TensorFlow.NET/src/TensorFlowNET.Core/Graphs/Graph.cs

/*****************************************************************************
   Copyright 2018 The TensorFlow.NET Authors. All Rights Reserved.

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.
******************************************************************************/

using System;
using System.Collections;
using System.Collections.Generic;
using System.Collections.Specialized;
using System.Linq;
using Tensorflow.Framework;
using Tensorflow.Functions;
using Tensorflow.Common.Extensions;
using Tensorflow.Graphs;
using static Tensorflow.Binding;

namespace Tensorflow
{
    /*
    A TensorFlow computation, represented as a dataflow graph.

    A `Graph` contains a set of
    `tf.Operation` objects,
    which represent units of computation; and
    `tf.Tensor` objects, which represent
    the units of data that flow between operations.

    A default `Graph` is always registered, and accessible by calling
    `tf.get_default_graph`.
    To add an operation to the default graph, simply call one of the functions
    that defines a new `Operation`:

    ```python
    c = tf.constant(4.0)
    assert c.graph is tf.get_default_graph()
    ```

    Another typical usage involves the
    `tf.Graph.as_default`
    context manager, which overrides the current default graph for the
    lifetime of the context:

    ```python
    g = tf.Graph()
    with g.as_default():
    # Define operations and tensors in `g`.
    c = tf.constant(30.0)
    assert c.graph is g
    ```

    Important note: This class *is not* thread-safe for graph construction. All
    operations should be created from a single thread, or external
    synchronization must be provided. Unless otherwise specified, all methods
    are not thread-safe.

    A `Graph` instance supports an arbitrary number of "collections"
    that are identified by name. For convenience when building a large
    graph, collections can store groups of related objects: for
    example, the `tf.Variable` uses a collection (named
    `tf.GraphKeys.GLOBAL_VARIABLES`) for
    all variables that are created during the construction of a graph. The caller
    may define additional collections by specifying a new name.     
 */

    /// <summary>
    ///     TensorFlow uses a dataflow graph to represent your computation in terms of the dependencies between individual operations. 
    ///     This leads to a low-level programming model in which you first define the dataflow graph, 
    ///     then create a TensorFlow session to run parts of the graph across a set of local and remote devices.
    /// </summary>
    /// <remarks>https://www.tensorflow.org/guide/graphs <br></br>https://www.tensorflow.org/api_docs/python/tf/Graph</remarks>
    public partial class Graph : IEnumerable<Operation>
    {
        protected new SafeGraphHandle _handle;
        private Dictionary<int, ITensorOrOperation> _nodes_by_id;
        public Dictionary<string, ITensorOrOperation> _nodes_by_name;
        private Dictionary<string, int> _names_in_use;
        public int _version;
        private int _next_id_counter;
        private List<Operation> _unfetchable_ops = new List<Operation>();
        private List<Tensor> _unfeedable_tensors = new List<Tensor>();
        private Dictionary<string, EagerDefinedFunction> _functions = new();
        internal Dictionary<string, Func<Operation, object[], Tensor[]>> _gradient_function_map = new();
        private VersionDef _graph_def_versions = new VersionDef()
        {
            Producer = versions.GRAPH_DEF_VERSION,
            MinConsumer = versions.GRAPH_DEF_VERSION_MIN_CONSUMER
        };

        public string _name_stack = "";
        protected string _graph_key;
        public string graph_key => _graph_key;
        public string _last_loss_reduction;
        public bool _is_loss_scaled_by_optimizer { get; set; }

        /// <summary>
        /// True if the graph is considered "finalized".  In that case no
        /// new operations can be added.
        /// </summary>
        private bool _finalized = false;

        /// <summary>
        /// Arbitrary collections of objects.
        /// </summary>
        private Dictionary<string, object> _collections = new Dictionary<string, object>();

        public bool building_function;

        string _container = "";
        public string Container => _container;

        int _seed;
        public int seed
        {
            get => _seed;
            set
            {
                _seed = value;
            }
        }

        internal Graph outer_graph;
        public Graph OuterGraph => outer_graph;
        public Dictionary<string, EagerDefinedFunction> Functions => _functions;
        public SafeGraphHandle c_graph => _handle;

        public Graph()
        {
            _handle = c_api.TF_NewGraph();
            _nodes_by_id = new Dictionary<int, ITensorOrOperation>();
            _nodes_by_name = new Dictionary<string, ITensorOrOperation>();
            _names_in_use = new Dictionary<string, int>();
            _graph_key = $"graph-{ops.GraphUniqueId()}/";
        }

        public ITensorOrOperation as_graph_element(object obj, bool allow_tensor = true, bool allow_operation = true)
        {
            return _as_graph_element_locked(obj, allow_tensor, allow_operation);
        }

        /// <summary>
        /// Returns a context manager that makes this `Graph` the default graph.
        /// Must call Exit() to pop graph
        /// </summary>
        /// <returns></returns>
        public virtual Graph as_default()
        {
            tf.Context.graph_mode(isFunc: false);
            return ops.set_default_graph(this);
        }

        public bool IsFunction(string name)
        {
            return _functions.ContainsKey(tf.compat.as_str(name));
        }

        internal void AddFunction(EagerDefinedFunction function)
        {
            _check_not_finalized();

            var name = function.Name;
            if(function._grad_func_name is not null && function.csharp_grad_func is not null)
            {
                throw new ValueError($"Gradient defined twice for function {name}");
            }

            var c_graph = this.c_graph;
            var func = function._c_func.Get();
            Status status = new();
            if (function._grad_func is not null)
            {
                var gradient = function._grad_func._c_func.Get();
                c_api.TF_GraphCopyFunction(c_graph, func, gradient, status);
                status.Check(true);
            }
            else
            {
                c_api.TF_GraphCopyFunction(c_graph, func, new SafeFuncGraphHandle(IntPtr.Zero), status);
                status.Check(true);
            }

            _functions[tf.compat.as_str(name)] = function;

            if(_graph_def_versions.MinConsumer < 12)
            {
                _graph_def_versions.MinConsumer = 12;
            }
        }

        private Tensor _as_graph_element(object obj)
        {
            if (obj is RefVariable var)
                return var._as_graph_element();
            else if (obj is ResourceVariable resVar)
                return resVar.GraphElement;

            return null;
        }

        private ITensorOrOperation _as_graph_element_locked(object obj, bool allow_tensor = true, bool allow_operation = true)
        {
            string types_str = "";

            if (allow_tensor && allow_operation)
            {
                types_str = "Tensor or Operation";
            }
            else if (allow_tensor)
            {
                types_str = "Tensor";
            }
            else if (allow_operation)
            {
                types_str = "Operation";
            }

            var temp_obj = _as_graph_element(obj);
            if (temp_obj != null)
                obj = temp_obj;

            // If obj appears to be a name...
            if (obj is string name)
            {
                if (name.Contains(":") && allow_tensor)
                {
                    string op_name = name.Split(':')[0];
                    int out_n = int.Parse(name.Split(':')[1]);

                    if (_nodes_by_name.ContainsKey(op_name))
                        return _nodes_by_name[op_name].outputs[out_n];
                    else
                        throw new KeyError($"The name {name} refers to a Tensor which does not " +
                            $"exist. The operation, {op_name}, does not exist in the " +
                            "graph.");
                }
                else if (!name.Contains(":") & allow_operation)
                {
                    if (!_nodes_by_name.ContainsKey(name))
                        throw new KeyError($"The name {name} refers to an Operation not in the graph.");
                    return _nodes_by_name[name];
                }
                else if (!name.Contains(":") & !allow_operation)
                {
                    // Looks like an Operation name but can't be an Operation.
                    if (_nodes_by_name.ContainsKey(name))
                        // Yep, it's an Operation name
                        throw new ValueError($"The name {name} refers to an Operation, not a {types_str}.");
                    else
                        throw new ValueError(
                            $"The name {name} looks like an (invalid) Operation name, not a {types_str}" +
                            " Tensor names must be of the form \"<op_name>:<output_index>\".");
                }
            }

            if (obj is Tensor tensor && allow_tensor)
            {
                if (tensor.graph.Equals(this))
                {
                    return tensor;
                }
                else
                {
                    throw new Exception($"Tensor {obj} is not an element of this graph.");
                }
            }
            else if (obj is Operation op && allow_operation)
            {
                if (op.graph.Equals(this))
                {
                    return op;
                }
                else
                {
                    throw new Exception($"Operation {obj} is not an element of this graph.");
                }
            }

            throw new Exception($"Can not convert a {obj.GetType().Name} into a {types_str}.");
        }

        public void add_to_collection<T>(string name, T value)
        {
            _check_not_finalized();
            if (_collections.ContainsKey(name))
                (_collections[name] as List<T>).Add(value);
            else
                _collections[name] = new List<T> { value };
        }

        public void add_to_collections<T>(List<string> names, T value)
        {
            foreach (string name in names)
                add_to_collection(name, value);
        }

        private void _check_not_finalized()
        {
            if (_finalized)
                throw new RuntimeError("Graph is finalized and cannot be modified.");
        }

        public virtual Operation create_op(string op_type, Tensor[] inputs, TF_DataType[] dtypes,
            TF_DataType[] input_types = null, string name = null,
            Dictionary<string, AttrValue> attrs = null, OpDef op_def = null,
            bool compute_device = true)
        {
            if (inputs == null)
                inputs = new Tensor[0];

            if (string.IsNullOrEmpty(name))
                name = op_type;

            // If a names ends with a '/' it is a "name scope" and we use it as-is,
            // after removing the trailing '/'.
            // This was causing duplicate graph node name errors, when testing a conv2d autoencoder
            // https://keras.io/guides/functional_api/#:~:text=keras.,graph%20(DAG)%20of%20layers.
            // name = name.EndsWith("/") ? ops.name_from_scope_name(name) : unique_name(name);
            name = name.EndsWith("/") ? ops.name_from_scope_name(name) : unique_name(name);
            var node_def = ops._NodeDef(op_type, name, attrs: attrs);

            var input_ops = inputs.Select(x => x.op).ToArray();
            var control_inputs = _control_dependencies_for_inputs(input_ops);

            var op = new Operation(node_def,
                this,
                inputs: inputs,
                output_types: dtypes,
                control_inputs: control_inputs,
                input_types: input_types,
                original_op: null,
                op_def: op_def);

            _create_op_helper(op, compute_device);

            return op;
        }

        public ITensorFlowObject device(string device_name)
        {
            return new GraphDeviceContext(this, device_name);
        }

        private void add_device_to_stack(string device_name, int offset = 0)
        {
            // TODO(Rinne): deal with device spec.
            int total_offset = offset + 1;
        }

        private void _create_op_helper(Operation op, bool compute_device = true)
        {
            // high priority
            // TODO(Rinne): complete the implementation.
            op._gradient_function = _gradient_function_map.GetOrDefault(op.type, null);
            _record_op_seen_by_control_dependencies(op);
        }

        public void _add_op(Operation op)
        {
            op._id_value = _next_id();
            _nodes_by_id[op._id] = op;
            _nodes_by_name[op.name] = op;
            _version = Math.Max(_version, op._id);
        }

        public int _next_id()
        {
            return ++_next_id_counter;
        }

        public bool is_fetchable<T>(T tensor_or_op)
        {
            if (tensor_or_op is Tensor tensor)
            {
                return !_unfetchable_ops.Contains(tensor); ;
            }
            else if (tensor_or_op is Operation op)
            {
                return !_unfetchable_ops.Contains(op);
            }

            return false;
        }

        public string get_name_scope()
        {
            return _name_stack;
        }

        public string name_scope(string name)
        {
            string new_stack = "";

            if (string.IsNullOrEmpty(name))
                new_stack = "";
            else if (name.EndsWith("/"))
                new_stack = ops.name_from_scope_name(name);
            else
                new_stack = unique_name(name);

            _name_stack = new_stack;

            return String.IsNullOrEmpty(new_stack) ? "" : new_stack + "/";
        }

        /// <summary>
        /// Return a unique operation name for `name`.
        /// 
        /// Note: You rarely need to call `unique_name()` directly.Most of
        /// the time you just need to create `with g.name_scope()` blocks to
        /// generate structured names.
        /// 
        /// `unique_name` is used to generate structured names, separated by
        /// `"/"`, to help identify operations when debugging a graph.
        /// Operation names are displayed in error messages reported by the
        /// TensorFlow runtime, and in various visualization tools such as
        /// TensorBoard.
        /// 
        /// If `mark_as_used` is set to `True`, which is the default, a new
        /// unique name is created and marked as in use.If it's set to `False`,
        /// the unique name is returned without actually being marked as used.
        /// This is useful when the caller simply wants to know what the name
        /// to be created will be.
        /// </summary>
        /// <param name="name">The name for an operation.</param>
        /// <param name="mark_as_used"> Whether to mark this name as being used.</param>
        /// <returns>A string to be passed to `create_op()` that will be used
        /// to name the operation being created.</returns>
        public string unique_name(string name, bool mark_as_used = true)
        {
            if (!String.IsNullOrEmpty(_name_stack))
                name = _name_stack + "/" + name;
            // For the sake of checking for names in use, we treat names as case
            // insensitive (e.g. foo = Foo).
            var name_key = name.ToLower();
            int i = 0;
            if (_names_in_use.ContainsKey(name_key))
                i = _names_in_use[name_key];
            // Increment the number for "name_key".
            if (mark_as_used)
                _names_in_use[name_key] = i + 1;
            if (i > 0)
            {
                // Make sure the composed name key is not already used.
                var base_name_key = name_key;
                while (_names_in_use.ContainsKey(name_key))
                {
                    name_key = $"{base_name_key}_{i}";
                    i += 1;
                }
                // Mark the composed name_key as used in case someone wants
                // to call unique_name("name_1").
                if (mark_as_used)
                    _names_in_use[name_key] = 1;

                // Return the new name with the original capitalization of the given name.
                name = $"{name}_{i - 1}";
            }
            return name;
        }

        public TF_Output[] ReturnOutputs(SafeImportGraphDefResultsHandle results)
        {
            IntPtr return_output_handle = IntPtr.Zero;
            int num_return_outputs = 0;
            c_api.TF_ImportGraphDefResultsReturnOutputs(results, ref num_return_outputs, ref return_output_handle);
            TF_Output[] return_outputs = new TF_Output[num_return_outputs];
            unsafe
            {
                var tf_output_ptr = (TF_Output*)return_output_handle;
                for (int i = 0; i < num_return_outputs; i++)
                    return_outputs[i] = *(tf_output_ptr + i);
                return return_outputs;
            }
        }

        public string[] get_all_collection_keys()
        {
            return _collections.Keys.Where(x => !x.StartsWith("__")).ToArray();
        }

        public object get_collection(string name, string scope = null)
        {
            return _collections.ContainsKey(name) ? _collections[name] : null;
        }

        public List<T> get_collection<T>(string name, string scope = null)
        {
            List<T> t = default;
            var collection = _collections.ContainsKey(name) ? _collections[name] : new List<T>();
            switch (collection)
            {
                case List<IVariableV1> list:
                    t = list.Select(x => (T)(object)x).ToList();
                    break;
                case List<ResourceVariable> list:
                    t = list.Select(x => (T)(object)x).ToList();
                    break;
                case List<RefVariable> list:
                    t = list.Select(x => (T)(object)x).ToList();
                    break;
                case List<Tensor> list:
                    t = list.Select(x => (T)(object)x).ToList();
                    break;
                case List<Operation> list:
                    t = list.Select(x => (T)(object)x).ToList();
                    break;
                default:
                    throw new NotImplementedException($"get_collection<{typeof(T).FullName}>");
            }
            return t;
        }

        public List<T> get_collection_ref<T>(string name)
        {
            if (!_collections.ContainsKey(name))
                _collections[name] = new List<T>();
            return _collections[name] as List<T>;
        }

        public void prevent_feeding(Tensor tensor)
        {
            _unfeedable_tensors.Add(tensor);
        }

        public void prevent_fetching(Operation op)
        {
            _unfetchable_ops.Add(op);
        }

        public Tensor get_tensor_by_tf_output(TF_Output tf_output)
        {
            var op = _get_operation_by_tf_operation(tf_output.oper);
            return op.outputs[tf_output.index];
        }

        /// <summary>
        /// Returns the <see cref="Tensor"/> with the given <paramref name="name"/>.
        /// This method may be called concurrently from multiple threads.
        /// </summary>
        /// <param name="name">The name of the `Tensor` to return.</param>
        /// <exception cref="KeyError">If <paramref name="name"/> does not correspond to a tensor in this graph.</exception>
        /// <returns>The `Tensor` with the given <paramref name="name"/>.</returns>
        public Tensor get_tensor_by_name(string name)
        {
            return (Tensor)this.as_graph_element(name, allow_tensor: true, allow_operation: false);
        }

        public Shape GetTensorShape(TF_Output output)
        {
            var status = tf.Status;
            var ndim = c_api.TF_GraphGetTensorNumDims(_handle, output, status);
            status.Check();

            if (ndim == -1)
                return Shape.Null;

            var dims = new long[ndim];
            c_api.TF_GraphGetTensorShape(_handle, output, dims, dims.Length, status);
            status.Check();

            return new Shape(dims.Select(x => (int)x).ToArray());
        }

        public virtual void Exit()
        {
            tf.Context.restore_mode();
            ops.pop_graph();
        }

        internal EagerDefinedFunction _get_function(string name)
        {
            return _functions.GetOrDefault(name, null);
        }

        string debugString = string.Empty;
        public override string ToString()
        {
            return $"{graph_key}, 0x{_handle.DangerousGetHandle().ToString("x16")}";
            /*if (string.IsNullOrEmpty(debugString))
            {
                int len = 0;
                debugString = c_api.TF_GraphDebugString(_handle, out len);
            }

            return debugString;*/
        }

        private IEnumerable<Operation> GetEnumerable()
            => c_api_util.tf_operations(this);

        IEnumerator<Operation> IEnumerable<Operation>.GetEnumerator()
            => GetEnumerable().GetEnumerator();

        IEnumerator IEnumerable.GetEnumerator()
            => throw new NotImplementedException();

        public static implicit operator SafeGraphHandle(Graph graph)
        {
            return graph._handle;
        }
    }
}