Merge pull request #7213 from JiayiFeng/dev_add_callback_for_backward

Error Clip Design Doc

Author: JiayiFeng

doc/design/error_clip.md

@@ -0,0 +1,87 @@
+# Error Clip
+
+## Overview
+
+Error clip is widely used in model training to prevent gradient exploding. It takes some specific rules to adjust variables' gradients and prevent them from being too large. With it, values of a gradient will be checked before they are taken by the next `grad_op` and be shrunk if necessary.
+## Usage
+
+Users are allowed to assign different error clip methods or attributes to different `Variable`s. Users can specify it as a parameter of `Variable`'s constructor:
+
+```python
+var = framework.Variable(..., error_clip=myErrorClip, ...)
+```
+
+The default value of `error_clip` is `None`, which means no error clip is employed. When it's not `None`, it should take an object of `BaseErrorClipAttr`'s derived class. So far, `BaseErrorClipAttr` has only one derived class: `ErrorClipByValue`, whose constructor is:
+
+```python
+ErrorClipByValue(max, min=None)
+```
+
+`max` and `min` represent the maximal and minimal clip threshold respectively. In backward pass, all values of `var`'s gradient greater than `max` or less than `min` will be clipped to `max` and `min` respectively. When the `min` is None, the minimal threshold will be assigned with `-max` automatically.
+
+So we can enable the error clip with threshold `[-5.0, 5.0]` for variable `var` by:
+
+```python
+var = framework.Variable(..., error_clip=ErrorClipByValue(max=5.0), ...)
+```
+
+## Implementation
+
+The `BaseErrorClipAttr` and its derived class `ErrorClipByValue` are defined in *clip.py*.
+
+```python
+class BaseErrorClipAttr(object):
+    def append_clip_op(self, block, grad_name):
+        raise NotImplementedError()
+
+
+class ErrorClipByValue(BaseErrorClipAttr):
+    def __init__(self, max, min=None):
+        max = float(max)
+        if min is None:
+            min = -max
+        else:
+            min = float(min)
+        self.max = max
+        self.min = min
+
+    def append_clip_op(self, block, grad_name):
+        block.append_op(
+            type="clip",
+            inputs={"X": grad_name},
+            outputs={"Out": grad_name},
+            attrs={"min": self.min,
+                   "max": self.max})
+```
+
+The `BaseErrorClipAttr` have one main member functions: `append_clip_op(self, block, grad_name)`.
+
+This function is used to create a `clip_op` and append it to the end of given `block`. For different error clip algorithm require different `clip_op`, the function is defined as virtual in the base class. All derived classes must implement their own versions of this function.
+
+These `clip_op`s should be inserted after `grad_op`s whose output gradients need to be clipped. It is equivalent to appending some `clip_op`s to the end of the target block every time a new `grad_op` is added.
+
+```python
+for op_desc in grad_op_descs:
+        new_op_desc = target_block.desc.append_op()
+        new_op_desc.copy_from(op_desc)
+        callback(block=target_block, context=grad_to_var)
+```
+
+Here we employ a callback function to complete this kind of jobs. In `_append_backward_ops_` function, each time after a `grad_op` is added to the `target_block`, a callback function is invoked. The logic of `clip_op` appending can be implemented inside the callback function.
+
+The callback function for `clip_op` appending is defined in *clip.py*:
+
+```python
+def error_clip_callback(block, context):
+    # the context is a grad_to_var map
+    grad_to_var = context
+    op_desc = block.desc.op(block.desc.op_size() - 1)
+    for grad_n in filter(lambda n: grad_to_var.has_key(n),
+                         op_desc.output_arg_names()):
+        fwd_var = block.var_recursive(grad_to_var[grad_n])
+        error_clip = getattr(fwd_var, "error_clip", None)
+        if error_clip is not None:
+            error_clip.append_clip_op(block, grad_n)
+```
+
+This function takes a `block` and a `context`(which is actually a grad\_to\_var map) as inputs. It checks each output of the last `OpDesc` in the `block`. Notice that the last `OpDesc` of the `block` must be a `grad_op` and its outputs must be some forward variables' gradients. If an output gradient's corresponding forward variable has an attribute of `error_clip`, `error_clip_callback` will call the `error_clip`'s `append_clip_op` function to append the required `clip_op` into the `block`.

python/paddle/v2/fluid/backward.py

@@ -188,7 +188,17 @@ def _append_backward_ops_(target,
         grad_to_var(dict)(output argument):
             key(str): grad variable name
             val(str): corresponding forward variable name
+        callback(callable object): a callable object used to decorate new generated grad ops
     """
+    if callback is None:
+
+        def empty_callback(block, context):
+            pass
+
+        callback = empty_callback
+    elif not hasattr(callback, '__call__'):
+        raise ValueError("'callback' must be a callable object.")
+
     # grad_op_descs holds created grad_op, and will be appended to target_block
     grad_op_descs = []
     program = block.program
@@ -226,6 +236,7 @@ def _append_backward_ops_(target,
     for op_desc in grad_op_descs:
         new_op_desc = target_block.desc.append_op()
         new_op_desc.copy_from(op_desc)
+        callback(block=target_block, context=grad_to_var)
 
 
 def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
@@ -268,7 +279,7 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
                 _infer_var_data_type_(arg, block)
 
 
-def append_backward(loss, parameter_list=None, no_grad_set=None):
+def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
     """
     Append backward part to main_program
 
@@ -312,7 +323,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None):
     grad_to_var = dict()
 
     _append_backward_ops_(loss, root_block, root_block, no_grad_dict,
-                          grad_to_var)
+                          grad_to_var, callback)
     _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
 
     program.current_block_idx = current_block_idx

python/paddle/v2/fluid/clip.py

@@ -1,7 +1,46 @@
 import functools
 import layers
+from . import core
 
-__all__ = ['GradientClipByValue', 'append_gradient_clip_ops']
+__all__ = [
+    'GradientClipByValue', 'append_gradient_clip_ops', 'error_clip_callback'
+]
+
+
+class BaseErrorClipAttr(object):
+    def append_clip_op(self, block, grad_name):
+        raise NotImplementedError()
+
+
+class ErrorClipByValue(BaseErrorClipAttr):
+    def __init__(self, max, min=None):
+        max = float(max)
+        if min is None:
+            min = -max
+        else:
+            min = float(min)
+        self.max = max
+        self.min = min
+
+    def append_clip_op(self, block, grad_name):
+        block.append_op(
+            type="clip",
+            inputs={"X": grad_name},
+            outputs={"Out": grad_name},
+            attrs={"min": self.min,
+                   "max": self.max})
+
+
+def error_clip_callback(block, context):
+    # the context is a grad_to_var map
+    grad_to_var = context
+    op_desc = block.desc.op(block.desc.op_size() - 1)
+    for grad_n in filter(lambda n: grad_to_var.has_key(n),
+                         op_desc.output_arg_names()):
+        fwd_var = block.var_recursive(grad_to_var[grad_n])
+        error_clip = getattr(fwd_var, "error_clip", None)
+        if error_clip is not None:
+            error_clip.append_clip_op(block, grad_n)
 
 
 class BaseGradientClipAttr(object):

python/paddle/v2/fluid/framework.py

@@ -143,9 +143,11 @@ def __init__(self,
                  dtype=None,
                  lod_level=None,
                  persistable=None,
+                 error_clip=None,
                  stop_gradient=False,
                  **kwargs):
         self.block = block
+        self.error_clip = error_clip
 
         if name is None:
             name = Variable._unique_var_name_()
@@ -622,6 +624,17 @@ def var(self, name):
             raise ValueError("var %s not in this block" % name)
         return v
 
+    def var_recursive(self, name):
+        if self.has_var(name):
+            return self.var(name)
+        else:
+            if self.idx == 0:
+                raise ValueError("var %s is not in block(%d) nor its parents." %
+                                 name, self.idx)
+            else:
+                parent_block = self.program.block(self.parent_idx)
+                return parent_block.var_recursive(name)
+
     def all_parameters(self):
         return list(self.iter_parameters())
 
@@ -740,6 +753,7 @@ def copy_param_info_from(self, other):
                 optimize_attr=p.optimize_attr,
                 regularizer=p.regularizer,
                 clip_attr=p.clip_attr,
+                error_clip=p.error_clip,
                 name=v.name)
             self.vars[new_p.name] = new_p
 

python/paddle/v2/fluid/optimizer.py

@@ -6,7 +6,7 @@
 from initializer import Constant
 from layer_helper import LayerHelper
 from regularizer import append_regularization_ops
-from clip import append_gradient_clip_ops
+from clip import append_gradient_clip_ops, error_clip_callback
 
 __all__ = ['SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad']
 
@@ -197,7 +197,8 @@ def minimize(self,
         This method combines interface `append_backward()` and
         `create_optimization_pass()` into one.
         """
-        params_grads = append_backward(loss, parameter_list, no_grad_set)
+        params_grads = append_backward(loss, parameter_list, no_grad_set,
+                                       error_clip_callback)
 
         params_grads = append_gradient_clip_ops(params_grads)