进行物体检测训练时使用tf-1.5或tf-2.1

Question

当我尝试运行时：

来自tensorflow / models / research /目录

PIPELINE_CONFIG_PATH=/home/paul/objdetect/models/research/object_detection/samples/configs/ssdlite_mobilenet_v3_small_320x320_kitti.config
MODEL_DIR=/home/paul/objdetect/models/research/object_detection/samples/sample_model_ckpt/
NUM_TRAIN_STEPS=5000
SAMPLE_1_OF_N_EVAL_EXAMPLES=1
python object_detection/model_main.py \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --model_dir=${MODEL_DIR} \
    --num_train_steps=${NUM_TRAIN_STEPS} \
    --sample_1_of_n_eval_examples=$SAMPLE_1_OF_N_EVAL_EXAMPLES \
    --alsologtostderr

如果我使用tensorflow-gpu == 1.5.0环境，则会出现以下错误：

    from nets import inception_resnet_v2
  File "/home/paul/objdetect/models/research/slim/nets/inception_resnet_v2.py", line 375, in <module>
    batch_norm_updates_collections=tf.compat.v1.GraphKeys.UPDATE_OPS,
AttributeError: 'module' object has no attribute 'v1'

然后我切换到tensorflow-gpu == 2.1.0环境，然后再次运行它，出现以下错误：

    from object_detection import eval_util
  File "/home/paul/objdetect/models/research/object_detection/eval_util.py", line 40, in <module>
    slim = tf.contrib.slim
AttributeError: module 'tensorflow' has no attribute 'contrib'

那么有没有办法运行它？我需要修复什么才能使其在tf-1.5或tf-2.1中运行？

有人可以帮忙吗？预先非常感谢。

真诚的，

-保罗

Answer 1

tf.contrib在tf2.x中已被弃用。类似地，在1.14.x和1.15.x中引入了tf.compat来解决tf 1.x和tf2.x之间的兼容性。尝试安装tensorflow 1.15.x版本并运行