Steve-2040
commented
3 years ago @cgeller if you have exported the model as saved model, then you can also use something like the flowing code for inference: (test images are labeled img0.jpg to img99.jpg in directory test_inf and saved model is in savedmodeldir)
import tensorflow as tf
from PIL import Image, ImageOps
import numpy as np
import os
# os.environ['CUDA_VISIBLE_DEVICES'] = "-1" # uncomment to run on CPU
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
tf.config.experimental.set_memory_growth(gpus[0], True)
print(len(gpus), "Physical GPUs")
imported = tf.saved_model.load('savedmodeldir/')
f = imported.signatures["serving_default"]
def get_image_array(im_id):
im = Image.open(f"test_inf/img{im_id}.jpg")
im_arr = np.frombuffer(im.tobytes(), dtype=np.uint8)
im_arr = im_arr.reshape((1, im.size[1], im.size[0], 3))
return im_arr
for i in range(0, 100):
im_arr = get_image_array(i)
detections = f(tf.constant(im_arr))['detections:0'].numpy()
detections = detections[0] # num detections:100 (Update comment)
for x in range(len(detections)):
y1 = detections[x][1]
x1 = detections[x][2]
y2 = detections[x][3]
x2 = detections[x][4]
pr = int(detections[x][5]*100.0) # probability
cl = int(detections[x][6]) # class_id
if pr > 50:
print(f"img{i}.jpg class_id:{cl} probability%:{pr} y1:{y1} x1:{x1} y2:{y2} x2:{x2}")to get the model inputs and output you use: (from what I have seen, it looks like the saved model always output 100 detections with the highest probability first)
saved_model_cli show --dir savedmodeldir --all
signature_def['serving_default']:
The given SavedModel SignatureDef contains the following input(s):
inputs['image_arrays:0'] tensor_info:
dtype: DT_UINT8
shape: (-1, -1, -1, -1)
name: image_arrays:0
The given SavedModel SignatureDef contains the following output(s):
outputs['detections:0'] tensor_info:
dtype: DT_FLOAT
shape: (1, 100, 7)
name: detections:0
Method name is: tensorflow/serving/predictI also have used TensorFlow Serving for inference using a EfficientDet exported saved model. https://www.tensorflow.org/tfx/guide/serving
Hope this helps.
ghost
aakash665
fitoule
kevinkvothe
achukhrov-ffr-team
mingxingtan
td43
.
Hello,
i trained a model on my own dataset. It worked pretty good so far.
After training my goal is inference on the test set, but according to the README there are two ways:
1) export frozen graph model + "model_inspect.py --runmode=saved_model_infer" 2) direct inference on checkpoint "model_inspect.py --runmode=saved_model_infer"
Both methods give me reasonable detections, BUT: 1) gives me less detections. Only detections with higher score (~>0.5) 2) gives me much more detections with score higher 0.1, which was me specified threshold. So method 2) leads to my expected result.
Why is method 1) different? I need the frozen graph as .pb file for further inference. But with a model which gives me all detections...
Thanks a lot in advance