ranandalon / mtl Goto Github PK

I tried tf 2.7, 2.1 with python 3.8 or 3.6 doesn't work.
Then try tf 1.14 with python 2.7 got stuck at config.colors has no attribute colors
So could you share your environment (or requirements.txt would be even better) such that we can duplicate your environment to run inference.py?
Thank you so much for your help.

Dear author, when realize the model I found that I do not have the package config so I can not run the inference

Hi,

thank you very much for your work. Could you please guide me on how I can convert the network to solve 3 semantic segmentation tasks instead of instance segmentation and depth map estimation?

Thank you for your time.

you do not implement 1/σ^2 and replace with 1/σ (loss_handler.py)?

In most script are the following lines

import config   
import user_config

The problem is there are no such files in the repo! So I don't know how to create them from scratch.
Could someone please share the correct files so I can try running "inference.py" on my computer? Thanks!

Dear author, when realize the model I found that I do not have the package config so I can not run the inference, and also there is a problem should we create a new folder inside the main folder caller 'trained_nets'?

Can you provide the config.py file? @ranandalon. thank you!

Thank you for sharing the code!
I have a question about the config/user_config, do you have another config file?
could you show an example for the config/user_config file?
or give an instruction for how to use the config/user_config module?

Thanks!

hi, the implementation is very attractive! could you share the train process so that I can train my own model? It'll be really helpful.

Does it equal to the multiplication of losses?
https://piccolboni.info/2018/03/a-simple-loss-function-for-multi-task-learning-with-keras-implementation.html

Hello sir,

How do you determine the centers of the instances? Do you just take the geometric center of each instance in your GT instance label?

Hope for reply, thanks.

First This is a great project which produce a nice code for multi task learning, however I notice this multi_task weigh loss has a detail not consistent with origin paper.
In paper, we should use exp for \theta to void this parameter become zero, but this is ignored in this code??? Is tha a bug which cause the result is different with paper??
`class MultiLossLayer():
def init(self, loss_list):
self._loss_list = loss_list
self._sigmas_sq = []
for i in range(len(self._loss_list)):
self.sigmas_sq.append(slim.variable('Sigma_sq' + str(i), dtype=tf.float32, shape=[], initializer=tf.initializers.random_uniform(minval=0.2, maxval=1)))

def get_loss(self):
factor = tf.div(1.0, tf.multiply(2.0, self._sigmas_sq[0]))
loss = tf.add(tf.multiply(factor, self._loss_list[0]), tf.log(self._sigmas_sq[0]))
for i in range(1, len(self._sigmas_sq)):
factor = tf.div(1.0, tf.multiply(2.0, self._sigmas_sq[i]))
loss = tf.add(loss, tf.add(tf.multiply(factor, self._loss_list[i]), tf.log(self._sigmas_sq[i])))
return loss`

Hi,

I am unable to run your code without config file. Can you please share the required files?

Thanks,
Aishwarya