Are the labels used for training and the ones used for validation the same? I thought they should be the same; however, there seem to be a discrepancy in the labels that are available online. When I downloaded the imagenet 2012 labels for its validation data from the official website, I get labels that start with kit_fox as the first label, which matches the exact 2012's dataset validation images I downloaded from the official website. This is the example of the labels: https://gist.github.com/aaronpolhamus/964a4411c0906315deb9f4a3723aac57
However, for almost all the pretrained models, including those trained by Google, the imagenet labels they use for training, actually start with tench, tinca tinca instead. See here: https://gist.github.com/yrevar/942d3a0ac09ec9e5eb3a
Why is there such a huge discrepancy? Where did the 'tinca tinca' kind of labels come from?
If we use the first label mapping that corresponds to the actual validation images, we face another problem: 2 classes ("Crane" and "maillot") are actually duplicated, i.e. they have the same name but refer to different kind of crane - the mechanical crane and the animal crane - resulting in 100 image in 2 of the classes instead of the supposed 50. If we do not use the first mapping, where is a reliable source of the validation images that correspond to the second label mapping?
I have the same problem in my finetuning. You solve your problem change the name of classes tench, tinca tinca to the synset number. You can find here the mapping
Related
I want to detect digits on a display. For doing that I am using a custom 19 classes dataset. The choosen model has been yolov5-X. The resolution is 640x640. Some of the objets are:
0-9 digits
Some text as objects
Total --> 17 classes
I am having problems to detect all the digits when I want to detect 23, 28, 22 for example. If they are very close to each other the model finds problems.
I am using roboflow to create diferent folders in which I add some prepcocessings to have a full control of what I am entering into the model. All are checked and entered in a new folder called TRAIN_BASE. In total I have 3500 images with digits and the majority of variance is with hue and brightness.
Any advice to make the model able to catch all the digits besides being to close from each other?
Here are the steps I follow:
First of all, The use of mosaic dataset was not a good choice the purpose of detecting digits on a display because in a real scenario I was never gonna find pieces of digits. That reason made the model not to recognize some digits if it was not shure.
example of the digits problem concept
Another big improvement was to change the anchor boxes of the yolo model to adapt them to small objects. To know which anchor boxes I needed. Just with adding this argument to train.py is enought in the script provided by ultralitics to print custom anchors and add them to your custom architecture.
To check which augmentations can be good and which not, the next article explains it quite visually.
P.D: Thanks for the fast response to help the comunity gave me.
Pretty much brand new to ML here. I'm trying to create a hand-detection CoreML model using turicreate.
The dataset I'm using is from https://github.com/aurooj/Hand-Segmentation-in-the-Wild , which provides images of hands from an egocentric perspective, along with masks for the images. I'm following the steps in turicreate's "Data Preparation" (https://github.com/apple/turicreate/blob/master/userguide/object_detection/data-preparation.md) step-by-step to create the SFrame. Checking the contents of the variables throughout this process, there doesn't appear to be anything wrong.
Following data preparation, I follow the steps in the "Introductory Example" section of https://github.com/apple/turicreate/tree/master/userguide/object_detection
I get the hint of an error when turicreate is performing iterations to create the model. There doesn't appear to be any Loss at all, which doesn't seem right.
After the model is created, I try to test it with a test_data portion of the SFrame. The results of these predictions are just empty arrays though, which is obviously not right.
After exporting the model as a CoreML .mlmodel and trying it out in an app, it is unable to recognize anything (not surprisingly).
Me being completely new to model creation, I can't figure out what might be wrong. The dataset seems quite accurate to me. The only changes I made to the dataset were that some of the masks didn't have explicit file extensions (they are PNGs), so I added the .png extension. I also renamed the images to follow turicreate's tutorial formats (i.e. vid4frame025.image.png and vid4frame025.mask.0.png. Again, the SFrame creation process using this data seems correct at each step. I was able to follow the process with turicreate's tutorial dataset (bikes and cars) successfully. Any ideas on what might be going wrong?
I found the problem, and it basically stemmed from my unfamiliarity with Python.
In one part of the Data Preparation section, after creating bounding boxes out of the mask images, each annotation is assigned a 'label' indicating the type of object the annotation is meant to be. My data had a different name format than the tutorial's data, so rather than each annotation having 'label': 'bike', my annotations had 'label': 'vid4frame25`, 'label': 'vid4frame26', etc.
Correcting this such that each annotation has 'label': 'hand' seems to have corrected this (or at least it's creating a legitimate-seeming model so far).
This is a bit of an abstract question.
I have a group of 28x28 px images from certain people, and I would like to label that data with each person who wrote it. How would I go about labeling it for training and testing? This is my first neural network, and I'm having difficulty finding any tutorials that suit my particular need. It feels like most Data, like MNIST/EMNIST, are already labeled.
Some more info is that I'm using Python 3, and Keras with Tensorflow backend.
I am assuming that you know who wrote each image. Then this is a matter of associating that information (the class label) with each image. There are several ways of doing this. Two common approaches are:
Folder structure
Make a folder for each class (person), and put the images inside.
Folder contents:
john/01.png
john/02.png
jane/03.png
susan/...
CSV file
In this case the images can be all in one folder, and then a dedicate Comma-Separated-Values file is used to contain
Folder contents:
dataset.csv
images/01.png
images/02.png
images/03.png
images/....
dataset.csv contents:
filename,person
images/01.png,john
images/02.png,john
images/03.png,jane
...
The CSV approach is nice if you have additional data about each file that you want to store. For instance metadata that could be relevant such as who recorded the file, when was it recorded, with what kind of equipment, what locations etc.
Combinations of the two are also possible, of course.
I run caffe using an image_data_layer and don't want to create an LMDB or LevelDB for the data, But The compute_image_mean tool only works with LMDB/LevelDB databases.
Is there a simple solution for creating a mean file from a list of files (the same format that image_data_layer is using)?
You may notice that recent models (e.g., googlenet) do not use a mean file the same size as the input image, but rather a 3-vector representing a mean value per image channel. These values are quite "immune" to the specific dataset used (as long as it is large enough and contains "natural images").
So, as long as you are working with natural images you may use the same values as e.g., GoogLenet is using: B=104, G=117, R=123.
The simplest solution is to create a LMDB or LevelDB database of the image set.
The complicated solution is to write a tool similar to compute_image_mean, which takes image inputs and do the transformations and find the mean!
I just got an interview question.
"Assume you want to build a statistical or machine learning model, but you have very limited data on hand. Your boss told you can duplicate original data several times, to make more data for building the model" Does it help?
Intuitively, it does not help, because duplicating original data doesn't create more "information" to feed the model.
But is there anyone can explain it more statistically? Thanks
Consider e.g. variance. The data set with the duplicated data will have the exact same variance - you don't have a more precise estimate of the distrbution afterwards.
There are, however, some exceptions. For example bootstrap validation helps when evaluating your model, but you have very little data.
Well, it depends on exactly what one means by "duplicating the data".
If one is exactly duplicating the whole data set a number of times, then methods based on maximum likelihood (as with many models in common use) must find exactly the same result since the log likelihood function of the duplicated data is exactly a multiple of the unduplicated data's log likelihood, and therefore has the same maxima. (This argument doesn't apply to methods which aren't based on the likelihood function; I believe that CART and other tree models, and SVM's, are such models. In that case you'll have to work out a different argument.)
However, if by duplicating, one means duplicating the positive examples in a classification problem (which is common enough, since there are often many more negative examples than positive), then that does make a difference, since the likelihood function is modified.
Also if one means bootstrapping, then that, too, makes a difference.
PS. Probably you'll get more interest in this question on stats.stackexchange.com.