Closed SuroshAhmadZobair closed 1 year ago
@SuroshAhmadZobair recently i have not changed back network to use [256,128,64,32] size, so please change it back using #53 , reinstall the package and try it again.
I think it would be nice if you fine-tune the model by retraining it from scratch using the tfrecord, and add your test image in the training loop to visualize it in every epoch using tensorboard
, please view train.py
. So it may show sth different.
The original training set (tfrecord) is quite small like around 200 images, so it does not always yield good result. I have no access to bigger dataset from original author, so thats why the performance might not be as good as the paper suggested .
@SuroshAhmadZobair I am guessing you are loading images with range [0,1] instead of [0,255]. Can you try editing the deploy.py
? with
img = mpimg.imread(config.image)[:, :, :3]
shp = img.shape
img = tf.convert_to_tensor(img, dtype=tf.uint8)
img = tf.image.resize(img, [512, 512])
img = tf.cast(img, dtype=tf.float32)
img = tf.reshape(img, [-1, 512, 512, 3])
if tf.math.reduce_max(img) > 1.0:
img /= 255
Hi
Thanks for this great repo.
I have tried your pre-trained model to post-process and generate the colorized images for my custom test_set. the image_shape also matches the image you have in the demo. Unfortunately, the output image is all black.
which pre-trained model did I use? https://github.com/zcemycl/TF2DeepFloorplan/issues/36
what command did I use?
python -m dfp.deploy --image floorplan.jpg --weight log/store/G --postprocess --colorize --save output.jpg --loadmethod log
Did I try Colab as well? yes, but the result is still the same.
Do I need to fine-tune your pre-trained model on the dataset that you have provided on the repo?
Any help is appreciated. Thanks in advance