Відео

Thank you :). I appreciate the sub!

Thanks so much for the kind words. glad it helped :)

Glad it helped!

You're welcome :)

hi, have you find a solution? because i've same error

Thank you Samit :) I appreciate the kind words.

Hi Gina, sorry for the late reply. I can't remember exactly how long it took to create the TFRecord. It was a few minutes. Your dataset seems large so I would recommend creating the TFRecord on a fast machine and then uploading the TFRecord somewhere that colab can download it from. (The upload file functionality for colab was quite slow last time I checked).

@@aiexpedition5314 Thank you for your answer. My current work-around is to create multiple TFRecord files so that I can run it in parallel on multiple machines. I will do the upload for Colab with the link to Google Drive, which I have only had good experiences with so far. You seem to have trained CNNs quite often. Can you estimate how long the training will take? (600.000 Images from 2 Classes, Transfer learning with a ResNet-50) Also, I'm going to start your Udemy course next week and I'm very excited about it!

@@ginanottenkaemper7960 Downloading from Google Drive works well. 600,000 images of size 2500x3000. That's quite a lot and will probably take days or even weeks to train if you perform a couple of training iterations and use all of the data. However, with transfer learning and only using 2 classes you should start to get good training accuracy even after your network has only seen a couple hundred of examples, so you could stop the training early even after a few minutes. Thanks for considering my course! I hope you learn a lot and have a good learning experience. Feel free to ask if you have any questions. Good luck

​@@aiexpedition5314 Hey Jeremy, Thanks for your quick reply. Hope you are well. Training the network worked perfectly thanks to you. now i'm having trouble evaluating my model. I would like to call the function model.predict() or model.evaluate(). I tried to create a batch of the test dataset (dataset.batch(128)) and pass it to the methods. The code has been running for over an hour with no result and I don't know if I should have done something differently or if I need to wait for longer.

@@ginanottenkaemper7960 glad your training is working. It could be that you just have many images in your testing dataset and it will take a while. Is there any indication of progress while predicting or evaluating? How many images are there in your test set? Something you could try is calling predict() or evaluate() on dataset.take(256).batch(128). This will just run prediction on the first 256 images over 2 batches and should be a lot quicker than running over the entire dataset. Then you can use that to give you an indication of how long the evaluation will take over the entire dataset.

Thank you :). I don't have a video on that just yet. If you want a quick solution you can run the code in the training script which will download and extract the Div2k dataset. Then you can replace the images in the 4 folders: DIV2K_train_HR, DIV2K_train_LR_bicubic, DIV2K_valid_HR and DIV2K_valid_LR_bicubic with the appropriate images and delete the "cache" folder. When you rerun it, it should train on your custom dataset. Hope this helps.

@@aiexpedition5314 thank you so much for your help 🙏🏻 I'm currently doing it 💪🏼

Hi Muhammad, which script are you running? And did you run this line: !pip install tensorflow==2.4.3

I'm run SRGAN script and it got me this error since i am already running on google colab so i Haven't run pip install tenserflow . why it is important to install this colab has already all thing preinstall ..? thanks for responding your video is very helpful for me

this is single video on UA-cam that has done some implementation with logic looking for more in-depth videos on this topic.

Hi @@MuhammadHussain-ub8nd, the problem is that the error is caused by the latest version of TensorFlow which is installed on colab. That line will install a previous version which is compatible with the code. I hope this helps :)

hm thanks do you have any social media platform where we can contact you .

And I think not even TFrecord can really speed up the process also the memory fitting U may not fit it inside to 8GB ram with that large dataset

Thanks Viraj, I will work on it :)

Hi Dang, an epoch in this code is defined as 1000 steps and the batch size is 16 images, so each epoch the model sees 16,000 images. We don't necessarily see all of the images during an epoch - it's possible we could even see the same image multiple times an epoch - I can't remember how many images the training dataset has.

i have confused because in the DIV2K it has only 800 for training and 100 for validating, when you mapping augmentation job it does same size 800-100 after mapping, and you take 16,000 images that's mean you mapping 20 times for the whole dataset to one, i dont see any "length" when load dataset you set up in your code, thank you. if you free, can i have your contact for discuss about my project?

@@DangNguyen-yh5mm The augmentation gets a random crop each time, so the model see's a different part of each of the 800 images.

@@aiexpedition5314 Oh thank you, now i understrand that, but one more question, do you know how they balanced the hyperparameters for each term of loss, for example: ( w1*content_loss + w2*adversarial_loss + w3*MSE_loss ) and in content_loss they rescaled feature maps by multiply for 1/12.75 that's mean W1 == 0.006 (in paper), W2==...., W3=.... how do they do that? in my case there is four term like that.

Thank you :)

Hi Muhammad, thanks for your comment. I'm not sure I understand the question, could you perhaps rephrase it or ask it differently :)

Can I use these algorithms for satellite imagery to get better results?

@@muhammadawon8164 I believe it should improve your results but remember that this will make your images larger, so if you feed your images into another network then it will use more memory and will probably result in slower training

@@aiexpedition5314 This is helpful. Thank you.

Late reply but I think if you want to apply it for satellite imagery, you should integrate other network architectures to SRRESNET like a network named RCAN(Residual channel attention network). You can check this article for further info: www.mdpi.com/2072-4292/14/12/2890. Hope this helps.

Thanks for letting me know. I've found that the problem seems to be that colab uses the latest version of Tensorflow (currently 2.7) which seems to break the code. I've fixed this by adding this line to the top: pip install tensorflow==2.4.3

Same problem even now not fix

@@aiexpedition5314 Seens that the problem persist. Would be a fault on our computer?

I think the problem is that we have python 3.8 instead 3.7, them te problem with that line

@@aiexpedition5314 I cant figure out. But its seens thar I don have thar path (/usr/local/lib/python3.7/dist-packages/keras/saving/hdf5_format.py) in my machine, but even if I create one, just dont work...

Absolutely, you can save the loss values in the train_step method and then visualize them using tensorboard or just by plotting them.

​@@aiexpedition5314 is it reducing along the time,, or just fluctuated up and down? because I've tried to train the model for 100 steps and I notice the loss value changed up and down from one step to another! is that right?

@@rusulamer2892 it should decrease over time. 100 steps is way to small considering there are already supposed to be 1000 steps per epoch. I've just opened the code in colab and changed my runtype to GPU and ran everything as default and I can see the loss decreasing during training. The loss starts off around 5000 at the beginning of training, epoch 1, step 36 it was around 2500, epoch 1 step 65 it was around 1960 and at epoch 1 step 100 it was around 1612. Yours probably isn't training because of this line: training_epochs = training_steps / steps_per_epoch. If your're only using 100 training steps and the steps per epoch is 1000 then you get 0 training epochs...

@@aiexpedition5314 no I've changed the steps per epoch to 10 ,, and I see this fluctuating, is it possible because you are generating a new image that should have different pixels from the original one ? so I think it's possible to have this fluctuation

@@rusulamer2892 hmm I'm not sure to be honest. Does it decrease if you leave it with default settings?

Thank you :)

Hi James, thanks for letting me know. The video is playing on HD on my side but the audio seems a bit poor near the end of the video. Is that what you meant?

Thanks OW :D

Glad it was helpful. It should work fine depending on your use case. You won't be able to zoom in infinitely and see images on the ground from satellite images but it should enhance the quality at least.

Thanks Steve, glad you enjoyed it!

Hi Olom, thanks for the question! I've done a search for this and found the following paper: www.mdpi.com/1424-8220/21/9/2998/pdf where they use LSGAN for image denoising. Unfortunately I didn't see a link to their code. My understanding of LSGAN is that it uses the least squared function for the loss instead of binary cross entropy. If you'd like to use the code in this video, you could switch out the generative part of the loss function for SRGAN. I do add jpeg noise to the images when training but to make it better you should also add other types of noise (maybe Gaussian noise). If you only want to remove noise without changing the scale of the image, then you can just send in the same image for both the low res and high res image, add noise to the high res image and use a scaling factor of 1. Although this will require some code changes.

Thanks Harry :)

Thank you so much 😀

Thanks Zenville :)

Thanks OW. Much appreciated :D

Thanks Ashwinee :)

Thanks Gopolang! I'll definitely do a follow up video explaining how these models work!