kshitizrimal / fast-scnn Goto Github PK

From what I understand, droupout can only benefit the layers after it, since they must predict the same output with a subset of the input. In your implementation, dropout appears after the final conv layer, and is only succeeded by an upsampling and softmax layer (which do not have parameters to learn). It seems more logical to place the dropout layer(s) between the Classification layers i.e at line 110 and 114.

Which value in the code refers to number of classes?? t or s??

dear author,
Great!
After that, how to train and what is data?

the number of trainable parameters is 12,240,761, while the paper said it is 1.11 million, who is wrong?

Hi
I trained the network using your code, with the same augmentation author have been used, but the result is very poor.
did you train your code yourself? what was the result?

Looks like nice work, could you please post your trained weights?

Hello sir.

I am testing your code as semantic segmentation algorithm.
As you mention,
On tensorflow 2.0.0 (beta version).
I run your code.

I met this errors.

I found that this error can occur on tensorflow 2.0.0.
(please check below site)
tensorflow/benchmarks#300

If you don't mind, please send your sample code (of Fast-SCNN on tensorflow 2.0.0) to me.
([email protected])

Hello, This looks great! Any chance for pre-trained weights?