Excellent explanation of Upsampling vs TransposeConv. Many blogs and channels does not mention that Unet may not have checkerboard effects but you do and also elaborated it.
Other than looking at IoU values do you have any suggestions for evaluating U-net results? I’ll definitely make a video if someone can suggest any methods. I use IoU as the primary metric.
@@DigitalSreeni MIoU, Precision, Recall, ROC-AUC, P-R-AUC, Dice Similarity Coefficient. These are the common ones they used in most published papers and benchmarks. They also use accuracy sometimes. TPR FPR
ROC-AUC is used to compare models and I’ve done a video on the topic. This is used during exploratory phase where you’re trying to compare various models. Researchers who publish papers use them to compare their model against other models. Similarly other parameters are used to compare results from various approaches. If your goal is to explore different models to find which one works for your dataset then you can do all that. But, when you have a dataset to segment your goal is to understand the efficacy of the model in segmenting your dataset and for that IoU is a great metric. I personally do not know how other metrics are offering any better insights compared to IoU. There are 10s of other metrics you can report but you need to consider whether they offer any useful insights about the segmentation. In my experience IoU turned out to be more than adequate. By the way other metrics such as Dice and precision will trend similarly to IoU.
Hi Sreeni, thank you for the video. Just wanted to know whether Upsampling has any learning to do ? Since it is basically the opposite of MaxPooling (If I'm not mistaken) and MaxPooling has no learning parameters.
15:05 - 15:20 why're you saying that the output shape is equal to the input shape? the input shape is 3x3 and the output shape is 6x6. still not clear why it added the last column and bottom row as padding
Hi Sreeni sir, Thanks for the informative video, and I have one doubt that in transfer learning (mainly in deep learning) we are using pre-trained weights and customizing the model for our data/microscopic images. If we have trained our customized model with traditional ML approaches like Xgboost or LightGBM whether it is possible to retrain the ML model/pickles by loading as we are doing in transfer learning (due to drifting problem in the test/input samples at production).
In your previous video on how to predict on large images, I would like to know whether we can do the same for rgb images if the model trained takes a rgb image as input.. If yes, could you please brief what are the changes to be made in the code(like reading image as rgb n what else....)
We are having problem with importing Unet from sm.. Also please mention what to change if mask images are colored. And explain in brief classes : parameter. Thank You
I saw blog articles where people offer their opinions on the best place to put batch normalization in your network but I haven't read any publications on this topic. I am not aware of any methodical work that offers theoretical or empirical proof to suggest the best placement for batch normalization.
Hi sreeni sir,I have a doubt regarding semantic segmentation. Sir i have a dataset with different classes. I have different masks for each class with each mask in each folder. can i do semantic segmentation with a model or should i go for classification. my aim is to do staging of different classes based on their texture and size
Hey, Did you get a solution for this? If so can you share the approach used. I faced this problem few years back and designed very bad approach of training 4 different UNET models for 4 different classes I had. But after looking at Sreeni's videos I understood we could stack masks into single image with different pixel values and create a single multi-class unet segmentation model.
This is awesome stuff as always! I’ve managed to train a UNET on some of my data - how should I best reference your UNET architecture? Is it in a paper somewhere?
There are many folks going through ML/DL stuff but you are the only person who explains fundamentals . . great work . . .
Excellent explanation of Upsampling vs TransposeConv. Many blogs and channels does not mention that Unet may not have checkerboard effects but you do and also elaborated it.
Gorgeous as always! The instances are accessible and the procedure is delineated exquisitely.
Thank you! Cheers!
Your contents are always Amazing…please keep uploading these type of stuff.
My concepts are now cleared. You're the best teacher.
Thanks for making such a detailed explanation and code implementation!
Glad it was helpful!
I wish you can provide a more in-depth lecture about evaluating the results of U-Nets.
Other than looking at IoU values do you have any suggestions for evaluating U-net results? I’ll definitely make a video if someone can suggest any methods. I use IoU as the primary metric.
@@DigitalSreeni MIoU, Precision, Recall, ROC-AUC, P-R-AUC, Dice Similarity Coefficient. These are the common ones they used in most published papers and benchmarks. They also use accuracy sometimes. TPR FPR
ROC-AUC is used to compare models and I’ve done a video on the topic. This is used during exploratory phase where you’re trying to compare various models. Researchers who publish papers use them to compare their model against other models. Similarly other parameters are used to compare results from various approaches. If your goal is to explore different models to find which one works for your dataset then you can do all that. But, when you have a dataset to segment your goal is to understand the efficacy of the model in segmenting your dataset and for that IoU is a great metric. I personally do not know how other metrics are offering any better insights compared to IoU. There are 10s of other metrics you can report but you need to consider whether they offer any useful insights about the segmentation. In my experience IoU turned out to be more than adequate. By the way other metrics such as Dice and precision will trend similarly to IoU.
@@DigitalSreeni ok thank you
Hi Sreeni, thank you for the video. Just wanted to know whether Upsampling has any learning to do ? Since it is basically the opposite of MaxPooling (If I'm not mistaken) and MaxPooling has no learning parameters.
Is it possible to define kernel_initializer so that the output looks just like that of upsampling?
Thanks for making videos on these small small things. It really helps.👍
Glad you like them!
Thank you very much for this clear explanation
You are a god send with your videos and knowledge.
So nice of you :)
15:05 - 15:20
why're you saying that the output shape is equal to the input shape?
the input shape is 3x3 and the output shape is 6x6.
still not clear why it added the last column and bottom row as padding
hey there, did you find the answer pls ?
This was very clear thank you very much
Hi Sreeni sir, Thanks for the informative video, and I have one doubt that in transfer learning (mainly in deep learning) we are using pre-trained weights and customizing the model for our data/microscopic images. If we have trained our customized model with traditional ML approaches like Xgboost or LightGBM whether it is possible to retrain the ML model/pickles by loading as we are doing in transfer learning (due to drifting problem in the test/input samples at production).
In your previous video on how to predict on large images, I would like to know whether we can do the same for rgb images if the model trained takes a rgb image as input.. If yes, could you please brief what are the changes to be made in the code(like reading image as rgb n what else....)
We are having problem with importing Unet from sm..
Also please mention what to change if mask images are colored.
And explain in brief classes : parameter.
Thank You
Great lectures
I found this very helpful
Do you have batch normalization between Conv2D and ReLU when going downwards and upwards? Most sources suggest it
I saw blog articles where people offer their opinions on the best place to put batch normalization in your network but I haven't read any publications on this topic. I am not aware of any methodical work that offers theoretical or empirical proof to suggest the best placement for batch normalization.
Awesome explanation like always
Hi sreeni sir,I have a doubt regarding semantic segmentation. Sir i have a dataset with different classes. I have different masks for each class with each mask in each folder. can i do semantic segmentation with a model or should i go for classification. my aim is to do staging of different classes based on their texture and size
Hey,
Did you get a solution for this? If so can you share the approach used. I faced this problem few years back and designed very bad approach of training 4 different UNET models for 4 different classes I had. But after looking at Sreeni's videos I understood we could stack masks into single image with different pixel values and create a single multi-class unet segmentation model.
i want to do lung ct segmentation using deep learning please guide me
let's do teamworks on kaggle?
keep going. great content. this helped me
Glad to hear it!
This is awesome stuff as always! I’ve managed to train a UNET on some of my data - how should I best reference your UNET architecture? Is it in a paper somewhere?
Thanks
So thank you