Added ResetVariableTensor #3298
Conversation
| TfLiteStatus MicroInterpreterGraph::ResetVariableTensor(int tensor_index, | ||
| int subgraph_index) { | ||
| const SubGraph* subgraph = (*subgraphs_)[subgraph_index]; | ||
| auto* tensor = subgraph->tensors()->Get(tensor_index); | ||
| if (tensor->is_variable()) { | ||
| size_t buffer_size; | ||
| TF_LITE_ENSURE_STATUS(TfLiteEvalTensorByteLength( | ||
| &subgraph_allocations_[subgraph_index].tensors[tensor_index], | ||
| &buffer_size)); | ||
|
|
||
| int value = 0; | ||
| if (tensor->type() == tflite::TensorType_INT8) { | ||
| value = tensor->quantization()->zero_point()->Get(0); | ||
| } | ||
| memset(subgraph_allocations_[subgraph_index].tensors[tensor_index].data.raw, | ||
| value, buffer_size); | ||
| } | ||
| return kTfLiteOk; | ||
| } | ||
|
|
There was a problem hiding this comment.
Since this only works for TensorFlow 1.0 variable type tensors, the default return value should be failure. Also TensorType_INT8 should not be used for quantization checks. Instead the existence of the quantization field in the schema should be used. Multi-channel quantization checks also need to be here, or alternatively fail this method.
There was a problem hiding this comment.
Since this only works for TensorFlow 1.0 variable type tensors, the default return value should be failure.
Done.
Also TensorType_INT8 should not be used for quantization checks. Instead the existence of the quantization field in the schema should be used. Multi-channel quantization checks also need to be here, or alternatively fail this method.
This reset logic is identical to the existing ResetVariableTensors. We can improve it in a separate PR if this is needed.
|
@veblush This will not work for RESOURCE variable tensors (TensorFlow 2.0) |
Noted. If we need a way to reset RESOURCE variable, then we can add a new api for that. |
There was a problem hiding this comment.
Only the use of data.data is allowed from the TfLitePtrUnion struct. All other members have been deprecated by the LiteRT team (formerly the TfLite folks).
There was a problem hiding this comment.
Please be advised that TF_LITE_MICRO_EXPECT_ macros do not end the test. This could cause the test to crash on null pointer checks, and the logs for the test will be lost.
Perhaps we should add a new macro:
#define TF_LITE_MICRO_CHECK_FAIL_AND_STOP if (micro_test::did_test_fail) continue
| TF_LITE_ENSURE_STATUS(TfLiteEvalTensorByteLength(eval_tensor, &buffer_size)); | ||
|
|
||
| int value = 0; | ||
| if (tensor->type() == tflite::TensorType_INT8) { |
There was a problem hiding this comment.
You're right, a multi-channel tensor variable probably doesn't exist. But this line would be better as:
if (tensor->type() == tflite::TensorType_INT8 && tensor->quantization() && tensor->quantization()->zero_point() && tensor->quantization->zero_point()->size() > 0) {
3 participants
Brainstroming.
BUG=N/a