Différences

Ci-dessous, les différences entre deux révisions de la page.

--- movenet [2021/10/09 06:20] – serge
+++ movenet [2021/12/16 09:07] (Version actuelle) – [Ressources] serge
@@ Ligne 1: / Ligne 1: @@
 ====== MoveNet ======
+=====Ressources=====
   * https://www.tensorflow.org/hub/tutorials/movenet
+  * https://tfhub.dev/s?q=movenet
+=====Optimisation de Tensorflow =====
+Movenet appelle Tensorflow Lite, qui n'est pas optimisé pour GPU. Inutile d'installer Cuda.\\
+Il ne tourne que sur CPU, et demande à avoir AVX2 et FMA.\\
+Pour compiler Tensorflow, voir **[[linux_compiler_tensorflow_avec_bazel|]]**
+Altrnatives, trouver un truc tout fait:
+  * https://www.intel.com/content/www/us/en/developer/articles/guide/optimization-for-tensorflow-installation-guide.html
+=====Exemple=====
+  * https://tfhub.dev/google/lite-model/movenet/singlepose/thunder/3
+<file python movenet_test.py>
+from time import time
+import cv2
+import numpy as np
+import tensorflow as tf
+def draw(frame, keypoints_with_scores):
+    for item in keypoints_with_scores[0][0]:
+        if item[2] > 0.3:  # confiance
+            x = int(item[1]*720)  # numpy array x est le 2ème
+            y = int(item[0]*720)
+            cv2.circle(frame, (x, y), 6, color=(0,0,255), thickness=-1)
+    return frame
+models = [  "lite-model_movenet_singlepose_thunder_3.tflite",
+            "lite-model_movenet_singlepose_thunder_tflite_float16_4.tflite",
+            "lite-model_movenet_singlepose_thunder_tflite_int8_4.tflite"]
+interpreter = tf.lite.Interpreter(model_path=models[0])
+interpreter.allocate_tensors()
+cv2.namedWindow('color', cv2.WND_PROP_FULLSCREEN)
+cam = cv2.VideoCapture(2)
+cam.set(cv2.CAP_PROP_FRAME_WIDTH, 1280)
+cam.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)
+t0 = time()
+nbr = 0
+while 1:
+    nbr += 1
+    ret, frame = cam.read()
+    if not ret:
+        continue
+    # Extraction d'une image carrée [rows, columns]
+    frame = frame[:, 280:1000]
+    image = tf.expand_dims(frame, axis=0)
+    # Resize and pad the image to keep the aspect ratio and fit the expected size.
+    image = tf.image.resize_with_pad(image, 256, 256)
+    # TF Lite format expects tensor type of float32.
+    input_image = tf.cast(image, dtype=tf.float32)
+    input_details = interpreter.get_input_details()
+    output_details = interpreter.get_output_details()
+    interpreter.set_tensor(input_details[0]['index'], input_image.numpy())
+    interpreter.invoke()
+    # Output is a [1, 1, 17, 3] numpy array.
+    keypoints_with_scores = interpreter.get_tensor(output_details[0]['index'])
+    frame = draw(frame, keypoints_with_scores)
+    # Affichage de l'image
+    cv2.imshow('color', frame)
+    # Calcul du FPS, affichage toutes les 10 s
+    if time() - t0 > 10:
+        print("FPS =", round(nbr/10, 2))
+        t0, nbr = time(), 0
+    # Pour quitter
+    if cv2.waitKey(1) == 27:
+        break
+cv2.destroyAllWindows()
+</file>
+=====Des essais amusants=====
+  * **[[https://github.com/sergeLabo/movenet|Essai de movenet par SergeLabo]]**
+{{tag>python sb skeleton tensorflow}}