How to run inference with dynamic shape model ?

[phamquyhai]

Hi bro, I have a model that supports dynamic shape, but tflite does not support it when converting, but it does support it from the code.

Doc from TFL https://www.tensorflow.org/lite/guide/inference#run_inference_with_dynamic_shape_model

If you want to run a model with dynamic input shape, resize the input shape before running inference. Otherwise, the None shape in Tensorflow models will be replaced by a placeholder of 1 in TFLite models.

The following examples show how to resize the input shape before running inference in different languages. All the examples assume that the input shape is defined as [1/None, 10], and need to be resized to [3, 10].

// Resize input tensors before allocate tensors
interpreter->ResizeInputTensor(/*tensor_index=*/0, std::vector<int>{3,10});
interpreter->AllocateTensors();

How to operate with react-native-fast-tflite

[mrousavy]

Hey - this is currently not supported in Fast TFLite. Shouldn't be too tricky to add though

[mbpictures]

Do you plan to add this or do you have a quick tip to self-implement it? I'm not that familiar with C++ and didn't find where the input tensors are fed into the interpreter/where the tensors are allocated. Thanks already so much!

[mrousavy]

The link from the docs should be a hint on how to get started, I personally don't have any plans to implement this right now unless someone pays me to :)

[mbpictures]

I managed to fixed this using this patch:

diff --git a/node_modules/react-native-fast-tflite/cpp/TensorflowPlugin.cpp b/node_modules/react-native-fast-tflite/cpp/TensorflowPlugin.cpp
index fbdc44f..81372c7 100644
--- a/node_modules/react-native-fast-tflite/cpp/TensorflowPlugin.cpp
+++ b/node_modules/react-native-fast-tflite/cpp/TensorflowPlugin.cpp
@@ -170,13 +171,6 @@ TensorflowPlugin::TensorflowPlugin(TfLiteInterpreter* interpreter, Buffer model,
                                    std::shared_ptr<react::CallInvoker> callInvoker)
     : _interpreter(interpreter), _delegate(delegate), _model(model), _callInvoker(callInvoker) {
   // Allocate memory for the model's input/output `TFLTensor`s.
-  TfLiteStatus status = TfLiteInterpreterAllocateTensors(_interpreter);
-  if (status != kTfLiteOk) {
-    [[unlikely]];
-    throw std::runtime_error("Failed to allocate memory for input/output tensors! Status: " +
-                             tfLiteStatusToString(status));
-  }
-
   log("Successfully created Tensorflow Plugin!");
 }
@@ -213,9 +207,17 @@ void TensorflowPlugin::copyInputBuffers(jsi::Runtime& runtime, jsi::Object input
   }

   for (size_t i = 0; i < count; i++) {
-    TfLiteTensor* tensor = TfLiteInterpreterGetInputTensor(_interpreter, i);
     auto value = array.getValueAtIndex(runtime, i);
     auto inputBuffer = getTypedArray(runtime, value.asObject(runtime));
+    int inputDimensions[] = {static_cast<int>(inputBuffer.length(runtime))};
+    TfLiteInterpreterResizeInputTensor(_interpreter, i, inputDimensions, 1);
+    TfLiteStatus status = TfLiteInterpreterAllocateTensors(_interpreter);
+    if (status != kTfLiteOk) {
+      [[unlikely]];
+      throw std::runtime_error("Failed to allocate memory for input/output tensors! Status: " +
+                               tfLiteStatusToString(status));
+    }
+    TfLiteTensor* tensor = TfLiteInterpreterGetInputTensor(_interpreter, i);
     TensorHelpers::updateTensorFromJSBuffer(runtime, tensor, inputBuffer);
   }
 }
@@ -230,6 +232,7 @@ jsi::Value TensorflowPlugin::copyOutputBuffers(jsi::Runtime& runtime) {
     TensorHelpers::updateJSBufferFromTensor(runtime, *outputBuffer, outputTensor);
     result.setValueAtIndex(runtime, i, *outputBuffer);
   }
+
   return result;
 }

However, this solution has some drawbacks: Only one-dimensional input data for each input tensor is supported, as I wasn't able to determine the dimensions of the inputBuffer.

[mrousavy]

Interesting - yea maybe we can expose this differently by making the tensors in inputs read-write as well? Not sure if that'd be a good API design as it's probably not what users expect:

const model = loadTensorFlowModel(..)
model.inputs[0].size = 4 // resize it

Or somthing like

const model = loadTensorFlowModel(..)
model.resizeInputTensors([4])

[RayanMoarkech]

any updates on this?