modules/py_ml: Return tensor references for post-processors.

Converting the output tensors into floats for the prost-processors
causes memory exhaustion when models become very large. Additionally,
it wastes processing time converting values which may not be used. By
moving the conversion step into the post-processors we avoid this issue.

If no callback is passed for post-processing the converted output to
a floating point ndarray is returned still.
This commit is contained in:
Kwabena W. Agyeman 2025-06-14 19:51:06 -07:00
parent ec33fd865b
commit 57c7fc5374
2 changed files with 61 additions and 34 deletions

View File

@ -65,7 +65,7 @@ static size_t py_ml_tuple_sum(mp_obj_tuple_t *o) {
return size; return size;
} }
static size_t pl_ml_dtype_size(char dtype) { static size_t py_ml_dtype_size(char dtype) {
switch (dtype) { switch (dtype) {
case 'f': case 'f':
return 4; return 4;
@ -92,7 +92,7 @@ static void py_ml_process_input(py_ml_model_obj_t *model, mp_obj_t arg) {
if (mp_obj_is_callable(input_arg)) { if (mp_obj_is_callable(input_arg)) {
// Input is a callable. Call the object and pass the tensor buffer and dtype. // Input is a callable. Call the object and pass the tensor buffer and dtype.
mp_obj_t fargs[3] = { mp_obj_t fargs[3] = {
mp_obj_new_bytearray_by_ref(input_size * pl_ml_dtype_size(input_dtype), input_buffer), mp_obj_new_bytearray_by_ref(input_size * py_ml_dtype_size(input_dtype), input_buffer),
MP_OBJ_FROM_PTR(input_shape), MP_OBJ_FROM_PTR(input_shape),
mp_obj_new_int(input_dtype) mp_obj_new_int(input_dtype)
}; };
@ -151,7 +151,7 @@ static void py_ml_process_input(py_ml_model_obj_t *model, mp_obj_t arg) {
} }
} }
static mp_obj_t py_ml_process_output(py_ml_model_obj_t *model) { static mp_obj_t py_ml_process_output(py_ml_model_obj_t *model, bool callback) {
mp_obj_list_t *output_list = MP_OBJ_TO_PTR(mp_obj_new_list(model->outputs_size, NULL)); mp_obj_list_t *output_list = MP_OBJ_TO_PTR(mp_obj_new_list(model->outputs_size, NULL));
for (size_t i = 0; i < model->outputs_size; i++) { for (size_t i = 0; i < model->outputs_size; i++) {
void *model_output = ml_backend_get_output(model, i); void *model_output = ml_backend_get_output(model, i);
@ -172,7 +172,12 @@ static mp_obj_t py_ml_process_output(py_ml_model_obj_t *model) {
shape[ulab_offset + j] = mp_obj_get_int(output_shape->items[j]); shape[ulab_offset + j] = mp_obj_get_int(output_shape->items[j]);
} }
ndarray_obj_t *ndarray = ndarray_new_dense_ndarray(output_shape->len, shape, NDARRAY_FLOAT); ndarray_obj_t *ndarray;
if (callback) {
ndarray = ndarray_new_ndarray(output_shape->len, shape, NULL, output_dtype, model_output);
} else {
ndarray = ndarray_new_dense_ndarray(output_shape->len, shape, NDARRAY_FLOAT);
if (output_dtype == 'f') { if (output_dtype == 'f') {
memcpy(ndarray->array, model_output, size * sizeof(float)); memcpy(ndarray->array, model_output, size * sizeof(float));
@ -197,6 +202,8 @@ static mp_obj_t py_ml_process_output(py_ml_model_obj_t *model) {
((float *) ndarray->array)[j] = v * output_scale; ((float *) ndarray->array)[j] = v * output_scale;
} }
} }
}
output_list->items[i] = MP_OBJ_FROM_PTR(ndarray); output_list->items[i] = MP_OBJ_FROM_PTR(ndarray);
} }
@ -254,6 +261,8 @@ static mp_obj_t py_ml_model_predict(size_t n_args, const mp_obj_t *pos_args, mp_
mp_raise_msg(&mp_type_ValueError, MP_ERROR_TEXT("Unsupported input type. Expected a list")); mp_raise_msg(&mp_type_ValueError, MP_ERROR_TEXT("Unsupported input type. Expected a list"));
} }
bool callback = args[ARG_callback].u_obj != mp_const_none;
OMV_PROFILE_START(preprocess); OMV_PROFILE_START(preprocess);
py_ml_process_input(model, pos_args[1]); py_ml_process_input(model, pos_args[1]);
OMV_PROFILE_PRINT(preprocess); OMV_PROFILE_PRINT(preprocess);
@ -263,10 +272,10 @@ static mp_obj_t py_ml_model_predict(size_t n_args, const mp_obj_t *pos_args, mp_
OMV_PROFILE_PRINT(inference); OMV_PROFILE_PRINT(inference);
OMV_PROFILE_START(postprocess); OMV_PROFILE_START(postprocess);
mp_obj_t output = py_ml_process_output(model); mp_obj_t output = py_ml_process_output(model, callback);
OMV_PROFILE_PRINT(postprocess); OMV_PROFILE_PRINT(postprocess);
if (args[ARG_callback].u_obj != mp_const_none) { if (callback) {
// Pass model, inputs, outputs to the post-processing callback. // Pass model, inputs, outputs to the post-processing callback.
mp_obj_t fargs[3] = { MP_OBJ_FROM_PTR(model), pos_args[1], output }; mp_obj_t fargs[3] = { MP_OBJ_FROM_PTR(model), pos_args[1], output };
OMV_PROFILE_START(postprocess_callback); OMV_PROFILE_START(postprocess_callback);

View File

@ -36,6 +36,12 @@ from ulab import numpy as np
_NO_DETECTION = const(()) _NO_DETECTION = const(())
def dequantize(value, dtype, zero_point, scale):
if dtype == 'f':
return value
return (value - zero_point) * scale
# FOMO generates an image per class, where each pixel represents the centroid # FOMO generates an image per class, where each pixel represents the centroid
# of the trained object. These images are processed with `find_blobs()` to # of the trained object. These images are processed with `find_blobs()` to
# extract centroids, and `get_stats()` is used to get their scores. Overlapping # extract centroids, and `get_stats()` is used to get their scores. Overlapping
@ -48,9 +54,12 @@ class fomo_postprocess:
def __call__(self, model, inputs, outputs): def __call__(self, model, inputs, outputs):
n, oh, ow, oc = model.output_shape[0] n, oh, ow, oc = model.output_shape[0]
s = model.output_scale[0]
zp = model.output_zero_point[0]
dt = model.output_dtype[0]
nms = NMS(ow, oh, inputs[0].roi) nms = NMS(ow, oh, inputs[0].roi)
for i in range(oc): for i in range(oc):
img = image.Image(outputs[0][0, :, :, i] * 255) img = image.Image(dequantize(outputs[0][0, :, :, i], dt, zp, s) * 255)
blobs = img.find_blobs( blobs = img.find_blobs(
self.threshold_list, x_stride=1, area_threshold=1, pixels_threshold=1 self.threshold_list, x_stride=1, area_threshold=1, pixels_threshold=1
) )
@ -89,6 +98,9 @@ class yolo_v2_postprocess:
def __call__(self, model, inputs, outputs): def __call__(self, model, inputs, outputs):
ob, oh, ow, oc = model.output_shape[0] ob, oh, ow, oc = model.output_shape[0]
s = model.output_scale[0]
zp = model.output_zero_point[0]
dt = model.output_dtype[0]
class_count = (oc // self.anchors_len) - _YOLO_V2_CLASSES class_count = (oc // self.anchors_len) - _YOLO_V2_CLASSES
def sigmoid(x): def sigmoid(x):
@ -106,13 +118,13 @@ class yolo_v2_postprocess:
_YOLO_V2_CLASSES + class_count)) _YOLO_V2_CLASSES + class_count))
# Threshold all the scores # Threshold all the scores
score_indices = sigmoid(row_outputs[:, _YOLO_V2_SCORE]) score_indices = sigmoid(dequantize(row_outputs[:, _YOLO_V2_SCORE], dt, zp, s))
score_indices = np.nonzero(score_indices > self.threshold)[0] score_indices = np.nonzero(score_indices > self.threshold)[0]
if not len(score_indices): if not len(score_indices):
return _NO_DETECTION return _NO_DETECTION
# Get the bounding boxes that have a valid score # Get the bounding boxes that have a valid score
bb = np.take(row_outputs, score_indices, axis=0) bb = dequantize(np.take(row_outputs, score_indices, axis=0), dt, zp, s)
# Extract rows, columns, and anchor indices # Extract rows, columns, and anchor indices
bb_rows = score_indices // (ow * self.anchors_len) bb_rows = score_indices // (ow * self.anchors_len)
@ -179,19 +191,22 @@ class yolo_v5_postprocess:
def __call__(self, model, inputs, outputs): def __call__(self, model, inputs, outputs):
oh, ow, oc = model.output_shape[0] oh, ow, oc = model.output_shape[0]
s = model.output_scale[0]
zp = model.output_zero_point[0]
dt = model.output_dtype[0]
class_count = oc - _YOLO_V5_CLASSES class_count = oc - _YOLO_V5_CLASSES
# Reshape the output to a 2D array # Reshape the output to a 2D array
row_outputs = outputs[0].reshape((oh * ow, _YOLO_V5_CLASSES + class_count)) row_outputs = outputs[0].reshape((oh * ow, _YOLO_V5_CLASSES + class_count))
# Threshold all the scores # Threshold all the scores
score_indices = row_outputs[:, _YOLO_V5_SCORE] score_indices = dequantize(row_outputs[:, _YOLO_V5_SCORE], dt, zp, s)
score_indices = np.nonzero(score_indices > self.threshold)[0] score_indices = np.nonzero(score_indices > self.threshold)[0]
if not len(score_indices): if not len(score_indices):
return _NO_DETECTION return _NO_DETECTION
# Get the bounding boxes that have a valid score # Get the bounding boxes that have a valid score
bb = np.take(row_outputs, score_indices, axis=0) bb = dequantize(np.take(row_outputs, score_indices, axis=0), dt, zp, s)
# Get the score information # Get the score information
bb_scores = bb[:, _YOLO_V5_SCORE] bb_scores = bb[:, _YOLO_V5_SCORE]
@ -233,19 +248,22 @@ class yolo_v8_postprocess:
def __call__(self, model, inputs, outputs): def __call__(self, model, inputs, outputs):
oh, ow, oc = model.output_shape[0] oh, ow, oc = model.output_shape[0]
s = model.output_scale[0]
zp = model.output_zero_point[0]
dt = model.output_dtype[0]
class_count = ow - _YOLO_V8_CLASSES class_count = ow - _YOLO_V8_CLASSES
# Reshape the output to a 2D array # Reshape the output to a 2D array
column_outputs = outputs[0].reshape((oh * (_YOLO_V8_CLASSES + class_count), oc)) column_outputs = outputs[0].reshape((oh * (_YOLO_V8_CLASSES + class_count), oc))
# Threshold all the scores # Threshold all the scores
score_indices = np.max(column_outputs[_YOLO_V8_CLASSES:, :], axis=0) score_indices = np.max(dequantize(column_outputs[_YOLO_V8_CLASSES:, :], dt, zp, s), axis=0)
score_indices = np.nonzero(score_indices > self.threshold)[0] score_indices = np.nonzero(score_indices > self.threshold)[0]
if not len(score_indices): if not len(score_indices):
return _NO_DETECTION return _NO_DETECTION
# Get the bounding boxes that have a valid score # Get the bounding boxes that have a valid score
bb = np.take(column_outputs, score_indices, axis=1) bb = dequantize(np.take(column_outputs, score_indices, axis=1), dt, zp, s)
# Get the score information # Get the score information
bb_scores = np.max(bb[_YOLO_V8_CLASSES:, :], axis=0) bb_scores = np.max(bb[_YOLO_V8_CLASSES:, :], axis=0)