When training a tf 2.0 YoloV3 model Chapter 5 mentions training using VGG16 or other pretrained NN as a backbone after the head is deactivated. How do you format the y_true value i.e. (batch_size, grid, grid, anchors, (x1, y1, x2, y2, obj, cls)). Batch size is obvious but I wanted to see an example of how the rest would look and how I should format the data to use it in a model.fit operation, for example. I have annotated the images and have the x and y values for the bounding boxes and obviously the class, but I'm not sure how to supply obj value ( and all the other values when transitioning from the backbone to the YoloV3 head). Is obj (objectness) always going to be 1 for y_true since there is always an object in the bounding box for y_true? Is the class simply going to be the corresponding text label used in the coco.names file or some other format. Is there a way to view a full example of the proposed implemention i.e VGG then YoloV3 for example.

I have been trying to debug this for a few days to no avail.
I have been running the notebooks as they are.
I'm getting
Epoch 0/90: loss = nan; v-loss = nan; acc = 0.389; v-acc = 0.366; mIoU = 0.020; v-mIoU = 0.019
as output from training and can't figure out what is the problem

Hi,
I have the same issue as previous "Unable to prepare cityscape-dataset #8" on Jun 29, 2019.

I have verified the environment has been se by:
import os
print(os.environ['CITYSCAPES_DATASET'])
it retuned
C:\Users\c5911\tensorflow_datasets\cityscapes.

However, two scripts
!csCreateTrainIdLabelImgs
or
!csViewer (ran under Jupyter notebook)
still returned:
ERROR: Did not find any files. Please consult the README.

I ran csViewer at prompt command, it also showed "The data was not found".

Under "C:\Users\c5911\tensorflow_datasets\cityscapes", there are two sub folders, getFine and leftImg8bit.

Anything could be wrong?

Thanks.

The number of filters for each macroblock is 64, but according to the diagrams and white paper, it should be increasing.

this is due to using np.random.rand which outputs always positive numbers.

what is the shape of ground truth numpy color map for each image ? chapter 6 "preparing data for smart car apps".

Hi,

running the script (Mac, Jupyter-notebook, Python 2), I am unable to prepare the dataset.
There is the message "ERROR: Did not find any files. Please consult the README.", which I can not interpret.
Running the script step by step on the terminal, there is the same problem.
Has anyone an idea how to solve this?

Searching the internet, I only found this link:
mcordts/cityscapesScripts#73
However, their solution does not work for me.

Unfortunately, without this data, I can not further work on the chapter of this nice book.

With kind regards

chapter 6 material: in fcn.py
in the model definition layer, f5_conv3 ,seems to not be included in the computational graph

Hi
I am running ch6. nb5 build and train.
At fcn8s_model.fit, I have a ValueError: Shapes (None, 512, 512, 19) and (None, 512, 512, 1) are incompatible (after "C:\Users\c5911\anaconda3\lib\site-packages\tensorflow\python\framework\tensor_shape.py:1134 assert_is_compatible_with
raise ValueError("Shapes %s and %s are incompatible" % (self, other))").
fcn8s_model.summary() showed
conv2d_transpose_2 (Conv2DTrans (None, 512, 512, 19) 92435 add_1[0][0]

Total params: 14,983,177
Trainable params: 14,983,177
Non-trainable params: 0

Why were there incompatible shape?

Thanks,
Gary

P.S.: I changed: visual_val_dataset = cityscapes_input_fn(
root_folder=CITYSCAPES_FOLDER, resize_to=image_size, batch_size=num_show,
shuffle=True, num_epochs=1, augment=False, seed=random_seed, blurred=False) #blurred changed from "True" to "False".

Summing the weights of the input and weight vector is not correct.
The inputs are indexed starting at 0, your x0 w0 in all text and pictures.
The conclusion that the dot product is x1 * w1 + x2 * w2 is therefore incorrect it should be
x0 * w0 + x1 * w1 .... etc

For other books reporting this kind or error gets you street cred as a programmer.
https://en.wikipedia.org/wiki/Knuth_reward_check

Hello!

I ran your notebook for chapter8 and I would like to use the trained model for new videos. I am aware this may sound like a super idiotic question, but would you help me using the trained model? how do i save it for later use? and then how would i 'run' it on new videos?

Thanks for your help, and apologies for the basic question,

Veronica.

I have downloaded weight file (yolov3.weights) and tried to convert it following the instructions. But convert.py program did not produce yolov3.tf. Instead, the program produced the following files:

yolov3.tf.index (24KB),
yolov3.tf.data-00000-of-00002 (62KB), and
yolov3.tf.data-00001-of-00002 (237MB).

Each number within the parenthesis is the size of each file.

I tested the program using the docker (tensorflow:latest-gpu) on Ubuntu 18.04.

Just in case, I attach the screen capture of the log that is written by convert.py.

screen_capture_of_convert_weight.txt