merge master

.pre-commit-config.yaml

+-   repo: https://github.com/PaddlePaddle/mirrors-yapf.git
+    sha: 0d79c0c469bab64f7229c9aca2b1186ef47f0e37
+    hooks:
+    -   id: yapf
+        files: \.py$
+-   repo: https://github.com/pre-commit/pre-commit-hooks
+    sha: a11d9314b22d8f8c7556443875b731ef05965464
+    hooks:
+    -   id: check-merge-conflict
+    -   id: check-symlinks
+    -   id: detect-private-key
+        files: (?!.*paddle)^.*$
+    -   id: end-of-file-fixer
+        files: \.(md|yml)$
+    -   id: trailing-whitespace
+        files: \.(md|yml)$
+-   repo: https://github.com/Lucas-C/pre-commit-hooks
+    sha: v1.0.1
+    hooks:
+    -   id: forbid-crlf
+        files: \.(md|yml)$
+    -   id: remove-crlf
+        files: \.(md|yml)$
+    -   id: forbid-tabs
+        files: \.(md|yml)$
+    -   id: remove-tabs
+        files: \.(md|yml)$

mnist.py

@@ -26,7 +26,7 @@ from paddle import fluid
 from paddle.fluid.optimizer import Momentum
 from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear
 
-from model import Model, CrossEntropy
+from model import Model, CrossEntropy, Input
 from metrics import Accuracy
 
 
@@ -79,7 +79,6 @@ class SimpleImgConvPool(fluid.dygraph.Layer):
 class MNIST(Model):
     def __init__(self):
         super(MNIST, self).__init__()
-
         self._simple_img_conv_pool_1 = SimpleImgConvPool(
             1, 20, 5, 2, 2, act="relu")
 
@@ -89,12 +88,13 @@ class MNIST(Model):
         pool_2_shape = 50 * 4 * 4
         SIZE = 10
         scale = (2.0 / (pool_2_shape**2 * SIZE))**0.5
-        self._fc = Linear(800,
-                          10,
-                          param_attr=fluid.param_attr.ParamAttr(
-                              initializer=fluid.initializer.NormalInitializer(
-                                  loc=0.0, scale=scale)),
-                          act="softmax")
+        self._fc = Linear(
+            800,
+            10,
+            param_attr=fluid.param_attr.ParamAttr(
+                initializer=fluid.initializer.NormalInitializer(
+                    loc=0.0, scale=scale)),
+            act="softmax")
 
     def forward(self, inputs):
         x = self._simple_img_conv_pool_1(inputs)
@@ -138,13 +138,15 @@ def main():
         paddle.batch(paddle.dataset.mnist.test(),
                      batch_size=FLAGS.batch_size, drop_last=True), 1, 1)
 
-    device_ids = list(range(FLAGS.num_devices))
-
     with guard:
         model = MNIST()
-        optim = Momentum(learning_rate=FLAGS.lr, momentum=.9,
-                         parameter_list=model.parameters())
-        model.prepare(optim, CrossEntropy(), metrics=Accuracy(topk=(1, 2)))
+        optim = Momentum(
+            learning_rate=FLAGS.lr,
+            momentum=.9,
+            parameter_list=model.parameters())
+        inputs = [Input([None, 1, 28, 28], 'float32', name='image')]
+        labels = [Input([None, 1], 'int64', name='label')]
+        model.prepare(optim, CrossEntropy(), Accuracy(topk=(1, 2)), inputs, labels)
         if FLAGS.resume is not None:
             model.load(FLAGS.resume)
 
@@ -153,8 +155,7 @@ def main():
             val_loss = 0.0
             print("======== train epoch {} ========".format(e))
             for idx, batch in enumerate(train_loader()):
-                losses, metrics = model.train(batch[0], batch[1], device='gpu',
-                                              device_ids=device_ids)
+                losses, metrics = model.train(batch[0], batch[1])
 
                 train_loss += np.sum(losses)
                 if idx % 10 == 0:
@@ -167,8 +168,7 @@ def main():
 
             print("======== eval epoch {} ========".format(e))
             for idx, batch in enumerate(val_loader()):
-                losses, metrics = model.eval(batch[0], batch[1], device='gpu',
-                                             device_ids=device_ids)
+                losses, metrics = model.eval(batch[0], batch[1])
 
                 val_loss += np.sum(losses)
                 if idx % 10 == 0:
@@ -188,14 +188,21 @@ if __name__ == '__main__':
     parser.add_argument(
         "-e", "--epoch", default=100, type=int, help="number of epoch")
     parser.add_argument(
-        '--lr', '--learning-rate', default=1e-3, type=float, metavar='LR',
+        '--lr',
+        '--learning-rate',
+        default=1e-3,
+        type=float,
+        metavar='LR',
         help='initial learning rate')
     parser.add_argument(
         "-b", "--batch_size", default=128, type=int, help="batch size")
     parser.add_argument(
-        "-n", "--num_devices", default=4, type=int, help="number of devices")
+        "-n", "--num_devices", default=1, type=int, help="number of devices")
     parser.add_argument(
-        "-r", "--resume", default=None, type=str,
+        "-r",
+        "--resume",
+        default=None,
+        type=str,
         help="checkpoint path to resume")
     FLAGS = parser.parse_args()
     main()

model.py

@@ -28,10 +28,12 @@ from paddle.fluid.io import is_belong_to_optimizer
 from paddle.fluid.dygraph.base import to_variable
 from metrics import Metric
 
-__all__ = ['shape_hints', 'Model', 'Loss', 'CrossEntropy']
+__all__ = ['Model', 'Loss', 'CrossEntropy', 'Input']
 
 
 def to_list(value):
+    if value is None:
+        return value
     if isinstance(value, (list, tuple)):
         return value
     return [value]
@@ -72,20 +74,14 @@ def extract_args(func):
         return inspect.getargspec(func)[0]
 
 
-def shape_hints(**hints):
-    assert hints, "hints can not be empty"
-    assert all(isinstance(h, (list, tuple)) for h in hints.values()), \
-        "shape hint must be a list or tuple"
+class Input(fluid.dygraph.Layer):
+    def __init__(self, shape=None, dtype=None, name=None):
+        self.shape = shape
+        self.dtype = dtype
+        self.name = name
 
-    def wrapper(func):
-        args = extract_args(func)
-        invalid = set(hints.keys()) - set(args)
-        assert not invalid, \
-            "shape hint for arguments that are not present in forward method" \
-            + ": ({})".format(", ".join(invalid))
-        func.shape_hints = hints
-        return func
-    return wrapper
+    def forward(self):
+        return fluid.data(self.name, shape=self.shape, dtype=self.dtype)
 
 
 class Loss(object):
@@ -93,12 +89,6 @@ class Loss(object):
         super(Loss, self).__init__()
         self.average = average
 
-    def infer_shape(self, outputs):
-        return [o.shape for o in outputs]
-
-    def infer_dtype(self, outputs):
-        return [o.dtype for o in outputs]
-
     def forward(self, outputs, labels):
         raise NotImplementedError()
 
@@ -107,24 +97,21 @@ class Loss(object):
         if in_dygraph_mode():
             labels = [to_variable(l) for l in labels]
         losses = to_list(self.forward(to_list(outputs), labels))
-        if not self.average:
-            return losses
-        return [fluid.layers.reduce_mean(l) for l in losses]
+        if self.average:
+            losses = [fluid.layers.reduce_mean(l) for l in losses]
+        else:
+            losses = [fluid.layers.reduce_sum(l) for l in losses]
+        return losses
 
 
 class CrossEntropy(Loss):
-    def __init__(self):
+    def __init__(self, average=True):
         super(CrossEntropy, self).__init__()
 
-    def infer_shape(self, outputs):
-        return [o.shape[:-1] + (1, ) for o in outputs]
-
-    def infer_dtype(self, outputs):
-        return ['int64' for _ in outputs]
-
     def forward(self, outputs, labels):
-        return [fluid.layers.cross_entropy(o, l) for o, l in zip(
-            outputs, labels)]
+        return [
+            fluid.layers.cross_entropy(o, l) for o, l in zip(outputs, labels)
+        ]
 
 
 class StaticGraphAdapter(object):
@@ -137,21 +124,13 @@ class StaticGraphAdapter(object):
         self._orig_prog = fluid.default_main_program()
 
         self._label_vars = {}  # label variables
+        self._input_vars = {}  # label variables
         self._endpoints = {}
         self._loss_endpoint = None
         self._executor = None
         self._progs = {}
         self._compiled_progs = {}
 
-        self._lazy_load_optimizer = None
-
-        # parse shape hints
-        self._input_desc = OrderedDict([
-            (n, None) for n in extract_args(self.model.forward) if n != 'self'
-        ])
-        if hasattr(self.model.forward, 'shape_hints'):
-            self._input_desc.update(self.model.forward.shape_hints)
-
     @property
     def mode(self):
         return self.model.mode
@@ -160,21 +139,19 @@ class StaticGraphAdapter(object):
     def mode(self, value):
         self.model.mode = value
 
-    def train(self, inputs, labels, device='CPU', device_ids=None):
-        assert self.model._optimizer and self.model._loss_function, \
+    def train(self, inputs, labels=None):
+        assert self.model._optimizer, \
             "model not ready, please call `model.prepare()` first"
         self.mode = 'train'
-        return self._run(inputs, labels, device, device_ids)
+        return self._run(inputs, labels)
 
-    def eval(self, inputs, labels, device='CPU', device_ids=None):
-        assert self.model._loss_function, \
-            "model not ready, please call `model.prepare()` first"
+    def eval(self, inputs, labels=None):
         self.mode = 'eval'
-        return self._run(inputs, labels, device, device_ids)
+        return self._run(inputs, labels)
 
-    def test(self, inputs, device='CPU', device_ids=None):
+    def test(self, inputs):
         self.mode = 'test'
-        return self._run(inputs, None, device, device_ids)
+        return self._run(inputs, None)
 
     def parameters(self, *args, **kwargs):
         return None
@@ -183,13 +160,18 @@ class StaticGraphAdapter(object):
         def _save(state, path):
             if not state:
                 return
-            state = {k: to_numpy(v) if isinstance(v, Variable) else v
-                     for k, v in state.items()}
+            state = {
+                k: to_numpy(v) if isinstance(v, Variable) else v
+                for k, v in state.items()
+            }
             with open(path, 'wb') as f:
                 pickle.dump(state, f)
 
         base = os.path.basename(path)
         assert base != "", "path should be of 'dirname/filename' format"
+        dir_name = os.path.dirname(path)
+        if dir_name and not os.path.exists(dir_name):
+            os.makedirs(dir_name)
         param_path = path + ".pdparams"
         _save(self.model.state_dict(), param_path)
         prog = self._progs.get('train', None)
@@ -197,13 +179,13 @@ class StaticGraphAdapter(object):
             return
         # XXX `optimizer.state_dict()` only work in dygraph mode
         optim_path = path + ".pdopt"
-        optim = {p.name: p for p in filter(
-            is_belong_to_optimizer, prog.list_vars())}
+        optim = {
+            p.name: p
+            for p in filter(is_belong_to_optimizer, prog.list_vars())
+        }
         if not optim:
             return
-        # HACK this is contrived, optimizer state is not the same for
-        # static/dynamic graph mode
-        optim['__static_graph_only__'] = True
+
         _save(optim, optim_path)
 
     def load(self, path):
@@ -238,27 +220,77 @@ class StaticGraphAdapter(object):
         optim_state = _load(optim_path)
         if optim_state is None:
             return
-        assert '__static_graph_only__' in optim_state, \
-            "optimizer saved in dygraph mode is not usable in static graph"
 
-        if self._executor is not None:
-           self._load_optimizer(optim_state)
-        else:
-           self._lazy_load_optimizer = optim_state
+        self._load_optimizer(optim_state, executor)
 
-    def _load_optimizer(self, state):
+    def _load_optimizer(self, state, executor):
         prog = self._progs.get('train', None)
         optim = list(filter(is_belong_to_optimizer, prog.list_vars()))
         if not optim:
             return
 
-        fluid.core._create_loaded_parameter(
-            optim, global_scope(), self._executor._default_executor)
+        fluid.core._create_loaded_parameter(optim, global_scope(), executor)
 
+        converted_state = dict(state)
         for var in optim:
-            assert var.name in state, \
+            if var.name in ["@LR_DECAY_COUNTER@", "global_step"]:
+                # When using learning rate scheduler, dygraph would name the
+                # global step var as "global_step" to save, while static-graph
+                # would has a state var named as "@LR_DECAY_COUNTER@".
+                # NOTE: dygraph saved global_step is 1 larger than that in
+                # static-graph, since the time of global_step to increase is
+                # different.
+                state_val = (
+                    np.array(converted_state.pop("global_step")) - 1
+                ) if "global_step" in converted_state else converted_state.pop(
+                    "@LR_DECAY_COUNTER@", None)
+                if state_val is not None:
+                    converted_state[var.name] = state_val
+            elif var.name.startswith("learning_rate_"):
+                # When using static learning rate, static-graph would make it
+                # a persistable var named 'unique_name.generate("learning_rate")',
+                # However, dygraph wouldn't save it.
+                if var.name not in state: continue
+            else:
+                # moment and other accumulators
+                if var.name not in converted_state:
+                    # try to convert from dygraph name
+                    opt_name = self.model._optimizer._name
+                    opt_cls_name = self.model._optimizer.__class__.__name__
+                    opt_unq_name = None
+                    for name in self.model._optimizer._accumulators.keys():
+                        accum_name = name if opt_name is None else name[len(
+                            opt_name) + 1:]
+                        for param_name, state_var in self.model._optimizer._accumulators[
+                                name].items():
+                            if opt_unq_name is None:
+                                # can not infer out the exact unique(opt_name),
+                                # thus try to extract rather than generate
+                                for state_key in sorted(
+                                        state.keys(),
+                                        key=lambda x: len(x),
+                                        reverse=True):
+                                    prefix = param_name + "_" + (
+                                        opt_cls_name if opt_name is None else
+                                        opt_name) + "_"
+                                    if state_key.startswith(prefix):
+                                        prefix_offset = state_key[len(
+                                            prefix):].find("_") + len(prefix)
+                                        opt_unq_name = state_key[len(
+                                            param_name + "_"):prefix_offset]
+                                        # TODO: assert
+                                        # assert opt_unq_name is None
+                                    # gen(param.name + "_" + gen(opt_name) + "_" + accum_name)
+                                    # always end with "_0" since the unique optimizer._name
+                            dy_state_name = (param_name + "_" + opt_unq_name +
+                                             "_" + accum_name + "_0")
+                            converted_state[
+                                state_var.name] = converted_state.pop(
+                                    dy_state_name)
+
+            assert var.name in converted_state, \
                 "variable [{}] is not in optimizer state file".format(var.name)
-            self._set_var(var, state[var.name])
+            self._set_var(var, converted_state[var.name])
 
     def _set_var(self, var, ndarray):
         t = global_scope().find_var(var.name).get_tensor()
@@ -274,21 +306,20 @@ class StaticGraphAdapter(object):
 
         t.set(ndarray, place)
 
-    def _run(self, inputs, labels=None, device='CPU', device_ids=None):
+    def _run(self, inputs, labels=None):
+        compiled_prog = self._compiled_progs.get(self.mode, None)
+        assert compiled_prog, \
+            "Model is not ready, please call `model.prepare()` first"
+
         inputs = to_list(inputs)
         if labels is not None:
             labels = to_list(labels)
-        assert len(inputs) == len(self._input_desc), "number of inputs" \
+        assert len(inputs) == len(self._input_vars[self.mode]), \
+            "number of inputs" \
             + " does not match number of arguments of `forward` method"
 
-        if self._progs.get(self.mode, None) is None:
-            self._make_program(self._infer_input_vars(inputs))
-
-        compiled_prog = self._compile_and_initialize(
-            self._progs[self.mode], device, device_ids)
-
         feed = {}
-        input_names = [name for name in self._input_desc.keys()]
+        input_names = [v.name for v in self._input_vars[self.mode]]
         for idx, n in enumerate(input_names):
             # train and test may take different arguments
             if inputs[idx] is not None:
@@ -319,64 +350,71 @@ class StaticGraphAdapter(object):
             metrics.append(metric.update(*state))
         return (losses, metrics) if len(metrics) > 0 else losses
 
-    def _make_program(self, inputs):
+    def prepare(self):
+        modes = ['train', 'eval', 'test']
+        for mode in modes:
+            self._make_program(mode)
+            self._compile_and_initialize(self._progs[mode], mode)
+
+    def _make_program(self, mode):
+        prog = self._progs.get(mode, None)
+        if prog is not None:
+            return
+
         prog = self._orig_prog.clone()
-        if self.mode == 'train' and self.model._optimizer._learning_rate_map:
+        # NOTE: When defining learning rate scheduling in static-graph, ops to
+        # increase the global step var and calculate learning rate would be
+        # prepended into _orig_prog. test program maked by `_orig_prog.clone`
+        # also would include these ops. Thus must prune these ops in test
+        # program, otherwise the global step would be changed in test.
+        if mode != 'train':
+            for op in list(prog.global_block().ops):
+                prog.global_block()._remove_op(0)
+        if mode == 'train' and self.model._optimizer \
+            and self.model._optimizer._learning_rate_map:
             # HACK workaround learning rate map issue
             lr_var = self.model._optimizer._learning_rate_map[self._orig_prog]
             self.model._optimizer._learning_rate_map[prog] = lr_var
         losses = []
+        metrics = []
         with fluid.program_guard(prog, self._startup_prog):
+            if isinstance(self.model._inputs, dict):
+                ins = [self.model._inputs[n] \
+                    for n in extract_args(self.model.forward) if n != 'self']
+            else:
+                ins = self.model._inputs
+            lbls = self.model._labels if self.model._labels else []
+            inputs = [k.forward() for k in to_list(ins)]
+            labels = [k.forward() for k in to_list(lbls)]
             outputs = to_list(self.model.forward(*inputs))
-            if self.mode != 'test':
-                label_vars = self._infer_label_vars(outputs)
-                self._label_vars[self.mode] = label_vars
-                losses = self.model._loss_function(outputs, label_vars)
-                metrics = []
-                for metric in self.model._metrics:
-                    metrics.append(to_list(metric.add_metric_op(outputs, label_vars)))
-                if self.mode == 'train':
+            if mode != 'test':
+                if self.model._loss_function:
+                    losses = self.model._loss_function(outputs, labels)
+                    for metric in self.model._metrics:
+                        metrics.append(to_list(metric.add_metric_op(outputs, labels)))
+                if mode == 'train' and self.model._optimizer:
                     self._loss_endpoint = fluid.layers.sum(losses)
                     self.model._optimizer.minimize(self._loss_endpoint)
-        if self.mode != 'train':  # clone again to put it in test mode
+        if mode != 'train':  # clone again to put it in test mode
             prog = prog.clone(for_test=True)
-        self._progs[self.mode] = prog
-        self._endpoints[self.mode] = {
-            "output": outputs,
-            "loss": losses,
-            "metric": metrics,
-        }
 
-    def _infer_input_vars(self, inputs):
-        input_vars = []
-        for idx, i in enumerate(inputs):
-            if i is None:  # train and test may take different arguments
-                input_vars.append(None)
-                continue
-            ndarray = np.array(i)
-            name = list(self._input_desc.keys())[idx]
-            shape = list(self._input_desc.values())[idx]
-            if shape is None:
-                shape = (None, ) + ndarray.shape[1:]
-            input_vars.append(fluid.data(name, shape, ndarray.dtype))
-        return input_vars
-
-    def _infer_label_vars(self, outputs):
-        shapes = self.model._loss_function.infer_shape(outputs)
-        dtypes = self.model._loss_function.infer_dtype(outputs)
-        label_vars = []
-        for idx, (shape, dtype) in enumerate(zip(shapes, dtypes)):
-            name = '__label{}'.format(idx)
-            label_vars.append(fluid.data(name, shape, dtype))
-        return label_vars
-
-    def _compile_and_initialize(self, prog, device='CPU', device_ids=None):
-        compiled_prog = self._compiled_progs.get(self.mode, None)
+        self._input_vars[mode] = inputs
+        self._label_vars[mode] = labels
+        self._progs[mode] = prog
+        self._endpoints[mode] = {"output": outputs, "loss": losses, "metric": metrics}
+
+    def _compile_and_initialize(self, prog, mode):
+        compiled_prog = self._compiled_progs.get(mode, None)
         if compiled_prog is not None:
             return compiled_prog
 
-        places = [device.lower() == 'gpu' and fluid.CUDAPlace(i)
-                  or fluid.CPUPlace() for i in device_ids]
+        device = self.model._device
+        device_ids = self.model._device_ids
+
+        if device.lower() == 'gpu':
+            places = fluid.cuda_places(device_ids)
+        else:
+            places = fluid.cpu_places(len(device_ids) if device_ids else None)
 
         # XXX *ALL WEIGHTS* should be initialized upon model construction
         # even if `forward()` may run different code path for different mode
@@ -394,31 +432,14 @@ class StaticGraphAdapter(object):
                 startup_prog = self._startup_prog._prune(uninitialized)
                 self._executor.run(startup_prog)
 
-            if self.mode == 'train' and self._lazy_load_optimizer:
-                self._load_optimizer(self._lazy_load_optimizer)
-                self._lazy_load_optimizer = None
-
         compiled_prog = fluid.CompiledProgram(prog)
-        if len(device_ids) > 1:
+        if len(places) > 1:
             loss_name = None
-            if self.mode == 'train' and self._loss_endpoint is not None:
+            if mode == 'train' and self._loss_endpoint is not None:
                 loss_name = self._loss_endpoint.name
-
-            share_vars_from = None
-            if self.mode == 'eval' and 'train' in self._compiled_progs:
-                share_vars_from = self._compiled_progs['train']
-            # HACK invalidate eval program if is compiled before train program
-            # quite hackish, OTOH, it is generally uncommon that the eval
-            # program will be run before the train program
-            if self.mode == 'train' and 'eval' in self._compiled_progs:
-                del self._compiled_progs['eval']
-
             compiled_prog = compiled_prog.with_data_parallel(
-                loss_name=loss_name, places=places,
-                share_vars_from=share_vars_from)
-
-        self._compiled_progs[self.mode] = compiled_prog
-        return compiled_prog
+                loss_name=loss_name, places=places)
+        self._compiled_progs[mode] = compiled_prog
 
 
 class DynamicGraphAdapter(object):
@@ -435,13 +456,14 @@ class DynamicGraphAdapter(object):
         self.model.mode = value
 
     # TODO multi device in dygraph mode not implemented at present time
-    def train(self, inputs, labels, device='CPU', device_ids=None):
-        assert self.model._optimizer and self.model._loss_function, \
+    def train(self, inputs, labels=None):
+        assert self.model._optimizer, \
             "model not ready, please call `model.prepare()` first"
         super(Model, self.model).train()
         self.mode = 'train'
         inputs = to_list(inputs)
-        labels = to_list(labels)
+        if labels is not None:
+            labels = to_list(labels)
         outputs = to_list(self.model.forward(*[to_variable(x) for x in inputs]))
         losses = self.model._loss_function(outputs, labels)
         final_loss = fluid.layers.sum(losses)
@@ -456,24 +478,31 @@ class DynamicGraphAdapter(object):
         return ([to_numpy(l) for l in losses], metrics) \
                 if len(metrics) > 0 else [to_numpy(l) for l in losses]
 
-    def eval(self, inputs, labels, device='CPU', device_ids=None):
-        assert self.model._loss_function, \
-            "model not ready, please call `model.prepare()` first"
+    def eval(self, inputs, labels=None):
         super(Model, self.model).eval()
         self.mode = 'eval'
         inputs = to_list(inputs)
-        labels = to_list(labels)
+        if labels is not None:
+            labels = to_list(labels)
         outputs = to_list(self.model.forward(*[to_variable(x) for x in inputs]))
-        losses = self.model._loss_function(outputs, labels)
+
+        if self.model._loss_function:
+            losses = self.model._loss_function(outputs, labels)
+        else:
+            losses = []
+
         metrics = []
         for metric in self.model._metrics:
             metric_outs = metric.add_metric_op(outputs, [to_variable(l) for l in labels])
             m = metric.update(*[to_numpy(m) for m in to_list(metric_outs)])
             metrics.append(m)
+
+        # To be consistent with static graph
+        # return empty loss if loss_function is None
         return ([to_numpy(l) for l in losses], metrics) \
                 if len(metrics) > 0 else [to_numpy(l) for l in losses]
 
-    def test(self, inputs, device='CPU', device_ids=None):
+    def test(self, inputs):
         super(Model, self.model).eval()
         self.mode = 'test'
         inputs = [to_variable(x) for x in to_list(inputs)]
@@ -497,15 +526,68 @@ class DynamicGraphAdapter(object):
         self.model.set_dict(params)
         if self.model._optimizer is None or optim is None:
             return
-        self.model._optimizer.set_dict(optim)
+
+        # If optimizer performs set_dict when state vars haven't been created,
+        # which would happen when set_dict before minimize, the state would be
+        # stored in optimizer._accumulators_holder and loaded lazily.
+        # To contrive this when loading from static-graph saved states, extend
+        # state dict to include keys named accoring to dygraph naming rules.
+        # TODO: if len(self.model._optimizer._accumulators) > 0
+        converted_state = dict(optim)
+        opt_unq_name = self.model._optimizer._name
+        opt_cls_name = self.model._optimizer.__class__.__name__
+        opt_name = opt_unq_name[:opt_unq_name.rfind("_")]  # remove suffix idx
+        param_names = [param.name for param in self.model.parameters()]
+        for var_name, state_var in sorted(
+                optim.items(), key=lambda x: len(x[0]), reverse=True):
+            if var_name in ["@LR_DECAY_COUNTER@", "global_step"]:
+                # NOTE: dygraph saved global_step is 1 larger than that in
+                # static-graph, since the time of global_step to increase is
+                # different.
+                if var_name == "@LR_DECAY_COUNTER@":
+                    converted_state["global_step"] = np.array(
+                        converted_state.pop("@LR_DECAY_COUNTER@")) + 1
+            else:
+                # moment and other accumulators
+                # extend state dict to include promising dygraph names
+                for param_name in param_names:
+                    if var_name.startswith(param_name + "_" + opt_name):
+                        # when init optimizer with name
+                        accum_name = var_name[len(param_name + "_" + opt_name +
+                                                  "_"):]
+                    elif var_name.startswith(param_name +
+                                             "_") and opt_name == opt_cls_name:
+                        # when init optimizer without name
+                        accum_name = var_name[len(param_name + "_"):]
+                    else:
+                        continue
+                    # remove suffix idx
+                    accum_name = accum_name[:accum_name.rfind("_")]
+                    # state names always end with "_0" in dygraph because of the
+                    # unique optimizer._name
+                    dy_state_name = (param_name + "_" + opt_unq_name + "_" +
+                                     accum_name + "_0")
+                    converted_state[dy_state_name] = state_var
+
+        self.model._optimizer.set_dict(converted_state)
 
 
 class Model(fluid.dygraph.Layer):
+    """
+    FIXME: add more comments and usage
+    """
+
     def __init__(self):
         super(Model, self).__init__(self.__class__.__name__)
         self.mode = 'train'
+        self._inputs = None
+        self._labels = None
         self._loss_function = None
         self._loss_weights = None
+        self._loss = None
+        self._optimizer = None
+        self._device = None
+        self._device_ids = None
         self._optimizer = None
         if in_dygraph_mode():
             self._adapter = DynamicGraphAdapter(self)
@@ -527,15 +609,75 @@ class Model(fluid.dygraph.Layer):
     def load(self, *args, **kwargs):
         return self._adapter.load(*args, **kwargs)
 
-    def prepare(self, optimizer, loss_function, metrics=[]):
+    def prepare(self,
+                optimizer=None,
+                loss_function=None,
+                metrics=None,
+                inputs=None,
+                labels=None,
+                device=None,
+                device_ids=None):
+        """
+        FIXME: add comments
+        Args:
+            optimizer (Optimizer|None): optimizer must be set in training
+                and should be a Optimizer instance. It can be None in eval
+                and test mode.
+            loss_function (Loss|None): loss function must be set in training
+                and should be a Loss instance. It can be None when there is
+                no loss.
+            metrics (Metric|list of Metric|None): if metrics is set, all
+                metric will be calculate and output in train/eval mode.
+            inputs (Input|list|dict|None): inputs, entry points of network,
+                could be a Input layer, or lits of Input layers,
+                or dict (name: Input), or None. For static graph,
+                inputs must be set. For dynamic graph, it could be None.
+            labels (Input|list|None): labels, entry points of network,
+                could be a Input layer or lits of Input layers, or None.
+                For static graph, if set loss_function in Model.prepare(), it
+                must be set. Otherwise, it could be None.
+            device (str|None): specify device type, 'CPU' or 'GPU'.
+                If None, automatically select device according to
+                installation package version.
+            device_ids (list[int]|None): specify device index. If None,
+                the available device will be obtained from the environment
+                variable when the model is executed: If the GPU is used, the
+                currently available device ID is obtained from the environment
+                variable FLAGS_selected_gpus or CUDA_VISIBLE_DEVICES when the
+                model is executed; CPU, when the model is executed,
+                the currently available CPU number is obtained from the
+                environment variable CPU_NUM. For example, export CPU_NUM=4,
+                if the environment variable is not set, the executor will add
+                the variable to the environment variable and set its value to 1.
+                The default is None.
+        """
         self._optimizer = optimizer
-        assert isinstance(loss_function, Loss), \
-            "'loss_function' must be sub classes of 'Loss'"
+        if loss_function:
+            if not isinstance(loss_function, Loss):
+                raise TypeError(
+                    "'loss_function' must be sub classes of 'Loss'")
         self._loss_function = loss_function
+        if not in_dygraph_mode():
+            if not isinstance(inputs, (list, dict, Input)):
+                raise TypeError(
+                    "'inputs' must be list or dict in static graph mode")
+            if loss_function and not isinstance(labels, (list, Input)):
+                raise TypeError("'labels' must be list in static graph mode")
+
+        metrics = metrics or []
         for metric in to_list(metrics):
             assert isinstance(metric, Metric), \
                 "{} is not sub class of Metric".format(metric.__class__.__name__)
         self._metrics = to_list(metrics)
 
+        self._inputs = inputs
+        self._labels = labels
+        self._device = device
+        if device is None:
+            self._device = 'GPU' if fluid.is_compiled_with_cuda() else 'CPU'
+        self._device_ids = device_ids
+        if not in_dygraph_mode():
+            self._adapter.prepare()
+
     def parameters(self, *args, **kwargs):
         return self._adapter.parameters(*args, **kwargs)

yolov3.py

@@ -33,7 +33,7 @@ from paddle.fluid.dygraph.nn import Conv2D
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.regularizer import L2Decay
 
-from model import Model, Loss, shape_hints
+from model import Model, Loss, Input
 from resnet import ResNet, ConvBNLayer
 
 import logging
@@ -152,7 +152,6 @@ class YOLOv3(Model):
                                 act='leaky_relu'))
                 self.route_blocks.append(route)
 
-    @shape_hints(inputs=[None, 3, None, None], img_info=[None, 3])
     def forward(self, inputs, img_info):
         outputs = []
         boxes = []
@@ -208,10 +207,9 @@ class YOLOv3(Model):
 
 
 class YoloLoss(Loss):
-    def __init__(self, num_classes=80, num_max_boxes=50):
+    def __init__(self, num_classes=80):
         super(YoloLoss, self).__init__()
         self.num_classes = num_classes
-        self.num_max_boxes = num_max_boxes
         self.ignore_thresh = 0.7
         self.anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45,
                         59, 119, 116, 90, 156, 198, 373, 326]
@@ -240,16 +238,6 @@ class YoloLoss(Loss):
             downsample //= 2
         return losses
 
-    def infer_shape(self, _):
-        return [
-            [None, self.num_max_boxes, 4],
-            [None, self.num_max_boxes],
-            [None, self.num_max_boxes]
-        ]
-
-    def infer_dtype(self, _):
-        return ['float32', 'int32', 'float32']
-
 
 def make_optimizer(parameter_list=None):
     base_lr = FLAGS.lr
@@ -470,8 +458,7 @@ def run(model, loader, mode='train'):
     start = time.time()
 
     for idx, batch in enumerate(loader()):
-        losses, _ = getattr(model, mode)(
-            batch[0], batch[1], device='gpu', device_ids=device_ids)
+        losses = getattr(model, mode)(batch[0], batch[1])
 
         total_loss += np.sum(losses)
         if idx > 1:  # skip first two steps
@@ -521,7 +508,8 @@ def main():
         os.mkdir('yolo_checkpoints')
 
     with guard:
-        NUM_CLASSES=7
+        NUM_CLASSES = 7
+        NUM_MAX_BOXES = 50
         model = YOLOv3(num_classes=NUM_CLASSES)
         # XXX transfer learning
         if FLAGS.pretrain_weights is not None:
@@ -530,12 +518,18 @@ def main():
             model.load(FLAGS.weights)
         optim = make_optimizer(parameter_list=model.parameters())
         anno_path = os.path.join(FLAGS.data, 'annotations', 'instances_val2017.json')
+        inputs = [Input([None, 3, None, None], 'float32', name='image'),
+                  Input([None, 3], 'int32', name='img_info')]
+        labels = [Input([None, NUM_MAX_BOXES, 4], 'float32', name='gt_bbox'),
+                  Input([None, NUM_MAX_BOXES], 'int32', name='gt_label'),
+                  Input([None, NUM_MAX_BOXES], 'float32', name='gt_score')]
         model.prepare(optim,
                       YoloLoss(num_classes=NUM_CLASSES),
                       # For YOLOv3, output variable in train/eval is different,
                       # which is not supported by metric, add by callback later?
                       # metrics=COCOMetric(anno_path, with_background=False)
-                      )
+                      inputs=inputs,
+                      labels = labels)
 
         for e in range(epoch):
             logger.info("======== train epoch {} ========".format(e))