The Google built-in object detection documentation/reference says the the num_classes argument should be set as follows:
E.g., for num_classes=5, the range of image/class/label in input tf.Example needs to be [0, 4].
Yet, most other resources (e.g., here) on how to create your own dataset in the object detection API world say that labels should start with 1, that is, for 5 classes they should be [1,5].
My questions are:
Is the example in the reference documentation correct, that is, should I use [0,4] for 5 classes?
Does it matter at all, i.e., can this break the training procedure?
Is the "built-in object detection" algorithm special in other ways or can I follow the "using your own dataset" function to create my TFrecord files?
Seems like the labels should be [1,5].
The documentation has changed :)
See the updated documnetation
under --> Hyperparameters --> num_classes
Related
Initially, I created an interactive map of the UK Postcode area where an individual area is color represented based on its value (e.g. population in that post code area) as following.
from bokeh.plotting import figure
from bokeh.palettes import Viridis256 as palette
from bokeh.models import LinearColorMapper
from bokeh.models import ColumnDataSource
import geopandas as gpd
shp = 'file_path_to_the_downloaded_shapefile'
#read shape file into dataframe using geopandas
df = gpd.read_file(shp)
def expandMultiPolygons(row, geometry):
if row[geometry].type = 'MultiPolygon':
row[geometry] = [p for p in row[geometry]]
return row
#Some rows were in MultiPolygons instead of Polygons.
#Expand MultiPolygons to multi rows of Polygons
df = df.apply(expandMultiPolygons, geometry='geometry', axis=1)
df = df.set_index('Area')['geometry'].apply(pd.Series).stack().reset_index()
#Visualize the polygons. To visualize different colors for different post areas, I added another column called 'value' which has some random integer value.
p = figure()
color_mapper = LinearColorMapper(palette=palette)
source = ColumnDataSource(df)
p.patches('x', 'y', source=source,\
fill_color={'field': 'value', 'transform': color_mapper},\
fill_alpha=1.0, line_color="black", line_width=0.05)
where df is a dataframe of four columns : post code area, x-coordinate, y-coordinate, value (i.e. population).
The above code creates an interactive map on a web browser which is great but I noticed the interactivity is not very smooth in speed. If I zoom in or move the map, it renders slowly. The size of the dataframe is only 1106 rows, so I'm quite confused why it is so slow.
As one of the possible solutions, I came across with datashader (https://datashader.readthedocs.io/en/latest/) but I find the example script is quite complicated and most of them are with holoview package on Jupyter notebook but I want to create a dashboard using bokeh.
Does anyone advise me in incorporating datashader into the above bokeh script? Do I need a different function within datashader to create the shape map instead of using bokeh's patches function?
Any suggestion would be highly appreciated!!!
Without the data file involved, I can't answer your question directly, but can offer some observations:
Datashader is unlikely to be of value for this purpose, because datashader does not currently have any support for rendering polygons. As a rule of thumb, Datashader is designed to aggregate your data, and if it's already aggregated, Datashader won't normally be of help. Here your data is aggregated by postcode, which datashader can't process, but if you had the original data per person it would be happy to render it.
If you prefer working with Bokeh directly rather than via the higher-level HoloViews/GeoViews interface, I'd recommend folllwing Matt Rocklin's work on accelerating geopandas; his approach should be very fast for your purpose.
All that said, HoloViews, and GeoViews should be a convenient way to work with Bokeh in general, whether or not you want to create a dashboard. E.g. the 2017 JupyterCon tutorial shows how to make a simple Bokeh dashboard using both libraries. It doesn't cover shape files, but those are covered in other GeoViews examples.
As mentioned in my comment, I believe that the complexity of your polygons might cause your problem. The file you linked to contains several shapefile of different sizes and complexities. You can simplify those, i.e. reduce the number of points for each polygon. This can change how they look. It can range from almost no difference over a bit more "edginess" to an angular appearance. This depends on the level of simplification you chose. Depending on your needs you can chose different levels of simplicity.
I know of three easy options to get this done:
GUI: Try QGis. It is a great opensource tool for geospatial data processing. Load your Shapefile as a new layer. Then use the "Simplify Geometries" tool under the Vector menu.
Command-Line: GDAL is an open-source library. It comes with an useful command-line tool. You can use it like this: ogr2ogr outfile.shp infile.shp -simplify 0.000001
Online: Visit mapshader. Import your file. Select simplify and chose your level. Then, export the result. What I really like here is that your file is rendered instantly. Hence, you can immediately see the result of your simplification.
Other than that, you should also update your bokeh version. It gets updated regularly and there have been some performance improvements since.
Using HoloViews or GeoViews will not positively affect your performance. Thus, it is not related to your issues. I guess #James A. Bednar was just giving some side advice there.
I found a way to speed up the interactive visualization of the UK map as I move the slider.
I created individual image (in 2D) for a different value of slider first and updated the map using the 2D images instead of using bokeh's patches function.
Since the images are in array format, it is much faster to update the image while changing the values in the slider. one downside in this method is that I can no longer use hover function on the UK map.
I referred to the following url to convert polygon information into arrays: https://gist.github.com/brendancol/db030013e981c46acb2886060dde607e#file-rasterio_datashader_polygons-py-L35
I run SegNet on my own dataset (by Segnet tutorial). I see great results via test_segmentation.py.
my problem is that I want to see the real net results and not test_segmentation own colorisation (via classes).
for example, if I have trained net with 2 classes, so after the train I will see not only 2 colors (as we see with the classes), but we will see the real net color segmentation ([0.22,0.19,0.3....) lighter and darker as the net see it]
I hope that I explained myself well. thanks for helping.
You could use a python script to achieve what you want. Take a look at this script.
The command out = out['argmax'], extracts the raw output, so you can get a segmentation map with 'lighter and darker' values as you wanted.
When you say the 'real' net color segmentation I will assume that you mean the probability maps. Effectively the last layer will have one map for every class; and if you check the function predict in inference.py, they take the argmax; that is the channel (which represents the class) with the highest probability. If you want to get these maps, you just have to get the data without computing the argmax; something like:
predicted = net.blobs['prob'].data
I solve it. the solution is to range cmin and cmax from 0 to 1 in the scipy saving method. for example: scipy.misc.toimage(output, cmin=0.0, amax=1).save(/path/.../image.png)
I'm trying to do binary LSTM classification using theano.
I have gone through the example code however I want to build my own.
I have a small set of "Hello" & "Goodbye" recordings that I am using. I preprocess these by extracting the MFCC features for them and saving these features in a text file. I have 20 speech files(10 each) and I am generating a text file for each word, so 20 text files that contains the MFCC features. Each file is a 13x56 matrix.
My problem now is: How do I use this text file to train the LSTM?
I am relatively new to this. I have gone through some literature on it as well but not found really good understanding of the concept.
Any simpler way using LSTM's would also be welcome.
There are many existing implementation for example Tensorflow Implementation, Kaldi-focused implementation with all the scripts, it is better to check them first.
Theano is too low-level, you might try with keras instead, as described in tutorial. You can run tutorial "as is" to understand how things goes.
Then, you need to prepare a dataset. You need to turn your data into sequences of data frames and for every data frame in sequence you need to assign an output label.
Keras supports two types of RNNs - layers returning sequences and layers returning simple values. You can experiment with both, in code you just use return_sequences=True or return_sequences=False
To train with sequences you can assign dummy label for all frames except the last one where you can assign the label of the word you want to recognize. You need to place input and output labels to arrays. So it will be:
X = [[word1frame1, word1frame2, ..., word1framen],[word2frame1, word2frame2,...word2framen]]
Y = [[0,0,...,1], [0,0,....,2]]
In X every element is a vector of 13 floats. In Y every element is just a number - 0 for intermediate frames and word ID for final frame.
To train with just labels you need to place input and output labels to arrays and output array is simpler. So the data will be:
X = [[word1frame1, word1frame2, ..., word1framen],[word2frame1, word2frame2,...word2framen]]
Y = [[0,0,1], [0,1,0]]
Note that output is vectorized (np_utils.to_categorical) to turn it to vectors instead of just numbers.
Then you create network architecture. You can have 13 floats for input, a vector for output. In the middle you might have one fully connected layer followed by one lstm layer. Do not use too big layers, start with small ones.
Then you feed this dataset into model.fit and it trains you the model. You can estimate model quality on heldout set after training.
You will have a problem with convergence since you have just 20 examples. You need way more examples, preferably thousands to train LSTM, you will only be able to use very small models.
I'm a little confused about online kmeans clustering. I know that it allows me to cluster with just one data at a time. But,is this all limited to one session? Suppose that I have a bunch of data clustered via this method and I get the clustered data result, would I be able to add more data to the cluster in the future?
I've also been looking for implementations of this code, and to no avail. Anyone know of any?
Update:
To clarify more. Here is how my code works right now:
Image is taken from live video feed, once enough pictures are saved, get kmeans of sift features.
Repeat step 1, a new batch of live feed pictures, get kmeans again. Combine the kmeans vectors with the previous kmeans like :[A B]
You can see that this is bad, because I quickly get too much clusters, and each batch of clusters will definitely have overlaps with another batch.
What I want:
Image taken from live video feed, once pics are saved, get kmeans
Repeat step 1, get kmeans again, which updates and adds new clusters to the previous cluster.
Nothing that I've seen could accommodate that, unless I'm just not understanding them correctly.
If you look at the original (!) publications, the method proposed by MacQueen - where the name k-means comes from - was in fact an online algorithm. I'm not sure if MacQueen did multiple passes over the data to improve the result. I believe he used a single pass, and objects would never be reassigned to a different cluster. If so, it was already an online algorithm!
Means are commonly computed as sum / count. This is not very sensible from a numerical point of view. E.g. in the classic Knuth book you can find a method for incrementally updating means. Wikipedia has it also.
Things get slightly more complicated once you actually want to reassign earlier points. But usually in a streaming context you do not know the previous points, so you cannot do that anyway.
I am using CERN's ROOT framework (required), and I want to take data from a TNtuple and graph it. I can either graph the data when I create the TNtuple, or after I write it to a .root file. Some of the support documentation suggested that I create a TTree, but that seemed like it might be overkill/roundabout since I wouldn't be using it for anything else (and the TNtuple fulfills all of my other requirements). Does anyone have a better suggestion for how to extract data from the TNtuple and graph it?
As TNtuple inherits from TTree, you can use all the methods presented in the support documentation for TTrees directly on the TNtuple.
This especially means that you can use TTree::Draw() which is typically more than sufficient for quickly graphing the data. This function is documented here.
For more elaborate plots you will have to read the data from the TNtuple event by event and feed it to your favorite graphing tool in ROOT. This again follows the basic principles from a tree. The best example I could find on the ROOT homepage is in the user manual, section trees in the paragraph "Reading the Tree".
The methods used to create histograms and plots for TNtuples is essentially the same as TTrees. The code:
ntuple->Draw("var");
will create a histogram of the variable var stored in the Ntuple. If you want to plot one variable in the Ntuple as a function of another, use
ntuple->Draw("xVar:yVar");
You can do fancier things such as creating plots only when a logical condition is satisfied. For example, suppose you want a histogram of var1 only when var2 is greater than 2 and var3 is less than 0.
ntuple->Draw("var","var2 > 2 && var3 < 0");
By plotting in this way, ROOT automatically sets the binning and range for the x-axis. If you wish to control these features yourself, use
ntuple->Draw("var >> hist(Nbins,xmin,xmax)");
This creates the object hist, which you treat as a usual histogram object in ROOT. As stated in the previous post, this is documented in the ROOT manual along with several other features and tools. Unfortunately, the manual doesn't always give clear explanations.
{
ntuple->Draw("py:px","px>py","goff");
TGraph *gr = new TGraph(ntuple->GetSelectedRows(),ntuple->GetV2(), ntuple->GetV1());
gr->Draw("AP");
}