Tensorflow implementation for ResNet38 dilated, with weights converted from the Torch implementation.
Weights were ported from mxnet and Torch using the conversion scripts in tools.
| # | Dataset | Source |
|---|---|---|
| 1 | Imagenet | itijyou/ademxapp |
| 2 | Pascal VOC 2012 | jiwoon-ahn/psa/res38_cls.pth |
Loading Imagenet weights:
import tensorflow as tf
from resnet38d import ResNet38d
input_tensor = tf.keras.Input([512, 512, 3], name='inputs')
rn38d = ResNet38d(input_tensor=input_tensor, weights='imagenet', include_top=False)Loading Pascal VOC 2012 weights:
import tensorflow as tf
from resnet38d import ResNet38d
input_tensor = tf.keras.Input([512, 512, 3], name='inputs')
rn38d = ResNet38d(input_tensor=input_tensor, weights='voc2012')For both cases, data must be preprocessed with
tf.keras.applications.imagenet_utils.preprocess_input(x, mode='torch').
In other words:
x = load_data()
x /= 255
x -= tf.convert_to_tensor([0.485, 0.456, 0.406])
x /= tf.convert_to_tensor([0.229, 0.224, 0.225])