
outputs = check_in_out(outputs, 'outputs')
......@@ -595,11 +586,6 @@ def grad(outputs,
for each_var in grad_outputs:
if each_var is not None:
if core._in_eager_mode():
assert isinstance(
each_var, core.eager.Tensor
), "grad_outputs must be None, a Variable or a list containing None or Variables"
else:
assert isinstance(
each_var, core.VarBase
), "grad_outputs must be None, a Variable or a list containing None or Variables"
......@@ -614,27 +600,14 @@ def grad(outputs,
no_grad_vars = []
elif isinstance(no_grad_vars, core.VarBase):
no_grad_vars = [no_grad_vars]
elif isinstance(no_grad_vars, core.eager.Tensor):
no_grad_vars = [no_grad_vars]
elif isinstance(no_grad_vars, (list, tuple, set)):
no_grad_vars = list(no_grad_vars)
for var in no_grad_vars:
if core._in_eager_mode():
assert isinstance(
var,
core.eager.Tensor), "no_grad_vars can only contains Tensor"
else:
assert isinstance(
var,
core.VarBase), "no_grad_vars can only contains Variable"
var, core.VarBase), "no_grad_vars can only contains Variable"
else:
if core._in_eager_mode():
raise AssertionError(
"no_grad_vars must be None, Tensor or list/tuple/set of Tensors")
else:
raise AssertionError(
"no_grad_vars must be None, Variable or list/tuple/set of Variables"
)
"no_grad_vars must be None, Variable or list/tuple/set of Variables")
assert isinstance(create_graph, bool), "create_graph must be True or False"
......@@ -649,11 +622,6 @@ def grad(outputs,
assert isinstance(only_inputs, bool), "only_inputs must be True or False"
assert only_inputs, "only_inputs=False is not supported yet"
if core._in_eager_mode():
return core.eager.run_partial_grad(
outputs, inputs, grad_outputs, retain_graph, create_graph,
only_inputs, allow_unused, no_grad_vars)
place = core.Place()
place.set_place(framework._current_expected_place())
return core.dygraph_partial_grad(inputs, outputs, grad_outputs,
......
python/paddle/fluid/tests/unittests/test_egr_python_api.py
浏览文件 @ 4b269baa
......@@ -52,7 +52,7 @@ class EagerScaleTestCase(unittest.TestCase):
out_eager = core.eager.scale(data_eager, 1.0, 0.9, True, True)
self.assertIsNone(data_eager.grad)
out_eager.backward(grad_eager, False)
self.assertIsNotNone(data_eager.grad)
self.assertTrue(data_eager.grad._is_initialized())
self.assertTrue(np.array_equal(data_eager.grad.numpy(), input_data))
def test_retain_grad_and_run_backward_raises(self):
......
python/paddle/fluid/tests/unittests/test_imperative_double_grad.py
浏览文件 @ 4b269baa
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -19,9 +19,6 @@ from paddle.vision.models import resnet50, resnet101
import unittest
from unittest import TestCase
import numpy as np
import paddle.compat as cpt
from paddle.fluid.framework import _test_eager_guard
import paddle.fluid.core as core
def _dygraph_guard_(func):
......@@ -43,80 +40,6 @@ def random_var(size, low=-1, high=1, dtype='float32'):
return fluid.dygraph.to_variable(x_np)
class TestEagerGrad(TestCase):
def func_simple_example_eager_grad(self):
np.random.seed(2021)
paddle.set_device('cpu')
np_x = np.random.random((3, 3))
np_y = np.random.random((3, 1))
x = paddle.to_tensor(np_x, dtype="float64", stop_gradient=False)
y = paddle.to_tensor(np_y, dtype="float64", stop_gradient=False)
out = paddle.matmul(x, y)
dx = fluid.dygraph.grad(out, x)
dout = np.ones_like(np_y)
expected_dx = np.matmul(dout, np.transpose(np_y))
# stop_gradient = !create_graph, create_graph default false
self.assertEqual(dx[0].stop_gradient, True)
self.assertTrue(np.allclose(dx[0].numpy(), expected_dx[0]))
def test_simple_example_eager_grad(self):
with _test_eager_guard():
self.func_simple_example_eager_grad()
self.func_simple_example_eager_grad()
def func_simple_example_eager_grad_allow_unused(self):
np.random.seed(2021)
paddle.set_device('cpu')
np_x = np.random.random((3, 3))
np_y = np.random.random((3, 1))
np_z = np.random.random((3, 1))
x = paddle.to_tensor(np_x, dtype="float64", stop_gradient=False)
y = paddle.to_tensor(np_y, dtype="float64", stop_gradient=False)
z = paddle.to_tensor(np_z, dtype="float64", stop_gradient=False)
out_z = paddle.nn.functional.sigmoid(z)
out = paddle.matmul(x, y)
dx = fluid.dygraph.grad(out, [x, z], allow_unused=True)
dout = np.ones_like(np_y)
expected_dx = np.matmul(dout, np.transpose(np_y))
self.assertTrue(np.allclose(dx[0].numpy(), expected_dx[0]))
# stop_gradient = !create_graph, create_graph default false
self.assertEqual(dx[0].stop_gradient, True)
# x is unused input in the graph
self.assertEqual(dx[1], None)
def test_simple_example_eager_grad_allow_unused(self):
with _test_eager_guard():
self.func_simple_example_eager_grad_allow_unused()
self.func_simple_example_eager_grad_allow_unused()
def func_simple_example_eager_grad_not_allow_unused(self):
np.random.seed(2021)
paddle.set_device('cpu')
np_x = np.random.random((3, 3))
np_y = np.random.random((3, 1))
np_z = np.random.random((3, 1))
x = paddle.to_tensor(np_x, dtype="float64", stop_gradient=False)
y = paddle.to_tensor(np_y, dtype="float64", stop_gradient=False)
z = paddle.to_tensor(np_z, dtype="float64", stop_gradient=False)
out_z = paddle.nn.functional.sigmoid(z)
out = paddle.matmul(x, y)
try:
# allow_unused is false in default
dx = fluid.dygraph.grad(out, [x, z])
except ValueError as e:
error_msg = cpt.get_exception_message(e)
assert error_msg.find("allow_unused") > 0
def test_simple_example_eager_grad_not_allow_unused(self):
with _test_eager_guard():
self.func_simple_example_eager_grad_not_allow_unused()
self.func_simple_example_eager_grad_not_allow_unused()
class TestDygraphDoubleGrad(TestCase):
def setUp(self):
self.sort_sum_gradient = False
......@@ -141,7 +64,7 @@ class TestDygraphDoubleGrad(TestCase):
allow_unused=allow_unused)
@dygraph_guard
def func_exception(self):
def test_exception(self):
with self.assertRaises(AssertionError):
self.grad(None, None)
......@@ -170,13 +93,8 @@ class TestDygraphDoubleGrad(TestCase):
with self.assertRaises(AssertionError):
self.grad([random_var(shape)], [random_var(shape)], no_grad_vars=1)
def test_exception(self):
with _test_eager_guard():
self.func_exception()
self.func_exception()
@dygraph_guard
def func_simple_example(self):
def test_simple_example(self):
x = random_var(self.shape)
x.stop_gradient = False
y = x + 1
......@@ -205,44 +123,8 @@ class TestDygraphDoubleGrad(TestCase):
self.assertNotEqual(grad_with_none_and_not_none.stop_gradient,
create_graph)
def test_simple_example(self):
with _test_eager_guard():
self.func_simple_example()
self.func_simple_example()
@dygraph_guard
def func_example_no_grad_vars(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
x.stop_gradient = False
y1 = fluid.layers.relu(x)
y2 = fluid.layers.relu(x)
z = y1 + y2
w = z * z
w_mean = fluid.layers.reduce_mean(w)
del y1, z, w
dx_actual, = self.grad(
[w_mean], [x], create_graph=True, no_grad_vars=[y2])
self.assertFalse(y2.stop_gradient)
self.assertFalse(dx_actual.stop_gradient)
dx_expected = (1.0 / float(numel) * (np.maximum(x_np, 0) + y2.numpy()) *
(x_np > 0) * 2).astype('float32')
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
def test_example_no_grad_vars(self):
with _test_eager_guard():
self.func_example_no_grad_vars()
self.func_example_no_grad_vars()
@dygraph_guard
def func_none_one_initial_gradient(self):
def test_none_one_initial_gradient(self):
numel = 1
for s in self.shape:
numel *= s
......@@ -308,13 +190,8 @@ class TestDygraphDoubleGrad(TestCase):
np.array_equal(grad_z.numpy(),
original_random_grad_z))
def test_none_one_initial_gradient(self):
with _test_eager_guard():
self.func_none_one_initial_gradient()
self.func_none_one_initial_gradient()
@dygraph_guard
def func_example_with_gradient_accumulation_and_create_graph(self):
def test_example_with_gradient_accumulation_and_create_graph(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
......@@ -337,15 +214,12 @@ class TestDygraphDoubleGrad(TestCase):
(x_np > 0) * 2).astype('float32')
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
if core._in_eager_mode():
pass
else:
loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
loss.backward(retain_graph=True)
x_grad_actual = x.gradient()
x_grad_expected = (2.0 / float(numel) * (
x_np + dx_expected *
x_grad_expected = (2.0 / float(numel) *
(x_np + dx_expected *
(x_np > 0) * 2 / float(numel))).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
......@@ -357,13 +231,8 @@ class TestDygraphDoubleGrad(TestCase):
(x_np > 0) * 2 / float(numel))).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
def test_example_with_gradient_accumulation_and_create_graph(self):
with _test_eager_guard():
self.func_example_with_gradient_accumulation_and_create_graph()
self.func_example_with_gradient_accumulation_and_create_graph()
@dygraph_guard
def func_example_with_gradient_accumulation_and_no_grad_vars(self):
def test_example_with_gradient_accumulation_and_no_grad_vars(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
......@@ -387,25 +256,17 @@ class TestDygraphDoubleGrad(TestCase):
(x_np > 0) * 2).astype('float32')
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
if core._in_eager_mode():
pass
else:
loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
loss.backward()
x_grad_actual = x.gradient()
x_grad_expected = (2.0 / float(numel) * (
x_np + dx_expected *
x_grad_expected = (2.0 / float(numel) *
(x_np + dx_expected *
(x_np > 0) * 4 / float(numel))).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
def test_example_with_gradient_accumulation_and_no_grad_vars(self):
with _test_eager_guard():
self.func_example_with_gradient_accumulation_and_no_grad_vars()
self.func_example_with_gradient_accumulation_and_no_grad_vars()
@dygraph_guard
def func_example_with_gradient_accumulation_and_not_create_graph(self):
def test_example_with_gradient_accumulation_and_not_create_graph(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
......@@ -428,9 +289,6 @@ class TestDygraphDoubleGrad(TestCase):
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
if core._in_eager_mode():
pass
else:
loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
loss.backward()
......@@ -438,11 +296,6 @@ class TestDygraphDoubleGrad(TestCase):
x_grad_expected = (2.0 * x_np / float(numel)).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
def test_example_with_gradient_accumulation_and_not_create_graph(self):
with _test_eager_guard():
self.func_example_with_gradient_accumulation_and_not_create_graph()
self.func_example_with_gradient_accumulation_and_not_create_graph()
class TestDygraphDoubleGradSortGradient(TestDygraphDoubleGrad):
def setUp(self):
......@@ -451,7 +304,7 @@ class TestDygraphDoubleGradSortGradient(TestDygraphDoubleGrad):
class TestDygraphDoubleGradVisitedUniq(TestCase):
def func_compare(self):
def test_compare(self):
value = np.random.uniform(-0.5, 0.5, 100).reshape(10, 2,
5).astype("float32")
......@@ -496,11 +349,6 @@ class TestDygraphDoubleGradVisitedUniq(TestCase):
self.assertTrue(np.array_equal(grad_1, grad_2))
def test_compare(self):
with _test_eager_guard():
self.func_compare()
self.func_compare()
class TestRaiseNoDoubleGradOp(TestCase):
def raise_no_grad_op(self):
......
python/paddle/fluid/tests/unittests/test_paddle_imperative_double_grad.py
浏览文件 @ 4b269baa
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -18,8 +18,6 @@ import unittest
from unittest import TestCase
import numpy as np
import paddle
from paddle.fluid.framework import _test_eager_guard
import paddle.fluid.core as core
def _dygraph_guard_(func):
......@@ -64,7 +62,7 @@ class TestDygraphDoubleGrad(TestCase):
allow_unused=allow_unused)
@dygraph_guard
def func_exception(self):
def test_exception(self):
with self.assertRaises(AssertionError):
self.grad(None, None)
......@@ -93,13 +91,8 @@ class TestDygraphDoubleGrad(TestCase):
with self.assertRaises(AssertionError):
self.grad([random_var(shape)], [random_var(shape)], no_grad_vars=1)
def test_exception(self):
with _test_eager_guard():
self.func_exception()
self.func_exception()
@dygraph_guard
def func_simple_example(self):
def test_simple_example(self):
x = random_var(self.shape)
x.stop_gradient = False
y = x + 1
......@@ -128,13 +121,8 @@ class TestDygraphDoubleGrad(TestCase):
self.assertNotEqual(grad_with_none_and_not_none.stop_gradient,
create_graph)
def test_simple_example(self):
with _test_eager_guard():
self.func_simple_example()
self.func_simple_example()
@dygraph_guard
def func_none_one_initial_gradient(self):
def test_none_one_initial_gradient(self):
numel = 1
for s in self.shape:
numel *= s
......@@ -200,13 +188,8 @@ class TestDygraphDoubleGrad(TestCase):
np.array_equal(grad_z.numpy(),
original_random_grad_z))
def test_none_one_initial_gradient(self):
with _test_eager_guard():
self.func_none_one_initial_gradient()
self.func_none_one_initial_gradient()
@dygraph_guard
def func_example_with_gradient_accumulation_and_create_graph(self):
def test_example_with_gradient_accumulation_and_create_graph(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
......@@ -229,25 +212,17 @@ class TestDygraphDoubleGrad(TestCase):
(x_np > 0) * 2).astype('float32')
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
if core._in_eager_mode():
pass
else:
loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
loss.backward()
x_grad_actual = x.gradient()
x_grad_expected = (2.0 / float(numel) * (
x_np + dx_expected *
x_grad_expected = (2.0 / float(numel) *
(x_np + dx_expected *
(x_np > 0) * 2 / float(numel))).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
def test_example_with_gradient_accumulation_and_create_graph(self):
with _test_eager_guard():
self.func_example_with_gradient_accumulation_and_create_graph()
self.func_example_with_gradient_accumulation_and_create_graph()
@dygraph_guard
def func_example_with_gradient_accumulation_and_no_grad_vars(self):
def test_example_with_gradient_accumulation_and_no_grad_vars(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
......@@ -271,25 +246,17 @@ class TestDygraphDoubleGrad(TestCase):
(x_np > 0) * 2).astype('float32')
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
if core._in_eager_mode():
pass
else:
loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
loss.backward()
x_grad_actual = x.gradient()
x_grad_expected = (2.0 / float(numel) * (
x_np + dx_expected *
x_grad_expected = (2.0 / float(numel) *
(x_np + dx_expected *
(x_np > 0) * 4 / float(numel))).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
def test_example_with_gradient_accumulation_and_no_grad_vars(self):
with _test_eager_guard():
self.func_example_with_gradient_accumulation_and_no_grad_vars()
self.func_example_with_gradient_accumulation_and_no_grad_vars()
@dygraph_guard
def func_example_with_gradient_accumulation_and_not_create_graph(self):
def test_example_with_gradient_accumulation_and_not_create_graph(self):
x = random_var(self.shape)
x_np = x.numpy()
numel = x_np.size
......@@ -312,9 +279,6 @@ class TestDygraphDoubleGrad(TestCase):
self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
if core._in_eager_mode():
pass
else:
loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
loss.backward()
......@@ -322,11 +286,6 @@ class TestDygraphDoubleGrad(TestCase):
x_grad_expected = (2.0 * x_np / float(numel)).astype('float32')
self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
def test_example_with_gradient_accumulation_and_not_create_graph(self):
with _test_eager_guard():
self.func_example_with_gradient_accumulation_and_not_create_graph()
self.func_example_with_gradient_accumulation_and_not_create_graph()
class TestDygraphDoubleGradSortGradient(TestDygraphDoubleGrad):
def setUp(self):
......
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