quantization

426_YOLOX-Body-Head-Hand/quantization/01_make_calib_data.py

@@ -0,0 +1,24 @@
+import cv2
+import glob
+import numpy as np
+
+H=128
+W=160
+# H=256
+# W=320
+# H=480
+# W=640
+
+files = glob.glob("data/*.jpg")
+img_datas = []
+for idx, file in enumerate(files):
+    bgr_img = cv2.imread(file)
+    resized_img = cv2.resize(bgr_img, (W, H))
+    extend_batch_size_img = resized_img[np.newaxis, :].astype(np.float32)
+    print(
+        f'{str(idx+1).zfill(2)}. extend_batch_size_img.shape: {extend_batch_size_img.shape}'
+    )
+    img_datas.append(extend_batch_size_img)
+calib_datas = np.vstack(img_datas)
+print(f'calib_datas.shape: {calib_datas.shape}')
+np.save(file=f'calibdata_bgr_no_norm_{H}x{W}.npy', arr=calib_datas)

426_YOLOX-Body-Head-Hand/quantization/02_int8.py

@@ -0,0 +1,59 @@
+import numpy as np
+import tensorflow as tf
+
+"""
+onnx2tf -i yolox_ti_body_head_hand_n_1x3x128x160.onnx -coion -osd
+onnx2tf -i yolox_ti_body_head_hand_n_1x3x256x320.onnx -coion -osd
+onnx2tf -i yolox_ti_body_head_hand_n_1x3x480x640.onnx -coion -osd
+"""
+
+RESOLUTIONS = [
+    [128,160],
+    [256,320],
+    [480,640],
+]
+
+def representative_dataset_128x160():
+    images = np.load('calibdata_bgr_no_norm_128x160.npy')
+    for image in images:
+        yield {
+            "input": image[np.newaxis, ...],
+        }
+
+def representative_dataset_256x320():
+    images = np.load('calibdata_bgr_no_norm_256x320.npy')
+    for image in images:
+        yield {
+            "input": image[np.newaxis, ...],
+        }
+
+def representative_dataset_480x640():
+    images = np.load('calibdata_bgr_no_norm_480x640.npy')
+    for image in images:
+        yield {
+            "input": image[np.newaxis, ...],
+        }
+
+
+for H, W in RESOLUTIONS:
+    print(f'@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ {H}x{W}')
+    converter = tf.lite.TFLiteConverter.from_saved_model(f'saved_model_{H}x{W}')
+    converter.optimizations = [tf.lite.Optimize.DEFAULT]
+    if H == 128:
+        converter.representative_dataset = representative_dataset_128x160
+    elif H == 256:
+        converter.representative_dataset = representative_dataset_256x320
+    elif H == 480:
+        converter.representative_dataset = representative_dataset_480x640
+    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+    converter.inference_input_type = tf.uint8
+    # converter.inference_output_type = tf.int8
+    tflite_quant_model = converter.convert()
+    with open(f'saved_model_{H}x{W}/yolox_ti_body_head_hand_n_1x3x{H}x{W}_bgr_uint8.tflite', 'wb') as w:
+        w.write(tflite_quant_model)
+
+"""
+tfliteiorewriter -i saved_model_128x160/yolox_ti_body_head_hand_n_1x3x128x160_bgr_uint8.tflite
+tfliteiorewriter -i saved_model_256x320/yolox_ti_body_head_hand_n_1x3x256x320_bgr_uint8.tflite
+tfliteiorewriter -i saved_model_480x640/yolox_ti_body_head_hand_n_1x3x480x640_bgr_uint8.tflite
+"""

426_YOLOX-Body-Head-Hand/url.txt

@@ -1,4 +1,5 @@
 https://github.com/Megvii-BaseDetection/YOLOX
+https://github.com/TexasInstruments/edgeai-yolox
 https://github.com/Kazuhito00/YOLOX-Colaboratory-Training-Sample
 
 https://github.com/PINTO0309/onnx2tf