Hope you could give some insight why the void class or background class should be ignore during confusion matrix calculation since during training void class is included?

The second part of the deep learning basics notebook mentions using CNNs for the MNIST dataset, but the code is just using dense layers. This may be misleading to students.

Hi. I tried to run DeepLearningBasics notebook on Colab.
Below are some of the issues I came across:

The code to download mnist_dream.mp4 video when using Colab doesn't work if ran as is:

mnist_dream_path = urllib.request.urlretrieve(os.path.join(deep_repo_url, 'mit_driveseg_sample.mp4')[0]

deep_repo_url is not declared anywhere and the video mentioned in the code is incorrect

I tried below two versions of code to download the video and both worked:

 !wget https://github.com/lexfridman/mit-deep-learning/tree/master/tutorial_deep_learning_basics/images/mnist_dream.mp4

import urllib
urllib.request.urlretrieve("https://github.com/lexfridman/mit-deep-learning/tree/master/tutorial_deep_learning_basics/images/mnist_dream.mp4","mnist_dream.mp4")

The code ret, img = cap.read() returns the tuple False, None. There seems to be some problem with VideoCapture() on Colab.

Since the img is not being returned as expected, the rest of the code to do inference on mnist_dream.mp4 frames isn't working.

I remember facing problems in the past when using OpenCV methods on Colab. I'm not aware of any fix or workaround for this issue atm.

Besides these issues on Colab, below are few things that can be modified:

The functiondef plot_history(history) works just fine without the parameter history since it's not being used anywhere inside the function

In the below code, rmse needs to be printed instead of rmse_final

test_features_norm = (test_features - train_mean) / train_std
mse, _, _ = model.evaluate(test_features_norm, test_labels)
rmse = np.sqrt(mse)
print('Root Mean Square Error on test set: {}'.format(round(rmse_final, 3)))

Hi there, In the part1 of deep_learning_basic lession, this cell i think you displayed the RMSE value in validation set. It should be the new 'rmse' variable instead.

X_test = scaler.transform(X_test)
mse, _, _ = model.evaluate(X_test, y_test)
rmse = np.sqrt(mse)
print('Root Mean Square Error on test set: {}'.format(round(rmse_final, 3)))

since we cannot iterate through a float its better we take the floor division of (len(train_x) / batch_size) ==> (len(train_x) // batch_size)

can't specify softmax_cross_entropy_with_logits without labels or logits
tf.nn.softmax_cross_entropy_with_logits( y, y_) ==> tf.nn.softmax_cross_entropy_with_logits(labels = y_, logits = y)

The problem is from this: sep_color= (yn+1)/2.0; line, since it return's a array of list of list. And the pl.scatter() cannot accept array of list as parameter.

I unlisted the sep_color.

# Computing the colors for the points from itertools import chain sep_color = (yn+1)/2.0; sep_color = list(chain(*sep_color))

path: tutorial_gans/tutorial_gans.ipynb

Some tf functions are deprecated

I added: !pip install tensorflow==1.15

and there was no video, so i changed the last code cell to: mp4 = open('gan.mp4','rb').read() data_url = "data:video/mp4;base64," + base64.b64encode(mp4).decode() HTML(""" <video width=400 controls autoplay loop> <source src="%s" type="video/mp4"> </video> """ % data_url)

minor tweak to get running w/current versions of toolset since initial checkin. Thanks-- awesome hands on primer. Now i can crank Hendrix - 'Are you Experienced' and do some deeeeep learning.

And Lex, goes w/out saying love the podcasts. On that note, beyond the scope of the code though maybe you can ask some big brains with whom <or which, to extent they're not of human origin> you have close encounters to share their thoughts on whether tic tac(s) imply success at implementation of the Alcubierre Warp Drive (such that >c travel is now possible across interstellar distances via inducing ‘inflation’ astern while experiencing neither acceleration , nor time dilation, within the warp bubble-- essentially warping SpaceTime). And if so, how they might explain working out some accompanying kinks... additionally facilitating 'low' earthly speeds, and mitigation for what would seem should be one hell of a 'wake'... <maybe i can be a fly-on-the-wall via a ClubHouse invite for me @W1ndH2O woot woot>

Thanks for all you're doing!

I've updated these tutorials to look for TensorFlow version 2.x and if so invoke the legacy mode to allow these models to run. I also updated a couple tutorials to work with python3.x (mainly issues with the zip object).