Add optimizer save and load (#16986)

* save optimizer related vars in dygraph

* test=develop, add optimizer save and load

* test=develop, add optimizer save and load

* test=develop, merge code and add multi-optimizer save and load

* test=develop, fix test_imperative_checkpoint

* test=develop, fix include error

* test=develop, fix include error

* test=develop, renew api spec

* test=develop, refine code

* test=develop, set default value for checkpoint

* test=develop, fix ci error

* test=develop, change API.spec and make api more readable

* test=develop, refine version and time stamp

* test=develop, add example code and refine code

* test=develop, refine doc

* test=develop, change version

paddle/fluid/API.spec

@@ -450,54 +450,63 @@ paddle.fluid.optimizer.SGDOptimizer.apply_gradients (ArgSpec(args=['self', 'para
 paddle.fluid.optimizer.SGDOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.SGDOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.SGDOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.SGDOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.SGDOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.MomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'use_nesterov', 'regularization', 'name'], varargs=None, keywords=None, defaults=(False, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.MomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.MomentumOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.MomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.MomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.MomentumOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.MomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.AdagradOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name', 'initial_accumulator_value'], varargs=None, keywords=None, defaults=(1e-06, None, None, 0.0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdagradOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdagradOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.AdagradOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.AdagradOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.AdagradOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.AdagradOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.AdamOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdamOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdamOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.AdamOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.AdamOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.AdamOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.AdamOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.AdamaxOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdamaxOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdamaxOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.AdamaxOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.AdamaxOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.AdamaxOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.AdamaxOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'decay', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.DecayedAdagradOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.FtrlOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.0, 0.0, -0.5, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.FtrlOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.FtrlOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.FtrlOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.FtrlOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.FtrlOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.FtrlOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.RMSPropOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum', 'centered', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, 0.0, False, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.RMSPropOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.RMSPropOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.RMSPropOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.RMSPropOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.RMSPropOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.RMSPropOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.AdadeltaOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, 0.95, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdadeltaOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.AdadeltaOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.AdadeltaOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.AdadeltaOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.AdadeltaOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.ModelAverage.__init__ (ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.ModelAverage.apply (ArgSpec(args=['self', 'executor', 'need_restore'], varargs=None, keywords=None, defaults=(True,)), ('document', '648010d0ac1fa707dac0b89f74b0e35c'))
@@ -505,6 +514,7 @@ paddle.fluid.optimizer.ModelAverage.apply_gradients (ArgSpec(args=['self', 'para
 paddle.fluid.optimizer.ModelAverage.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.ModelAverage.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.ModelAverage.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.ModelAverage.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.ModelAverage.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.ModelAverage.restore (ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None), ('document', '5f14ea4adda2791e1c3b37ff327f6a83'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -512,18 +522,21 @@ paddle.fluid.optimizer.LarsMomentumOptimizer.apply_gradients (ArgSpec(args=['sel
 paddle.fluid.optimizer.LarsMomentumOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.LarsMomentumOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.DGCMomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'rampup_begin_step', 'rampup_step', 'sparsity', 'use_nesterov', 'local_grad_clip_norm', 'num_trainers', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1, [0.999], False, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.DGCMomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.DGCMomentumOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.DGCMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.DGCMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.DGCMomentumOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.DGCMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.LambOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'lamb_weight_decay', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.01, 0.9, 0.999, 1e-06, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.LambOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.LambOptimizer.apply_optimize (ArgSpec(args=['self', 'loss', 'startup_program', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', '5c46d1926a40f1f873ffe9f37ac89dae'))
 paddle.fluid.optimizer.LambOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
 paddle.fluid.optimizer.LambOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.optimizer.LambOptimizer.load (ArgSpec(args=['self', 'stat_dict'], varargs=None, keywords=None, defaults=None), ('document', '649a92cf7f1ea28666fd00c4ea01acde'))
 paddle.fluid.optimizer.LambOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'grad_clip'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'b15cffad0903fc81af77a0580ceb2a9b'))
 paddle.fluid.optimizer.ExponentialMovingAverage.__init__ (ArgSpec(args=['self', 'decay', 'thres_steps', 'name'], varargs=None, keywords=None, defaults=(0.999, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.ExponentialMovingAverage.apply (ArgSpec(args=['self', 'executor', 'need_restore'], varargs=None, keywords=None, defaults=(True,)), ('document', '30f494752ac8921dc5835a63637f453a'))

python/paddle/fluid/dygraph/base.py

@@ -92,3 +92,6 @@ def to_variable(value, block=None, name=None):
         return py_var
     elif isinstance(value, framework.Variable):
         return value
+    else:
+        raise TypeError(
+            "to_variable only accepts 'ndarray' and 'Variable' as value's input")

python/paddle/fluid/dygraph/checkpoint.py

-# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+# Copyright (c) 2019 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.
@@ -16,13 +16,18 @@ from __future__ import print_function
 
 import os
 import collections
-from .. import core
 from ..framework import Variable, default_main_program
+import pickle
+from . import learning_rate_scheduler
+import warnings
 
 __all__ = ['save_persistables', 'load_persistables']
 
 
-def save_persistables(vardict, dirname, filename=None):
+def save_persistables(model_dict,
+                      optimizer=None,
+                      dirname='save_dir',
+                      filename=None):
     """
     This function filters out all variables in layer.parameters from the
     give `layer` and then trys to load these variables from the folder
@@ -34,12 +39,12 @@ def save_persistables(vardict, dirname, filename=None):
     the file name.
 
     Args:
-        vardict(dict of Parameters): The parameters will
+        model_dict(dict of Parameters): The parameters will
                                     be saved. If it is None, nothing
                                     will be deal.
         dirname(str): The directory path.
         filename(str|None): The file which saved all variables. If variables were
-                            saved in differnet files, set it to None.
+                            saved in different files, set it to None.
                             Default: None
 
     Returns:
@@ -71,11 +76,11 @@ def save_persistables(vardict, dirname, filename=None):
             fluid.dygraph.save_persistables(ptb_model.state_dict(), dirname=param_path,
                                        layer=ptb_model)
     """
-    if isinstance(vardict, collections.OrderedDict):
-        _save_var_to_file(vardict, dirname, filename)
+    if isinstance(model_dict, collections.OrderedDict):
+        _save_var_to_file(model_dict, optimizer, dirname, filename)
 
 
-def load_persistables(dirname):
+def load_persistables(dirname='save_dir'):
     """
     This function trys to load persistable variables from the folder
     `dirname` or the file `filename`.
@@ -86,7 +91,8 @@ def load_persistables(dirname):
     the file name.
 
     Args:
-        dirname(str): The directory path.
+        dirname(str): The directory path. default is save_dir
+        optimizer(Optimizer): Optimizer to be saved
 
     Returns:
         dict: The parameter-dict resumed from file
@@ -103,7 +109,7 @@ def load_persistables(dirname):
     return _load_var_from_file(dirname)
 
 
-def _save_var_to_file(stat_dict, file_dir, file_name):
+def _save_var_to_file(stat_dict, optimizers, file_dir, file_name):
     save_block = default_main_program().global_block()
     save_var_map = {}
     for var_key, each_var in stat_dict.items():
@@ -117,6 +123,32 @@ def _save_var_to_file(stat_dict, file_dir, file_name):
                     'file_path': os.path.join(file_dir,
                                               os.path.normpath(each_var.name))
                 })
+    if isinstance(optimizers, (list, tuple)):
+        optimizers = optimizers
+    else:
+        optimizers = [optimizers]
+    if os.path.exists(os.path.join(file_dir, os.path.normpath("optimizers"))):
+        pass
+    else:
+        os.mkdir(os.path.join(file_dir, os.path.normpath("optimizers")))
+    for optimizer in optimizers:
+        if isinstance(optimizer._learning_rate,
+                      learning_rate_scheduler.LearningRateDecay):
+            try:
+                f = open(
+                    os.path.join(file_dir, "optimizers",
+                                 os.path.normpath(str(optimizer._name))), "wb")
+                pickle.dump(optimizer._learning_rate, f, 2)
+                f.close()
+            except ():
+                raise IOError("Can't load %s",
+                              os.path.join(
+                                  file_dir, "optimizers",
+                                  os.path.normpath(str(optimizer._name))))
+        else:
+            warnings.warn(
+                "Optimizer not saved, Only optimizer with 'LearningRateDecay' under DyGraph mode need to be saved"
+            )
 
     if file_name is not None:
         save_var_list = []
@@ -138,6 +170,8 @@ def _load_var_from_file(file_dir):
         var_name_list = []
         if os.path.exists(base_path):
             for dirpath, dirnames, filenames in os.walk(base_path):
+                if "optimizers" in dirpath:
+                    continue
                 pt = dirpath.replace(base_path, "", 1)
                 if pt.startswith("/") or pt.startswith("\\"):
                     pt = pt[1:]
@@ -152,6 +186,7 @@ def _load_var_from_file(file_dir):
 
     load_block = default_main_program().global_block()
     load_var_map = {}
+    load_optimizer_map = {}
     file_var_list = walk_filename(file_dir)
     for var_name in file_var_list:
         new_var = Variable(block=load_block, name=var_name)
@@ -165,8 +200,22 @@ def _load_var_from_file(file_dir):
             })
 
         load_var_map[new_var.name] = new_var
-
-    return load_var_map
+    opt_path = os.path.join(file_dir, "optimizers")
+    for _, _, optimizers in os.walk(opt_path):
+        for optimizer in optimizers:
+            try:
+                f = open(os.path.join(opt_path, optimizer), "rb")
+                load_optimizer_map[optimizer] = pickle.load(f)
+                f.close()
+            except IOError:
+                raise IOError("Can't load %s",
+                              os.path.join(
+                                  file_dir, "optimizers",
+                                  os.path.normpath(str(optimizer._name))))
+    if len(load_optimizer_map) == 0:
+        warnings.warn("No optimizer loaded")
+
+    return load_var_map, load_optimizer_map
 
 
 def _clone_var_in_block_(block, var):

python/paddle/fluid/dygraph/learning_rate_scheduler.py

@@ -63,13 +63,13 @@ class PiecewiseDecay(LearningRateDecay):
 
         self.vars = []
         for value in values:
-            self.vars.append(self.create_lr_var(value))
+            self.vars.append(value)
 
     def step(self):
         for i in range(len(self.boundaries)):
             if self.step_num < self.boundaries[i]:
                 return self.vars[i]
-        return self.vars[len(self.values) - 1]
+        return self.create_lr_var(self.vars[len(self.values) - 1])
 
 
 class NaturalExpDecay(LearningRateDecay):

python/paddle/fluid/optimizer.py

-# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+# Copyright (c) 2019 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.
@@ -16,25 +16,25 @@ from __future__ import print_function
 
 import numpy as np
 from collections import defaultdict
-from functools import reduce
 
-from paddle.fluid import core
 from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table
 from paddle.fluid.framework import Program, Variable, name_scope, default_main_program, default_startup_program
-from paddle.fluid.layers import tensor
 
 from . import framework
 from . import layers
 from . import unique_name
 from .backward import append_backward
 from .clip import append_gradient_clip_ops, error_clip_callback
-from .dygraph import base as imperative_base
-from .dygraph.learning_rate_scheduler import LearningRateDecay
 from .framework import program_guard
 from .initializer import Constant
 from .layer_helper import LayerHelper
 from .layers import ops
 from .regularizer import append_regularization_ops
+from .dygraph import base as imperative_base
+from .dygraph.learning_rate_scheduler import LearningRateDecay
+from paddle.fluid import core
+from paddle.fluid.layers import tensor
+from functools import reduce
 from .wrapped_decorator import signature_safe_contextmanager
 
 __all__ = [
@@ -63,14 +63,18 @@ class Optimizer(object):
                 raise TypeError(
                     "learning rate should be float or LearningRateDecay, got %s here"
                     % type(learning_rate))
+            if name is not None:
+                self._name = unique_name.generate(name)
+            else:
+                self._name = unique_name.generate(self.__class__.__name__)
         else:
             if not isinstance(learning_rate, float) and \
                     not isinstance(learning_rate, framework.Variable):
                 raise TypeError(
                     "learning rate should be float or Variable, got %s here" %
                     type(learning_rate))
+            self._name = name
 
-        self._name = name
         self.regularization = regularization
         self._learning_rate = learning_rate
         # the learning rate type should be inferenced from loss
@@ -89,6 +93,90 @@ class Optimizer(object):
         self.helper = None
         self._opti_name_list = []
 
+    def load(self, stat_dict):
+        """
+        load optimizer with learning rate decay in dygraph mode
+        :return: None
+
+        Args:
+            stat_dict: the dict load by load_persistable method
+
+        Examples:
+
+        .. code-block:: python
+
+            from __future__ import print_function
+            import numpy as np
+            import paddle
+            import paddle.fluid as fluid
+            from paddle.fluid.optimizer import SGDOptimizer
+            from paddle.fluid.dygraph.nn import FC
+            from paddle.fluid.dygraph.base import to_variable
+
+            class MLP(fluid.Layer):
+                def __init__(self, name_scope):
+                    super(MLP, self).__init__(name_scope)
+
+                    self._fc1 = FC(self.full_name(), 10)
+                    self._fc2 = FC(self.full_name(), 10)
+
+                def forward(self, inputs):
+                    y = self._fc1(inputs)
+                    y = self._fc2(y)
+                    return y
+
+            with fluid.dygraph.guard():
+                mlp = MLP('mlp')
+                optimizer2 = SGDOptimizer(
+                    learning_rate=fluid.layers.natural_exp_decay(
+                    learning_rate=0.1,
+                    decay_steps=10000,
+                    decay_rate=0.5,
+                    staircase=True))
+
+                train_reader = paddle.batch(
+                        paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
+
+                for batch_id, data in enumerate(train_reader()):
+                    dy_x_data = np.array(
+                            [x[0].reshape(1, 28, 28) for x in data]).astype('float32')
+
+                    y_data = np.array([x[1] for x in data]).astype('int64').reshape(
+                            128, 1)
+
+                    img = to_variable(dy_x_data)
+                    label = to_variable(y_data)
+                    label._stop_gradient = True
+                    cost = mlp(img)
+                    avg_loss = fluid.layers.reduce_mean(cost)
+                    avg_loss.backward()
+                    optimizer.minimize(avg_loss)
+                    mlp.clear_gradients()
+                    fluid.dygraph.save_persistables(
+                            mlp.state_dict(), [optimizer, optimizer2], "save_dir_2")
+                    if batch_id == 2:
+                            break
+
+            with fluid.dygraph.guard():
+                mlp_load = MLP('mlp')
+                optimizer_load2 = SGDOptimizer(
+                        learning_rate=fluid.layers.natural_exp_decay(
+                        learning_rate=0.1,
+                        decay_steps=10000,
+                        decay_rate=0.5,
+                        staircase=True))
+                parameters, optimizers = fluid.dygraph.load_persistables(
+                    "save_dir_2")
+                mlp_load.load_dict(parameters)
+                optimizer_load2.load(optimizers)
+            self.assertTrue(optimizer2._learning_rate.__dict__ == optimizer_load2._learning_rate.__dict__)
+
+        """
+        if framework.in_dygraph_mode():
+            self._learning_rate = stat_dict[self._name]
+        else:
+            raise TypeError("load can only be used under DyGraph mode")
+
     def get_opti_var_name_list(self):
         return self._opti_name_list
 

python/paddle/fluid/tests/unittests/test_imperative_checkpoint.py

@@ -144,14 +144,14 @@ class TestDygraphCheckpoint(unittest.TestCase):
 
                     avg_loss.backward()
                     sgd.minimize(avg_loss)
-                    fluid.dygraph.save_persistables(mnist.state_dict(),
+                    fluid.dygraph.save_persistables(mnist.state_dict(), [sgd],
                                                     "save_dir")
                     mnist.clear_gradients()
 
                     for param in mnist.parameters():
                         dy_param_init_value[param.name] = param.numpy()
 
-                    restore = fluid.dygraph.load_persistables("save_dir")
+                    restore, _ = fluid.dygraph.load_persistables("save_dir")
                     mnist.load_dict(restore)
 
                     self.assertEqual(len(dy_param_init_value), len(restore))

python/paddle/fluid/tests/unittests/test_imperative_save_load_optimizer.py

+# Copyright (c) 2019 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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.optimizer import SGDOptimizer, Adam
+from paddle.fluid.dygraph.nn import FC
+from paddle.fluid.dygraph.base import to_variable
+
+
+class MLP(fluid.Layer):
+    def __init__(self, name_scope):
+        super(MLP, self).__init__(name_scope)
+
+        self._fc1 = FC(self.full_name(), 10)
+        self._fc2 = FC(self.full_name(), 10)
+
+    def forward(self, inputs):
+        y = self._fc1(inputs)
+        y = self._fc2(y)
+        return y
+
+
+class TestImperativeOptimizerBase(unittest.TestCase):
+    def setUp(self):
+        self.batch_num = 20
+
+    def get_optimizer(self):
+        raise NotImplementedError()
+
+    def _check_mlp(self):
+        seed = 90
+        with fluid.dygraph.guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+
+            mlp = MLP('mlp')
+            optimizer = self.get_optimizer()
+            optimizer2 = SGDOptimizer(
+                learning_rate=fluid.layers.natural_exp_decay(
+                    learning_rate=0.1,
+                    decay_steps=10000,
+                    decay_rate=0.5,
+                    staircase=True))
+            train_reader = paddle.batch(
+                paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
+
+            for batch_id, data in enumerate(train_reader()):
+                dy_x_data = np.array(
+                    [x[0].reshape(1, 28, 28) for x in data]).astype('float32')
+                y_data = np.array([x[1] for x in data]).astype('int64').reshape(
+                    128, 1)
+
+                img = to_variable(dy_x_data)
+                label = to_variable(y_data)
+                label._stop_gradient = True
+
+                cost = mlp(img)
+                avg_loss = fluid.layers.reduce_mean(cost)
+
+                avg_loss.backward()
+                optimizer.minimize(avg_loss)
+                optimizer2.minimize(avg_loss)
+                mlp.clear_gradients()
+                fluid.dygraph.save_persistables(
+                    mlp.state_dict(), [optimizer, optimizer2], "save_dir_2")
+                if batch_id == 2:
+                    break
+
+        with fluid.dygraph.guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+
+            mlp_load = MLP('mlp')
+            optimizer_load1 = self.get_optimizer()
+            optimizer_load2 = SGDOptimizer(
+                learning_rate=fluid.layers.natural_exp_decay(
+                    learning_rate=0.1,
+                    decay_steps=10000,
+                    decay_rate=0.5,
+                    staircase=True))
+            parameters, optimizers = fluid.dygraph.load_persistables(
+                "save_dir_2")
+            mlp_load.load_dict(parameters)
+            optimizer_load1.load(optimizers)
+            optimizer_load2.load(optimizers)
+
+        self.assertTrue(optimizer._learning_rate.__dict__ ==
+                        optimizer_load1._learning_rate.__dict__)
+        self.assertTrue(optimizer2._learning_rate.__dict__ ==
+                        optimizer_load2._learning_rate.__dict__)
+
+
+class TestImperativeOptimizerPiecewiseDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        bd = [3, 6, 9]
+        optimizer = SGDOptimizer(learning_rate=fluid.layers.piecewise_decay(
+            boundaries=bd, values=[0.1 * (0.1**i) for i in range(len(bd) + 1)]))
+        return optimizer
+
+    def test_sgd(self):
+        self._check_mlp()
+
+
+class TestImperativeOptimizerNaturalExpDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        optimizer = SGDOptimizer(learning_rate=fluid.layers.natural_exp_decay(
+            learning_rate=0.1,
+            decay_steps=10000,
+            decay_rate=0.5,
+            staircase=True))
+        return optimizer
+
+    def test_sgd(self):
+        self._check_mlp()
+
+
+class TestImperativeOptimizerExponentialDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        optimizer = SGDOptimizer(learning_rate=fluid.layers.exponential_decay(
+            learning_rate=0.1,
+            decay_steps=10000,
+            decay_rate=0.5,
+            staircase=True))
+        return optimizer
+
+    def test_sgd(self):
+        self._check_mlp()
+
+
+class TestImperativeOptimizerInverseTimeDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        optimizer = Adam(learning_rate=fluid.layers.inverse_time_decay(
+            learning_rate=0.1,
+            decay_steps=10000,
+            decay_rate=0.5,
+            staircase=True))
+        return optimizer
+
+    def test_adam(self):
+        self._check_mlp()
+
+
+class TestImperativeOptimizerPolynomialDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        optimizer = SGDOptimizer(learning_rate=fluid.layers.polynomial_decay(
+            learning_rate=0.1, decay_steps=5, cycle=self.cycle))
+        return optimizer
+
+    def test_sgd_cycle(self):
+        self.cycle = True
+        self._check_mlp()
+
+    def test_sgd(self):
+        self.cycle = False
+        self._check_mlp()
+
+
+class TestImperativeOptimizerCosineDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        optimizer = SGDOptimizer(learning_rate=fluid.layers.cosine_decay(
+            learning_rate=0.1, step_each_epoch=10000, epochs=120))
+        return optimizer
+
+    def test_sgd(self):
+        self._check_mlp()
+
+
+class TestImperativeOptimizerNoamDecay(TestImperativeOptimizerBase):
+    def get_optimizer(self):
+        optimizer = SGDOptimizer(learning_rate=fluid.layers.noam_decay(
+            d_model=512, warmup_steps=8000))
+        return optimizer
+
+    def test_sgd(self):
+        self._check_mlp()
+
+
+if __name__ == '__main__':
+    unittest.main()