Lesson 14 in-class

Part 2 (2017)
41

Are these kinds of architectures useful for classification as well?

42

Do we have pretrained models for Densenet (trained on imagenet for example)?

43

@thejaswi.hr the 100 layers tiramisu architecture is designed for 224x224 input

44

@layla.tadjpour this is the 100 layers tiramisu paper

45

The tiramisu model scores seems to indicate overfitting, is it a problem ?

46

Sorry! something weird was happening with my streaming video. I had to refresh my page !!

1 Like
47

re Densnet Tiramisu: If no hand annotated segmentation maps are available - is there another way to create rasonable performance using e.g. some unupervised segmentation and using this to train the model

2 Likes
48

@jeremy and @rachel thank you both for this wonderful well thought out course. I have benefited a lot from the course and thanks for doing the remote international fellowship and selecting me for this group.

1 Like
49

I don’t think I made a strong recommendation one way or the other. GRU is a somewhat simpler architecture and therefore may be a little faster. But either one is fine really.

50

Hi there,

I am trying to understand the def statements. Which one removes the log transformations on the predictions?

  1. def log_max_inv(preds, mx = max_log_y):
    return np.exp(preds * mx)

  2. def normalize_inv(preds):
    return preds * ystd + ymean

Given other code in the notebook it makes me think it is (1) def log_max_in

preds = np.squeeze(model_pkl.predict(map_valid_pkl, 1024))
y_orig_pkl_val = log_max_inv(y_pkl_val, max_log_y_pkl)

If I am correct then what does normalize_inv even do?

51

are you able to call model.predict with the code from the rossman lesson using the xgboost model? i tried and all the predictions are the same.

pred=model.predict(Xdata_val