split crop model into backbone and head then predict, the res is not right

[yunshangyue71]

Thank you for your outstanding work. I encountered some problems during my study and would like to ask you for advice.

your is right, split and saved is wrong

model_fold is download. pred is right

model = tf.saved_model.load(os.path.join(model_fold, "metrabs_rn18_y4"))
pred = model.estimate_poses(np.array(frames), boxes=np.array([[0,0,255,255]], dtype=np.float32))

wrong one, preds is img above. model_folder is I first split the crop model into two models: backbone and head. and saved

model_bbone_folder = r"----------\resnet18\bbone"
model_bbone = tf.saved_model.load(model_bbone_folder)

model_head_folder = r"---------------\resnet18\metrab_head"
model_head = tf.saved_model.load(model_head_folder)
feat = model_bbone(frames)
feat = feat.numpy()

feat = feat.astype(np.float16)
pred = model_head(feat, False)

1. split the crop model into two models: backbone and head.

2. save them into two models respectively.

3. two split models, load the overall model you provided, and verify it.

all the abs is 0.0
model_bbone_folder = r"-------------\resnet18\bbone"
model_bbone = tf.saved_model.load(model_bbone_folder)
for i in range(len(model_backbone_vars)):
     string = model_backbone_vars[i].name + "-" + model_make.variables[i].name
     abs =  model_backbone_vars[i].numpy()-model_bbone.variables[i].numpy()
     abs = np.max(abs)
     print(string, abs)

4. loaded separately and inferred.

import cv2
import numpy as np
import tensorflow as tf

imgpath = r"C:\Users\20567\Desktop\ticao2.jpg"
frames = cv2.imread(imgpath)
frames = cv2.cvtColor(frames, cv2.COLOR_BGR2RGB)
frames= cv2.resize(frames, (256, 256))

plot_img = np.copy(frames)

frames = frames/ 255.
frames = frames[np.newaxis, :]
frames = frames.astype(np.float32)

model_bbone_folder = r"....................resnet18\bbone"
model_bbone = tf.saved_model.load(model_bbone_folder)

model_head_folder = r"........................\resnet18\metrab_head"
model_head = tf.saved_model.load(model_head_folder)
feat = model_bbone(frames)
feat = feat.numpy()

feat = feat.astype(np.float16)
pred = model_head(feat, False)

kpt2d = pred[0][0]
for i in range(kpt2d.shape[0]):
cv2.circle(plot_img, (int(kpt2d[i][0]), int(kpt2d[i][1])), 1, (0, 255, 0), -1)

cv2.imshow("", plot_img)
cv2.waitKey()

print(pred)

print("done")