ufal / npfl114

Materials for the Deep Learning -- ÚFAL course NPFL114

Deep Learning – ÚFAL Course NPFL114

This repository contains materials to the Deep Learning (ÚFAL course NPFL114).

All created content is available under CC BY-SA 4.0 license, while all existing materials (mostly images and excerpts from papers) are properly referenced and are subject to original licensing.

npfl114's Issues

mnist package

In pca_first.py there is a line

from mnist import MNIST

but this mnist package is nowhere.
Even after installing tf and gym, it gives errors.

lemmatizer BOW/EOW confusion

This comment (and a few others, in the other files as well) mention [EOW], but it is actually [BOW] in the results. This is because BOW and EOW are both defined as 1 in the MorphoDataset. I don't know if this is done on purpose or not, but I am guessing that it's not because both have characters added here.

Uppercase.py

args.logdir not in base template, making people lose data by not saving it.
It's masked by the
os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "2") # Report only TF errors by default
which i think should NOT by used.

Change the code to fetch data from the .tfrecord files

https://github.com/ufal/npfl114/blob/master/labs03/1-mnist.py

Hello, I tried to change your code to use the .tfrecord as data source file. but it doesm't work, can you help solve my problem? I put a question on the stackoverflow

https://stackoverflow.com/questions/46985411/tensorflow-read-tfrecord-without-a-graph

site bboxes_utils.py issue

Sites hyperlink for bboxes_utils.py doesn't work. It gives out 404 error as lab06 doesn't contain the script.

Wildly different example outputs in `sequence_classification`

Hi,

outputs in the 'Examples' section of the task sequence_classification are wildly different to what I'm getting locally. The outputs in the 'Tests' section are identical to mine and my solution also passed the tests in ReCodEx. Perhaps you used a different seed? I doubt there's such a huge difference between macOS and Linux.

Btw. it seems you swapped the order of 'Examples' and 'Tests' on the webpage :D

Example of the output I'm getting

$ python3 sequence_classification.py --rnn=LSTM --epochs=5 --hidden_layer=50

Epoch 1/5 loss: 0.6828 - accuracy: 0.5167 - val_loss: 0.6590 - val_accuracy: 0.5178
Epoch 2/5 loss: 0.6441 - accuracy: 0.5408 - val_loss: 0.6303 - val_accuracy: 0.5264
Epoch 3/5 loss: 0.6227 - accuracy: 0.5565 - val_loss: 0.6145 - val_accuracy: 0.5573
Epoch 4/5 loss: 0.6108 - accuracy: 0.5579 - val_loss: 0.6020 - val_accuracy: 0.5617
Epoch 5/5 loss: 0.5932 - accuracy: 0.5699 - val_loss: 0.5876 - val_accuracy: 0.5717

Cheers.

Non-binary masks for CAGS dataset

The expected result of the model in cags_segmentation.py is a probabilistic mask of shape (batch_size, 224, 224, 1) with values of elements being tf.float32 in 0, 1 range (i.e. probabilities).
Therefore, the true mask should have the corresponding form of (batch_size, 224, 224, 1) with elements being binary 0/1 for presence of the mask and comparison using the 0.5 threshold.
The comments were updated to warn that the mask's dtype is tf.uint8, however I would still expect the masks to have binary 0/255 values. Nevertheless, when I plotted one of the masks, the tensor was full of various unique numbers and the mask was not binary. This might be hard to discover at first, since the masks at https://ufal.mff.cuni.cz/~straka/courses/npfl114/2223/demos/cags_train.html seem to be binary. I propose to binarize the masks accordingly inside the dataset and not only inside MaskIoUMetric, since otherwise, in order for the model training to work, the users have to correctly binarize the masks themselves, which, in turn, breaks the MaskIoUMetric (at least it did for me, I have not yet discovered why that was the case (see the row at the bottom))

Site redirects to old year

In the assignment tab, in the section for sgd_manual, hyper link for sgd_backpropagation redirects to former year of this course
( the link: https://ufal.mff.cuni.cz/courses/npfl114/2021-summer#sgd_backpropagation)
It should redirect to current year course so to https://ufal.mff.cuni.cz/courses/npfl114/2022-summer#sgd_backpropagation
or something in that nature.

pca_first.py throws IndentationError

Hello,

in file https://github.com/ufal/npfl114/blob/master/labs/01/pca_first.py, section:

    # TODO: Now run `args.iterations` of the power iteration algorithm.
    # Start with a vector of `cov.shape[0]` ones of type tf.float32 using `tf.ones`.
    v = None
    for i in range(args.iterations):
        # TODO: In the power iteration algorithm, we compute
        # 1. v = cov * v
        #    The matrix-vector multiplication can be computed using `tf.linalg.matvec`.
        # 2. s = l2_norm(v)
        #    The l2_norm can be computed using `tf.linalg.norm`.
        # 3. v = v / s

    # The `v` is now the eigenvector of the largest eigenvalue, `s`. We now
    # compute the explained variance, which is a ration of `s` and `total_variance`.
    explained_variance = s / total_variance

would be nice to add something to the code block inside the for loop, for example:

    # TODO: Now run `args.iterations` of the power iteration algorithm.
    # Start with a vector of `cov.shape[0]` ones of type tf.float32 using `tf.ones`.
    v = None
    for i in range(args.iterations):
        # TODO: In the power iteration algorithm, we compute
        # 1. v = cov * v
        #    The matrix-vector multiplication can be computed using `tf.linalg.matvec`.
        # 2. s = l2_norm(v)
        #    The l2_norm can be computed using `tf.linalg.norm`.
        # 3. v = v / s
        ...

    # The `v` is now the eigenvector of the largest eigenvalue, `s`. We now
    # compute the explained variance, which is a ration of `s` and `total_variance`.
    explained_variance = s / total_variance

will be valid code that can be executed without throwing an error.

Thanks :)