Add description of Nudging logic to FakeQuant documentation.

PiperOrigin-RevId: 238716177

tensorflow/core/api_def/base_api/api_def_FakeQuantWithMinMaxArgs.pbtxt

@@ -8,6 +8,15 @@ when `narrow_range` is false and `[1; 2^num_bits - 1]` when it is true) and
 then de-quantized and output as floats in `[min; max]` interval.
 `num_bits` is the bitwidth of the quantization; between 2 and 16, inclusive.
 
+Before quantization, `min` and `max` values are adjusted with the following
+logic.
+It is suggested to have `min <= 0 <= max`. If `0` is not in the range of values,
+the behavior can be unexpected:
+If `0 < min < max`: `min_adj = 0` and `max_adj = max - min`.
+If `min < max < 0`: `min_adj = min - max` and `max_adj = 0`.
+If `min <= 0 <= max`: `scale = (max - min) / (2^num_bits - 1) `,
+`min_adj = scale * round(min / scale)` and `max_adj = max + min_adj - min`.
+
 Quantization is called fake since the output is still in floating point.
 END
 }

tensorflow/core/api_def/base_api/api_def_FakeQuantWithMinMaxVars.pbtxt

@@ -10,6 +10,15 @@ when `narrow_range` is false and `[1; 2^num_bits - 1]` when it is true) and
 then de-quantized and output as floats in `[min; max]` interval.
 `num_bits` is the bitwidth of the quantization; between 2 and 16, inclusive.
 
+Before quantization, `min` and `max` values are adjusted with the following
+logic.
+It is suggested to have `min <= 0 <= max`. If `0` is not in the range of values,
+the behavior can be unexpected:
+If `0 < min < max`: `min_adj = 0` and `max_adj = max - min`.
+If `min < max < 0`: `min_adj = min - max` and `max_adj = 0`.
+If `min <= 0 <= max`: `scale = (max - min) / (2^num_bits - 1) `,
+`min_adj = scale * round(min / scale)` and `max_adj = max + min_adj - min`.
+
 This operation has a gradient and thus allows for training `min` and `max`
 values.
 END

tensorflow/core/api_def/base_api/api_def_FakeQuantWithMinMaxVarsPerChannel.pbtxt

@@ -11,6 +11,15 @@ when `narrow_range` is false and `[1; 2^num_bits - 1]` when it is true) and
 then de-quantized and output as floats in `[min; max]` interval.
 `num_bits` is the bitwidth of the quantization; between 2 and 16, inclusive.
 
+Before quantization, `min` and `max` values are adjusted with the following
+logic.
+It is suggested to have `min <= 0 <= max`. If `0` is not in the range of values,
+the behavior can be unexpected:
+If `0 < min < max`: `min_adj = 0` and `max_adj = max - min`.
+If `min < max < 0`: `min_adj = min - max` and `max_adj = 0`.
+If `min <= 0 <= max`: `scale = (max - min) / (2^num_bits - 1) `,
+`min_adj = scale * round(min / scale)` and `max_adj = max + min_adj - min`.
+
 This operation has a gradient and thus allows for training `min` and `max`
 values.
 END