Skip to content

Commit fe4d943

Browse files
hcurEsther2013
hcur
authored and
Esther2013
committed
add constant_value_as_shape
tests will fail until next commit
1 parent 0b71c96 commit fe4d943

File tree

1 file changed

+198
-0
lines changed

1 file changed

+198
-0
lines changed

‎src/TensorFlowNET.Core/Tensors/tensor_util.cs‎

Lines changed: 198 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -216,6 +216,204 @@ public static TensorProto make_tensor_proto(object values, TF_DataType dtype = T
216216
return tensor_proto;
217217
}
218218

219+
public static TensorShape constant_value_as_shape(Tensor tensor)
220+
{
221+
bool hasattr(Graph property, string attr)
222+
{
223+
var t = property.GetType().GetProperties();
224+
foreach (System.Reflection.PropertyInfo pi in t)
225+
{
226+
if (pi.Name == attr)
227+
return true;
228+
}
229+
return false;
230+
}
231+
232+
if (tensor.GetType() == typeof(EagerTensor))
233+
{
234+
int[] dims = {};
235+
foreach (int dim in tensor.numpy())
236+
if (dim != 1)
237+
{
238+
dims[dims.Length] = dim;
239+
} else
240+
{
241+
// -1 == Unknown
242+
dims[dims.Length] = -1;
243+
}
244+
return new TensorShape(dims);
245+
}
246+
247+
if (tensor.TensorShape.ndim == 0)
248+
{
249+
var value_ = constant_value(tensor);
250+
if (value_ == null)
251+
throw new ValueError(
252+
@"Received a scalar with unknown value as shape; require a statically
253+
known scalar with value '-1' to describe an unknown shape.");
254+
if (value_ != -1)
255+
throw new ValueError(
256+
String.Format(@"Received a scalar value {0} as shape; require a statically known
257+
scalar with value '-1' to describe an unknown shape.", value_));
258+
return tensor.TensorShape.unknown_shape(-1);
259+
}
260+
261+
var shape = tensor.TensorShape.with_rank(1);
262+
if (shape == new TensorShape(new int[] {1}))
263+
{
264+
return new TensorShape(new int[] {});
265+
} else if (tensor.op.type == "Cast")
266+
{
267+
var pre_cast = constant_value_as_shape(tensor.op.inputs[0]);
268+
if (pre_cast.dims == null)
269+
return pre_cast;
270+
var cast_dtype = dtypes.as_dtype((Type)tensor.op.get_attr("DstT"));
271+
if (!Array.Exists(new [] {dtypes.int32, dtypes.int64}, cast_dtype_ => cast_dtype_ == cast_dtype))
272+
return tensor.TensorShape.unknown_shape(shape.dims[0]);
273+
274+
int[] x_ = {};
275+
foreach (var x in pre_cast.as_list())
276+
if (x != -1)
277+
x_[x_.Length] = x;
278+
else
279+
x_[x_.Length] = -1;
280+
var dest_dtype_shape_array = np.array(x_).astype(cast_dtype.as_numpy_dtype());
281+
282+
int[] y_ = {};
283+
foreach(int y in dest_dtype_shape_array)
284+
if (y >= 0)
285+
y_[y_.Length] = y;
286+
else
287+
y_[y_.Length] = -1;
288+
return new TensorShape(y_);
289+
} else if (tensor.op.type == "Shape")
290+
{
291+
return tensor.op.inputs[0].shape;
292+
} else if (tensor.op.type == "Pack")
293+
{
294+
var ret_ = new TensorShape(new int[] {});
295+
if ((int)tensor.op.get_attr("axis") != 0)
296+
throw new ValueError(String.Format(
297+
@"Since rank 1 inputs are expected, Pack's axis: {0} must be 0, otherwise it
298+
would not be rank 1.", tensor.op.get_attr("axis")));
299+
foreach (Tensor pack_input in tensor.op.inputs)
300+
{
301+
var pack_input_val = constant_value(pack_input);
302+
Dimension new_dim;
303+
if (pack_input_val < 0)
304+
{
305+
new_dim = new Dimension(-1);
306+
} else if (pack_input_val == null)
307+
{
308+
new_dim = new Dimension(-1);
309+
} else
310+
{
311+
new_dim = new Dimension(pack_input_val);
312+
}
313+
ret_ = ret_.concatenate(new int[] {new_dim});
314+
}
315+
return ret_;
316+
} else if (tensor.op.type == "Concat")
317+
{
318+
var ret_ = new TensorShape(new int[] {});
319+
320+
var inputlist_ = new ArraySegment<Tensor>(tensor.op.inputs, 1,
321+
tensor.op.inputs.Length - 1);
322+
foreach (var concat_input in inputlist_)
323+
{
324+
ret_ = ret_.concatenate(constant_value_as_shape(concat_input));
325+
}
326+
return ret_;
327+
} else if (tensor.op.type == "StridedSlice")
328+
{
329+
try
330+
{
331+
var begin = constant_value(tensor.op.inputs[1]);
332+
var end = constant_value(tensor.op.inputs[2]);
333+
var strides = constant_value(tensor.op.inputs[3]);
334+
if (new [] {begin, end, strides}.All(x => x == null))
335+
{
336+
begin = begin[0];
337+
end = end[0];
338+
strides = strides[0];
339+
var begin_mask = tensor.op.get_attr("begin_mask");
340+
if ((int)begin_mask == 1)
341+
{
342+
begin = null;
343+
}
344+
var end_mask = tensor.op.get_attr("end_mask");
345+
if ((int)end_mask == 1)
346+
{
347+
end = null;
348+
}
349+
350+
var ellipsis_mask = tensor.op.get_attr("ellipsis_mask");
351+
var new_axis_mask = tensor.op.get_attr("new_axis_mask");
352+
var shrink_axis_mask = tensor.op.get_attr("shrink_axis_mask");
353+
354+
bool valid_attributes;
355+
if (!(bool)ellipsis_mask && !(bool)new_axis_mask &&
356+
!(bool)shrink_axis_mask && !((bool)begin_mask || (int)begin_mask == 1) &&
357+
!((bool)end_mask || (int)end_mask == 1))
358+
{
359+
valid_attributes = true;
360+
} else {valid_attributes = false;}
361+
if (valid_attributes)
362+
{
363+
// sorry for the mess here, but this hacky solution was the best way
364+
// i could come up with to implement the things done in python in c#
365+
var prev_ = constant_value_as_shape(tensor.op.inputs[0]).dims;
366+
var prev = prev_.Skip(begin).Take(end - begin).ToArray();
367+
// 100 being the comparison doesn't really matter here; it's going to break anyway
368+
for (int iter = 0; iter != 100; iter = iter + strides)
369+
{
370+
prev[prev.Length] = prev_[iter];
371+
if ((iter + strides) > prev_.Length)
372+
break;
373+
}
374+
var ret_ = new TensorShape(prev);
375+
return ret_;
376+
}
377+
}
378+
} catch (Exception ex)
379+
{
380+
if (ex is ValueError || ex is TypeError) {}
381+
}
382+
} else if (tensor.op.type == "Placeholder" &&
383+
tensor.op.graph.building_function &&
384+
hasattr(tensor.op.graph, "internal_captures"))
385+
{
386+
int i = 0;
387+
foreach (Tensor capture in tensor.op.graph.internal_captures())
388+
{
389+
if (capture.GetType() == typeof(Tensor))
390+
{
391+
var external_capture = tensor.op.graph.external_captures()[i];
392+
return constant_value_as_shape(external_capture);
393+
}
394+
395+
i++;
396+
}
397+
}
398+
399+
var ret = tensor.TensorShape.unknown_shape(shape.dims[0]);
400+
var value = constant_value(tensor);
401+
if (value != null)
402+
{
403+
int[] d_ = {};
404+
foreach (int d in value)
405+
{
406+
if (d >= 0)
407+
d_[d_.Length] = d;
408+
else
409+
d_[d_.Length] = -1; // None
410+
}
411+
ret = ret.merge_with(new TensorShape(d_));
412+
413+
}
414+
return ret;
415+
}
416+
219417
public static NDArray convert_to_numpy_ndarray(object values)
220418
{
221419
NDArray nd;

0 commit comments

Comments
 (0)