TFLite: 输出模型结构，tensors, node 输入数据、预测、得到结果
1. keras 训练时得到的模型结构
...: model.summary()
...:
Model: "mnist_model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
img (InputLayer) [(None, 784)] 0
_________________________________________________________________
dense (Dense) (None, 64) 50240 (64*784 + 64)
_________________________________________________________________
dense_1 (Dense) (None, 64) 4160 (64*64 + 64)
_________________________________________________________________
dense_2 (Dense) (None, 10) 650 (64*10 + 10)
=================================================================
Total params: 55,050
Trainable params: 55,050
Non-trainable params: 0
_________________________________________________________________
h5转换为pb然后通过tensorboard分析
https://github.com/amir-abdi/keras_to_tensorflow/blob/master/keras_to_tensorflow.py
有问题，分析下python代码应该可以解决
通过log打印分析
lite/optional_debug_tools.h lite/optional_debug_tools.cc
// Prints a dump of what tensors and what nodes are in the interpreter.
void PrintInterpreterState(Interpreter* interpreter) {
printf("Interpreter has %zu tensors and %zu nodes\n",
interpreter->tensors_size(), interpreter->nodes_size());
printf("Inputs:");
PrintIntVector(interpreter->inputs());//内容是tensor index组成的vector<int>
printf("Outputs:");
PrintIntVector(interpreter->outputs());
printf("\n");
for (int tensor_index = 0; tensor_index < interpreter->tensors_size();
tensor_index++) {
TfLiteTensor* tensor = interpreter->tensor(tensor_index);
printf("Tensor %3d %-20s %10s %15s %10zu bytes (%4.1f MB) ", tensor_index,
tensor->name, TensorTypeName(tensor->type),
AllocTypeName(tensor->allocation_type), tensor->bytes,
(static_cast<float>(tensor->bytes) / (1 << 20)));
PrintTfLiteIntVector(tensor->dims);
}
printf("\n");
for (int node_index = 0; node_index < interpreter->nodes_size();
node_index++) {
const std::pair<TfLiteNode, TfLiteRegistration>* node_and_reg =
interpreter->node_and_registration(node_index);
const TfLiteNode& node = node_and_reg->first;
const TfLiteRegistration& reg = node_and_reg->second;
if (reg.custom_name != nullptr) {
printf("Node %3d Operator Custom Name %s\n", node_index, reg.custom_name);
} else {
printf("Node %3d Operator Builtin Code %3d\n", node_index, reg.builtin_code);
}
printf(" Inputs:");
PrintTfLiteIntVector(node.inputs);
printf(" Outputs:");
PrintTfLiteIntVector(node.outputs);
}
}
void PrintTfLiteIntVector(const TfLiteIntArray* v) {
if (!v) {
printf(" (null)\n");
return;
}
for (int k = 0; k < v->size; k++) {
printf(" %d", v->data[k]); //data的Index表示轴的索引，每个轴上元素的形状(个数)就是data的内容
}
printf("\n");
}
模型文件包含的内容
In [1]: import tensorflow as tf
...: from tensorflow import keras
...: from tensorflow.keras import layers
...:
...: inputs = keras.Input(shape=(784,), name='img')
...: x = layers.Dense(64, activation='relu')(inputs)
...: x = layers.Dense(64, activation='relu')(x)
...: outputs = layers.Dense(10, activation='softmax')(x)
...:
In [2]: model = keras.Model(inputs=inputs, outputs=outputs, name='mnist_model')
...: model.summary()
Layer (type) Output Shape Param #
=================================================================
img (InputLayer) [(None, 784)] 0
_________________________________________________________________
dense (Dense) (None, 64) 50240 (64*784 + 64)
_________________________________________________________________
dense_1 (Dense) (None, 64) 4160 (64*64 + 64)
_________________________________________________________________
dense_2 (Dense) (None, 10) 650 (64*10 + 10)
Interpreter
Interpreter has 17 tensors and 4 nodes
tensor
Inputs: 1
Outputs: 0
Tensor 0 Identity kTfLiteFloat32 kTfLiteArenaRw 40 bytes ( 0.0 MB) 1 10
Tensor 1 img kTfLiteFloat32 kTfLiteArenaRw 3136 bytes ( 0.0 MB) 1 784
Tensor 2 mnist_model/dense/MatMul/ReadVariableOp/transpose kTfLiteFloat32 kTfLiteMmapRo 200704 bytes ( 0.2 MB) 64 784
Tensor 3 mnist_model/dense/MatMul_bias kTfLiteFloat32 kTfLiteMmapRo 256 bytes ( 0.0 MB) 64
Tensor 4 mnist_model/dense/Relu kTfLiteFloat32 kTfLiteArenaRw 256 bytes ( 0.0 MB) 1 64
Tensor 5 mnist_model/dense_1/MatMul/ReadVariableOp/transpose kTfLiteFloat32 kTfLiteMmapRo 16384 bytes ( 0.0 MB) 64 64
Tensor 6 mnist_model/dense_1/MatMul_bias kTfLiteFloat32 kTfLiteMmapRo 256 bytes ( 0.0 MB) 64
Tensor 7 mnist_model/dense_1/Relu kTfLiteFloat32 kTfLiteArenaRw 256 bytes ( 0.0 MB) 1 64
Tensor 8 mnist_model/dense_2/BiasAdd kTfLiteFloat32 kTfLiteArenaRw 40 bytes ( 0.0 MB) 1 10
Tensor 9 mnist_model/dense_2/MatMul/ReadVariableOp/transpose kTfLiteFloat32 kTfLiteMmapRo 2560 bytes ( 0.0 MB) 10 64
Tensor 10 mnist_model/dense_2/MatMul_bias kTfLiteFloat32 kTfLiteMmapRo 40 bytes ( 0.0 MB) 10
Tensor 11 (null) kTfLiteNoType kTfLiteMemNone 0 bytes ( 0.0 MB) (null)
Tensor 12 (null) kTfLiteNoType kTfLiteMemNone 0 bytes ( 0.0 MB) (null)
Tensor 13 (null) kTfLiteNoType kTfLiteMemNone 0 bytes ( 0.0 MB) (null)
Tensor 14 (null) kTfLiteNoType kTfLiteMemNone 0 bytes ( 0.0 MB) (null)
Tensor 15 (null) kTfLiteNoType kTfLiteMemNone 0 bytes ( 0.0 MB) (null)
Tensor 16 (null) kTfLiteNoType kTfLiteMemNone 0 bytes ( 0.0 MB) (null)
Node
Node 0 Operator Builtin Code 9 FULLY_CONNECTED
Inputs: 1 2 3 : input/weight/bias,为什么有3个inputs
Outputs: 4
Node 1 Operator Builtin Code 9 FULLY_CONNECTED
Inputs: 4 5 6 : input/weight/bias
Outputs: 7
Node 2 Operator Builtin Code 9 FULLY_CONNECTED
Inputs: 7 9 10: input/weight/bias
Outputs: 8
Node 3 Operator Builtin Code 25 SOFTMAX
Inputs: 8
Outputs: 0
怎么给输入赋值然后进行预测
以简单的float feature_test[784] = {};为例
1. 获得input tensor
TfLiteTensor* input = interpreter->tensor(interpreter->inputs()[0]);
2. Tensor的数据保存在哪里？
typedef struct {
// The data type specification for data stored in `data`. This affects
// what member of `data` union should be used.
TfLiteType type;
// A union of data pointers. The appropriate type should be used for a typed
// tensor based on `type`.
TfLitePtrUnion data;
}
// A union of pointers that points to memory for a given tensor.
typedef union {
int* i32;
int64_t* i64;
float* f;
char* raw;
const char* raw_const;
uint8_t* uint8;
bool* b;
int16_t* i16;
TfLiteComplex64* c64;
} TfLitePtrUnion;
2.1 如果保存的数据是uint8
const uint8_t* yes_features_data = g_yes_f2e59fea_nohash_1_data;
for (int i = 0; i < input->bytes; ++i) {
input->data.uint8[i] = yes_features_data[i];
}
2.2 如果保存的数据是float类型
for (int i = 0; i < input->bytes/4; ++i) { //注意下float是4个字节
input->data.f[i] = test_features_data[i];
}
获得推断后的结果
// Read output buffers
// TODO(user): Insert getting data out code.
TfLiteTensor* output = interp->tensor(interp->outputs()[0]);
for (int i=0; i < output->bytes/4; i++) {
printf("%f, ", output->data.f[i]);
}
printf("\n");