features = extract_features('cat.jpg', model)
is assuming it is run in colab. I suggest features = extract_features(IMG_PATH, model) so it also works locally off colab and uses IMG_PATH which was defined in the cell right before

IMG_PATH = '../../sample-images/cat.jpg'
if IS_COLAB_ENV:
  !curl https://raw.githubusercontent.com/PracticalDL/Practical-Deep-Learning-Book/master/sample-images/cat.jpg --output cat.jpg
  IMG_PATH = 'cat.jpg'

Same for the %timeit features = extract_features('cat.jpg', model) line, to change to %timeit features = extract_features(IMG_PATH, model)

Hi,
Thanks for the amazing book. I recently finished the book and started trying out exercises.

In this notebook [1], the predictions like "10% cardigan" are not plotted in the -output.jpg image. What am I missing?

Thank you.

[1] https://github.com/PracticalDL/Practical-Deep-Learning-Book/blob/master/code/chapter-2/2-what-does-my-neural-network-think.ipynb

Before I open an issue, I have to admit that this is a great piece of work that cuts to the chase.
I got the book today. And, I already finished first chapter. It is the best intro I ever read on AI books. Koul Cool trick with Dr. May Carson 😉.

Issue:
Unless I missed something, the readers will have minor hiccup in installing the requirements.txt. I spent about half an hour, and I found some of the pinned packages are either not on PyPi or versions are stale. Please try steps below to reproduce:

1. System running Ubuntu 16.04 LTS
2. Python version 3.7.4
3. Created venv > then activate it
4. Then issued this command: pip install -r requirements.txt

Outcome:
Some packages will be either unable to install or can not be found in PyPi server (e.g. apturl).
There are few, but I mentioned one as example. It seems the requirement file may need some maintenance

Thanks.

u = AnnoyIndex(2048)
%timeit u.get_nns_by_vector(feature_list[random_image_index], 5, include_distances=True)

The above code appeared a few times.
u.load('data/caltech101index.ann') is missing. Without it all returned indexes are empty. Performance measures are for an empty index, and are inaccurate (faster compared to with index loaded) by 10x on my CPU.

batch_size = 128
datagen = tensorflow.keras.preprocessing.image.ImageDataGenerator(preprocessing_function=preprocess_input)

generator = datagen.flow_from_directory(root_dir,
                                        target_size=(224, 224),
                                        class_mode=None,
                                        shuffle=False)

uses default batch_size = 32.

num_epochs = int(math.ceil(num_images / batch_size))
calculates 68 epochs based on 128 batch size.
So generator only generates 68*32 = 2176 images for feature_list, mismatching the 8677 in standard_feature_list with features extracted from the non generator method.

Congratulations on an amazing book. Excellent so far!

I noticed in the image search project in Chapter 4 that the feature embeddings for the out of the box ResNet50 model are calculated from an intermediate layer (GlobalMaxPooling2D) (None, 2048).

But after fine tuning based on our own dataset the features are calculated based on the final output layer and hence the feature dimensions become more like (1, 10) based on the number of classes in our dataset.

I'm not sure is this an error/oversight or if both methodologies are correct?
Would love to get your opinion on this. And if a fix is needed I can offer to help also.
Thanks again for the great book.
Niall

Hello to every one, Am training a bird specie object detection. I have 7(with 525 image, 75 pictures per class) classes and using:

software and libraire

• tensorflow1.15
• python3.6

model:

• faster_rcnn_resnet152_pets.config

Laptop:

• OMEN 15 which have RTX graphic card (6Go) on Ubuntu 18.04.

The level:
loss:1.4247278, step: 75185
The loss still decreasing but slowly.

I started the training since yesterday and I stopped the training and continued at this morning at the saved checkpoint as you can know.
But at this stage, I saved and export the model.*-7108 to test test prediction, but i did not get prediction and no boxes was drawn, only if I decrease the treshold from 0.6 to 0.1, i get some false and true prediction.

My question is: does that means my model was not learned very well? or should I continue the training until I reach 200k step.

Chapter 1 mentioned in README is missing in the code repo.

It is taking ~40 minutes on my CPU which is unbearable.

Why not use numpy?
class_ids was pickled but not used.

Here's the full code, takes < 1 second (i didn't wrap in function or multiply accuracy by 100)

class_ids = pickle.load(open('features/class_ids-caltech101.pickle', 'rb'))

neighbors = NearestNeighbors(n_neighbors=5,
                             algorithm='brute',
                             metric='euclidean').fit(feature_list_compressed)
    
distances, np_indices = neighbors.kneighbors(feature_list_compressed)

neighbors_class_ids = class_ids[np_indices]
class_equalities = np.equal(neighbors_class_ids[:, [0]], neighbors_class_ids[:, 1:5])
class_equalities.mean()

I'm trying to make tfrecords for my dataset. How can I create tfrecords for a multilabel data with thousands of files for training and validation?
I tried this code storing-data-as-tfrecord.ipynb. This works fine for a single image.
I tried using this code also build_image_data.py, but this doesn't work for TensorFlow 2.x .

whille training or running this particular code snippet, showing this error.
Please help me out.

There is no chapter 9 code...what h5_to_tf.ipynb that is supposed to be in this repository under chapter-9 directory

i want to make a image search engine using deep learning and flask framework and try to load a pickle files but confused how to build a flask app to browse as a web based search engine.

Hi, do you happen to have the code for finetuning Resnet on Caltech 101, and the code for extracting on calnet256 ?
Thanks !

Hi,
In the notebook 2-similarity-search-level-2.ipynb, would you recommend running on a subset of the features / classnames / filenames ?
The benchmarks takes a very long time, maybe it would be better to add some code to select a subpart of the files only.

1.cant visualize this:

#color_map = plt.cm.get_cmap('coolwarm')
color_map = matplotlib.colormaps['coolwarm']
scatter_plot = plt.scatter(tsne_results[:, 0],
                           tsne_results[:, 1],
                           c=selected_class_ids,
                           cmap=color_map)
plt.colorbar(scatter_plot)
plt.show()
# To save the plot in a high definition format i.e. PDF, uncomment the following line:
#plt.savefig('results/' + str(ADD_NAME_HERE)+'.pdf', format='pdf', dpi=1000)

gives:

ValueError                                Traceback (most recent call last)
File ~/.local/lib/python3.8/site-packages/matplotlib/axes/_axes.py:4439, in Axes._parse_scatter_color_args(c, edgecolors, kwargs, xsize, get_next_color_func)
   4438 try:  # Is 'c' acceptable as PathCollection facecolors?
-> 4439     colors = mcolors.to_rgba_array(c)
   4440 except (TypeError, ValueError) as err:

File ~/.local/lib/python3.8/site-packages/matplotlib/colors.py:487, in to_rgba_array(c, alpha)
    486 else:
--> 487     rgba = np.array([to_rgba(cc) for cc in c])
    489 if alpha is not None:

File ~/.local/lib/python3.8/site-packages/matplotlib/colors.py:487, in <listcomp>(.0)
    486 else:
--> 487     rgba = np.array([to_rgba(cc) for cc in c])
    489 if alpha is not None:

File ~/.local/lib/python3.8/site-packages/matplotlib/colors.py:299, in to_rgba(c, alpha)
    298 if rgba is None:  # Suppress exception chaining of cache lookup failure.
--> 299     rgba = _to_rgba_no_colorcycle(c, alpha)
    300     try:

File ~/.local/lib/python3.8/site-packages/matplotlib/colors.py:381, in _to_rgba_no_colorcycle(c, alpha)
    380 if not np.iterable(c):
--> 381     raise ValueError(f"Invalid RGBA argument: {orig_c!r}")
    382 if len(c) not in [3, 4]:

ValueError: Invalid RGBA argument: 0.0

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Cell In[62], line 3
      1 #color_map = plt.cm.get_cmap('coolwarm')
      2 color_map = matplotlib.colormaps['coolwarm']
----> 3 scatter_plot = plt.scatter(tsne_results[:, 0],
      4                            tsne_results[:, 1],
      5                            c=selected_class_ids,
      6                            cmap=color_map)
      7 plt.colorbar(scatter_plot)
      8 plt.show()

File ~/.local/lib/python3.8/site-packages/matplotlib/pyplot.py:2862, in scatter(x, y, s, c, marker, cmap, norm, vmin, vmax, alpha, linewidths, edgecolors, plotnonfinite, data, **kwargs)
   2857 @_copy_docstring_and_deprecators(Axes.scatter)
   2858 def scatter(
   2859         x, y, s=None, c=None, marker=None, cmap=None, norm=None,
   2860         vmin=None, vmax=None, alpha=None, linewidths=None, *,
   2861         edgecolors=None, plotnonfinite=False, data=None, **kwargs):
-> 2862     __ret = gca().scatter(
   2863         x, y, s=s, c=c, marker=marker, cmap=cmap, norm=norm,
   2864         vmin=vmin, vmax=vmax, alpha=alpha, linewidths=linewidths,
   2865         edgecolors=edgecolors, plotnonfinite=plotnonfinite,
   2866         **({"data": data} if data is not None else {}), **kwargs)
   2867     sci(__ret)
   2868     return __ret

File ~/.local/lib/python3.8/site-packages/matplotlib/__init__.py:1442, in _preprocess_data.<locals>.inner(ax, data, *args, **kwargs)
   1439 @functools.wraps(func)
   1440 def inner(ax, *args, data=None, **kwargs):
   1441     if data is None:
-> 1442         return func(ax, *map(sanitize_sequence, args), **kwargs)
   1444     bound = new_sig.bind(ax, *args, **kwargs)
   1445     auto_label = (bound.arguments.get(label_namer)
   1446                   or bound.kwargs.get(label_namer))

File ~/.local/lib/python3.8/site-packages/matplotlib/axes/_axes.py:4602, in Axes.scatter(self, x, y, s, c, marker, cmap, norm, vmin, vmax, alpha, linewidths, edgecolors, plotnonfinite, **kwargs)
   4599 if edgecolors is None:
   4600     orig_edgecolor = kwargs.get('edgecolor', None)
   4601 c, colors, edgecolors = \
-> 4602     self._parse_scatter_color_args(
   4603         c, edgecolors, kwargs, x.size,
   4604         get_next_color_func=self._get_patches_for_fill.get_next_color)
   4606 if plotnonfinite and colors is None:
   4607     c = np.ma.masked_invalid(c)

File ~/.local/lib/python3.8/site-packages/matplotlib/axes/_axes.py:4445, in Axes._parse_scatter_color_args(c, edgecolors, kwargs, xsize, get_next_color_func)
   4443 else:
   4444     if not valid_shape:
-> 4445         raise invalid_shape_exception(c.size, xsize) from err
   4446     # Both the mapping *and* the RGBA conversion failed: pretty
   4447     # severe failure => one may appreciate a verbose feedback.
   4448     raise ValueError(
   4449         f"'c' argument must be a color, a sequence of colors, "
   4450         f"or a sequence of numbers, not {c!r}") from err

ValueError: 'c' argument has 8677 elements, which is inconsistent with 'x' and 'y' with size 2176.

I'm trying to figure out why the call to validation_generator.class_indicies is always showing a 1 next to dog. Regardless of whether I use a cat or dog. Prediction is also showing 2 float values rather than one. During training, my loss and accuracy are both looking good across the board (approx 97/98%), but when I proceed to test I get results similar to those shown below :

[[9.9989903e-01 1.0094591e-04]]
{'cat': 0, 'dog': 1}

I am using Tensorflow-gpu 2.2 under Ubuntu 18.04 Cuda 10.1. I originally was following with the code in the book, but then verified with your code on git and all seems to look ok to me. Does anything jump out at you as to what I am doing wrong?

I have attached a screenshot showing the test result.

When trying to run the AnnoyIndex code of 2-similarity-search-level-2, I get the following error:

IndexError: Vector has wrong length (expected 2048, got 500)

Should the length of features passed be 2048 ?

Seems that the folder containing the script.py is missing

The data passes to the explain method (when using ExtractActivations) must be fixed to

data_for_activation = (np.array([img]), None)
grid = explainer.explain(data_for_activation, model, ['conv1'])

https://github.com/PracticalDL/Practical-Deep-Learning-Book/blob/master/code/chapter-5/3-tf-explain.ipynb

After using the instructions mentioned in https://github.com/PracticalDL/Practical-Deep-Learning-Book/tree/master/code/chapter-2 (updating the .json file) and running the script on my own environment. I get the following error:

2019-12-29 21:33:16.538110: I tensorflow/core/common_runtime/process_util.cc:115] Creating new thread pool with default inter op setting: 2. Tune using inter_op_parallelism_threads for best performance.
Traceback (most recent call last):
File "visualization.py", line 162, in
process_image(image_path, output_prefix + "_output.jpg")
File "visualization.py", line 126, in process_image heatmap = explainer.explain(data, model, "block5_conv3", class_index)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tf_explain/core/grad_cam.py", line 50, inexplain model, images, layer_name, class_index
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 457, in call result = self._call(*args, **kwds)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 503, in _call
self._initialize(args, kwds, add_initializers_to=initializer_map)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 408, in _initialize
*args, **kwds))
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/function.py", line 1848, in _get_concrete_function_internal_garbage_collected
graph_function, _, _ = self._maybe_define_function(args, kwargs)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/function.py", line 2150, in _maybe_define_function
graph_function = self._create_graph_function(args, kwargs)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/function.py", line 2041, in _create_graph_function
capture_by_value=self._capture_by_value),
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/framework/func_graph.py", line 915, in func_graph_from_py_func
func_outputs = python_func(*func_args, **func_kwargs)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 358, in wrapped_fn
return weak_wrapped_fn().wrapped(*args, **kwds)
File "/home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/framework/func_graph.py", line 905, in wrapper
raise e.ag_error_metadata.to_exception(e)
TypeError: in converted code:
relative to /home/ec2-user/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages:

tf_explain/core/grad_cam.py:102 get_gradients_and_filters  *
    grad_model = tf.keras.models.Model(
tensorflow_core/python/keras/engine/network.py:539 get_layer
    raise ValueError('No such layer: ' + name)

TypeError: must be str, not numpy.int64

please add back the requirements.txt the various package used here have many conflicts if we just update to the latest version. I am having a lot of difficulty setting up to start working in chapter 5 for example

gzip: stdin: not in gzip format
tar: Child returned status 1
tar: Error is not recoverable: exiting now

I tried all the ways but its not getting unzipped

The image-to-json.py conversion script used in chapter 13 does not preprocess the data, instead leaving all inputs as values between 0-255. If used with the model from chapter 3, the predictions will be more or less random.

An easy solution is to convert the value to floating-point numbers. Here's the script I use:

import json
import numpy as np
import tensorflow as tf
from PIL import Image

model = tf.keras.models.load_model('cats_vs_dogs_01')
# The layer name of the input needs to be present in the JSON file.
input_layer_name = model.layers[0].name

img = Image.open('data/cats-vs-dogs/cat.9827.jpg')
# Because the values are floating-point, don't make the image to large
img.thumbnail((128, 128), Image.ANTIALIAS)
# Preprocess the data
data = np.asarray(img) / 255.
with open('cat_image.json', 'w') as fp:
    json.dump({input_layer_name: data.tolist()}, fp)

This script can now be used as such:

export AI_MODEL='your_model_name'
export AI_JSON_FILE='cat_image.json'
export AI_REGION='europe-west4'
gcloud ai-platform predict \
    --model $AI_MODEL \
    --json-instances $AI_JSON_FILE \
    --region $AI_REGION

There is a final finetuning throwing away dense and dropout layers that was expected according to the book but is missing in the notebook

when train model (transfer learning)

# select the percentage of layers to be trained while using the transfer learning
# technique. The selected layers will be close to the output/final layers.
unfreeze_percentage = 0

learning_rate = 0.001

if training_format == "scratch":
    print("Training a model from scratch")
    model = scratch(train, val, learning_rate)
elif training_format == "transfer_learning":
    print("Fine Tuning the MobileNet model")
    model = transfer_learn(train, val, unfreeze_percentage, learning_rate)

i see messages like that:
Corrupt JPEG data: 65 extraneous bytes before marker 0xd9

i googled that and the problem is with corrupted images in dataset (cats and dogs). to fix that one needs to use code to clean dataset:

import os
num_skipped = 0
for folder_name in ("Cat", "Dog"):
    folder_path = os.path.join("PetImages", folder_name)
    for fname in os.listdir(folder_path):
        fpath = os.path.join(folder_path, fname)
        try:
            fobj = open(fpath, "rb")
            is_jfif = tf.compat.as_bytes("JFIF") in fobj.peek(10)
        finally:
            fobj.close()

        if not is_jfif:
            num_skipped += 1
            # Delete corrupted image
            os.remove(fpath)
print("Deleted %d images" % num_skipped)

before training.
see also - https://discuss.tensorflow.org/t/first-steps-in-keras-error/8049/11

but i have no idea how to clean it if i have dataset presented as tfrecords:

cats_vs_dogs-train.tfrecord-00000-of-00008
cats_vs_dogs-train.tfrecord-00001-of-00008
cats_vs_dogs-train.tfrecord-00002-of-00008
cats_vs_dogs-train.tfrecord-00003-of-00008
cats_vs_dogs-train.tfrecord-00004-of-00008
cats_vs_dogs-train.tfrecord-00005-of-00008
cats_vs_dogs-train.tfrecord-00006-of-00008
cats_vs_dogs-train.tfrecord-00007-of-00008

how you fix that and if is not will the model will be correct ?

chapter ends with MobileNetV2 example. i extend a bit using MobileNetV3, as it is a bit fresher.
but perfomance results shows strange things.

def predict2(img_path):
    img = image.load_img(img_path, target_size=(224, 224))
    model = tf.keras.applications.**MobileNetV2**()
    img_array = image.img_to_array(img)
    img_batch = np.expand_dims(img_array, axis=0)
    img_preprocessed = preprocess_input(img_batch)
    prediction = model.predict(img_preprocessed)
    print(decode_predictions(prediction, top=3)[0])

%timeit -r 3 predict2(IMG_PATH) 

MobileNetV2 gives 912 ms ± 5.02 ms per loop

while as MobileNetV3 gives 1.04 s ± 19.9 ms per loop
from tensorflow.keras.applications import MobileNetV3Small

def predict3(img_path):
    img = image.load_img(img_path, target_size=(224, 224))
    model = tf.keras.applications.MobileNetV3Small()
    img_array = image.img_to_array(img)
    img_batch = np.expand_dims(img_array, axis=0)
    img_preprocessed = preprocess_input(img_batch)    
    prediction = model.predict(img_preprocessed)
    #print(prediction)
    print(decode_predictions(prediction, top=3)[0]) 
%timeit -r 3 predict3(IMG_PATH) 

very strange. MobileNetV3Small should be faster.

When using the Non-Colab-Code in a Jupiter Notebook I receive the following error:

"2020-01-11 20:34:01.993022: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-01-11 20:34:02.030317: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7fee2133a110 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-01-11 20:34:02.030342: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version
Traceback (most recent call last):
File "visualization.py", line 165, in
process_image(image_path, output_prefix + "_output.jpg")
File "visualization.py", line 128, in process_image
class_index, class_name, prob_value = get_predictions(img, model)
File "visualization.py", line 50, in get_predictions
return decode_predictions_modified(preds, top=1)
File "visualization.py", line 35, in decode_predictions_modified
CLASS_INDEX = json.load(open(fpath))
File "/usr/local/Cellar/python/3.7.6_1/Frameworks/Python.framework/Versions/3.7/lib/python3.7/json/init.py", line 293, in load
return loads(fp.read(),
File "/Users/olgawillner/venv2/bin/../lib/python3.7/codecs.py", line 322, in decode
(result, consumed) = self._buffer_decode(data, self.errors, final)
UnicodeDecodeError: 'utf-8' codec can't decode byte 0xff in position 0: invalid start byte"

The Code for the Colab Workbook works completely fine. Does anyone have an idea why I receive the error "UnicodeDecodeError: 'utf-8' codec can't decode byte 0xff in position 0: invalid start byte"?

2 occurences of feature_list should be dataset instead, also 2048 should be 100.
These are defined earlier in

num_items = 100000
num_dimensions = 100
dataset = np.random.randn(num_items, num_dimensions)
dataset /= np.linalg.norm(dataset, axis=1).reshape(-1, 1)

annoy_training_time = []
annoy_test_time = []
annoy_trees = [
    1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 200, 300
]
for num_trees in annoy_trees:
    t = AnnoyIndex(2048)  # Length of item vector that will be indexed
    for i in range(num_images):
        feature = feature_list[i]
        t.add_item(i, feature)
    start_time = time.time()
    t.build(num_trees)  #50 trees
    end_time = time.time()
    annoy_training_time.append(end_time - start_time)
    start_time = time.time()
    indices = t.get_nns_by_vector(feature_list[random_image_index],
                                  5,
                                  include_distances=True)
    end_time = time.time()
    annoy_test_time.append(end_time - start_time)
    print("For number of trees = ", num_trees, ",\tTime to train = ",
          annoy_training_time[-1], ",\tTime to test = ", annoy_test_time[-1])

none of chapter 3 works on google colab

Hi, in the chapter5 4-keras-tuner notebook there is a missing import

from kerastuner.tuners import RandomSearch

receiving this error message in ch14 generate_tfRecord.py

File "generate_tfRecord.py", line 20, in
flags = tf.app.flags
AttributeError: module 'tensorflow' has no attribute 'app'

For example on creating TFRecords - Consider adding the following functions to store the packed BytesList and Int64List as per the documentation to assist the reader, throws an error otherwise.

def _int64_feature(value):
  return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))

def _bytes_feature(value):
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

If I execute the same code as mentioned in the notebook https://github.com/PracticalDL/Practical-Deep-Learning-Book/blob/master/code/chapter-5/5-autokeras.ipynb, then for the line
from autokeras.image.image_supervised import ImageClassifier ,
it is throwing the following error
ModuleNotFoundError: No module named 'autokeras.image'

And if I import just
from autokeras import ImageClassifier ,
then while executing the file, it is throwing an error saying
TypeError: __init__() got an unexpected keyword argument 'path'
on the following part of the code
clf = ImageClassifier(path=".",verbose=True, augment=True, \ searcher_args={'trainer_args':{'max_iter_num':7}} )

AttributeError Traceback (most recent call last)
in
6
7 from tensorflow.keras.applications.vgg16 import VGG16, preprocess_input
----> 8 from tf_explain.core.grad_cam import GradCAM
9
10 import PIL

c:\python\lib\site-packages\tf_explain_init_.py in
8 version = "0.2.1"
9
---> 10 from . import core
11 from . import callbacks
12 from . import utils

c:\python\lib\site-packages\tf_explain\core_init_.py in
5 """
6 from .activations import ExtractActivations
----> 7 from .grad_cam import GradCAM
8 from .gradients_inputs import GradientsInputs
9 from .vanilla_gradients import VanillaGradients

c:\python\lib\site-packages\tf_explain\core\grad_cam.py in
10
11
---> 12 class GradCAM:
13
14 """

c:\python\lib\site-packages\tf_explain\core\grad_cam.py in GradCAM()
90
91 @staticmethod
---> 92 @tf.function
93 def get_gradients_and_filters(model, images, layer_name, class_index):
94 """

AttributeError: module 'tensorflow' has no attribute 'function'

The finetuned model generated 101 columns.
It should have been truncated at the end to only predict on the feature extraction layers to produce a 2048 dimension vector per image, which makes it consistent with the model before finetuning.

All subsequent notebooks have drastically different accuracy results maybe due to this issue

Hi,

This happened when running the 2-colab-what-does-my-neural-network-think.ipynb notebook in colabs.

When tensorflow 2.0 is installed in step 2 (!pip install tensorflow==2.0.0), it seems like you have to restart the runtime and run the code again. That is not entirely clear in the notebook but you can see a notice at the end of the installation.

If you do not restart, the process_image give and errorOperatorNotAllowedInGraphError(See https://stackoverflow.com/questions/57888872/how-to-fix-operatornotallowedingrapherror-error-in-tensorflow-2-0).

Not sure but maybe the code or instructions need an update to make this work from scratch. Hope this helps.

Seems like there is an issue reported in tensorflow 2.0, see tensorflow/tensorflow#32546.

The issue mentions also a gist which reproduces the exact same error, see https://colab.research.google.com/gist/oanush/0983d6584c1c687c0456fff142db4ad4/32546.ipynb

I have not made any changes to the code but when I run the below code

train, info_train, val, info_val, IMG_H, IMG_W, IMG_CHANNELS, NUM_CLASSES, NUM_EXAMPLES = get_dataset(dataset_name, 100)

I am getting an error saying
AssertionError: Unrecognized instruction format: NamedSplit('train')(tfds.percent[0:90])

Please help. Thank You

It seems to be missing in the repo?

Issue

https://github.com/PracticalDL/Practical-Deep-Learning-Book/blob/master/code/chapter-2/visualization.py

Run the following command to run the program:
python visualization.py --process image --path ../../sample-images/dog.jpg

The program will fail if you are running this locally (not on Colab) for the first time:

data_utils.py", line 251, in get_file
    urlretrieve(origin, fpath, dl_progress)
  File "/usr/lib/python3.7/urllib/request.py", line 247, in urlretrieve
    with contextlib.closing(urlopen(url, data)) as fp:
  File "/usr/lib/python3.7/urllib/request.py", line 222, in urlopen
    return opener.open(url, data, timeout)
  File "/usr/lib/python3.7/urllib/request.py", line 510, in open
    req = Request(fullurl, data)
  File "/usr/lib/python3.7/urllib/request.py", line 328, in __init__
    self.full_url = url
  File "/usr/lib/python3.7/urllib/request.py", line 354, in full_url
    self._parse()
  File "/usr/lib/python3.7/urllib/request.py", line 383, in _parse
    raise ValueError("unknown url type: %r" % self.full_url)
ValueError: unknown url type: '../imagenet_class_index.json'

Code
https://github.com/PracticalDL/Practical-Deep-Learning-Book/blob/master/code/chapter-2/visualization.py

CLASS_INDEX_PATH = '../imagenet_class_index.json'
...
 if CLASS_INDEX is None:
        fpath = get_file('imagenet_class_index.json',
                         CLASS_INDEX_PATH, cache_subdir='models')
        CLASS_INDEX = json.load(open(fpath))

Issue
TensorFlow's get_file() API lists the second parameter as URL not as relative path: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/keras/utils/data_utils.py#L164

Under the hood, it uses following code:

try:
        urlretrieve(origin, fpath, dl_progress)
      except HTTPError as e:
        raise Exception(error_msg.format(origin, e.code, e.msg))

The following one-liner will fix the issue:
CLASS_INDEX_PATH = 'https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json'

Note

1. If someone runs this example once using the one-liner fix, the current code as is would work because it uses cached data imagenet_class_index.json. Probably that is how the bug snuck in.

2. Easy fix, but I could not make pull request due to permission. You can technically enable PR by protecting your branch. This way any merge will be blocked until code owners approve the PR. This will help you to keep the code up-to-date by leveraging external PR. Just food for thought.

def plot_images(filenames, distances):
    images = []
    for filename in filenames:
        images.append(mpimg.imread(filename))
    plt.figure(figsize=(20, 10))
    columns = 4
    for i, image in enumerate(images):
        ax = plt.subplot(len(images) / columns + 1, columns, i + 1)

ax = plt.subplot(len(images) / columns + 1, columns, i + 1) should be ax = plt.subplot(len(images) // columns + 1, columns, i + 1), subplot requires integers for specifying number of rows/cols

the link under Pix2Pix goes to PosetNet instead of Pix2Pix

chapter 2 shows the dog with code
predict(img)

the github code doesn't include this line, but if we implement the code the following error is received.

NameError: name 'predict' is not defined

Hi, when trying to run the script to convert the video frames to heatmaps, I get the following OOM error on a P100 GPU:

tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[25088,4096] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc [Op:Add] name: fc1_24/kernel/Initializer/random_uniform/

Not sure why it happens, since it is just one image..
Seems memory previously allocated doesn't get deallocated.
Any recommendation ? Using tensorflow 2.0 inside Docker.

By the way, looks like a great book :)

Getting this message when running the code ch4-1
/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:6: TqdmDeprecationWarning: This function will be removed in tqdm==5.0.0
Please use tqdm.notebook.tqdm instead of tqdm.tqdm_notebook

the model doesn't seem to run with the current tqdm, i changed to tqdm.notebook.tqdm and it errors out. if i find a way to get through this I will post.

could not open the ipynb in Colab; please check