def _assert_in_default_state(t): t.assertIs(distribute._default_tower_context, distribute.get_tower_context()) t.assertIs(None, distribute.get_cross_tower_context()) t.assertIs(distribute._default_distribution_strategy, distribute.get_distribution_strategy()) t.assertFalse(distribute.has_distribution_strategy())
def _assign_func(self, *args, **kwargs): f = kwargs.pop("f") if distribute_lib.get_cross_tower_context(): update_device = distribute_lib.get_update_device() # We are calling update on the mirrored variable in cross tower context. if update_device is not None: # We are calling an assign function on the mirrored variable in cross # tower context. v = self.get(device=update_device) return f(v, *args, **kwargs) return distribute_lib.get_distribution_strategy().update( self, f, *args, **kwargs) else: _assert_tower_context() # We are calling an assign function on the mirrored variable in tower # context. # We reduce the value we want to assign/add/sub. More details about how we # handle the different use cases can be found in the _reduce method. # We call the function on each of the mirrored variables with the reduced # value. if self._aggregation == vs.VariableAggregation.NONE: raise ValueError("You must specify an aggregation method to update a " "MirroredVariable in Tower Context.") def merge_fn(strategy, value, *other_args, **other_kwargs): return strategy.update( self, f, strategy.reduce( aggregation=self._aggregation, value=value, destinations=self), *other_args, **other_kwargs) return distribute_lib.get_tower_context().merge_call(merge_fn, *args, **kwargs)
def merge_fn(dist, s): self.assertIs(distribute._default_distribution_strategy, dist) self.assertIs(None, distribute.get_tower_context()) self.assertIs(dist, distribute.get_cross_tower_context()) self.assertIs(dist, distribute.get_distribution_strategy()) self.assertFalse(distribute.has_distribution_strategy()) return "foo_" + s
def run_fn(): tower_context = distribute.get_tower_context() self.assertTrue(tower_context is not None) self.assertIs(None, distribute.get_cross_tower_context()) self.assertTrue(distribute.has_distribution_strategy()) self.assertIs(dist, distribute.get_distribution_strategy()) self.assertEqual("foo", tower_context.merge_call(None, test_arg="foo")) self.assertEqual("bar", variable_scope.variable(1.0, name="bar"))
def op(self): # We want cross-tower code that does some var.op.X calls # to work (even if the current device isn't in self.devices), but # other uses of var.op in a cross-tower context to fail. if distribute_lib.get_cross_tower_context(): return DistributedVarOp(self._primary_var.op.name, self._primary_var.op.graph, self._primary_var.op.type) return self.get().op
def testScope(self): _assert_in_default_state(self) dist = _TestStrategy() with dist.scope(): self.assertIs(None, distribute.get_tower_context()) self.assertIs(dist, distribute.get_cross_tower_context()) self.assertTrue(distribute.has_distribution_strategy()) self.assertIs(dist, distribute.get_distribution_strategy()) self.assertEqual("baz", variable_scope.variable(1.0, name="baz")) _assert_in_default_state(self)
def set_non_tensor_output(self, name, output): """Set `output` with `name` to be captured as a non tensor output.""" if distribute_lib.get_cross_tower_context(): self._non_tensor_outputs[name] = output else: def merge_fn(distribution, value): # NOTE(priyag): For non tensor outputs, we simply return all the values # in a list as aggregation doesn't make sense on non tensors. self._non_tensor_outputs[name] = distribution.unwrap(value) distribute_lib.get_tower_context().merge_call(merge_fn, output)
def run_fn(): tower_context = distribute.get_tower_context() self.assertTrue(tower_context is not None) self.assertIs(None, distribute.get_cross_tower_context()) self.assertTrue(distribute.has_distribution_strategy()) self.assertIs(dist, distribute.get_distribution_strategy()) self.assertEqual("foo", tower_context.merge_call(None, test_arg="foo")) expected_value = _get_test_variable( "bar", variable_scope.VariableSynchronization.AUTO, variable_scope.VariableAggregation.NONE) self.assertDictEqual(expected_value, variable_scope.variable(1.0, name="bar"))
def _set_checkpoint_initializer(variable, ckpt_file, tensor_name, slice_spec, name="checkpoint_initializer"): """Overrides given variable's initialization op. Sets variable initializer to assign op that initializes variable from tensor's value in the checkpoint. Args: variable: `tf.Variable` object. ckpt_file: string, full path of the checkpoint. tensor_name: Name of the tensor to load from the checkpoint. slice_spec: Slice specification for loading partitioned tensors. name: Name of the operation. """ base_type = variable.dtype.base_dtype # Do not colocate with variable since RestoreV2 op only runs on CPU and # colocation will force variable (and other ops that colocate with variable) # to be on CPU as well. It is okay to place the variable's initializer op on # CPU since it will only be run once at the start. with ops.device(variable.device), ops.device("/cpu:0"): restore_op = io_ops.restore_v2(ckpt_file, [tensor_name], [slice_spec], [base_type], name=name)[0] # TODO(priyag, allenl): Use `SaveableObject.restore` instead here. if resource_variable_ops.is_resource_variable(variable): init_op = variable.assign(restore_op, read_value=False) else: init_op = state_ops.assign(variable, restore_op) # pylint:disable=protected-access # We need special handling for `DistributedVariable`s as they contain # mutliple actual variables. `assign` on a `DistributedVariable` returns a # combined `init_op` which contains initializers for all the contained # variables. We then set each underlying variable's `_initializer_op` using # the corresponding `init_op`. # TODO(priyag): Use `isinstance` checks when `DistributedVariable` class # moves out of contrib. if any(base.__name__ == "DistributedVariable" for base in variable.__class__.__bases__): assert distribute_lib.get_cross_tower_context() assert hasattr(variable, "_index") for (d, v) in six.iteritems(variable._index): v._initializer_op = init_op._index[d] restore_op.set_shape(v.shape) v._initial_value = restore_op else: variable._initializer_op = init_op restore_op.set_shape(variable.shape) variable._initial_value = restore_op
def testScope(self): _assert_in_default_state(self) dist = _TestStrategy() with dist.scope(): self.assertIs(None, distribute.get_tower_context()) self.assertIs(dist, distribute.get_cross_tower_context()) self.assertTrue(distribute.has_distribution_strategy()) self.assertIs(dist, distribute.get_distribution_strategy()) expected_value = _get_test_variable( "baz", variable_scope.VariableSynchronization.AUTO, variable_scope.VariableAggregation.NONE) self.assertDictEqual(expected_value, variable_scope.variable(1.0, name="baz")) _assert_in_default_state(self)
def assign(self, *args, **kwargs): if distribute_lib.get_cross_tower_context(): # To preserve the sum across save and restore, we have to divide the # total across all devices when restoring a variable that was summed # when saving. tensor = args[0] if self._aggregation == vs.VariableAggregation.SUM: tensor *= 1. / len(self.devices) return control_flow_ops.group( [_assign_on_device(d, v, tensor) for d, v in six.iteritems(self._index)]) else: _assert_tower_context() return self.get().assign(*args, **kwargs)
def set_last_step_output(self, name, output, aggregation=variables_lib.VariableAggregation.NONE): """Set `output` with `name` to be outputted from the last step. Args: name: String, name to identify the output. Doesn't need to match tensor name. output: The tensors that should be outputted with `name`. See below for actual types supported. aggregation: Aggregation method to use to aggregate outputs from multiple towers. Required if `set_last_step_output` is called in a tower context. Optional in cross_tower_context. When present, the outputs from all the towers are aggregated using the current distribution strategy's `reduce` method. Hence, the type of `output` must be what's supported by the corresponding `reduce` method. For e.g. if using MirroredStrategy and aggregation is set, output must be a `PerDevice` value. The aggregation method is also recorded in a dictionary `_last_step_outputs_aggregations` for later interpreting of the outputs as already reduced or not. """ if distribute_lib.get_cross_tower_context(): self._last_step_outputs_aggregations[name] = aggregation if aggregation is variables_lib.VariableAggregation.NONE: self._last_step_outputs[name] = output else: distribution = distribute_lib.get_distribution_strategy() self._last_step_outputs[name] = distribution.reduce( aggregation, output, destinations="/device:CPU:0") else: assert aggregation is not variables_lib.VariableAggregation.NONE def merge_fn(distribution, value): self._last_step_outputs[name] = distribution.reduce( aggregation, value, destinations="/device:CPU:0") # Setting this inside the `merge_fn` because all towers share the same # context object, so it's more robust to set it only once (even if all # the towers are trying to set the same value). self._last_step_outputs_aggregations[name] = aggregation distribute_lib.get_tower_context().merge_call(merge_fn, output)
def _as_graph_element(self): # pylint: disable=protected-access if distribute_lib.get_cross_tower_context(): return self._get_cross_tower() return self.get()._as_graph_element()
def apply_gradients(self, grads_and_vars, global_step=None, name=None): """Apply gradients to variables. This is the second part of `minimize()`. It returns an `Operation` that applies gradients. Args: grads_and_vars: List of (gradient, variable) pairs as returned by `compute_gradients()`. global_step: Optional `Variable` to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the `Optimizer` constructor. Returns: An `Operation` that applies the specified gradients. If `global_step` was not None, that operation also increments `global_step`. Raises: TypeError: If `grads_and_vars` is malformed. ValueError: If none of the variables have gradients. RuntimeError: If you should use `_distributed_apply()` instead. """ # This is a default implementation of apply_gradients() that can be shared # by most optimizers. It relies on the subclass implementing the following # methods: _create_slots(), _prepare(), _apply_dense(), and _apply_sparse(). # Handle DistributionStrategy case. if distribute_lib.get_cross_tower_context(): raise RuntimeError("Use `_distributed_apply()` instead of " "`apply_gradients()` in a cross-tower context.") # TODO(isaprykin): Get rid of `has_distribution_strategy()` check by # always calling _distributed_apply(), using the default distribution # as needed. if distribute_lib.has_distribution_strategy(): grads_and_vars = get_filtered_grad_fn(lambda _: grads_and_vars)() return distribute_lib.get_tower_context().merge_call( self._distributed_apply, grads_and_vars, global_step, name) # No DistributionStrategy case. grads_and_vars = tuple(grads_and_vars) # Make sure repeat iteration works. if not grads_and_vars: raise ValueError("No variables provided.") converted_grads_and_vars = [] for g, v in grads_and_vars: if g is not None: try: # Convert the grad to Tensor or IndexedSlices if necessary. g = ops.convert_to_tensor_or_indexed_slices(g) except TypeError: raise TypeError( "Gradient must be convertible to a Tensor" " or IndexedSlices, or None: %s" % g) if not isinstance(g, (ops.Tensor, ops.IndexedSlices)): raise TypeError( "Gradient must be a Tensor, IndexedSlices, or None: %s" % g) p = _get_processor(v) converted_grads_and_vars.append((g, v, p)) converted_grads_and_vars = tuple(converted_grads_and_vars) var_list = [v for g, v, _ in converted_grads_and_vars if g is not None] if not var_list: raise ValueError("No gradients provided for any variable: %s." % ([str(v) for _, _, v in converted_grads_and_vars],)) with ops.init_scope(): self._create_slots(var_list) update_ops = [] with ops.name_scope(name, self._name) as name: self._prepare() for grad, var, processor in converted_grads_and_vars: if grad is None: continue # We colocate all ops created in _apply_dense or _apply_sparse # on the same device as the variable. # TODO(apassos): figure out how to get the variable name here. if context.executing_eagerly() or isinstance( var, resource_variable_ops.ResourceVariable) and not var._in_graph_mode: # pylint: disable=protected-access scope_name = "" else: scope_name = var.op.name with ops.name_scope("update_" + scope_name), ops.colocate_with(var): update_ops.append(processor.update_op(self, grad)) if global_step is None: apply_updates = self._finish(update_ops, name) else: with ops.control_dependencies([self._finish(update_ops, "update")]): with ops.colocate_with(global_step): if isinstance(global_step, resource_variable_ops.ResourceVariable): # TODO(apassos): the implicit read in assign_add is slow; consider # making it less so. apply_updates = resource_variable_ops.assign_add_variable_op( global_step.handle, ops.convert_to_tensor(1, dtype=global_step.dtype), name=name) else: apply_updates = state_ops.assign_add(global_step, 1, name=name) if not context.executing_eagerly(): if isinstance(apply_updates, ops.Tensor): apply_updates = apply_updates.op train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP) if apply_updates not in train_op: train_op.append(apply_updates) return apply_updates
def init_from_checkpoint(ckpt_dir_or_file, assignment_map): """Initializes current variables with tensors loaded from given checkpoint. Note: This overrides default initialization ops of specified variables and redefines dtype. Assignment map supports following syntax: * `'checkpoint_scope_name/': 'scope_name/'` - will load all variables in current `scope_name` from `checkpoint_scope_name` with matching tensor names. * `'checkpoint_scope_name/some_other_variable': 'scope_name/variable_name'` - will initialize `scope_name/variable_name` variable from `checkpoint_scope_name/some_other_variable`. * `'scope_variable_name': variable` - will initialize given `tf.Variable` object with tensor 'scope_variable_name' from the checkpoint. * `'scope_variable_name': list(variable)` - will initialize list of partitioned variables with tensor 'scope_variable_name' from the checkpoint. * `'/': 'scope_name/'` - will load all variables in current `scope_name` from checkpoint's root (e.g. no scope). Supports loading into partitioned variables, which are represented as `'<variable>/part_<part #>'`. Example: ```python # Say, '/tmp/model.ckpt' has the following tensors: # -- name='old_scope_1/var1', shape=[20, 2] # -- name='old_scope_1/var2', shape=[50, 4] # -- name='old_scope_2/var3', shape=[100, 100] # Create new model's variables with tf.variable_scope('new_scope_1'): var1 = tf.get_variable('var1', shape=[20, 2], initializer=tf.zeros_initializer()) with tf.variable_scope('new_scope_2'): var2 = tf.get_variable('var2', shape=[50, 4], initializer=tf.zeros_initializer()) # Partition into 5 variables along the first axis. var3 = tf.get_variable(name='var3', shape=[100, 100], initializer=tf.zeros_initializer(), partitioner=lambda shape, dtype: [5, 1]) # Initialize all variables in `new_scope_1` from `old_scope_1`. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/': 'new_scope_1'}) # Use names to specify which variables to initialize from checkpoint. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/var1': 'new_scope_1/var1', 'old_scope_1/var2': 'new_scope_2/var2'}) # Or use tf.Variable objects to identify what to initialize. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/var1': var1, 'old_scope_1/var2': var2}) # Initialize partitioned variables using variable's name init_from_checkpoint('/tmp/model.ckpt', {'old_scope_2/var3': 'new_scope_2/var3'}) # Or specify the list of tf.Variable objects. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_2/var3': var3._get_variable_list()}) ``` Args: ckpt_dir_or_file: Directory with checkpoints file or path to checkpoint. assignment_map: Dict, where keys are names of the variables in the checkpoint and values are current variables or names of current variables (in default graph). Raises: tf.errors.OpError: If missing checkpoints or tensors in checkpoints. ValueError: If missing variables in current graph. """ if distribute_lib.get_cross_tower_context(): _init_from_checkpoint(None, ckpt_dir_or_file, assignment_map) else: distribute_lib.get_tower_context().merge_call( _init_from_checkpoint, ckpt_dir_or_file, assignment_map)
def init_from_checkpoint(ckpt_dir_or_file, assignment_map): """Initializes current variables with tensors loaded from given checkpoint. Note: This overrides default initialization ops of specified variables and redefines dtype. Assignment map supports following syntax: * `'checkpoint_scope_name/': 'scope_name/'` - will load all variables in current `scope_name` from `checkpoint_scope_name` with matching tensor names. * `'checkpoint_scope_name/some_other_variable': 'scope_name/variable_name'` - will initialize `scope_name/variable_name` variable from `checkpoint_scope_name/some_other_variable`. * `'scope_variable_name': variable` - will initialize given `tf.Variable` object with tensor 'scope_variable_name' from the checkpoint. * `'scope_variable_name': list(variable)` - will initialize list of partitioned variables with tensor 'scope_variable_name' from the checkpoint. * `'/': 'scope_name/'` - will load all variables in current `scope_name` from checkpoint's root (e.g. no scope). Supports loading into partitioned variables, which are represented as `'<variable>/part_<part #>'`. Example: ```python # Say, '/tmp/model.ckpt' has the following tensors: # -- name='old_scope_1/var1', shape=[20, 2] # -- name='old_scope_1/var2', shape=[50, 4] # -- name='old_scope_2/var3', shape=[100, 100] # Create new model's variables with tf.variable_scope('new_scope_1'): var1 = tf.get_variable('var1', shape=[20, 2], initializer=tf.zeros_initializer()) with tf.variable_scope('new_scope_2'): var2 = tf.get_variable('var2', shape=[50, 4], initializer=tf.zeros_initializer()) # Partition into 5 variables along the first axis. var3 = tf.get_variable(name='var3', shape=[100, 100], initializer=tf.zeros_initializer(), partitioner=lambda shape, dtype: [5, 1]) # Initialize all variables in `new_scope_1` from `old_scope_1`. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/': 'new_scope_1'}) # Use names to specify which variables to initialize from checkpoint. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/var1': 'new_scope_1/var1', 'old_scope_1/var2': 'new_scope_2/var2'}) # Or use tf.Variable objects to identify what to initialize. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/var1': var1, 'old_scope_1/var2': var2}) # Initialize partitioned variables using variable's name init_from_checkpoint('/tmp/model.ckpt', {'old_scope_2/var3': 'new_scope_2/var3'}) # Or specify the list of tf.Variable objects. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_2/var3': var3._get_variable_list()}) ``` Args: ckpt_dir_or_file: Directory with checkpoints file or path to checkpoint. assignment_map: Dict, where keys are names of the variables in the checkpoint and values are current variables or names of current variables (in default graph). Raises: tf.errors.OpError: If missing checkpoints or tensors in checkpoints. ValueError: If missing variables in current graph. """ if distribute_lib.get_cross_tower_context(): _init_from_checkpoint(None, ckpt_dir_or_file, assignment_map) else: distribute_lib.get_tower_context().merge_call(_init_from_checkpoint, ckpt_dir_or_file, assignment_map)