/// <summary>
/// map on the list of tensors unpacked from `elems` on dimension 0.
/// </summary>
/// <param name="fn"></param>
/// <param name="elems"></param>
/// <param name="dtype"></param>
/// <param name="parallel_iterations"></param>
/// <param name="back_prop"></param>
/// <param name="swap_memory"></param>
/// <param name="infer_shape"></param>
/// <param name="name"></param>
/// <returns>A tensor or (possibly nested) sequence of tensors.</returns>
public static Tensor map_fn(Func <Tensor, Tensor> fn,
Tensor elems,
TF_DataType dtype = TF_DataType.DtInvalid,
int parallel_iterations = 10,
bool back_prop = true,
bool swap_memory = false,
bool infer_shape = true,
string name = null)
{
bool input_is_sequence = nest.is_sequence(elems);
Tensor[] input_flatten(Tensor x) => input_is_sequence?nest.flatten(x).ToArray() : new [] { x };
Tensor input_pack(Tensor[] x) => input_is_sequence ? (Tensor)nest.pack_sequence_as(elems, x) : x[0];
bool output_is_sequence;
Func <Tensor, Tensor[]> output_flatten;
Func <Tensor[], Tensor> output_pack;
if (dtype == TF_DataType.DtInvalid)
{
output_is_sequence = input_is_sequence;
output_flatten = input_flatten;
output_pack = input_pack;
}
else
{
output_is_sequence = nest.is_sequence(dtype);
output_flatten = (x) => output_is_sequence?nest.flatten(x).ToArray() : new [] { x };
output_pack = (x) => output_is_sequence ? (Tensor)nest.pack_sequence_as(dtype, x) : x[0];
}
var elems_flat = input_flatten(elems);
return(tf_with(ops.name_scope(name, "map", elems_flat), delegate
{
//if in_graph_mode:
//# Any get_variable calls in fn will cache the first call locally
//# and not issue repeated network I/O requests for each iteration.
//varscope = vs.get_variable_scope()
//varscope_caching_device_was_none = False
//if varscope.caching_device is None:
// # TODO(ebrevdo): Change to using colocate_with here and in other
// # methods.
// varscope.set_caching_device(lambda op: op.device)
// varscope_caching_device_was_none = True
elems_flat = elems_flat.Select(elem => ops.convert_to_tensor(elem, name: "elem"))
.ToArray();
dtype = elems_flat.Select(elem => elem.dtype).First();
var dtype_flat = new[] { dtype };
// Convert elems to tensor array. n may be known statically.
var static_shape = elems_flat[0].shape;
var n = static_shape[0];
// TensorArrays are always flat
var elems_ta = elems_flat.Select(elem => new TensorArray(dtype: elem.dtype,
size: ops.convert_to_tensor(n),
dynamic_size: false,
infer_shape: true)).ToArray();
// Unpack elements
var elems_ta_1 = new List <TensorArray>();
foreach (var(elem_ta, elem) in zip(elems_ta, elems_flat))
{
elems_ta_1.Add(elem_ta.unstack(elem));
}
elems_ta = elems_ta_1.ToArray();
var i = constant_op.constant(0);
var accs_ta = dtype_flat.Select(dt => new TensorArray(dtype: dt,
size: ops.convert_to_tensor(n),
dynamic_size: false,
infer_shape: infer_shape)).ToArray();
BodyItem compute(BodyItem item)
{
var packed_values = input_pack(elems_ta.Select(elem_ta => elem_ta.read(item.I)).ToArray());
var packed_fn_values = fn(packed_values);
//nest.assert_same_structure(dtype or elems, packed_fn_values)
var flat_fn_values = output_flatten(packed_fn_values);
for (int j = 0; j < item.Accs_ta.Length; j++)
{
item.Accs_ta[j].write(item.I, flat_fn_values[j]);
}
return new BodyItem(item.I + 1, item.Accs_ta);
}
var r_a = control_flow_ops.while_loop(
(x) => x.I < n,
compute,
new BodyItem(i, accs_ta),
parallel_iterations: parallel_iterations,
back_prop: back_prop,
swap_memory: swap_memory,
maximum_iterations: tf.constant(n));
var results_flat = r_a.Accs_ta.Select(r => r.stack()).ToArray();
var n_static = new Dimension(tensor_shape.dimension_value(elems_flat[0].TensorShape.with_rank_at_least(1).dims[0]));
foreach (var elem in elems_flat.Skip(1))
{
n_static.merge_with(new Dimension(tensor_shape.dimension_value(elem.TensorShape.with_rank_at_least(1).dims[0])));
}
foreach (Tensor r in results_flat)
{
r.set_shape(new TensorShape(n_static).concatenate(r.dims.Skip(1).ToArray()));
}
// todo get working when the above caching_device is fixed
//if (in_graph_mode && varscope_caching_device_was_none) {
// varscope.set_caching_device(None);
//}
return output_pack(results_flat);
}));
public static TensorShape constant_value_as_shape(Tensor tensor)
{
bool hasattr(Graph property, string attr)
{
var t = property.GetType().GetProperties();
foreach (System.Reflection.PropertyInfo pi in t)
{
if (pi.Name == attr)
{
return(true);
}
}
return(false);
}
if (tensor.GetType() == typeof(EagerTensor))
{
int[] dims = {};
foreach (int dim in tensor.numpy())
{
if (dim != 1)
{
dims[dims.Length] = dim;
}
else
{
// -1 == Unknown
dims[dims.Length] = -1;
}
}
return(new TensorShape(dims));
}
if (tensor.TensorShape.ndim == 0)
{
var value_ = constant_value(tensor);
if (value_ == null)
{
throw new ValueError(
@"Received a scalar with unknown value as shape; require a statically
known scalar with value '-1' to describe an unknown shape.");
}
if (value_ != -1)
{
throw new ValueError(
String.Format(@"Received a scalar value {0} as shape; require a statically known
scalar with value '-1' to describe an unknown shape.", value_));
}
return(tensor.TensorShape.unknown_shape(-1));
}
var shape = tensor.TensorShape.with_rank(1);
if (shape == new TensorShape(new int[] { 1 }))
{
return(new TensorShape(new int[] {}));
}
else if (tensor.op.type == "Cast")
{
var pre_cast = constant_value_as_shape(tensor.op.inputs[0]);
if (pre_cast.dims == null)
{
return(pre_cast);
}
var cast_dtype = dtypes.as_dtype((Type)tensor.op.get_attr("DstT"));
if (!Array.Exists(new [] { dtypes.int32, dtypes.int64 }, cast_dtype_ => cast_dtype_ == cast_dtype))
{
return(tensor.TensorShape.unknown_shape(shape.dims[0]));
}
int[] x_ = {};
foreach (var x in pre_cast.as_list())
{
if (x != -1)
{
x_[x_.Length] = x;
}
else
{
x_[x_.Length] = -1;
}
}
var dest_dtype_shape_array = np.array(x_).astype(cast_dtype.as_numpy_dtype());
int[] y_ = {};
foreach (int y in dest_dtype_shape_array)
{
if (y >= 0)
{
y_[y_.Length] = y;
}
else
{
y_[y_.Length] = -1;
}
}
return(new TensorShape(y_));
}
else if (tensor.op.type == "Shape")
{
return(tensor.op.inputs[0].shape);
}
else if (tensor.op.type == "Pack")
{
var ret_ = new TensorShape(new int[] {});
if ((int)tensor.op.get_attr("axis") != 0)
{
throw new ValueError(String.Format(
@"Since rank 1 inputs are expected, Pack's axis: {0} must be 0, otherwise it
would not be rank 1.", tensor.op.get_attr("axis")));
}
foreach (Tensor pack_input in tensor.op.inputs)
{
var pack_input_val = constant_value(pack_input);
Dimension new_dim;
if (pack_input_val < 0)
{
new_dim = new Dimension(-1);
}
else if (pack_input_val == null)
{
new_dim = new Dimension(-1);
}
else
{
new_dim = new Dimension(pack_input_val);
}
ret_ = ret_.concatenate(new int[] { new_dim });
}
return(ret_);
}
else if (tensor.op.type == "Concat")
{
var ret_ = new TensorShape(new int[] {});
var inputlist_ = new ArraySegment <Tensor>(tensor.op.inputs, 1,
tensor.op.inputs.Length - 1);
foreach (var concat_input in inputlist_)
{
ret_ = ret_.concatenate(constant_value_as_shape(concat_input));
}
return(ret_);
}
else if (tensor.op.type == "StridedSlice")
{
try
{
var begin = constant_value(tensor.op.inputs[1]);
var end = constant_value(tensor.op.inputs[2]);
var strides = constant_value(tensor.op.inputs[3]);
if (new [] { begin, end, strides }.All(x => x == null))
{
begin = begin[0];
end = end[0];
strides = strides[0];
var begin_mask = tensor.op.get_attr("begin_mask");
if ((int)begin_mask == 1)
{
begin = null;
}
var end_mask = tensor.op.get_attr("end_mask");
if ((int)end_mask == 1)
{
end = null;
}
var ellipsis_mask = tensor.op.get_attr("ellipsis_mask");
var new_axis_mask = tensor.op.get_attr("new_axis_mask");
var shrink_axis_mask = tensor.op.get_attr("shrink_axis_mask");
bool valid_attributes;
if (!(bool)ellipsis_mask && !(bool)new_axis_mask &&
!(bool)shrink_axis_mask && !((bool)begin_mask || (int)begin_mask == 1) &&
!((bool)end_mask || (int)end_mask == 1))
{
valid_attributes = true;
}
else
{
valid_attributes = false;
}
if (valid_attributes)
{
// sorry for the mess here, but this hacky solution was the best way
// i could come up with to implement the things done in python in c#
var prev_ = constant_value_as_shape(tensor.op.inputs[0]).dims;
var prev = prev_.Skip(begin).Take(end - begin).ToArray();
// 100 being the comparison doesn't really matter here; it's going to break anyway
for (int iter = 0; iter != 100; iter = iter + strides)
{
prev[prev.Length] = prev_[iter];
if ((iter + strides) > prev_.Length)
{
break;
}
}
var ret_ = new TensorShape(prev);
return(ret_);
}
}
} catch (Exception ex)
{
if (ex is ValueError || ex is TypeError)
{
}
}
}
else if (tensor.op.type == "Placeholder" &&
tensor.op.graph.building_function &&
hasattr(tensor.op.graph, "internal_captures"))
{
int i = 0;
foreach (Tensor capture in tensor.op.graph.internal_captures())
{
if (capture.GetType() == typeof(Tensor))
{
var external_capture = tensor.op.graph.external_captures()[i];
return(constant_value_as_shape(external_capture));
}
i++;
}
}
var ret = tensor.TensorShape.unknown_shape(shape.dims[0]);
var value = constant_value(tensor);
if (value != null)
{
int[] d_ = {};
foreach (int d in value)
{
if (d >= 0)
{
d_[d_.Length] = d;
}
else
{
d_[d_.Length] = -1; // None
}
}
ret = ret.merge_with(new TensorShape(d_));
}
return(ret);
}
public static Tensor scan(
Func <Tensor, Tensor, Tensor> fn,
Tensor elems,
Tensor initializer = null,
int parallel_iterations = 10,
bool back_prop = true,
bool swap_memory = false,
bool infer_shape = true,
bool reverse = false,
string name = null)
{
bool input_is_sequence = nest.is_sequence(elems);
Tensor[] input_flatten(Tensor x) => input_is_sequence?nest.flatten(x).ToArray() : new [] { x };
Tensor input_pack(Tensor[] x) => input_is_sequence ? (Tensor)nest.pack_sequence_as(elems, x) : x[0];
bool output_is_sequence;
Func <Tensor, Tensor[]> output_flatten;
Func <Tensor[], Tensor> output_pack;
if (initializer == null)
{
output_is_sequence = input_is_sequence;
output_flatten = input_flatten;
output_pack = input_pack;
}
else
{
output_is_sequence = nest.is_sequence(initializer);
output_flatten = (x) => output_is_sequence?nest.flatten(x).ToArray() : new [] { x };
output_pack = (x) => output_is_sequence ? (Tensor)nest.pack_sequence_as(initializer, x) : x[0];
}
var elems_flat = input_flatten(elems);
bool in_graph_mode = tf.context.executing_eagerly();
return(tf_with(ops.name_scope(name, "scan", new { elems_flat }), scope =>
{
if (in_graph_mode)
{
// todo tf.net doesn't expose .caching_device
//// Any get_variable calls in fn will cache the first call locally
//// and not issue repeated network I/O requests for each iteration.
//var varscope = variable_scope.get_variable_scope();
//bool varscope_caching_device_was_none = false;
//if (varscope.caching_device = null)
//{
// // varscope.set_caching_device(lambda op: op.device)
// // varscope_caching_device_was_none = True
//}
}
elems_flat = elems_flat.Select(elem => ops.convert_to_tensor(elem, name: "elem")).ToArray();
var n = tensor_shape.dimension_value(elems_flat[0].shape[0]);
// todo python had the below but dimension_value returns int which can't be null
//if (n == null)
//{
// n = array_ops.shape(elems_flat[0])[0];
//}
var elems_ta = elems_flat.Select(elem => new TensorArray(
elem.dtype,
size: tf.constant(n),
dynamic_size: false,
element_shape: elem.shape.Skip(1).ToArray(),
infer_shape: true)).ToList();
for (int index = 0; index < elems_ta.Count; index++)
{
elems_ta[index].unstack(elems_flat[index]);
}
Tensor[] a_flat;
int i;
if (initializer == null)
{
a_flat = elems_ta.Select(elem => elem.read(tf.constant(reverse ? n - 1 : 0))).ToArray();
i = 1;
}
else
{
Tensor[] initializer_flat = output_flatten(initializer);
a_flat = initializer_flat.Select(init => ops.convert_to_tensor(init)).ToArray();
i = 0;
}
var accs_ta = a_flat.Select(init => new TensorArray(
dtype: init.dtype,
size: tf.constant(n),
element_shape: infer_shape ? init.shape : null,
dynamic_size: false,
infer_shape: infer_shape)).ToArray();
if (initializer == null)
{
for (int index = 0; index < accs_ta.Length; index++)
{
accs_ta[index].write(tf.constant(reverse ? n - 1 : 0), a_flat[index]);
}
}
BodyItem compute(BodyItem item)
{
var packed_elems = input_pack(elems_ta.Select(elem_ta => elem_ta.read(item.I)).ToArray());
var packed_a = output_pack(item.A_Flat);
var a_out = fn(packed_a, packed_elems);
var flat_a_out = output_flatten(a_out);
for (int j = 0; j < item.Accs_ta.Length; j++)
{
item.Accs_ta[j].write(item.I, flat_a_out[j]);
}
var next_i = reverse ? item.I - 1 : item.I + 1;
return new BodyItem(next_i, flat_a_out, item.Accs_ta);
}
int initial_i;
Func <BodyItem, Tensor> condition;
if (reverse)
{
initial_i = n - 1 - i;
condition = x => x.I >= 0;
}
else
{
initial_i = i;
condition = x => x.I < n;
}
BodyItem bodyItem =
control_flow_ops.while_loop(
condition,
compute,
new BodyItem(tf.constant(initial_i), a_flat, accs_ta),
parallel_iterations: parallel_iterations,
back_prop: back_prop,
swap_memory: swap_memory,
maximum_iterations: tf.constant(n));
var results_flat = bodyItem.Accs_ta.Select(r => r.stack()).ToArray();
var n_static = new Dimension(tensor_shape.dimension_value(elems_flat[0].TensorShape.with_rank_at_least(1).dims[0]));
foreach (var elem in elems_flat.Skip(1))
{
n_static.merge_with(new Dimension(tensor_shape.dimension_value(elem.TensorShape.with_rank_at_least(1).dims[0])));
}
foreach (Tensor r in results_flat)
{
r.set_shape(new TensorShape(n_static).concatenate(r.dims.Skip(1).ToArray()));
}
// todo get working when the above caching_device is fixed
//if (in_graph_mode && varscope_caching_device_was_none) {
// varscope.set_caching_device(None);
//}
return output_pack(results_flat);
}));
}
