| # Copyright 2022-2024 NXP |
| |
| |
| |
| |
| @@ -8,11 +8,15 @@ import torch |
| def post_processing(reg_list, cls_list, num_classes, image_size, feature_map_wh_list, min_boxes, |
| center_variance, size_variance, |
| conf_threshold=0.6, nms_max_output_size=100, nms_iou_threshold=0.3, top_k=100): |
| - reg_list = [tf.keras.layers.Reshape([-1, 4])(reg) for reg in reg_list] |
| - cls_list = [tf.keras.layers.Reshape([-1, num_classes])(cls) for cls in cls_list] |
| |
| - reg = tf.keras.layers.Concatenate(axis=1)(reg_list) |
| - cls = tf.keras.layers.Concatenate(axis=1)(cls_list) |
| + shapes = [3600, 600, 160, 60] |
| + input_shapes = [(30,40),(15,20),(8,10),(4,5)] |
| + |
| + reg_list = [tf.reshape(reg, shape=(shape,4)) for reg,shape, input_shape in zip(reg_list, shapes, input_shapes)] |
| + cls_list = [tf.reshape(cls, shape=(shape, num_classes)) for cls,shape, input_shape in zip(cls_list, shapes, input_shapes)] |
| + |
| + reg = tf.keras.layers.Concatenate(axis=0)(reg_list) |
| + cls = tf.keras.layers.Concatenate(axis=0)(cls_list) |
| |
| # post process |
| cls = tf.keras.layers.Softmax(axis=-1)(cls) |
| @@ -21,23 +25,14 @@ def post_processing(reg_list, cls_list, num_classes, image_size, feature_map_wh_ |
| |
| result = tf.keras.layers.Concatenate(axis=-1)([cls, loc]) |
| |
| - # confidence thresholding |
| - mask = conf_threshold < cls[..., 1] |
| - result = tf.boolean_mask(tensor=result, mask=mask) |
| - |
| # non-maximum suppression |
| - mask = tf.image.non_max_suppression(boxes=result[..., -4:], |
| - scores=result[..., 1], |
| + mask = tf.image.non_max_suppression(boxes=loc, |
| + scores=cls[:,1], |
| max_output_size=nms_max_output_size, |
| iou_threshold=nms_iou_threshold, |
| name='non_maximum_suppresion') |
| result = tf.gather(params=result, indices=mask, axis=0) |
| |
| - # top-k filtering |
| - top_k_value = tf.math.minimum(tf.constant(top_k), tf.shape(result)[0]) |
| - mask = tf.nn.top_k(result[..., 1], k=top_k_value, sorted=True).indices |
| - result = tf.gather(params=result, indices=mask, axis=0) |
| - |
| return result |
| |
| |
| |
| |
| |
| |
| @@ -1,9 +1,9 @@ |
| import argparse |
| import sys |
| |
| -from tf.backend.utils import load_weight |
| -from tf.model.rfb_320 import create_rfb_net |
| -from tf.model.slim_320 import create_slim_net |
| +from backend.utils import load_weight |
| +from model.rfb_320 import create_rfb_net |
| +from model.slim_320 import create_slim_net |
| |
| parser = argparse.ArgumentParser( |
| description='convert model') |
| |
| |
| |
| |
| @@ -1,7 +1,7 @@ |
| import tensorflow as tf |
| |
| -from tf.backend.op import conv_bn, conv_dw, basic_rfb, separable_conv |
| -from tf.backend.utils import post_processing |
| +from backend.op import conv_bn, conv_dw, basic_rfb, separable_conv |
| +from backend.utils import post_processing |
| |
| conf_threshold = 0.6 |
| nms_iou_threshold = 0.3 |
| |
| |
| |
| |
| @@ -1,7 +1,7 @@ |
| import tensorflow as tf |
| |
| -from tf.backend.op import conv_bn, conv_dw, separable_conv |
| -from tf.backend.utils import post_processing |
| +from backend.op import conv_bn, conv_dw, separable_conv |
| +from backend.utils import post_processing |
| |
| conf_threshold = 0.6 |
| nms_iou_threshold = 0.3 |
|
|
