In code block 7.26, If the sub-block alone is executed, it generates NameError: name 'model' is not defined. The subclass CustomModel uses the optimizer and trainable weights that were defined in the previous model. By accessing the actual optimizer and trainable weights that passed to CustomModel's compile(), the subclass can be made independent of the previously defined model and can eliminate the mentioned error.

Hello,

I am currently getting the error: AttributeError: module 'numpy.random' has no attribute 'randit' when running model.fit_generator(). However, the numpy.random.randit() works fine when we call generator() directly.

I have run pip3 install numpy as well as python3 -m pip install numpy to make sure numpy is update.

I am currently running a jupyter notebook on a Ubuntu 16.04 machine.

The instructions in 6.3 say the data is

recorded at the Weather Station at the Max-Planck-Institute for Biogeochemistry in Jena, Germany: http://www.bgc-jena.mpg.de/wetter/.

The data file name in the code is jena_climate_2009_2016.csv. The website allows us to download data in 6-month increments, so the period 2009-2016 is split into 16 files.

Were these files concatenated to form the file used in the notebook? If so, a sentence to that effect might clarify where the data came from. Alternatively, if jena_climate_2009_2016.csv isn't the concatenation of these files, I think it's unclear where a reader would find the data.

Hi @fchollet, thanks for this amazing book (and the corresponding colab notebooks).

I tried to run this file https://colab.research.google.com/github/fchollet/deep-learning-with-python-notebooks/blob/master/chapter11_part01_introduction.ipynb#scrollTo=xdw1FYamgsP9

But I am getting the following error

from tensorflow.keras.layers import TextVectorization
text_vectorization = TextVectorization(
    output_mode="int",
)
---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
<ipython-input-1-0642862b90e9> in <module>()
----> 1 from tensorflow.keras.layers import TextVectorization
      2 text_vectorization = TextVectorization(
      3     output_mode="int",
      4 )

ImportError: cannot import name 'TextVectorization' from 'tensorflow.keras.layers' (/usr/local/lib/python3.7/dist-packages/tensorflow/keras/layers/__init__.py)

Do you know what the issue is? Thanks!

I am trying some codes from your fantastic book, but got unconventional error . Please give an eye on it . Thank you model.compile(optimizer=optimizers.RMSprop(learning_rate=0.001), loss='binary_crossentropy' , metrics=['accuracy']) (on p-73) Error : module 'keras.optimizers' has no attribute 'RMSprop'

The author foget deleting the temperature column in the float_data which is used as a pool for sampling. Thus the outcome looks perfect due to the target leakage.

Delete the target response when try the code yourself, though it will let u see how bad the Network performs. 😄

temp = float_data[:, 1] a = float_data[:, 0] a = np.reshape(a, newshape=(len(a), 1)) b = float_data[:, 2:] print(a.shape, b.shape) float_data = np.concatenate([a, b], axis=1)

Using Keras 2.2.4, I'm working my way though this notebook 5.4-visualizing-what-convnets-learn , except I switched the model with a unet one provided by Kaggle-Carvana-Image-Masking-Challenge . The first layer of the Kaggle model looks like this, followed by the rest of the example code.

def get_unet_512(input_shape=(512, 512, 3),
                 num_classes=1):
    inputs = Input(shape=input_shape)

...

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_13 (InputLayer)           (None, 512, 512, 3)  0    
...

from keras import models
layer_outputs = [layer.output for layer in model.layers[:8]]
activation_model = models.Model(inputs=model.input, outputs=layer_outputs)
activations = activation_model.predict(img_tensor)

Now the error I am getting is

InvalidArgumentError: input_13:0 is both fed and fetched.

Does anyone have any suggestions on how to work around this?

Having trouble getting listing 6.37 to work. The model stalls on the first epoch. Getting the following output, but it never reaches the end of epoch 1:

Epoch 1/20
496/500 [============================>.] - ETA: 0s - loss: 1.2985

Prior code in this chapter and code in prior chapters works fine. Any suggestions?

Thanks,

• title above code block: Writing a step-by-step training loop: the loop itself

• error output:

• TF, TF-base v2.4.1; TF-estimator v2.6.0

• conda v4.11.0

• Ubuntu 21.10

On the book:

acc = history.history['acc']
val_acc = history.history['val_acc']

But if I typed so, it would result in an error where keys of history are incorrect. But when I did history.history.keys(), it showed "accuracy" and "val_accuracy" instead. Is this a mistake on my part?

Here's the full code:

from keras.datasets import reuters
from tensorflow.keras import models
from tensorflow.keras import layers
import numpy as np
import matplotlib.pyplot as plt

def vectorize_sequences(sequences, dimension = 10000):
	results = np.zeros((len(sequences),dimension))
	for i, sequence in enumerate(sequences):
		results[i, sequence] = 1
	return results

def to_one_hot(labels, dimension=46):
	results = np.zeros((len(labels), dimension))
	for i, label in enumerate(labels):
		results[i, label] = 1
	return results

(train_data, train_labels), (test_data, test_labels) = reuters.load_data(num_words=10000)

x_train = vectorize_sequences(train_data)
x_test = vectorize_sequences(test_data)

one_hot_train_label = to_one_hot(train_labels)
one_hot_test_label = to_one_hot(test_labels)

# one_hot_train_label = to_categorical(train_labels)
# one_hot_test_label = to_categorical(test_labels)

model = models.Sequential()
model.add(layers.Dense(64, activation='relu', input_shape=(10000,)))
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(46, activation='softmax'))

model.compile(optimizer='rmsprop',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

x_val = x_train[:1000]
partial_x_train = x_train[1000:]

y_val = one_hot_train_label[:1000]
partial_y_train = one_hot_train_label[1000:]

history = model.fit(partial_x_train,
                    partial_y_train,
                    epochs=20,
                    batch_size=512,
                    validation_data=(x_val, y_val))

print("History key: ", history.history.keys())

loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(1, len(loss) + 1)

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()

plt.show()

plt.clf()   # clear figure

acc = history.history['accuracy']
val_acc = history.history['val_accuracy']

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()

plt.show()

Hi,

I tried to use vgg16 using keras for the first time. I loaded model offline in weights parameter using file path. on model.summary it is not showing flatten, fc1, fc2 layers . it shows layers up to bock5_pool . why so ? on proceeding further, I am getting model output predictions of the type (1,512,7,7) which shows error in decode_predictions ..Pls help

When trying to run the code for preparing Glove word-embeddings matrix, towards the end of the notebook for chapter 11, part 2, I get an error: embedding_dim = 100

Retrieve the vocabulary indexed by our previous TextVectorization layer.

vocabulary = text_vectorization.get_vocabulary()

Use it to create a mapping from words to their index in the vocabulary.

word_index = dict(zip(vocabulary, range(len(vocabulary))))

Prepare a matrix that will be filled with the GloVe vectors.

embedding_matrix = np.zeros((max_tokens, embedding_dim)) for word, i in word_index.items(): if i < max_tokens: embedding_vector = embeddings_index.get(word) # Fill entry i in the matrix with the word vector for index i. # Words not found in the embedding index will be all zeros. if embedding_vector is not None: embedding_matrix[i] = embedding_vector UnicodeDecodeError Traceback (most recent call last) Input In [29], in <cell line: 4>() 1 embedding_dim = 100 3 # Retrieve the vocabulary indexed by our previous TextVectorization layer. ----> 4 vocabulary = text_vectorization.get_vocabulary() 5 # Use it to create a mapping from words to their index in the vocabulary. 6 word_index = dict(zip(vocabulary, range(len(vocabulary))))

File C:\ProgramData\Anaconda3\envs\tf-gpu\lib\site-packages\keras\layers\preprocessing\text_vectorization.py:448, in TextVectorization.get_vocabulary(self, include_special_tokens) 439 def get_vocabulary(self, include_special_tokens=True): 440 """Returns the current vocabulary of the layer. 441 442 Args: (...) 446 vocabulary will not include any padding or OOV tokens. 447 """ --> 448 return self._lookup_layer.get_vocabulary(include_special_tokens)

File C:\ProgramData\Anaconda3\envs\tf-gpu\lib\site-packages\keras\layers\preprocessing\index_lookup.py:336, in IndexLookup.get_vocabulary(self, include_special_tokens) 334 keys, values = self.lookup_table.export() 335 vocab, indices = (values, keys) if self.invert else (keys, values) --> 336 vocab, indices = (self._tensor_vocab_to_numpy(vocab), indices.numpy()) 337 lookup = collections.defaultdict(lambda: self.oov_token, 338 zip(indices, vocab)) 339 vocab = [lookup[x] for x in range(self.vocabulary_size())]

File C:\ProgramData\Anaconda3\envs\tf-gpu\lib\site-packages\keras\layers\preprocessing\string_lookup.py:401, in StringLookup._tensor_vocab_to_numpy(self, vocabulary) 399 def _tensor_vocab_to_numpy(self, vocabulary): 400 vocabulary = vocabulary.numpy() --> 401 return np.array([tf.compat.as_text(x, self.encoding) for x in vocabulary])

File C:\ProgramData\Anaconda3\envs\tf-gpu\lib\site-packages\keras\layers\preprocessing\string_lookup.py:401, in (.0) 399 def _tensor_vocab_to_numpy(self, vocabulary): 400 vocabulary = vocabulary.numpy() --> 401 return np.array([tf.compat.as_text(x, self.encoding) for x in vocabulary])

File C:\ProgramData\Anaconda3\envs\tf-gpu\lib\site-packages\tensorflow\python\util\compat.py:110, in as_text(bytes_or_text, encoding) 108 return bytes_or_text 109 elif isinstance(bytes_or_text, bytes): --> 110 return bytes_or_text.decode(encoding) 111 else: 112 raise TypeError('Expected binary or unicode string, got %r' % bytes_or_text)

UnicodeDecodeError: 'utf-8' codec can't decode byte 0xc3 in position 0: unexpected end of data

FutureWarning: Option --id was deprecated in version 4.3.1 and will be removed in 5.0. You don't need to pass it anymore to use a file ID.

https://github.com/fchollet/deep-learning-with-python-notebooks/blob/master/chapter12_part05_gans.ipynb****

In the 2nd edition notebook of chapter 12 part 4, the latent space sampling layer is defined as the following class:

class Sampler(layers.Layer):
    def call(self, z_mean, z_log_var):
        batch_size = tf.shape(z_mean)[0]
        z_size = tf.shape(z_mean)[1]
        epsilon = tf.random.normal(shape=(batch_size, z_size))
        return z_mean + tf.exp(0.5 * z_log_var) * epsilon

In the 1st edition, it's defined as the following fucntion:

def sampling(args):
    z_mean, z_log_var = args
    epsilon = K.random_normal(shape=(K.shape(z_mean)[0], 2),
                                          mean=0., stddev=1.)
    return z_mean + K.exp(z_log_var) * epsilon

Most changes come from following best practices and API changes, but I do have the folowing doubt: Why is z_log_var multiplied by 0.5 in the 2nd edition class? Is it because a difference between tf.exp() and tf.keras.backend.exp()? Is it becasue reducing it by half yields better results? Any other reason?

In Deep Learning with Python 2e, @fchollet says:

Multi-hot encode your lists to turn them into vectors of 0s and 1s. This would mean, for instance, turning the sequence [8, 5] into a 10,000-dimensional vector that would be all 0s except for indices 8 and 5, which would be 1s.

Wouldn’t this encoding essentially reduce the sequence to a set, which means losing information? For instance, how would the encoded representation of [8, 5] differ from [5, 8] or [8, 5, 8] (or any longer sequence of 8 and 5 elements)?

(This example arises in a tutorial about classifying movie reviews by whether they’re positive or negative. To make the example concrete, if word 5 is pretty and word 8 is awful, then there’s going to be a classification difference between pretty awful and awful pretty!)

Following the example in the book, I increased the sample size from 200 to 3000 and added a dropout regularization, but the final validation accuracy is only 50%.

In the chapter11_part04_sequence-to-sequence-learning.ipynb, the TransformerDecoder receives the mask from the PositionalEmbedding layer of the target sequence:

x = PositionalEmbedding(sequence_length, vocab_size, embed_dim)(decoder_inputs)
x = TransformerDecoder(embed_dim, dense_dim, num_heads)(x, encoder_outputs)

Shouldn’t the mask be the one created from encoding the source sequence?

For example, I have seen that in this TF tutorial the mask from the source sequence is used instead.

Any clarification would be greatly appreciated.