Skip to content

Latest commit

 

History

History
159 lines (107 loc) · 9.7 KB

Lab.md

File metadata and controls

159 lines (107 loc) · 9.7 KB
Raw

Lab challenge

Udacity PyTorch Final Lab Guides

-https://medium.com/@ml_kid/tips-to-help-you-get-started-with-the-final-lab-project-1e3ab1737b81 by @Avinash

Q1: What are the tips and tricks that could be helpful in the Lab challenge?

Q2: Is there any Google Colab setup that I can use for the Lab challenge?

Q3: Can I use transfer learning?

  • This is the recommended approach. The supplied notebook Image Classifier Project.ipynb states “you should use one of the pretrained models from torchvision.models to get the image features.”

Q4: How do I get the best possible model?

  • You need to select:
    • model for transfer learning
    • definition of your classifier
    • batch size
    • optimizer
    • learning rate(s)
    • epochs.

Completing the course, reading the Slack channels and reading around the subject will help with these decisions. Note the accuracy and loss results you get when varying these parameters and choose a combination that works well.

Q5: What is the deadline for submission of the project? The Udacity page for this program says "This term ends on Jan 23rd".

  • The deadline is Jan 9th.

Q6: What does "Duration 1 hour" mean on the Final Challenge page?

  • This is simply an estimate of the time needed to load and test your model. There is no limit on duration; you can start the lab and test your model as many times as you wish.

Q7: After submission of the Final Project will there be a second chance if I don’t pass?

  • You can submit as many times as you like. The best result will be considered for your final score.

Q8: Can I see my accuracy on the test set after passing?

  • No, there is no further information available.

Q9: I got this error: "RuntimeError: cuda runtime error (2) : out of memory

  • Restart your kernel and try decreasing your batch size and/or using a simpler model.

Q10: Will I have to save my model in CPU instead of GPU to load it properly?

  • Answered by @Vittorio Nardone

    If you have already saved model checkpoint from GPU, you can try this to load it in CPU env: checkpoint = torch.load(filename, map_location=lambda storage, loc: storage)

Alternatively, you can put your model onto the CPU before saving by using model.cpu()

Q11: Will I have to train my model on pytorch 0.4.0 to pass the test?

  • No, if you train your model on a later version of pytorch then to pass you should include strict=False when loading the state_dict, i.e. model.load_state_dict(checkpoint[‘state_dict’], strict=False). This ignores items added in later versions, e.g. num_batches_tracked.

Q12: Is it possible to see how the test set accuracy is being calculated (not the value)? There's a lot of confusion if someone uses a non-convential way to do the assignments...

  • Unfortunately, these details are not shared.

Q13: How can I use ResNet given that it does not have a classifier?

  • Put your classifier on the final layer of the model (the fully connected layer) model.fc = classifier

Q14: Why don’t my image labels match the names of the flowers?

  • Note that when using ImageFolderthe folder names are treated as character strings, so their order is 1, 10, 100, 101, 102, 11, 12, … and the images in these respective folders are given labels 0, 1, 2, 3, 4, 5, 6, … The folder names (which here happen to be numbers) are treated as characters. The code in the notebook creates a cat_to_name dictionary of folder names and flower names. We then need a look-up from categories/folder names to labels, for example, by using your_dataset.classes.

    class_to_idx = {sorted(your_dataset.classes)[i]: i for i in range(len(your_dataset.classes))}

    {k: class_to_idx[k] for k in list(class_to_idx)[:7]}

    {'1': 0, '10': 1, '100': 2, '101': 3, '102': 4, '11': 5, '12': 6}

Q15: How can I save bandwidth at the expense of drive storage when training my model using Google Colab?

  • Answer by @ecdrid

With your filename as xyz.pth, after some training:

# This only needs to be done once in a notebook.
!pip install -U -q PyDrive

from pydrive.auth import GoogleAuth
from pydrive.drive import GoogleDrive
from google.colab import auth
from oauth2client.client import GoogleCredentials

# Which file to send?
file_name = "xyz.pth" #make sure you always change this..

from googleapiclient.http import MediaFileUpload
from googleapiclient.discovery import build

auth.authenticate_user()
drive_service = build('drive', 'v3')

def save_file_to_drive(name, path):
  file_metadata = {'name': name, 'mimeType': 'application/octet-stream'}
  media = MediaFileUpload(path, mimetype='application/octet-stream', resumable=True)
  created = drive_service.files().create(body=file_metadata, media_body=media, fields='id').execute()
  return created

save_file_to_drive(file_name, file_name)

Now you can just connect your drive and start training further if you wish.

Q16: In case we want to see the feedback from the evaluation of the project, will it be a problem if we upload the project to get feedback from the evaluator/mentor?

  • You are welcome to receive feedback from your classmates and alumni volunteers. Mentors per se are only available in the full ND program. Nonetheless, our alumni volunteers have the skills to help to out, so don't hesistate.

Q17: Is there a public leaderboard where I can compare my final score?

  • There is no official leaderboard but some students have created a public spreadsheet where students can enter their scores and compare model architectures and validation scores.

Q18: In the final project, it says: Here, you'll build an image classifier from scratch that will identify different species of flowers. So, do I have to build a classifier or I can use Transfer Learning?

  • You can use Transfer Learning. It is stated in the Jupyter notebook of lab project that you can start with VGG.

Q19: Since the model will run on CPUs on the Udacity server, is there any technique that we can exploit to optimize the model for CPUs instead of GPUs?

  • You could train your model faster on GPU on Colab or Kaggle or another account. Training on CPU will be very slow and boring. Then, when you want to use the same model on CPU, you will need some variable casting like checkpoint = torch.load(filepath, map_location='cpu').

Q20: How could I run my model on Google Colab, and leave it running for hours without bothering about terminations?

  • You can save checkpoints to Google Drive and load them back to continue the model training later.

Q21: Even with GPU, my model is taking more than 3 hours to train. Did anyone's model take longer?

  • It depends on the GPU being used and yes, this lab is a relatively computationally expensive problem in the first place. You can also check this page for some info.

Q22: Anyone encountered this error when running the assessment code for the final challenge? RuntimeError: cuda runtime error (35) : CUDA driver version is insufficient for CUDA runtime version at torch/csrc/cuda/Module.cpp:51**

Answer by @Clement

  • This happens because the checkpoint file you saved contains parameters that was on the GPU. Udacity workspace does not allow the usage of GPU! To resolve this issue: In the torch.load function add an additional parameter, map_location=“cpu”. In other words, torch.load(“filepath”, map_location=“cpu”)

Q23: How to rectify this error: element 0 of tensors does not require grad and does not have a grad_fn ?

Answer by @Mohamed Shawky

Assign your classifier to last layer correctly, for resnet it's model.fc=your_classifier and for vgg and densenet it's model.classifier=your_classifier