Thanks for the tutorial. It might be worthwhile to show intermediate results of what different parts do earlier in the video to show exactly what certain code snippets do
In Custom Dataset class one should not add more augmentations or processes. It makes the training very very slow. Do you know any hack to fix this? Here you open the image file, numericalize the text which makes the dataloading process very slow.
Amazing Tutorial. Thanks for it! I am missing the need .unsqueeze(0) for each item in the batch while assigning it to the imgs. Any input on that would be much appreciated. Thanks!
Great video Aladdin. Thanks. I have one question: at the last of the video, sequence lengths seems different. Why they do not equal to [26,32], isn't that a mistake?
Link to the loader_customtext.py file in Alladdin's repo: github.com/aladdinpersson/Machine-Learning-Collection/blob/master/ML/Pytorch/Basics/custom_dataset_txt/loader_customtext.py
I think you can do very similar thing as we did in the video just that we would have to do it for the two languages used in the translation. To make it easier you could also check out torchtext which could make it a lot easier to load datasets
The video is great, really. Just 1 thing that (personally) would make everything literally perfect: could you explain literally everything? Like when you mention transform at 5:50 and you said that you put it as None, explain why etc. As well as for the rest. Basically what you did at 6:40 for the "csv" function explanation. Again, this is only my personal opinion and it would personally help me so much Keep up the great work!
amazing content, learnt so much from your channel...however your coding style is a bit strange....I am no one to judge you...but you code in inverted fashion which makes it difficult to follow the stuff...for ex. you write the function calls first and then you go on defining the classes and methods..so what happens is that the structure of your code is not clear initially....however everything comes together by the end of the code...anyways, thanks for all the efforts that you have put and all the best!
Thanks Alot for your videos it is helping me alot to learn pytorch , I am trying out to build an Image Captitioning model on a Custom Dataset , Your Videos on Image Captitioning will be useful alot :) , Thanks alot again
what if i have two text files , meaning the input to the model is image as well as some text and the output is also text?((Image+text_data)-->(text_data))? Do I have to create two vocabularies??
If we are doing machine translation for example then we would need two different vocabularies one for each language, However if both the input text and output text are in the same language then you can reuse the vocabulary
Hello Sir, Can you explain what are these 'pin memory' and 'collate function' actually means? I did go to documentation. But I didn't understand fully. Can you explain in easier way? That would be helpful. Rest of them were understood very well. Thanks
pin_memory=True should be set as default as is going to speed up the model by pinning the video memory for the model computations but the internals of how that works I'm as clueless as you. The collate function is for additional processing you want to do on the batch you've collected, so in this case we setup how to load all of these captions but when we actually have the batch we need to make sure they are all padded to be of equal number of time steps, this is done using the collate function
Hey Aladdin, is there any advantage of doing this over using let's say the Vocab that torchtext provides? I'm currently exploring PyTorch for a project.
I prefer to build things myself whenever possible so there is no lacking in my understanding but torchtext is great too. Perhaps this is more low level and isn't actually what you would use for larger projects but can be useful for understanding
@@AladdinPersson That's a great mindset to have! I'm actually pretty early in my deep learning journey. Currently just started a small project using BERT. Thanks for the video, will be trying to take away what I got here for my own project :D