improve optimizer support (PaddlePaddle#641) (PaddlePaddle#649)

gglin001 · web-flow · commit d1aa750dbf53 · 2022-04-20T10:49:30.000+08:00
* update GetOptPrePostfix
diff --git a/paddle/fluid/platform/device/ipu/ipu_compiler.cc b/paddle/fluid/platform/device/ipu/ipu_compiler.cc
@@ -285,42 +285,43 @@ void Compiler::LowerConstants(const Scope* scope) {
 
 void Compiler::LowerWeights(const Scope* scope) {
   VLOG(10) << "enter Compiler::LowerWeights";
-  // at this step, the graph doesn't contains optimizer related states
+  // At this step, the graph doesn't contains optimizer related states
   for (auto id : graph_helper_->sorted_vars_id) {
     auto* node = graph_helper_->nodes_id_map[id];
-    if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
-      if (node->Var()->Persistable() && node->inputs.empty()) {
-        auto var_name = node->Var()->Name();
-        if (resources_->tensors.count(var_name) != 0) {
-          VLOG(10) << "found existed one, skip lowering Weight: " << var_name;
-          continue;
-        }
-        if (var_name.rfind("learning_rate", 0) == 0) {
-          VLOG(10) << "skip learning_rate_var: " << var_name;
-          continue;
-        }
-        VLOG(10) << "lowering weight: " << var_name;
-
-        auto var = scope->FindVar(var_name);
-        if (var) {
-          auto tensor = var->Get<framework::LoDTensor>();
-          auto dtype = PdDataType2PopartType(tensor.dtype());
-          auto shape = std::vector<int64_t>();
-          for (size_t i = 0; i < tensor.dims().size(); ++i) {
-            shape.push_back(tensor.dims().at(i));
-          }
-          popart::TensorInfo tensor_info(dtype, shape);
-          popart::ConstVoidData const_data{tensor.data(), tensor_info};
-          if (!node->outputs.empty()) {
-            auto op_node = node->outputs[0];
-            PushNameScope(op_node->Op());
-            popart::TensorId result =
-                builder_->addInitializedInputTensor(const_data, var_name);
-            PopNameScope(op_node->Op());
-            resources_->tensors.emplace(var_name, result);
-            resources_->weights.push_back(var_name);
-          }
-        }
+    // Weights are var node and Persistable
+    if (node->IsVar() && !node->IsCtrlVar() && node->Var() &&
+        node->Var()->Persistable()) {
+      // Weights are Parameter in training mode
+      if (ipu_strategy_->is_training && !node->Var()->IsParameter()) {
+        continue;
+      }
+      auto var_name = node->Var()->Name();
+      // Some op has same input and output tensor, like batchnorm
+      if (resources_->tensors.count(var_name) != 0) {
+        VLOG(10) << "found existed one, skip lowering Weight: " << var_name;
+        continue;
+      }
+      VLOG(10) << "lowering weight: " << var_name;
+      auto var = scope->FindVar(var_name);
+      PADDLE_ENFORCE_NOT_NULL(
+          var, platform::errors::NotFound("Tensor %s is not found in the scope",
+                                          var_name));
+      auto tensor = var->Get<framework::LoDTensor>();
+      auto dtype = PdDataType2PopartType(tensor.dtype());
+      auto shape = std::vector<int64_t>();
+      for (size_t i = 0; i < tensor.dims().size(); ++i) {
+        shape.push_back(tensor.dims().at(i));
+      }
+      popart::TensorInfo tensor_info(dtype, shape);
+      popart::ConstVoidData const_data{tensor.data(), tensor_info};
+      if (!node->outputs.empty()) {
+        auto op_node = node->outputs[0];
+        PushNameScope(op_node->Op());
+        popart::TensorId result =
+            builder_->addInitializedInputTensor(const_data, var_name);
+        PopNameScope(op_node->Op());
+        resources_->tensors.emplace(var_name, result);
+        resources_->weights.push_back(var_name);
       }
     }
   }
diff --git a/paddle/fluid/platform/device/ipu/ipu_executor.cc b/paddle/fluid/platform/device/ipu/ipu_executor.cc
@@ -20,6 +20,40 @@ namespace paddle {
 namespace platform {
 namespace ipu {
 
+// Get paddle prefix and popart postfix of weight states
+// Format: {popart_postfix, paddle_prefix}
+std::vector<std::pair<std::string, std::string>> GetOptPrePostfix(
+    const std::string &opt_type) {
+  std::vector<std::pair<std::string, std::string>> pre_post_fix;
+  // Weight self
+  pre_post_fix.push_back(std::make_pair("", ""));
+
+  // Weight states
+  // TODO(alleng) support pair("Accl1___", "_moment1_{id!=0}")
+  if (opt_type == "adam" || opt_type == "lamb" || opt_type == "adamw") {
+    pre_post_fix.push_back(std::make_pair("Accl1___", "_moment1_0"));
+    pre_post_fix.push_back(std::make_pair("Accl2___", "_moment2_0"));
+    pre_post_fix.push_back(std::make_pair("Step___", "_beta1_pow_acc_0"));
+  } else if (opt_type == "momentum") {
+    pre_post_fix.push_back(std::make_pair("Accl___", "_velocity_0"));
+  } else if (opt_type == "adamax") {
+    pre_post_fix.push_back(std::make_pair("Accl1___", "_moment_0"));
+    pre_post_fix.push_back(std::make_pair("Accl2___", "_inf_norm__0"));
+    pre_post_fix.push_back(std::make_pair("Step___", "_beta1_pow_acc_0"));
+  } else if (opt_type == "adagrad") {
+    pre_post_fix.push_back(std::make_pair("Accl1___", "_moment_0"));
+  } else if (opt_type == "adadelta") {
+    pre_post_fix.push_back(std::make_pair("Accl1___", "__avg_squared_grad_0"));
+    pre_post_fix.push_back(
+        std::make_pair("Accl2___", "__avg_squared_update_0"));
+  } else if (opt_type == "rmsprop") {
+    pre_post_fix.push_back(std::make_pair("Accl1___", "_mean_square_0"));
+    pre_post_fix.push_back(std::make_pair("Accl2___", "_mean_grad_0"));
+    pre_post_fix.push_back(std::make_pair("Accl3___", "_momentum__0"));
+  }
+  return pre_post_fix;
+}
+
 Executor::~Executor() {
   Detach();
   session_.reset();
diff --git a/paddle/fluid/platform/device/ipu/ipu_utils.cc b/paddle/fluid/platform/device/ipu/ipu_utils.cc
@@ -184,27 +184,6 @@ bool GetBoolEnv(std::string str) {
   }
 }
 
-std::vector<std::pair<std::string, std::string>> GetOptPrePostfix(
-    const std::string& opt_type) {
-  // format: {popart_tensor_id, paddle_tensor_id}, ...
-  std::vector<std::pair<std::string, std::string>> pre_post_fix;
-
-  if (opt_type == "adam" || opt_type == "lamb") {
-    pre_post_fix.push_back(std::make_pair("", ""));
-    pre_post_fix.push_back(std::make_pair("Accl1___", "_moment1_0"));
-    pre_post_fix.push_back(std::make_pair("Accl2___", "_moment2_0"));
-    pre_post_fix.push_back(std::make_pair("Step___", "_beta1_pow_acc_0"));
-  } else if (opt_type == "sgd" || opt_type == "momentum") {
-    // sgd
-    pre_post_fix.push_back(std::make_pair("", ""));
-  } else {
-    pre_post_fix.push_back(std::make_pair("", ""));
-    //
-  }
-
-  return pre_post_fix;
-}
-
 int RequestIpus(const int num_ipus) {
   // num_ipus must be pow(2, n);
   return std::pow(2, ceil(log2(num_ipus)));
diff --git a/paddle/fluid/platform/device/ipu/ipu_utils.h b/paddle/fluid/platform/device/ipu/ipu_utils.h
@@ -229,9 +229,6 @@ struct ConstantOpAttrVisitor : public boost::static_visitor<void> {
   void operator()(boost::blank) const { RaiseError(); }
 };
 
-std::vector<std::pair<std::string, std::string>> GetOptPrePostfix(
-    const std::string& opt_type);
-
 int RequestIpus(const int num_ipus);
 
 }  // namespace ipu
diff --git a/python/paddle/fluid/tests/unittests/ipu/CMakeLists.txt b/python/paddle/fluid/tests/unittests/ipu/CMakeLists.txt
@@ -11,4 +11,5 @@ if(WITH_IPU)
     set_tests_properties(test_conv_op_ipu PROPERTIES TIMEOUT 300)
     set_tests_properties(test_elemetwise_x_op_ipu PROPERTIES TIMEOUT 300)
     set_tests_properties(test_reduce_x_op_ipu PROPERTIES TIMEOUT 600)
+    set_tests_properties(test_save_load_ipu PROPERTIES TIMEOUT 600)
 endif()
diff --git a/python/paddle/fluid/tests/unittests/ipu/test_save_load_ipu.py b/python/paddle/fluid/tests/unittests/ipu/test_save_load_ipu.py
@@ -14,9 +14,11 @@
 
 import tempfile
 import unittest
+from functools import partial
 
 import numpy as np
 import paddle
+import paddle.optimizer
 import paddle.static
 from paddle.fluid.tests.unittests.ipu.op_test_ipu import IPUOpTest
 
@@ -28,7 +30,8 @@ def setUp(self):
         self.set_atol()
         self.set_data_feed()
         self.set_feed_attr()
-        self.set_op_attrs()
+        self.set_attrs()
+        self.set_optimizer()
 
     def set_data_feed(self):
         data = np.random.uniform(size=[1, 3, 10, 10])
@@ -39,15 +42,16 @@ def set_feed_attr(self):
         self.feed_shape = [x.shape for x in self.feed_fp32.values()]
         self.feed_list = list(self.feed_fp32.keys())
 
-    def set_op_attrs(self):
+    def set_attrs(self):
         self.attrs = {}
         self.attrs['steps'] = 100
         self.attrs['save_at_step'] = 20
-        self.attrs['is_training'] = True
-        self.attrs['opt_type'] = 'sgd'
         self.attrs['enable_fp16'] = False
         self.attrs['model_path'] = tempfile.TemporaryDirectory()
 
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.SGD, learning_rate=1e-1)
+
     def _test_base(self, save_otherwise_load):
         scope = paddle.static.Scope()
         main_prog = paddle.static.Program()
@@ -71,16 +75,8 @@ def _test_base(self, save_otherwise_load):
                         name='conv2d')
                     loss = paddle.mean(conv1)
 
-                    if self.attrs['is_training']:
-                        if self.attrs['opt_type'] == 'sgd':
-                            sgd = paddle.optimizer.SGD(learning_rate=1e-2)
-                            sgd.minimize(loss)
-                        elif self.attrs['opt_type'] == 'adam':
-                            adam = paddle.optimizer.Adam(learning_rate=1e-2)
-                            adam.minimize(loss)
-                        elif self.attrs['opt_type'] == 'lamb':
-                            lamb = paddle.optimizer.Lamb(learning_rate=1e-2)
-                            lamb.minimize(loss)
+                    # apply optimizer
+                    self.optimizer().minimize(loss)
                     fetch_list = [loss.name]
 
                 place = paddle.IPUPlace()
@@ -91,8 +87,7 @@ def _test_base(self, save_otherwise_load):
                     paddle.static.load(main_prog, self.attrs['model_path'].name)
 
                 ipu_strategy = paddle.static.IpuStrategy()
-                ipu_strategy.set_graph_config(
-                    is_training=self.attrs['is_training'])
+                ipu_strategy.set_graph_config(is_training=True)
                 ipu_strategy.set_precision_config(
                     enable_fp16=self.attrs['enable_fp16'])
                 ipu_program = paddle.static.IpuCompiledProgram(
@@ -131,62 +126,109 @@ def test_base(self):
         self.attrs['model_path'].cleanup()
 
 
+class TestMomentum(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Momentum, learning_rate=1e-1)
+
+
 class TestAdam(TestBase):
-    def set_op_attrs(self):
-        self.attrs = {}
-        self.attrs['steps'] = 100
-        self.attrs['save_at_step'] = 20
-        self.attrs['is_training'] = True
-        self.attrs['opt_type'] = 'adam'
-        self.attrs['enable_fp16'] = False
-        self.attrs['model_path'] = tempfile.TemporaryDirectory()
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adam, learning_rate=1e-1)
 
 
 class TestLamb(TestBase):
-    def set_op_attrs(self):
-        self.attrs = {}
-        self.attrs['steps'] = 100
-        self.attrs['save_at_step'] = 20
-        self.attrs['is_training'] = True
-        self.attrs['opt_type'] = 'lamb'
-        self.attrs['enable_fp16'] = False
-        self.attrs['model_path'] = tempfile.TemporaryDirectory()
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Lamb, learning_rate=1e-1)
+
+
+class TestAdamW(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.AdamW, learning_rate=1e-1)
+
+
+class TestAdamax(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adamax, learning_rate=1e-1)
+
+
+class TestAdagrad(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adagrad, learning_rate=1e-1)
+
+
+class TestAdadelta(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adagrad, learning_rate=1e-1)
+
+
+class TestRMSProp(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.RMSProp, learning_rate=1e-1)
+
+
+class TestCenteredRMSProp(TestBase):
+    def set_optimizer(self):
+        self.optimizer = partial(
+            paddle.optimizer.RMSProp, learning_rate=1e-1, centered=True)
 
 
 @unittest.skipIf(IPUOpTest.use_ipumodel(), "skip for ipumodel")
 class TestSGDFP16(TestBase):
-    def set_op_attrs(self):
+    def set_attrs(self):
         self.attrs = {}
         self.attrs['steps'] = 100
         self.attrs['save_at_step'] = 20
-        self.attrs['is_training'] = True
-        self.attrs['opt_type'] = 'sgd'
         self.attrs['enable_fp16'] = True
         self.attrs['model_path'] = tempfile.TemporaryDirectory()
 
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.SGD, learning_rate=1e-1)
 
-@unittest.skipIf(IPUOpTest.use_ipumodel(), "skip for ipumodel")
-class TestAdamFP16(TestBase):
-    def set_op_attrs(self):
-        self.attrs = {}
-        self.attrs['steps'] = 100
-        self.attrs['save_at_step'] = 20
-        self.attrs['is_training'] = True
-        self.attrs['opt_type'] = 'adam'
-        self.attrs['enable_fp16'] = True
-        self.attrs['model_path'] = tempfile.TemporaryDirectory()
 
+class TestMomentumFp16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Momentum, learning_rate=1e-1)
 
-@unittest.skipIf(IPUOpTest.use_ipumodel(), "skip for ipumodel")
-class TestLambFP16(TestBase):
-    def set_op_attrs(self):
-        self.attrs = {}
-        self.attrs['steps'] = 100
-        self.attrs['save_at_step'] = 20
-        self.attrs['is_training'] = True
-        self.attrs['opt_type'] = 'lamb'
-        self.attrs['enable_fp16'] = True
-        self.attrs['model_path'] = tempfile.TemporaryDirectory()
+
+class TestAdamFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adam, learning_rate=1e-1)
+
+
+class TestLambFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Lamb, learning_rate=1e-1)
+
+
+class TestAdamWFP16FP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.AdamW, learning_rate=1e-1)
+
+
+class TestAdamaxFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adamax, learning_rate=1e-1)
+
+
+class TestAdagradFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adagrad, learning_rate=1e-1)
+
+
+class TestAdadeltaFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.Adagrad, learning_rate=1e-1)
+
+
+class TestRMSPropFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(paddle.optimizer.RMSProp, learning_rate=1e-1)
+
+
+class TestCenteredRMSPropFP16(TestSGDFP16):
+    def set_optimizer(self):
+        self.optimizer = partial(
+            paddle.optimizer.RMSProp, learning_rate=1e-1, centered=True)
 
 
 if __name__ == "__main__":
