Setting up DeepLabV3 in colab - tensorflow

Question

So I am trying to set up deeplab in colab.
I am running:
[1]
from google.colab import drive
drive.mount('/content/drive')
%cd /content/drive/My\ Drive/deeplab_files
[2]
%env PYTHONPATH=/content/drive/My\ Drive/deeplab_files/:/content/drive/My\ Drive/deeplab_files/slim
!echo $PYTHONPATH
[3]
!python deeplab/vis.py \
--logtostderr \
--vis_split="val" \
--model_variant="xception_65" \
--atrous_rates=6 \
--atrous_rates=12 \
--atrous_rates=18 \
--output_stride=16 \
--decoder_output_stride=4 \
--vis_crop_size=360 \
--vis_crop_size=480 \
--dataset="camvid" \
--colormap_type="pascal" \
--checkpoint_dir='/content/drive/My\ Drive/deeplab_files/deeplab/datasets/PQR/exp/train_on_trainval_set/train' \
--vis_logdir='/content/drive/My\ Drive/deeplab_files/deeplab/datasets/PQR/exp/train_on_trainval_set/vis' \
--dataset_dir='/content/drive/My\ Drive/deeplab_files/deeplab/datasets/PQR/tfrecord'
The last command, however, returns
sh: 1: export: Drive/deeplab_files/slim:/content/drive/My Drive/deeplab_files/:/content/drive/My Drive/deeplab_files/slim: bad variable name
Traceback (most recent call last):
File "deeplab/vis.py", line 28, in <module>
from deeplab import common
ModuleNotFoundError: No module named 'deeplab'
Anyone have any idea how I can set up deeplab? I have it set up on my personal machine, but it is much too slow. I uploaded the entire folder to my gdrive.
The odd thing is that I can do
from deeplab import common
from the notebook and that imports successfully

Answer 1

Here is a Github repo containing a Colab notebook running deeplab.
I have not tested it but the way you have uploaded your entire directory to Google Drive is not the right way to run things on Colab.
Think of Colab as a separate machine and you are mounting your Google Drive on this machine. Anything available on your Google Drive is not necessarily available to the Colab machine. You will have to add path of your Google Drive folder (say '\content\drive\My Drive\<path_to_your_folder>') to the sys.path for Colab machine using sys.path.insert(0, <path_of_your_drive_folder>) to make that path available to python environment running on the Colab machine.

Answer 2

Solved mt question. The linked repo that abggcv gave, unfortunately, runs into the same issue this question was citing.
You should clone the repo as normal, and run everything as normal. The only change is that before you run train.py, eval.py, or vis.py you'll need to run the following block:
%cd /root/deeplabvc/models/research/
import sys
sys.path.extend(['/root/deeplabvc/models/research/', '/root/deeplab/models/research/slim/'])
Note that /root/deeplab/ is the path to where I cloned the repo. You'll need to change this if the directory where you cloned the repo is different.
Furthermore, for some reason, you wont be able to run train.py/eval.py/vis.py successively. Even clearing the flags will give you an error about a duplicate flag. To fix this, just restart the runtime (wont lose your files).
Happy segmenting!

Answer 3

Deeplab import error occurs mostly when the PYTHONPATH is not setup properly. The installation instruction given does not work with COLAB environment. The Following has worked for me
%cd /content/deeplab/models/research/
!mkdir -p deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/train
!mkdir -p deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/eval
!mkdir -p deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/vis
!echo ${PYTHONPATH}
%env PATH_TO_TRAIN_DIR=/content/deeplab/models/research/deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/train
%env PATH_TO_DATASET=/content/deeplab/models/research/deeplab/datasets/pascal_voc_seg/tfrecord
%env PYTHONPATH=/content/deeplab/models/research:/content/deeplab/models/research/deeplab:/content/deeplab/models/research/slim:/env/python
!echo ${PYTHONPATH}
Here is my COLAB notebook for Training of deeplab that worked