Improve symbol bindings (apache#5870)

* Initial checkin

Add init functions for simple bind in graph_executor

Add simple_bind c_api

Add simple bind c-api

Assign zeros to in_args, arg_grads, and aux_states

Add simple_bind2 python interface

Fix python interface bugs

Interface changes

Fix

Fix core dump

Add bind_ith_exec c_api

Change simple_bind2

Fix seg fault

Finish simple_bind

Change _bind_ith_exec

Refactor simple_bind initialization flow for bind

Consolidate bind and simple_bind graph init flow

Fix bug

Clean up

Add comments

Clean up

Clean up

Minor correction

Rename APIs in graph executor

Refactor

Rebase

Delete deprecated functions

Move more front-end work to backend

Bug fix

Fix failed tests

Minor fix

Fix lint

Fix lint

Revert unnecessary changes

Revert

Revert

Clean up

Fix lint

Fix bind_ith_exec calling simple_bind

Fix bugs for _bind_ith_exec

* Add unit test (apache#1)

* Add unit test

* Fix

* Small fix

* Fix lint

* Fix lint

* Fix bugs of missing ndarrays in shared_buffer

* Fix lint

* Simple bind (apache#3)

* Add bucketing test

* Skip pylint

* Use cpu to train

* Fix bug

* Remove merge message

* Fix lint

* Add logging to test_bucketing.py

* Reduce model size (apache#4)

* Add checks for shape/type inferences

* Add printing error messages for shape/type inference failure

include/mxnet/c_api.h

@@ -1149,6 +1149,38 @@ MXNET_DLL int MXExecutorBindEX(SymbolHandle symbol_handle,
                                NDArrayHandle *aux_states,
                                ExecutorHandle shared_exec,
                                ExecutorHandle *out);
+
+MXNET_DLL int MXExecutorSimpleBind(SymbolHandle symbol_handle,
+                         int dev_type,
+                         int dev_id,
+                         const mx_uint num_g2c_keys,
+                         const char** g2c_keys,
+                         const int* g2c_dev_types,
+                         const int* g2c_dev_ids,
+                         const mx_uint provided_grad_req_list_len,
+                         const char** provided_grad_req_names,
+                         const char** provided_grad_req_types,
+                         const mx_uint num_provided_arg_shapes,
+                         const char** provided_arg_shape_names,
+                         const mx_uint* provided_arg_shape_data,
+                         const mx_uint* provided_arg_shape_idx,
+                         const mx_uint num_provided_arg_dtypes,
+                         const char** provided_arg_dtype_names,
+                         const int* provided_arg_dtypes,
+                         const mx_uint num_shared_arg_names,
+                         const char** shared_arg_name_list,
+                         int* shared_buffer_len,
+                         const char** shared_buffer_name_list,
+                         NDArrayHandle* shared_buffer_handle_list,
+                         const char*** updated_shared_buffer_name_list,
+                         NDArrayHandle** updated_shared_buffer_handle_list,
+                         mx_uint* num_in_args,
+                         NDArrayHandle** in_args,
+                         NDArrayHandle** arg_grads,
+                         mx_uint* num_aux_states,
+                         NDArrayHandle** aux_states,
+                         ExecutorHandle shared_exec_handle,
+                         ExecutorHandle* out);
 /*!
  * \brief set a call back to notify the completion of operation
  */

include/mxnet/executor.h

@@ -69,6 +69,21 @@ class Executor {
    * \return array of outputs in the executor.
    */
   virtual const std::vector<NDArray> &outputs() const = 0;
+  /*!
+   * \brief get input argument map, key is arg name, value is arg's NDArray.
+   * \return input argument map in the executor.
+   */
+  virtual const std::unordered_map<std::string, NDArray>& in_arg_map() const = 0;
+  /*!
+   * \brief get input argument graident map, key is arg name, value is gradient's NDArray.
+   * \return input argument gradient map in the executor.
+   */
+  virtual const std::unordered_map<std::string, NDArray>& arg_grad_map() const = 0;
+  /*!
+   * \brief get aux state map, key is arg name, value is aux state's NDArray.
+   * \return aux state map in the executor.
+   */
+  virtual const std::unordered_map<std::string, NDArray>& aux_state_map() const = 0;
   /*!
    * \brief Create an operator by bind symbol with context and arguments.
    *  If user do not want to compute the gradients of i-th argument, grad_req_type[i] can be kNullOp.
@@ -91,6 +106,23 @@ class Executor {
                         const std::vector<OpReqType> &grad_req_type,
                         const std::vector<NDArray> &aux_states,
                         Executor* shared_exec = NULL);
+
+  static Executor* SimpleBind(nnvm::Symbol symbol,
+                              const Context& default_ctx,
+                              const std::map<std::string, Context>& group2ctx,
+                              const std::vector<Context>& in_arg_ctxes,
+                              const std::vector<Context>& arg_grad_ctxes,
+                              const std::vector<Context>& aux_state_ctxes,
+                              const std::unordered_map<std::string, TShape>& arg_shape_map,
+                              const std::unordered_map<std::string, int>& arg_dtype_map,
+                              const std::vector<OpReqType>& grad_req_types,
+                              const std::unordered_set<std::string>& param_names,
+                              std::vector<NDArray>* in_args,
+                              std::vector<NDArray>* arg_grads,
+                              std::vector<NDArray>* aux_states,
+                              std::unordered_map<std::string, NDArray>*
+                                shared_data_arrays = nullptr,
+                              Executor* shared_exec = nullptr);
   /*!
    * \brief the prototype of user-defined monitor callback
    */

python/mxnet/module/executor_group.py

@@ -5,8 +5,6 @@
 import logging
 from collections import OrderedDict
 
-import numpy as np
-
 from .. import context as ctx
 from .. import ndarray as nd
 from ..io import DataDesc
@@ -564,6 +562,7 @@ def update_metric(self, eval_metric, labels):
 
     def _bind_ith_exec(self, i, data_shapes, label_shapes, shared_group):
         """Internal utility function to bind the i-th executor.
+        This function utilizes simple_bind python interface.
         """
         shared_exec = None if shared_group is None else shared_group.execs[i]
         context = self.contexts[i]
@@ -573,85 +572,14 @@ def _bind_ith_exec(self, i, data_shapes, label_shapes, shared_group):
         if label_shapes is not None:
             input_shapes.update(dict(label_shapes))
 
-        arg_shapes, _, aux_shapes = self.symbol.infer_shape(**input_shapes)
-        assert arg_shapes is not None, "shape inference failed"
-
         input_types = {x.name: x.dtype for x in data_shapes}
         if label_shapes is not None:
             input_types.update({x.name: x.dtype for x in label_shapes})
-        arg_types, _, aux_types = self.symbol.infer_type(**input_types)
-        assert arg_types is not None, "type inference failed"
-
-        arg_arrays = []
-        grad_arrays = {} if self.for_training else None
-
-        def _get_or_reshape(name, shared_data_arrays, arg_shape, arg_type, context, logger):
-            """Internal helper to get a memory block or re-use by re-shaping."""
-            if name in shared_data_arrays:
-                arg_arr = shared_data_arrays[name]
 
-                if np.prod(arg_arr.shape) >= np.prod(arg_shape):
-                    # nice, we can directly re-use this data blob
-                    assert arg_arr.dtype == arg_type
-                    arg_arr = arg_arr.reshape(arg_shape)
-                else:
-                    logger.warning(('bucketing: data "%s" has a shape %s' % (name, arg_shape)) +
-                                   (', which is larger than already allocated ') +
-                                   ('shape %s' % (arg_arr.shape,)) +
-                                   ('. Need to re-allocate. Consider putting ') +
-                                   ('default_bucket_key to') +
-                                   (' be the bucket taking the largest input for better ') +
-                                   ('memory sharing.'))
-                    arg_arr = nd.zeros(arg_shape, context, dtype=arg_type)
-
-                    # replace existing shared array because the new one is bigger
-                    shared_data_arrays[name] = arg_arr
-            else:
-                arg_arr = nd.zeros(arg_shape, context, dtype=arg_type)
-                shared_data_arrays[name] = arg_arr
-
-            return arg_arr
-
-        # create or borrow arguments and gradients
-        for j in range(len(self.arg_names)):
-            name = self.arg_names[j]
-            if name in self.param_names: # model parameters
-                if shared_exec is None:
-                    arg_arr = nd.zeros(arg_shapes[j], context, dtype=arg_types[j])
-                    if self.grad_req[name] != 'null':
-                        grad_arr = nd.zeros(arg_shapes[j], context, dtype=arg_types[j])
-                        grad_arrays[name] = grad_arr
-                else:
-                    arg_arr = shared_exec.arg_dict[name]
-                    assert arg_arr.shape == arg_shapes[j]
-                    assert arg_arr.dtype == arg_types[j]
-                    if self.grad_req[name] != 'null':
-                        grad_arrays[name] = shared_exec.grad_dict[name]
-            else: # data, label, or states
-                arg_arr = _get_or_reshape(name, shared_data_arrays, arg_shapes[j], arg_types[j],
-                                          context, self.logger)
-
-                # data might also need grad if inputs_need_grad is True
-                if self.grad_req[name] != 'null':
-                    grad_arrays[name] = _get_or_reshape('grad of ' + name, shared_data_arrays,
-                                                        arg_shapes[j], arg_types[j], context,
-                                                        self.logger)
-
-            arg_arrays.append(arg_arr)
-
-        # create or borrow aux variables
-        if shared_exec is None:
-            aux_arrays = [nd.zeros(s, context, dtype=t) for s, t in zip(aux_shapes, aux_types)]
-        else:
-            for j, arr in enumerate(shared_exec.aux_arrays):
-                assert aux_shapes[j] == arr.shape
-                assert aux_types[j] == arr.dtype
-            aux_arrays = shared_exec.aux_arrays[:]
-
-        executor = self.symbol.bind(ctx=context, args=arg_arrays,
-                                    args_grad=grad_arrays, aux_states=aux_arrays,
-                                    grad_req=self.grad_req, shared_exec=shared_exec)
-        # Get the total bytes allocated for this executor
+        executor = self.symbol.simple_bind(ctx=context, grad_req=self.grad_req,
+                                           type_dict=input_types, shared_arg_names=self.param_names,
+                                           shared_exec=shared_exec,
+                                           shared_buffer=shared_data_arrays, **input_shapes)
         self._total_exec_bytes += int(executor.debug_str().split('\n')[-3].split()[1])
         return executor