SCM tf_run(SCM scm_session, SCM scm_input, SCM scm_output)
{
SCM retval;
if (scm_is_true(scm_list_p(scm_output))) {
struct tf_session_t *session = get_tf_session(scm_session);
int ninputs = scm_ilength(scm_input);
TF_Output *inputs = scm_gc_malloc(sizeof(TF_Output) * ninputs, "tf-run");
TF_Tensor **input_values = scm_gc_malloc(sizeof(TF_Tensor *) * ninputs, "tf-run");
for (int i=0; i<ninputs; i++) {
memcpy(&inputs[i], &get_tf_output(scm_caar(scm_input))->output, sizeof(TF_Output));
input_values[i] = get_tf_tensor(scm_cdar(scm_input))->tensor;
scm_input = scm_cdr(scm_input);
};
int noutputs = scm_ilength(scm_output);
TF_Output *output = scm_gc_malloc(sizeof(TF_Output) * noutputs, "tf-run");
TF_Tensor **output_values = scm_gc_malloc(sizeof(TF_Tensor *) * noutputs, "tf-run");
for (int i=0; i<noutputs; i++) {
output[i] = get_tf_output(scm_car(scm_output))->output;
scm_output = scm_cdr(scm_output);
};
TF_SessionRun(session->session, NULL, inputs, input_values, ninputs, output, output_values, noutputs, NULL, 0, NULL, status());
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("tf-run", TF_Message(_status), SCM_EOL);
retval = SCM_EOL;
for (int i=noutputs-1; i>=0; i--) {
SCM element;
struct tf_tensor_t *result = (struct tf_tensor_t *)scm_gc_calloc(sizeof(struct tf_tensor_t), "make-tensor");
SCM_NEWSMOB(element, tf_tensor_tag, result);
result->tensor = output_values[i];
retval = scm_cons(element, retval);
};
} else
retval = scm_car(tf_run(scm_session, scm_input, scm_list_1(scm_output)));
return retval;
}
SCM tf_add_gradient_(SCM scm_graph, SCM scm_expression, SCM scm_variables)
{
SCM retval;
if (scm_is_true(scm_list_p(scm_variables))) {
struct tf_graph_t *graph = get_tf_graph(scm_graph);
struct tf_output_t *expression = get_tf_output(scm_expression);
int nvariables = scm_ilength(scm_variables);
TF_Output *variables = scm_gc_calloc(sizeof(TF_Output) * nvariables, "tf-add-gradient_");
for (int i=0; i<nvariables; i++) {
variables[i] = get_tf_output(scm_car(scm_variables))->output;
scm_variables = scm_cdr(scm_variables);
};
TF_Output *output = scm_gc_calloc(sizeof(TF_Output) * nvariables, "tf-add-gradient_");
TF_AddGradients(graph->graph, &expression->output, 1, variables, nvariables, NULL, status(), output);
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("tf-add-gradient_", TF_Message(_status), SCM_EOL);
retval = SCM_EOL;
for (int i=nvariables-1; i>=0; i--) {
SCM element;
struct tf_output_t *result = scm_gc_calloc(sizeof(struct tf_output_t), "tf-add-gradient_");
SCM_NEWSMOB(element, tf_output_tag, result);
result->output = output[i];
retval = scm_cons(element, retval);
};
} else
retval = scm_car(tf_add_gradient_(scm_graph, scm_expression, scm_list_1(scm_variables)));
return retval;
}
SCM tf_graph_import_(SCM scm_graph, SCM scm_file_name)
{
struct tf_graph_t *graph = get_tf_graph(scm_graph);
char *file_name = scm_to_locale_string(scm_file_name);
FILE *file = fopen(file_name, "r");
free(file_name);
if (!file)
scm_misc_error("tf-graph-import_", strerror(errno), SCM_EOL);
int fd = fileno(file);
struct stat st;
fstat(fd, &st);
size_t size = st.st_size;
TF_Buffer *buffer = TF_NewBuffer();
void *data = scm_gc_malloc(size, "tf-graph-import_");
fread(data, size, 1, file);
buffer->data = data;
buffer->length = size;
fclose(file);
TF_ImportGraphDefOptions* opts = TF_NewImportGraphDefOptions();
TF_GraphImportGraphDef(graph->graph, buffer, opts, status());
TF_DeleteImportGraphDefOptions(opts);
TF_DeleteBuffer(buffer);
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("tf-graph-import_", TF_Message(_status), SCM_EOL);
return SCM_UNDEFINED;
}
static TF_Operation *add_pad_op(TFModel *tf_model, TF_Operation *input_op, int32_t pad)
{
TF_OperationDescription *op_desc;
TF_Operation *op;
TF_Tensor *tensor;
TF_Output input;
int32_t *pads;
int64_t pads_shape[] = {4, 2};
op_desc = TF_NewOperation(tf_model->graph, "Const", "pads");
TF_SetAttrType(op_desc, "dtype", TF_INT32);
tensor = TF_AllocateTensor(TF_INT32, pads_shape, 2, 4 * 2 * sizeof(int32_t));
pads = (int32_t *)TF_TensorData(tensor);
pads[0] = 0; pads[1] = 0;
pads[2] = pad; pads[3] = pad;
pads[4] = pad; pads[5] = pad;
pads[6] = 0; pads[7] = 0;
TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
op_desc = TF_NewOperation(tf_model->graph, "MirrorPad", "mirror_pad");
input.oper = input_op;
input.index = 0;
TF_AddInput(op_desc, input);
input.oper = op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrType(op_desc, "Tpaddings", TF_INT32);
TF_SetAttrString(op_desc, "mode", "SYMMETRIC", 9);
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
return op;
}
SCM tf_set_attr_tensor(SCM scm_description, SCM scm_name, SCM scm_value)
{
struct tf_description_t *self = get_tf_description(scm_description);
struct tf_tensor_t *value = get_tf_tensor(scm_value);
char *name = scm_to_locale_string(scm_name);
TF_SetAttrTensor(self->description, name, value->tensor, status());
free(name);
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("tf-set-attr-tensor", TF_Message(_status), SCM_EOL);
return SCM_UNDEFINED;
}
SCM make_tf_session(SCM scm_graph)
{
SCM retval;
struct tf_session_t *self = (struct tf_session_t *)scm_gc_calloc(sizeof(struct tf_session_t), "make-tf-session");
SCM_NEWSMOB(retval, tf_session_tag, self);
self->graph = get_tf_graph(scm_graph);
TF_SessionOptions *options = TF_NewSessionOptions();
self->session = TF_NewSession(self->graph->graph, options, status());
TF_DeleteSessionOptions(options);
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("make-tf-session", TF_Message(_status), SCM_EOL);
return retval;
}
SCM make_tensor(SCM scm_type, SCM scm_shape, SCM scm_size, SCM scm_source)
{
SCM retval;
struct tf_tensor_t *self = (struct tf_tensor_t *)scm_gc_calloc(sizeof(struct tf_tensor_t), "make-tensor");
SCM_NEWSMOB(retval, tf_tensor_tag, self);
int type = scm_to_int(scm_type);
int num_dims = scm_to_int(scm_length(scm_shape));
int64_t *dims = scm_gc_malloc_pointerless(sizeof(int64_t) * num_dims, "make-tensor");
int count = 1;
for (int i=0; i<num_dims; i++) {
dims[i] = scm_to_int(scm_car(scm_shape));
count = count * dims[i];
scm_shape = scm_cdr(scm_shape);
};
if (type == TF_STRING) {
SCM* pointer = scm_to_pointer(scm_source);
size_t encoded_size = 0;
for (int i=0; i<count; i++) {
encoded_size += TF_StringEncodedSize(scm_c_string_length(*pointer)) + 8;
pointer++;
};
self->tensor = TF_AllocateTensor(type, dims, num_dims, encoded_size);
int64_t *offsets = TF_TensorData(self->tensor);
int offset = 0;
void *result = offsets + count;
pointer = scm_to_pointer(scm_source);
encoded_size = encoded_size - count * sizeof(int64_t);
for (int i=0; i<count; i++) {
char *str = scm_to_locale_string(*pointer);
int len = TF_StringEncodedSize(scm_c_string_length(*pointer));
*offsets++ = offset;
TF_StringEncode(str, scm_c_string_length(*pointer), result, encoded_size, status());
free(str);
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("make-tensor", TF_Message(_status), SCM_EOL);
offset += len;
encoded_size -= len;
result += len;
pointer++;
};
} else {
self->tensor = TF_AllocateTensor(type, dims, num_dims, scm_to_int(scm_size));
memcpy(TF_TensorData(self->tensor), scm_to_pointer(scm_source), scm_to_int(scm_size));
};
return retval;
}
DNNModel *ff_dnn_load_model_tf(const char *model_filename)
{
DNNModel *model = NULL;
TFModel *tf_model = NULL;
TF_Buffer *graph_def;
TF_ImportGraphDefOptions *graph_opts;
model = av_malloc(sizeof(DNNModel));
if (!model){
return NULL;
}
tf_model = av_malloc(sizeof(TFModel));
if (!tf_model){
av_freep(&model);
return NULL;
}
tf_model->session = NULL;
tf_model->input_tensor = NULL;
tf_model->output_data = NULL;
graph_def = read_graph(model_filename);
if (!graph_def){
av_freep(&tf_model);
av_freep(&model);
return NULL;
}
tf_model->graph = TF_NewGraph();
tf_model->status = TF_NewStatus();
graph_opts = TF_NewImportGraphDefOptions();
TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
TF_DeleteImportGraphDefOptions(graph_opts);
TF_DeleteBuffer(graph_def);
if (TF_GetCode(tf_model->status) != TF_OK){
TF_DeleteGraph(tf_model->graph);
TF_DeleteStatus(tf_model->status);
av_freep(&tf_model);
av_freep(&model);
return NULL;
}
model->model = (void *)tf_model;
model->set_input_output = &set_input_output_tf;
return model;
}
SCM tf_graph_export_(SCM scm_graph, SCM scm_file_name)
{
struct tf_graph_t *graph = get_tf_graph(scm_graph);
TF_Buffer *buffer = TF_NewBuffer();
TF_GraphToGraphDef(graph->graph, buffer, status());
if (TF_GetCode(_status) != TF_OK) {
TF_DeleteBuffer(buffer);
scm_misc_error("tf-graph-export_", TF_Message(_status), SCM_EOL);
};
char *file_name = scm_to_locale_string(scm_file_name);
FILE *file = fopen(file_name, "w");
free(file_name);
if (!file)
scm_misc_error("tf-graph-export_", strerror(errno), SCM_EOL);
fwrite(buffer->data, buffer->length, 1, file);
fclose(file);
TF_DeleteBuffer(buffer);
return SCM_UNDEFINED;
}
SCM tf_finish_operation(SCM scm_description, SCM scm_n_outputs)
{
SCM retval = SCM_EOL;
struct tf_description_t *self = get_tf_description(scm_description);
int n_outputs = scm_to_int(scm_n_outputs);
TF_Operation *operation = TF_FinishOperation(self->description, status());
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("tf-finish-operation", TF_Message(_status), SCM_EOL);
for (int i=n_outputs-1; i>=0; i--) {
SCM element;
struct tf_output_t *output = (struct tf_output_t *)scm_gc_calloc(sizeof(struct tf_output_t), "tf-finish-operation");
SCM_NEWSMOB(element, tf_output_tag, output);
output->output.oper = operation;
output->output.index = i;
retval = scm_cons(element, retval);
};
if (n_outputs == 1)
retval = scm_car(retval);
return retval;
}
static DNNReturnType add_depth_to_space_layer(TFModel *tf_model, TF_Operation **cur_op,
DepthToSpaceParams *params, const int layer)
{
TF_OperationDescription *op_desc;
TF_Output input;
char name_buffer[NAME_BUFFER_SIZE];
snprintf(name_buffer, NAME_BUFFER_SIZE, "depth_to_space%d", layer);
op_desc = TF_NewOperation(tf_model->graph, "DepthToSpace", name_buffer);
input.oper = *cur_op;
input.index = 0;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrInt(op_desc, "block_size", params->block_size);
*cur_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
return DNN_SUCCESS;
}
static TF_Operation *add_const_op(TFModel *tf_model, const float *values, const int64_t *dims, int dims_len, const char *name)
{
int dim;
TF_OperationDescription *op_desc;
TF_Tensor *tensor;
size_t len;
op_desc = TF_NewOperation(tf_model->graph, "Const", name);
TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
len = sizeof(float);
for (dim = 0; dim < dims_len; ++dim){
len *= dims[dim];
}
tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, len);
memcpy(TF_TensorData(tensor), values, len);
TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
return TF_FinishOperation(op_desc, tf_model->status);
}
DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model)
{
TFModel *tf_model = (TFModel *)model->model;
TF_Tensor *output_tensor;
TF_SessionRun(tf_model->session, NULL,
&tf_model->input, &tf_model->input_tensor, 1,
&tf_model->output, &output_tensor, 1,
NULL, 0, NULL, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
else{
memcpy(tf_model->output_data->data, TF_TensorData(output_tensor),
tf_model->output_data->height * tf_model->output_data->width *
tf_model->output_data->channels * sizeof(float));
TF_DeleteTensor(output_tensor);
return DNN_SUCCESS;
}
}
static DNNReturnType load_tf_model(TFModel *tf_model, const char *model_filename)
{
TF_Buffer *graph_def;
TF_ImportGraphDefOptions *graph_opts;
graph_def = read_graph(model_filename);
if (!graph_def){
return DNN_ERROR;
}
tf_model->graph = TF_NewGraph();
tf_model->status = TF_NewStatus();
graph_opts = TF_NewImportGraphDefOptions();
TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
TF_DeleteImportGraphDefOptions(graph_opts);
TF_DeleteBuffer(graph_def);
if (TF_GetCode(tf_model->status) != TF_OK){
TF_DeleteGraph(tf_model->graph);
TF_DeleteStatus(tf_model->status);
return DNN_ERROR;
}
return DNN_SUCCESS;
}
SCM tf_from_tensor(SCM scm_self)
{
struct tf_tensor_t *self = get_tf_tensor(scm_self);
int type = TF_TensorType(self->tensor);
int num_dims = TF_NumDims(self->tensor);
int count = 1;
SCM scm_shape = SCM_EOL;
for (int i=num_dims - 1; i>=0; i--) {
scm_shape = scm_cons(scm_from_int(TF_Dim(self->tensor, i)), scm_shape);
count = count * TF_Dim(self->tensor, i);
};
size_t size = TF_TensorByteSize(self->tensor);
void *data;
if (type == TF_STRING) {
int64_t *offsets = TF_TensorData(self->tensor);
void *pointer = offsets + count;
size_t str_len;
data = scm_gc_malloc(sizeof(SCM) * count, "from-tensor");
SCM *result = data;
for (int i=0; i<count; i++) {
const char *str;
size_t len;
TF_StringDecode(pointer + *offsets, size - *offsets, &str, &len, status());
if (TF_GetCode(_status) != TF_OK)
scm_misc_error("from-tensor", TF_Message(_status), SCM_EOL);
*result++ = scm_from_locale_stringn(str, len);
offsets++;
};
} else {
data = scm_gc_malloc_pointerless(size, "from-tensor");
memcpy(data, TF_TensorData(self->tensor), size);
};
return scm_list_3(scm_from_int(type),
scm_shape,
scm_from_pointer(data, NULL));
}
static DNNReturnType set_input_output_tf(void *model, DNNData *input, DNNData *output)
{
TFModel *tf_model = (TFModel *)model;
int64_t input_dims[] = {1, input->height, input->width, input->channels};
TF_SessionOptions *sess_opts;
const TF_Operation *init_op = TF_GraphOperationByName(tf_model->graph, "init");
TF_Tensor *output_tensor;
// Input operation should be named 'x'
tf_model->input.oper = TF_GraphOperationByName(tf_model->graph, "x");
if (!tf_model->input.oper){
return DNN_ERROR;
}
tf_model->input.index = 0;
if (tf_model->input_tensor){
TF_DeleteTensor(tf_model->input_tensor);
}
tf_model->input_tensor = TF_AllocateTensor(TF_FLOAT, input_dims, 4,
input_dims[1] * input_dims[2] * input_dims[3] * sizeof(float));
if (!tf_model->input_tensor){
return DNN_ERROR;
}
input->data = (float *)TF_TensorData(tf_model->input_tensor);
// Output operation should be named 'y'
tf_model->output.oper = TF_GraphOperationByName(tf_model->graph, "y");
if (!tf_model->output.oper){
return DNN_ERROR;
}
tf_model->output.index = 0;
if (tf_model->session){
TF_CloseSession(tf_model->session, tf_model->status);
TF_DeleteSession(tf_model->session, tf_model->status);
}
sess_opts = TF_NewSessionOptions();
tf_model->session = TF_NewSession(tf_model->graph, sess_opts, tf_model->status);
TF_DeleteSessionOptions(sess_opts);
if (TF_GetCode(tf_model->status) != TF_OK)
{
return DNN_ERROR;
}
// Run initialization operation with name "init" if it is present in graph
if (init_op){
TF_SessionRun(tf_model->session, NULL,
NULL, NULL, 0,
NULL, NULL, 0,
&init_op, 1, NULL, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK)
{
return DNN_ERROR;
}
}
// Execute network to get output height, width and number of channels
TF_SessionRun(tf_model->session, NULL,
&tf_model->input, &tf_model->input_tensor, 1,
&tf_model->output, &output_tensor, 1,
NULL, 0, NULL, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
else{
output->height = TF_Dim(output_tensor, 1);
output->width = TF_Dim(output_tensor, 2);
output->channels = TF_Dim(output_tensor, 3);
output->data = av_malloc(output->height * output->width * output->channels * sizeof(float));
if (!output->data){
return DNN_ERROR;
}
tf_model->output_data = output;
TF_DeleteTensor(output_tensor);
}
return DNN_SUCCESS;
}
static DNNReturnType add_conv_layer(TFModel *tf_model, TF_Operation *transpose_op, TF_Operation **cur_op,
ConvolutionalParams* params, const int layer)
{
TF_Operation *op;
TF_OperationDescription *op_desc;
TF_Output input;
int64_t strides[] = {1, 1, 1, 1};
TF_Tensor *tensor;
int64_t dims[4];
int dims_len;
char name_buffer[NAME_BUFFER_SIZE];
int32_t size;
size = params->input_num * params->output_num * params->kernel_size * params->kernel_size;
input.index = 0;
snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_kernel%d", layer);
op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
dims[0] = params->output_num;
dims[1] = params->kernel_size;
dims[2] = params->kernel_size;
dims[3] = params->input_num;
dims_len = 4;
tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, size * sizeof(float));
memcpy(TF_TensorData(tensor), params->kernel, size * sizeof(float));
TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
snprintf(name_buffer, NAME_BUFFER_SIZE, "transpose%d", layer);
op_desc = TF_NewOperation(tf_model->graph, "Transpose", name_buffer);
input.oper = op;
TF_AddInput(op_desc, input);
input.oper = transpose_op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrType(op_desc, "Tperm", TF_INT32);
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
snprintf(name_buffer, NAME_BUFFER_SIZE, "conv2d%d", layer);
op_desc = TF_NewOperation(tf_model->graph, "Conv2D", name_buffer);
input.oper = *cur_op;
TF_AddInput(op_desc, input);
input.oper = op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrIntList(op_desc, "strides", strides, 4);
TF_SetAttrString(op_desc, "padding", "VALID", 5);
*cur_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
snprintf(name_buffer, NAME_BUFFER_SIZE, "conv_biases%d", layer);
op_desc = TF_NewOperation(tf_model->graph, "Const", name_buffer);
TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
dims[0] = params->output_num;
dims_len = 1;
tensor = TF_AllocateTensor(TF_FLOAT, dims, dims_len, params->output_num * sizeof(float));
memcpy(TF_TensorData(tensor), params->biases, params->output_num * sizeof(float));
TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
snprintf(name_buffer, NAME_BUFFER_SIZE, "bias_add%d", layer);
op_desc = TF_NewOperation(tf_model->graph, "BiasAdd", name_buffer);
input.oper = *cur_op;
TF_AddInput(op_desc, input);
input.oper = op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
*cur_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
snprintf(name_buffer, NAME_BUFFER_SIZE, "activation%d", layer);
switch (params->activation){
case RELU:
op_desc = TF_NewOperation(tf_model->graph, "Relu", name_buffer);
break;
case TANH:
op_desc = TF_NewOperation(tf_model->graph, "Tanh", name_buffer);
break;
case SIGMOID:
op_desc = TF_NewOperation(tf_model->graph, "Sigmoid", name_buffer);
break;
default:
return DNN_ERROR;
}
input.oper = *cur_op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
*cur_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return DNN_ERROR;
}
return DNN_SUCCESS;
}
static TF_Operation* add_conv_layers(TFModel *tf_model, const float **consts, const int64_t **consts_dims,
const int *consts_dims_len, const char **activations,
TF_Operation *input_op, int layers_num)
{
int i;
TF_OperationDescription *op_desc;
TF_Operation *op;
TF_Operation *transpose_op;
TF_Output input;
int64_t strides[] = {1, 1, 1, 1};
int32_t *transpose_perm;
TF_Tensor *tensor;
int64_t transpose_perm_shape[] = {4};
#define NAME_BUFF_SIZE 256
char name_buffer[NAME_BUFF_SIZE];
op_desc = TF_NewOperation(tf_model->graph, "Const", "transpose_perm");
TF_SetAttrType(op_desc, "dtype", TF_INT32);
tensor = TF_AllocateTensor(TF_INT32, transpose_perm_shape, 1, 4 * sizeof(int32_t));
transpose_perm = (int32_t *)TF_TensorData(tensor);
transpose_perm[0] = 1;
transpose_perm[1] = 2;
transpose_perm[2] = 3;
transpose_perm[3] = 0;
TF_SetAttrTensor(op_desc, "value", tensor, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
transpose_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
input.index = 0;
for (i = 0; i < layers_num; ++i){
snprintf(name_buffer, NAME_BUFF_SIZE, "conv_kernel%d", i);
op = add_const_op(tf_model, consts[i << 1], consts_dims[i << 1], consts_dims_len[i << 1], name_buffer);
if (TF_GetCode(tf_model->status) != TF_OK || op == NULL){
return NULL;
}
snprintf(name_buffer, NAME_BUFF_SIZE, "transpose%d", i);
op_desc = TF_NewOperation(tf_model->graph, "Transpose", name_buffer);
input.oper = op;
TF_AddInput(op_desc, input);
input.oper = transpose_op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrType(op_desc, "Tperm", TF_INT32);
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
snprintf(name_buffer, NAME_BUFF_SIZE, "conv2d%d", i);
op_desc = TF_NewOperation(tf_model->graph, "Conv2D", name_buffer);
input.oper = input_op;
TF_AddInput(op_desc, input);
input.oper = op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrIntList(op_desc, "strides", strides, 4);
TF_SetAttrString(op_desc, "padding", "VALID", 5);
input_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
snprintf(name_buffer, NAME_BUFF_SIZE, "conv_biases%d", i);
op = add_const_op(tf_model, consts[(i << 1) + 1], consts_dims[(i << 1) + 1], consts_dims_len[(i << 1) + 1], name_buffer);
if (TF_GetCode(tf_model->status) != TF_OK || op == NULL){
return NULL;
}
snprintf(name_buffer, NAME_BUFF_SIZE, "bias_add%d", i);
op_desc = TF_NewOperation(tf_model->graph, "BiasAdd", name_buffer);
input.oper = input_op;
TF_AddInput(op_desc, input);
input.oper = op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
input_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
snprintf(name_buffer, NAME_BUFF_SIZE, "activation%d", i);
op_desc = TF_NewOperation(tf_model->graph, activations[i], name_buffer);
input.oper = input_op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
input_op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
return NULL;
}
}
return input_op;
}
DNNModel *ff_dnn_load_default_model_tf(DNNDefaultModel model_type)
{
DNNModel *model = NULL;
TFModel *tf_model = NULL;
TF_OperationDescription *op_desc;
TF_Operation *op;
TF_Output input;
static const int64_t input_shape[] = {1, -1, -1, 1};
static const char tanh[] = "Tanh";
static const char sigmoid[] = "Sigmoid";
static const char relu[] = "Relu";
static const float *srcnn_consts[] = {
srcnn_conv1_kernel,
srcnn_conv1_bias,
srcnn_conv2_kernel,
srcnn_conv2_bias,
srcnn_conv3_kernel,
srcnn_conv3_bias
};
static const long int *srcnn_consts_dims[] = {
srcnn_conv1_kernel_dims,
srcnn_conv1_bias_dims,
srcnn_conv2_kernel_dims,
srcnn_conv2_bias_dims,
srcnn_conv3_kernel_dims,
srcnn_conv3_bias_dims
};
static const int srcnn_consts_dims_len[] = {
4,
1,
4,
1,
4,
1
};
static const char *srcnn_activations[] = {
relu,
relu,
relu
};
static const float *espcn_consts[] = {
espcn_conv1_kernel,
espcn_conv1_bias,
espcn_conv2_kernel,
espcn_conv2_bias,
espcn_conv3_kernel,
espcn_conv3_bias
};
static const long int *espcn_consts_dims[] = {
espcn_conv1_kernel_dims,
espcn_conv1_bias_dims,
espcn_conv2_kernel_dims,
espcn_conv2_bias_dims,
espcn_conv3_kernel_dims,
espcn_conv3_bias_dims
};
static const int espcn_consts_dims_len[] = {
4,
1,
4,
1,
4,
1
};
static const char *espcn_activations[] = {
tanh,
tanh,
sigmoid
};
input.index = 0;
model = av_malloc(sizeof(DNNModel));
if (!model){
return NULL;
}
tf_model = av_malloc(sizeof(TFModel));
if (!tf_model){
av_freep(&model);
return NULL;
}
tf_model->session = NULL;
tf_model->input_tensor = NULL;
tf_model->output_data = NULL;
tf_model->graph = TF_NewGraph();
tf_model->status = TF_NewStatus();
#define CLEANUP_ON_ERROR(tf_model, model) { \
TF_DeleteGraph(tf_model->graph); \
TF_DeleteStatus(tf_model->status); \
av_freep(&tf_model); \
av_freep(&model); \
return NULL; \
}
op_desc = TF_NewOperation(tf_model->graph, "Placeholder", "x");
TF_SetAttrType(op_desc, "dtype", TF_FLOAT);
TF_SetAttrShape(op_desc, "shape", input_shape, 4);
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
CLEANUP_ON_ERROR(tf_model, model);
}
switch (model_type){
case DNN_SRCNN:
op = add_pad_op(tf_model, op, 6);
if (!op){
CLEANUP_ON_ERROR(tf_model, model);
}
op = add_conv_layers(tf_model, srcnn_consts,
srcnn_consts_dims, srcnn_consts_dims_len,
srcnn_activations, op, 3);
if (!op){
CLEANUP_ON_ERROR(tf_model, model);
}
break;
case DNN_ESPCN:
op = add_pad_op(tf_model, op, 4);
if (!op){
CLEANUP_ON_ERROR(tf_model, model);
}
op = add_conv_layers(tf_model, espcn_consts,
espcn_consts_dims, espcn_consts_dims_len,
espcn_activations, op, 3);
if (!op){
CLEANUP_ON_ERROR(tf_model, model);
}
op_desc = TF_NewOperation(tf_model->graph, "DepthToSpace", "depth_to_space");
input.oper = op;
TF_AddInput(op_desc, input);
TF_SetAttrType(op_desc, "T", TF_FLOAT);
TF_SetAttrInt(op_desc, "block_size", 2);
op = TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
CLEANUP_ON_ERROR(tf_model, model);
}
break;
default:
CLEANUP_ON_ERROR(tf_model, model);
}
op_desc = TF_NewOperation(tf_model->graph, "Identity", "y");
input.oper = op;
TF_AddInput(op_desc, input);
TF_FinishOperation(op_desc, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK){
CLEANUP_ON_ERROR(tf_model, model);
}
model->model = (void *)tf_model;
model->set_input_output = &set_input_output_tf;
return model;
}
