Python graph_editor.get_backward_walk_ops方法代碼示例

# 需要導入模塊: from tensorflow.contrib import graph_editor [as 別名]
# 或者: from tensorflow.contrib.graph_editor import get_backward_walk_ops [as 別名]
def dependency_of_targets(targets, op):
    """
    Check that op is in the subgraph induced by the dependencies of targets.
    The result is memoized.

    This is useful if some SessionRunHooks should be run only together with certain ops.

    Args:
        targets: a tuple of ops or tensors. The targets to find dependencies of.
        op (tf.Operation or tf.Tensor):

    Returns:
        bool
    """
    # TODO tensorarray? sparsetensor?
    if isinstance(op, tf.Tensor):
        op = op.op
    assert isinstance(op, tf.Operation), op

    from tensorflow.contrib.graph_editor import get_backward_walk_ops
    # alternative implementation can use graph_util.extract_sub_graph
    dependent_ops = get_backward_walk_ops(targets, control_inputs=True)
    return op in dependent_ops 

# 需要導入模塊: from tensorflow.contrib import graph_editor [as 別名]
# 或者: from tensorflow.contrib.graph_editor import get_backward_walk_ops [as 別名]
def fast_backward_ops(within_ops, seed_ops, stop_at_ts):
    bwd_ops = set(ge.get_backward_walk_ops(seed_ops, stop_at_ts=stop_at_ts))
    ops = bwd_ops.intersection(within_ops).difference(
        [t.op for t in stop_at_ts])
    return list(ops) 

# 需要導入模塊: from tensorflow.contrib import graph_editor [as 別名]
# 或者: from tensorflow.contrib.graph_editor import get_backward_walk_ops [as 別名]
def fast_backward_ops(within_ops, seed_ops, stop_at_ts):
    bwd_ops = set(ge.get_backward_walk_ops(seed_ops, stop_at_ts=stop_at_ts))
    ops = bwd_ops.intersection(within_ops).difference([t.op for t in stop_at_ts])
    return list(ops) 

# 需要導入模塊: from tensorflow.contrib import graph_editor [as 別名]
# 或者: from tensorflow.contrib.graph_editor import get_backward_walk_ops [as 別名]
def export_subgraph(checkpoint, output_tensors, saveto):
    """
    For the current graph, export the subgraph connected to output_tensors to a new graph_def file
    :param checkpoint: path to checkpoint
    :param output_tensors: output tensor names
    :param saveto: path to save graph_def file to
    :return:
    """
    saver = tf.train.import_meta_graph(checkpoint + '.meta', clear_devices=True)
    graph = tf.get_default_graph()
    if isinstance(output_tensors, str):
        output_tensors = [graph.get_tensor_by_name(output_tensors)]
    else:
        assert all([isinstance(out, str) for out in output_tensors])
        output_tensors = [graph.get_tensor_by_name(out) for out in output_tensors]

    def _var_ops(var_op):  # get operations one step ahead of variable ops: read/assign/etc.
        return [var_op.name] + [op.name for t in var_op.outputs for op in t.consumers()]

    keep_op_names = [out.op.name for out in output_tensors]
    var_ops = list({op for out in output_tensors for op in ge.get_backward_walk_ops(out) if op.type == 'VariableV2'})
    keep_op_names += [opname for op in var_ops for opname in _var_ops(op)]
    keep_op_names = [opname for opname in keep_op_names if 'save/' not in opname and 'save_' not in opname]
    graph_def = tf.graph_util.extract_sub_graph(graph.as_graph_def(), keep_op_names)

    with tf.Session() as sess:
        saver.restore(sess, checkpoint)
        new_graph_def = tf.graph_util.convert_variables_to_constants(
            sess, graph_def, [out.op.name for out in output_tensors])
    tf.reset_default_graph()
    tf.train.export_meta_graph(saveto, graph_def=new_graph_def, clear_devices=True) 

# 需要導入模塊: from tensorflow.contrib import graph_editor [as 別名]
# 或者: from tensorflow.contrib.graph_editor import get_backward_walk_ops [as 別名]
def create_session(self):
        sess = tf.Session(target=self.target, config=self.config)

        def blocking_op(x):
            """
            Whether an op is possibly blocking.
            """
            if x.op_def is not None and not x.op_def.is_stateful:
                return False
            if "Dequeue" in x.type or "Enqueue" in x.type:
                return True
            if "Unstage" in x.type:
                return True
            if x.type in ["ZMQPull"]:
                return True
            return False

        def run(op):
            if not is_tfv2():
                from tensorflow.contrib.graph_editor import get_backward_walk_ops

                deps = get_backward_walk_ops(op, control_inputs=True)
                for dep_op in deps:
                    if blocking_op(dep_op):
                        logger.warn(
                            "Initializer '{}' depends on a blocking op '{}'. "
                            "This initializer is likely to hang!".format(
                                op.name, dep_op.name))
            sess.run(op)

        run(tf.global_variables_initializer())
        run(tf.local_variables_initializer())
        run(tf.tables_initializer())
        return sess