我在设备上使用一个.tflite模型,最后一层是ConditionalRandomField层,我需要该层的权重来做预测。最后一层是ConditionalRandomField层,我需要该层的权重来做预测。如何用c++ api获取权重?
Netron或者flatc不能满足我的需求,对设备太重了。
似乎TfLiteNode将权重存储在void* user_data或void* builtin_data中。我如何读取它们?
更新。
结论:.tflite不存储CRF权重,而.h5则存储。(也许是因为它们不影响输出。)
我做了什么。
// obtain from model.
Interpreter *interpreter;
// get the last index of nodes.
// I'm not sure if the index sequence of nodes is the direction which tensors or layers flows.
const TfLiteNode *node = &((interpreter->node_and_registration(interpreter->nodes_size()-1))->first);
// then follow the answer of @yyoon
在TFLite节点中,权重应该存储在 inputs
数组,其中包含相应的 TfLiteTensor*
.
所以,如果你已经获得了 TfLiteNode*
的最后一层,你可以这样做来读取权重值。
TfLiteContext* context; // You would usually have access to this already.
TfLiteNode* node; // <obtain this from the graph>;
for (int i = 0; i < node->inputs->size; ++i) {
TfLiteTensor* input_tensor = GetInput(context, node, i);
// Determine if this is a weight tensor.
// Usually the weights will be memory-mapped read-only tensor
// directly baked in the TFLite model (flatbuffer).
if (input_tensor->allocation_type == kTfLiteMmapRo) {
// Read the values from input_tensor, based on its type.
// For example, if you have float weights,
const float* weights = GetTensorData<float>(input_tensor);
// <read the weight values...>
}
}