smenon8 / animalwildlifeestimator Goto Github PK

The getAnnotID() method still does a one-to-one comparison.
Fix the method to get a faster performing method to extract all annotations at once and give a map file directly as an output and test for any impacts across code base.

Bottlenecks:
Annotation ID extraction phase of JobsMapResultsFilesToContainerObjs.py

Issue reported on stack-overflow.

Can be tracked at: http://stackoverflow.com/questions/39339047/plotly-need-to-reset-the-color-scheme-of-charts-to-original

While bar charts seems to be working just fine, histograms plots are the real issue.

To be sure, test for line charts and scatter plots.

The new modifications to GetPropertiesAPI.py for supporting multiple instance query to IBEIS returns flat objects instead of embedding it inside a list object.

The est_age_months property passed through getAgeFeatureReadableFmt always returns unknown. Possibly because of flat objects.

The changed output json format has also to be tested in all scripts to ensure it is working fine and no changes are needed.

Remove all hardcoded code in VisualizeResults.ipynb and replace them with reusable methods. There should be a workflow ideally which should be followed for any task.

• Creating experiments using Mechanical Turk.
• Getting data from IBEIS for statistical analysis.
• Create data files required for analysis. Parameterize functions to accept the required attributes.
• Create a machine learning task, specify training data and return a classifier.
• Have an API for pulling image recognition from MS AI tool in Python. Currently in R - developed by Lucas.
• Apply the trained classifier on the test data.
• Completely parameterized method for calculating population estimates using Mark-Recapture on predicted shared data.

Commit #4d04ef9
This is a temporary fix to the issue where the adult attribute was repeated for multiple times in train_x.

To recreate the issue,

def trainTestSplitter(gidAttribDict,allAttribs,trainTestSplit):
    df = pd.DataFrame(gidAttribDict).transpose()
    df = df[allAttribs + ["TARGET"]] # Rearranging the order of the columns

    attributes = df.columns[:len(allAttribs)] # all attributes

    dataFeatures = df[list(set(attributes))] # remove the type-casting to recreate the issue
    targetVar = df['TARGET']

    return train_test_split(dataFeatures, targetVar, test_size=trainTestSplit,random_state=0)

The number of training attributes is 88 while testing attributes are just 86. It is caused by repeating adults.

This was verified.

Asked Jason for comments on impacts on way of access and any API namespace changes.

The method getCountingLogiccurrently resides in the script DeriveFinalResultSet.py.

This method is highly functional in other scripts and it doesn't really fit into the context of deriving final result.

Another issue, to use the getCountingLogic() method entire DeriveFinalResultSet.py has to be imported posing minor performance issues.

MarkRecapHelper.py currently handles the case of excluding zoo animals by hardcoding area parameters.

handle_zoo_animals(bottom_left_lat, bottom_left_long, top_right_lat, top_right_long)
should return animals that are found within this rectangular area in the map

It seems some of the performance comparison charts only extracts the last column slice of the data-frame.

For instance,
Mix and match regression results for beauty attributes

In the diagram, there should have been a cross comparison between all the regressors using beauty attributes, instead only elastic net is shown.

There is no current parameter support to the method genNumIndsRankList().
Currently this method will generate the rank list of number of individuals in an image by share proportion.

Marked for future improvements.

There is a major chance that the order of attributes in the clfObj.test_x data frame and clfObj.train_x is not the same.
There has to be some form of external trigger that will help counter this effect. This does mean that the results of the synthetic experiments may be incorrect.

Script: PopulationEstimatorFromClf.py > trainTestClf() and trainTestRgrs()

train_x.head()
	ground	arthropod	junk	dirt	plain	outdoor	ok	unknown	plant	tree	...	
test_x.head()
        zebra_plains	giraffe_masai	UNKNOWN SEX	Female	UNKNOWN NAME	Male	unknown	infant	juveniles- two year old	adult	...

The EXIF and FTR files for GGR and GZC are in different cases. The headers of the fields need to match.

Preferably, everything should be in UPPER CASE

This only happens in the case where the top k images scored by the predicted classifier probabilities are used.

The way to go about will be to actually check if all the classification methods really predict the same result for images irrespective of the matter what algorithm is being used.

The chances of a code bug is low since the same thing cannot be recreated using "bottom k shares" or "random k shares".

createDataFlDict()
While creating data for regression, presence of individual name attribute poses a problem.
INDIVIDUAL_NAME is not a part of the NON_SPARSE attribute list, but even then the result set contains the individual names. This is because of the hard-coded ftrs list in the method.

• ClassifierHelperAPI.py > getMasterData()
• ClassifierHelperAPI.py > createDataFlDict()

Comes down to writing good comments and generating pydoc for every script file