75 - Image Segmentation using U-Net - Part 3 (What are trainable parameters?)

This short video tutorial explains the meaning of trainable parameters using a simple example calculation. In summary, trainable parameters are the weights and biases that get updated during the training process.
The code from this video is available at: github.com/bnsreenu/python_fo...

[@hridaanshugusain474]

I know im about 2 years late, but as a high schooler learning neural networks for a research project, this series of videos has helped me so so much where other videos havent. Thank you for your work.

[@DigitalSreeni]

Glad it was helpful!

[@pullrequest1296]

The best tuitor for counting CNN parameters.

[@kernelguardian]

Words cannot describe how much your content has helped me. Your videos should have more views. Keep posting please

[@DigitalSreeni]

Thank you, I will

[@ad_academy]

Awesome way of teaching. Great

[@AshutoshSahuMRM]

You are a great teacher ❤

[@andresbergsneider6644]

@DigitalScreeni I'm a little late in joining the comments sections here. First off, thanks again for sharing this. Is so educational and informative!
I have a question in regards to the initial # of feature maps and been looking to get some guidance/clarification. Where are the 16 initial feature maps coming from? Is it from the original image after running it through 16 different "randomized" filters/kernels? I've seen multiple variations of U-Net and this number varies depending on developers preference. Thanks in advance!

[@sarabruschi7947]

Hello! Thank you so much for this videos that are really helping me with a university project. I'm sorry for the really late question but I would be grateful if you could help me. I'm using UNet for Semantic Segmentation but I need to reduce the number of parameters because the aim of the project is to upload my net in a embedded system. How could I reduce the parameters of UNet? Thank you again!

[@maxinteltech3321]

great videos.tnx

[@DigitalSreeni]

Glad you like them!

[@rajeshwarsehdev2318]

How we calculated 2320 weights? in second layer

[@lakshmitk7227]

Sir can u please help me out i have run code and got executed but I couldn't see prediction and complete image is black. I tried in all possible ways to solve the issue but no use

[@tilkesh]

Thanks

[@Irfankhan-jt9ug]

Great work......which tool creates Image masks?

[@adityanjsg99]

Make Sense ai

[@adityanjsg99]

One more VGG image annotator

[@RethinkerMedia]

What does 16 mean in 256x256x16? Is it like channels?

[@DigitalSreeni]

You can think of them as channels but effectively they are filtered images - variations of your input image.

[@zeeshanpatel6358]

I am not sure why, but my code compiles and shows me that I have 1,879,665 total parameters. There are 0 Non-trainable parameters. Is there a reason why my number is less than yours?

[@DigitalSreeni]

Not sure how to answer without knowing more details. I wouldn't worry about it though.

[@kevinchen3860]

Hello， how can i get the picture of model structure? Appreciate!

[@adityanjsg99]

Use Tensorboard.

[@adityanjsg99]

tf.keras.utils.plot_model(model, show_shapes=True)

[@anthonybalaraju1920]

can u explain how 448 parameters are coming for first layer?
16*9 + 16 ??

[@bhawnahanda4277]

3(channels)*16(features)*9(kernel)+16(bias) = 448

[@DigitalSreeni]

Thanks.

[@anthonybalaraju1920]

@@bhawnahanda4277 Thank you.
but why 3(channels)* ___*..
it should be=> input_units * hidden_units + bias
Input size is 128x128x3
so cant that be something like 128 * 16*9 + 16

[@viniciusceccon7755]

Hi Anthony. It might be a bit late for an answer, but I'll clarify just the same.
The calculation you are making (input_units * hidden_units + bias) refers to FULLY CONNECTED layers, or Dense layers, which are the ones used in Multilayer Perceptron. In these layers, EVERY neuron is CONNECTED to every neuron in adjacent (previous and next) layers. These connections are the weights or parameters of the network.
For U-Net, which uses Convolutions, we limit the weights to a receptive field, which is called the filter (or kernel). The input (say, a 3x3 image section) is ONLY multiplied by the filter parameters (using kernel 3x3, it's 9 parameters).
So, if we have, e.g., 16 filters, our number of parameters is = (size of filter) * (number of filters) + (bias terms). Using RGB images, we train separate filters for each channel, so we also multiply by the number of channels, as pointed out by Bhawna Handa.

[@RethinkerMedia]

I think you got the biases wrong at 2:30

[@DigitalSreeni]

Just watched it, looks fine to me.

[@RethinkerMedia]

@@DigitalSreeni Yeah. Actually, you said, "... 5 biases plus 2 biases" which put me off wrong. It's actually 5 bias connections plus 2 bias connections. Sorry for being a pain. I don't mean to point out mistakes I was confused that's why I asked. :)

[@DigitalSreeni]

@@RethinkerMedia Ahhh. You know how brain operates while recording, talking and coding :) I appreciate when people point out mistakes, gives me a chance to not make the same mistakes again.

[@RethinkerMedia]

@@DigitalSreeni :)

[@vamshisamineniz5905]

what is bias

[@DigitalSreeni]

May be this helps: czcams.com/video/2eQVKZFOHpI/video.html

[@azamatjonmalikov9553]

The best tuitor for counting CNN parameters.

[@DigitalSreeni]

Thanks :)

[@azamatjonmalikov9553]

great videos.tnx

[@DigitalSreeni]

Glad you like them!

[@azamatjonmalikov9553]

The best tuitor for counting CNN parameters.

[@DigitalSreeni]

I'm glad you think so.