Lecture 2: Working with Images & Logistic Regression

Deep Learning with PyTorch: Zero to GANs Lecture 2 - Working with Images & Logistic Regress
#1

Session Links
English: https://youtu.be/uuuzvOEC0zw
Hindi: https://youtu.be/qmzgJvVP9XA

Lecture Date and Time: November 28, 2020,
English: 9 PM IST/8.30 AM PST | Add to calendar (Google)
Hindi: 11 AM IST | Add to calendar (Google)

Notebooks :

  1. Logistic Regression: https://jovian.ai/aakashns/03-logistic-regression
  2. Logistic Regression minimal starter: https://jovian.ai/aakashns/mnist-logistic-minimal
  3. Linear Regression minimal starter: https://jovian.ai/aakashns/housing-linear-minimal

What to do after the lecture?

  • Run the Jupyter notebooks shared above (try other datasets)
  • Ask and answer questions on this topic (scroll down)
  • Start working on Assignment 2 - Train Your First Model

Asking/Answering Questions:

Reply on this thread to ask questions during and after the lecture. Before asking, scroll through the thread and check if your question (or a similar one) is already present. If yes, just like it. During the lecture, we’ll answer 8-10 questions with the most likes. The rest will be answered on the forum. If you see a question you know the answer to, please post your answer as a reply to that question. Let’s help each other learn!

Assignment 2: Train Your First Model
#2
#3

too fast for beginner … but that’s ok … I need to refresh maths and revisit session multiple times :slight_smile:

5 Likes
#4

Is there a method to choosebatch size?

#5

what are things to note when we do for other image datsets

1 Like
#6

how to do lebeling for unlabel image datasets

1 Like
#7

How can we change the learning rate at the point where the model’s accuracy starts becoming constant? Something like adaptive learning rate?

2 Likes
#8

Why during evaluation time we don’t call model.eval() and torch.no_grad?

#9

Helo aakash…
Can I kindly have the recorded video for lecture 2…??

1 Like
#10

@hemanth, if we want to post our project on linear regression and ask for people’s opinion and advises on it,for this in which group do i post my project?

#11

Thank you for pointing this out, @sathi-satb!
I have created a new thread which can be found here:

#12

Hey @agatha-l, the session links are given in the description above!

#13

as @aakashns was calling fit function for 5 5 epochs, and was printing result of each iteration, from there we can figure out whether our accuracy is getting constant, then we decrease learning rate for next call of fit function.

#14

the kernel dies as soon as i do plt.imshow() in my jupyter notebook

2 Likes
#15

Yeah, this sounds good however, we will have to manually keep a check on this stagnating accuracy and call a fit function. I was wondering if there’s any function pre built in torch that keeps a check on the accuracy for it to be constant in n epochs and reduces/ increases the learning rate. We can surely write one of our own, but just want to know it there’s one already. This will help when there are 50+ epochs.

#16

class MnistModel(nn.Module):

def __init__(self):

    super().__init__()

    self.linear = nn.Linear(input_size, num_classes)

    

def forward(self, xb):

    xb = xb.reshape(-1, 784)

    out = self.linear(xb) 

    # self.linear will call self.linear function which is available in __init__

    # 

    return out

model = MnistModel()

for images, labels in train_loader:
print(images.shape)
outputs = model(images)
break

how do we ensure that the below line, automatically call forward function?

outputs = model(images)

we know, a function wont execute until we call it, with above line, how forward function will execute? please help me to understand

2 Likes
#17

There are differnet learning rate schedulers, one of this is one cycle learning rate scheculer in which the learning rate gradually increases upto 30% of epoch and then continue to decrease till the last epoch. Here I link a blog post related to it.
https://towardsdatascience.com/learning-rate-scheduler-d8a55747dd90
I think this topic about learning rate scheduler might also be covered later in the course.

2 Likes
#18

Don’t worry @iamsusmita if you need any help or facing any doubt feel free to work together with me and many others here. :slight_smile: At the beginning, it seems tough but as you mould yourself into the process it becomes easier and remembers you are not alone

#19

I don’t there is official support for this. It’s good to do it manually, changes in batchsize can also affect the performance of the model.

1 Like
#20

There can be many points to note, you will have to gain them by developing intuition based on experiments, might be difficult to convey without context . I’ll list some down below

  • Image size
  • Object size
  • Model limitations/ Hardware limitations which can lead to Out of memory errors
  • Noisy images in datasets
  • Certain Loss functions for certain type of applications.
    …
  • If its for a classification dataset you can just simply organize data in a different folder having the label name.
  • If its for multi-object classification, you can consider having a CSV with 2 columns [image_path, classes]
    …so on. Check them out based on the application you can find of a lot of examples in Kaggle datasets and how others do the labeling.
1 Like