tflite model trainiert und quantisiert

.gitignore

@@ -114,3 +114,4 @@ dmypy.json
 runs/
 Test/
 yolo_dataset/
+calib_images/

quantisierung.py

@@ -0,0 +1,42 @@
+import tensorflow as tf
+import numpy as np
+import os
+from PIL import Image
+
+dir = os.path.dirname(os.path.realpath(__file__))
+
+# --- Pfad zum SavedModel ---
+saved_model_dir = dir + "/yolo_training/NAO_detector9/weights/best_saved_model"
+
+# --- Optional: Pfad zu Beispielbildern ---
+image_dir = dir + "/calib_images"  # z. B. 50–100 JPGs aus deinem Datensatz (z. B. NAO-Roboter-Bilder)
+input_size = (320, 320)  # oder (640, 640) je nach deinem Modellinput
+
+# --- Repräsentative Datenfunktion ---
+def representative_data_gen():
+    for filename in os.listdir(image_dir):
+        if filename.endswith(".jpg") or filename.endswith(".png"):
+            img = Image.open(os.path.join(image_dir, filename)).convert("RGB")
+            img = img.resize(input_size)
+            img = np.array(img, dtype=np.float32) / 255.0  # normalisieren falls nötig
+            img = np.expand_dims(img, axis=0)
+            yield [img]
+
+# --- Konverter konfigurieren ---
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
+converter.optimizations = [tf.lite.Optimize.DEFAULT]
+converter.representative_dataset = representative_data_gen
+converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+
+# Je nach Ziel-Hardware (z. B. NAO): INT8 in/out oder UINT8
+converter.inference_input_type = tf.uint8
+converter.inference_output_type = tf.uint8
+
+# --- Konvertieren ---
+quant_model = converter.convert()
+
+# --- Speichern ---
+with open("best_int8.tflite", "wb") as f:
+    f.write(quant_model)
+
+print("✅ INT8-Quantisierung abgeschlossen.")

train_yolo.py

@@ -99,14 +99,8 @@ def train_yolo():
     # OpenVINO Format (optimiert für Intel Hardware)
     model.export(format='openvino', imgsz=320)
 
-    model.export(
-        format="tflite",
-        dynamic=False,
-        inference_type="uint8",
-        quantize=True,
-        imgsz=320,
-        calibration_data=os.path.join(dir, "calib_images")
-    )
+    # tflite Format  
+    model.export(format="tflite",imgsz=320,)
     
     print("\nTraining abgeschlossen. Die Modelle wurden im 'yolo_training/NAO_detector' Ordner gespeichert.")
 

yolo_training/NAO_detector9/args.yaml

@@ -0,0 +1,105 @@
+task: detect
+mode: train
+model: yolov8n.pt
+data: yolo_dataset/dataset.yaml
+epochs: 50
+time: null
+patience: 5
+batch: 16
+imgsz: 320
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: yolo_training
+name: NAO_detector9
+exist_ok: false
+pretrained: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: false
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+half: false
+dnn: false
+plots: true
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+int8: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.1
+box: 7.5
+cls: 0.5
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: botsort.yaml
+save_dir: yolo_training/NAO_detector9

yolo_training/NAO_detector9/results.csv

@@ -0,0 +1,31 @@
+epoch,time,train/box_loss,train/cls_loss,train/dfl_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/dfl_loss,lr/pg0,lr/pg1,lr/pg2
+1,24.5837,1.38337,1.59497,1.13299,0.77381,0.63415,0.79174,0.51402,1.07247,1.59311,1.10924,0.000661064,0.000661064,0.000661064
+2,44.1935,1.37648,1.1011,1.12629,0.47912,0.56098,0.48856,0.29811,1.65133,2.146,1.48548,0.00130144,0.00130144,0.00130144
+3,63.5209,1.37521,1.00512,1.14473,0.75871,0.76707,0.85137,0.51276,1.45374,1.14832,1.33055,0.00191542,0.00191542,0.00191542
+4,83.0181,1.39248,0.97729,1.1521,0.72427,0.90244,0.83962,0.55562,1.1063,1.1655,1.1848,0.0018812,0.0018812,0.0018812
+5,102.099,1.34909,0.9365,1.1294,0.80313,0.82927,0.87456,0.59494,1.10444,0.98597,1.10162,0.0018416,0.0018416,0.0018416
+6,120.495,1.30565,0.85466,1.10696,0.78937,0.90244,0.88324,0.6495,1.01957,0.99015,1.0765,0.001802,0.001802,0.001802
+7,139.265,1.29615,0.83925,1.10297,0.81199,0.84285,0.91349,0.69203,0.9179,0.87651,1.0519,0.0017624,0.0017624,0.0017624
+8,158.55,1.23532,0.78344,1.08741,0.83294,0.8513,0.90222,0.70301,0.93553,0.8022,1.00482,0.0017228,0.0017228,0.0017228
+9,177.315,1.24022,0.77411,1.07758,0.89615,0.80488,0.92216,0.66249,0.98643,0.82604,1.02006,0.0016832,0.0016832,0.0016832
+10,195.778,1.22761,0.76174,1.09093,0.77696,0.9347,0.94549,0.74916,0.89332,0.75543,1.02896,0.0016436,0.0016436,0.0016436
+11,214.175,1.20899,0.74909,1.08335,0.80922,0.87805,0.91758,0.68197,0.89423,0.88144,1.04086,0.001604,0.001604,0.001604
+12,232.891,1.20618,0.74583,1.07317,0.9041,0.90244,0.94701,0.72502,0.86928,0.74318,0.9874,0.0015644,0.0015644,0.0015644
+13,251.636,1.18134,0.71034,1.05533,0.8864,0.95162,0.9707,0.75873,0.84101,0.64098,0.99018,0.0015248,0.0015248,0.0015248
+14,270.239,1.17983,0.70683,1.06267,0.79901,0.90244,0.92739,0.74632,0.74994,0.70678,0.96296,0.0014852,0.0014852,0.0014852
+15,288.895,1.17521,0.69821,1.06029,0.82494,0.87805,0.90711,0.68548,0.81982,0.72279,0.95689,0.0014456,0.0014456,0.0014456
+16,307.27,1.15779,0.68632,1.04943,0.92729,0.95122,0.96588,0.76468,0.82413,0.6346,0.99141,0.001406,0.001406,0.001406
+17,325.407,1.16109,0.6758,1.04592,0.92107,0.95122,0.96739,0.7732,0.7881,0.62989,0.9065,0.0013664,0.0013664,0.0013664
+18,343.35,1.12661,0.66474,1.04166,0.97356,0.97561,0.98827,0.81068,0.69881,0.55234,0.88231,0.0013268,0.0013268,0.0013268
+19,361.278,1.12698,0.64788,1.03084,0.94956,0.92683,0.96542,0.7717,0.7068,0.63099,0.92191,0.0012872,0.0012872,0.0012872
+20,379.226,1.13069,0.65455,1.03885,0.90112,0.97561,0.97418,0.78791,0.70749,0.60267,0.89277,0.0012476,0.0012476,0.0012476
+21,397.629,1.11647,0.63365,1.0299,0.92627,0.97561,0.9702,0.81019,0.67045,0.62768,0.8665,0.001208,0.001208,0.001208
+22,416.662,1.09492,0.62911,1.02721,0.95114,0.94962,0.98024,0.81772,0.69665,0.538,0.87337,0.0011684,0.0011684,0.0011684
+23,435.458,1.11042,0.63233,1.02942,0.91751,0.95122,0.9419,0.80744,0.6747,0.58275,0.86386,0.0011288,0.0011288,0.0011288
+24,453.303,1.09507,0.61695,1.01568,0.97617,0.99934,0.98929,0.84059,0.60503,0.48919,0.84754,0.0010892,0.0010892,0.0010892
+25,472.178,1.0891,0.61445,1.01644,0.92973,1,0.97934,0.8638,0.62973,0.48138,0.85107,0.0010496,0.0010496,0.0010496
+26,491.604,1.08137,0.60952,1.01945,0.93175,0.99902,0.97946,0.83194,0.63044,0.51215,0.87336,0.00101,0.00101,0.00101
+27,510.376,1.07061,0.59402,1.01434,0.9511,0.94891,0.96358,0.83484,0.60126,0.50725,0.86608,0.0009704,0.0009704,0.0009704
+28,528.926,1.07173,0.58656,1.00666,0.85579,0.97561,0.95489,0.8313,0.61699,0.53136,0.8611,0.0009308,0.0009308,0.0009308
+29,547.212,1.0585,0.59106,1.01413,0.90673,1,0.96285,0.80755,0.70054,0.54551,0.89977,0.0008912,0.0008912,0.0008912
+30,565.478,1.04424,0.57113,1.00423,0.88856,0.97245,0.9692,0.85572,0.61174,0.4804,0.83299,0.0008516,0.0008516,0.0008516

yolo_training/NAO_detector9/weights/best_openvino_model/metadata.yaml

@@ -0,0 +1,22 @@
+description: Ultralytics best model trained on yolo_dataset/dataset.yaml
+author: Ultralytics
+date: '2025-07-31T01:20:11.832379'
+version: 8.3.154
+license: AGPL-3.0 License (https://ultralytics.com/license)
+docs: https://docs.ultralytics.com
+stride: 32
+task: detect
+batch: 1
+imgsz:
+- 320
+- 320
+names:
+  0: NAO-Roboter
+args:
+  batch: 1
+  fraction: 1.0
+  half: false
+  int8: false
+  dynamic: false
+  nms: false
+channels: 3

yolo_training/NAO_detector9/weights/best_saved_model/fingerprint.pb

@@ -0,0 +1 @@
+╥═кш°ЙЬ╗s─·С╝©°∙и■
ы═²В├⌠б╠Q У╧■в╣У▀у≥
(⌠ИЯ∙╡жЫьН
2
:5233161973706285970

yolo_training/NAO_detector9/weights/best_saved_model/metadata.yaml

@@ -0,0 +1,21 @@
+description: Ultralytics best model trained on yolo_dataset/dataset.yaml
+author: Ultralytics
+date: '2025-07-31T01:21:01.673821'
+version: 8.3.154
+license: AGPL-3.0 License (https://ultralytics.com/license)
+docs: https://docs.ultralytics.com
+stride: 32
+task: detect
+batch: 1
+imgsz:
+- 320
+- 320
+names:
+  0: NAO-Roboter
+args:
+  batch: 1
+  fraction: 1.0
+  half: false
+  int8: false
+  nms: false
+channels: 3