Can't run vis.py on the given checkpoints without further training

1. The entire URL of the file you are using

https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/cityscapes.md

2. Describe the bug

I want to run inference on DeepLab model pretrained on Cityscapes, to be more specific, I want to use xception65_cityscapes_trainfine checkpoint which has Xception_65 as Network Backbone, and its pretrained dataset are ImageNet and Cityscapes train_fine set. When I download this checkpoint and give it as the --checkpoint_dir for vis.py script, it doesn't load the model and produce outputs at all. Instead, it waits for a new checkpoint.

3. Steps to reproduce

• Install DeepLab as described here
• Download Cityscapes dataset and converting it to TFRecord using the script mentioned here
• Download the checkpoint file from here.
• Run the script below:

PATH_TO_CHECKPOINT="./deeplab/datasets/cityscapes/init_folder/deeplabv3_cityscapes_train/model.ckpt"
PATH_TO_VIS_DIR="./deeplab/datasets/cityscapes/exp/train_on_train_set/vis"
PATH_TO_DATASET="./deeplab/datasets/cityscapes/tfrecord"

# From tensorflow/models/research/
python deeplab/vis.py \
    --logtostderr \
    --vis_split="val_fine" \
    --model_variant="xception_65" \
    --atrous_rates=6 \
    --atrous_rates=12 \
    --atrous_rates=18 \
    --output_stride=16 \
    --decoder_output_stride=4 \
    --vis_crop_size="1025,2049" \
    --dataset="cityscapes" \
    --colormap_type="cityscapes" \
    --checkpoint_dir=${PATH_TO_CHECKPOINT} \
    --vis_logdir=${PATH_TO_VIS_DIR} \
    --dataset_dir=${PATH_TO_DATASET}

4. Expected behavior

Instead of waiting for a new checkpoint, I'm expecting this script to run inference with the test images on the pretrained network and save output images to ${PATH_TO_VIS_DIR}.

5. Additional context

I tried mobilenet_v2 variant as well, but same thing happened.

The main reason I'm trying to run the model without further training is, even when I try further training, miou I get is 0 for most of the classes, so I started to think that maybe I'm not able to use pretrained weights at all, so I wanted make sure, even without further training I'm getting some results with the pretrained weights, but I can't even run eval.py or vis.py.

6. System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): pip install tensorflow-gpu==1.15.2
• Python version: Python3
• CUDA/cuDNN version: CUDA Version: 11.0 cuDNN 7.5.1
• GPU model and memory: Tesla M60 7618GiB