sess控制
TensorFlow默认占满显卡，需要按需启动sess。
1
2
3
4
5
6
7
8
9
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
# allow_soft_placement=True, log_device_placement=True
with tf.Session(config = config) as sess:
do something
代码升级
当遇到较老的tf代码时，需要对其进行升级。 比如caffe-TensorFlow的代码(api 0.8)中:
- 使用python2，不需要conda caffe。要升级
caffepb为caffe_pb2 - File “/home/zbh/Desktop/caffe-tensorflow-master/kaffe/tensorflow/network.py”, line 180, in concat
return tf.concat(
concat_dim=axis, values=inputs, name=name)
新格式 tf.concat(axis=axis, values=inputs, name=name) - File “/home/zbh/Desktop/caffe-tensorflow-master/examples/imagenet/dataset.py”, line 22, in process_image
new_shape =
tf.pack([scale, scale])
新格式tf.stack - File “/home/zbh/Desktop/caffe-tensorflow-master/examples/imagenet/dataset.py”, line 23, in process_image
img = tf.image.resize_images(img,
new_shape[0], new_shape[1]) File “/home/zbh/.local/lib/python2.7/site-packages/tensorflow/python/ops/image_ops_impl.py”, line 808, in resize_images raise ValueError(‘'size' must be a 1-D Tensor of 2 elements: ‘
新格式 tf.image.resize_images(img,(new_shape[0], new_shape[1]))
