add things resources from the last 5 years over machines

data_science/run_eval.py

@@ -0,0 +1,226 @@
+#!/usr/bin/env python
+
+"""
+    Run svm_light, parse its stdout, calculate
+    ML scores, HDFS copy data to local.
+"""
+
+import sys
+import os
+import getpass
+import subprocess
+import shutil
+import math
+
+
+def delete_dir(dir_path):
+    '''
+        Remove a directory.
+
+        Args:
+            dir_path: full path to the directory.
+    '''
+    if os.path.isdir(dir_path):
+        shutil.rmtree(dir_path)
+
+
+def usage():
+    '''
+        Handle the CLI arguments.
+    '''
+    args = sys.argv
+    if len(args) != 3:
+        print("Usage: ./runEval <method> <version>")
+        sys.exit(2)
+    return args[1], args[2]
+
+
+def create_dir(dir_path):
+    '''
+        Create a a directory.
+
+        Args:
+            dir_path: full path to the directory.
+    '''
+    if not os.path.exists(dir_path):
+        os.makedirs(dir_path)
+
+
+def run_svm_classify(test_data, svml_model, svml_eval):
+    '''
+        Spawn a subprocess to run svm_classify binary.
+
+        From svm_classify.c, svm_light usage requires the following
+        arguments: example_file model_file output_file.
+
+        Args:
+            test_data: path_to_feature/test.dat
+            svml_model: something like ~/data/models/svmlight/method/version/model
+            svml_eval: something like ~/data/models/svmlight/method/version/eval
+
+        Returns:
+            Strings with stdout and stderr so that it can be parsed later.
+    '''
+    p = subprocess.Popen(['./models/svm_classify', \
+            '{0}'.format(test_data), \
+            '{0}'.format(svml_model),\
+            '{0}'.format(svml_eval)],\
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE)
+    out, err = p.communicate()
+    return out, err
+
+
+def paste_data(test_data, svml_eval, final_eval, svml_alpha, final_alphas, out):
+    '''
+        Copy all eval and alpha data from results to local files.
+
+        Args:
+            src and dst paths.
+    '''
+    # Copy all eval data.
+    with open(test_data, 'r') as ft:
+        test_data = ft.readlines()
+
+    with open(svml_eval, 'r') as fe:
+        eval_data = fe.readlines()
+
+    with open(final_eval, 'a') as f:
+        for line in test_data:
+            f.write('{0}\n'.format(line))
+        for line in eval_data:
+            f.write('{0}\n'.format(line))
+
+    # Copy all alpha data.
+    with open(svml_alpha, 'r') as fa:
+        alpha_data = fa.readlines()
+
+    with open(final_alphas, 'a') as f:
+        for line in alpha_data:
+            f.write('{0}     {1}\n'.format(line, out))
+
+
+def parse_svmlight_output(out):
+    '''
+        Parse the svm_light stdout string for an example
+
+        Returns:
+            c: counts
+            p: precision
+            r: recall
+    '''
+    c = out.split('OK. (')[1].split(' support')[0]
+    pr = out.split('Precision/recall on test set: ')[1].split(' support')[0].strip()
+    p, r = pr.split('/')
+    p = float(p.strip('%').strip()) / 100
+    r = float(r.strip('%').strip()) / 100
+
+    return c, p, r
+
+
+def hdfs_copy_data(home_dir, method, version):
+    '''
+        Run CLI HDFS commands to clean up and save data.
+    '''
+    os.system('hdfs dfs -rm /data/shared/structdata/modelOutput/{0}/{1}/scores'.format(method, version))
+    os.system('hdfs dfs -rm /data/shared/structdata/modelOutput/{0}/{1}/alphas'.format(method, version))
+
+    os.system('hdfs dfs -mkdir /data/shared/structdata/modelOutput/{0}/{1}'.format(method, version))
+
+    os.system('hdfs dfs -copyFromLocal {0}/data/eval/{1}/{2}/alphas \
+            /data/shared/structdata/modelOutput/{3}/{4}/alphas'.format(home_dir, version, method, method, version))
+
+    os.system('hdfs dfs -copyFromLocal {0}/data/eval/{1}/{2}/eval \
+            /data/shared/structdata/modelOutput/{3}/{4}/scores'.format(home_dir, version, method, method, version))
+
+def calculate_scores(list_of_scores):
+    '''
+        Calculate the mean of a given list of scores,
+        taking care of any nan or 0 division.
+    '''
+    c, score = 0, 0
+    for i in list_of_scores:
+        if not math.isnan(i):
+            c += 1
+            score += i
+    if c > 0:
+        return score / c
+    else:
+        return 0
+
+
+def calculate_f1(precision, recall):
+    '''
+        Calculates the f1-score as the harmonic
+        mean of precision and recall.
+    '''
+    if precision + recall < 1:
+        return 0
+    else:
+        return  2 / (1/precision + 1/recall)
+
+
+if __name__ == '__main__':
+
+    # Grab the CLI arguments.
+    METHOD, VERSION = usage()
+
+    # Setup output dirs.
+    home_dir = os.path.join('/home', getpass.getuser())
+    final_dir = os.path.join(home_dir, 'data/eval', VERSION, METHOD)
+    final_alphas = os.path.join(final_dir, 'alphas')
+    final_eval = os.path.join(final_dir, 'eval')
+
+    delete_dir(final_alphas)
+    delete_dir(final_eval)
+    create_dir(final_dir)
+
+    # Loop over the attributes and features.
+    training_data_dir = os.path.join(home_dir, 'data/training_data/', VERSION, METHOD)
+
+    for attribute in os.listdir(training_data_dir):
+
+        attribute_path = os.path.join(training_data_dir, attribute)
+        counts = 0
+        precision, recall = [], []
+
+        for feature in os.listdir(attribute_path):
+
+            # Create all the paths in use.
+            out = os.path.join(VERSION, METHOD, attribute, feature)
+            svmlight = os.path.join(home_dir,'data/models/svmlight', out)
+            svml_model =  os.path.join(svmlight, 'model')
+            svml_eval = os.path.join(svmlight, 'eval')
+            svml_alpha = os.path.join(svmlight, 'alphas')
+            test_data = os.path.join(attribute_path, feature, 'test.dat')
+
+            # Run svm_classify.
+            out, err = run_svm_classify(test_data, svml_model, svml_eval)
+
+            # Save current results.
+            paste_data(test_data, svml_eval, final_eval, svml_alpha, final_alphas, out)
+
+            # Parse output from svm_classify to print to stdout.
+            if err:
+                print('Error: {0}'.format(err))
+
+            # Get Train counts, Test counts, Accuracy, Precision, Recall.
+            c, p ,r = parse_svmlight_output(out)
+
+            counts += int(c)
+            precision.append(p)
+            recall.append(r)
+
+        attribute_precision = calculate_scores(precision)
+        attribute_recall = calculate_scores(recall)
+        attribute_f1 = calculate_f1(attribute_precision, attribute_recall)
+
+        print("{: <20} Counts: {: <20} Precision: {: <20} Recall: {: <20} F1-score: {: <20}".format(attribute.title(), \
+            counts, round(attribute_precision, 4), round(attribute_recall, 4), round(attribute_f1, 4)))
+
+
+    # Copying results from remote hdfs.
+    print("\nCopying results to hdfs")
+    hdfs_copy_data(home_dir, METHOD, VERSION)
+    print("\nDone!".format())
+