def optimize(classifier, feature_dict):
# full feature set
all_features = ['eng', 'chroma', 'keystrength', 'brightness', 'zerocross',
'mfc', 'roughness', 'inharmonic', 'tempo', 'key']
# get the power set of all_features
all_features = list(powerset(all_features))
random.shuffle(all_features)
# initialize
idx = 0
results = {}
max_acc = 0
max_feat = []
start = time.clock()
# loop through each possible feature set
for feature_set in all_features:
if len(feature_set) == 0:
continue
p1, r1, = leave_one_out(feature_dict, glob, classifier, classifier.__name__)
acc = vis.present_results(p1, r1, classifier.__name__, print_results=False, show_results=True)
# print the running best combination
if acc > max_acc:
max_acc = acc
max_feat = feature_set
print()
print(max_acc, max_feat)
results[feature_set] = acc
idx += 1
t = time.clock() - start
remaining = t * (len(all_features) / idx) - t
sys.stdout.write("\r%d / %d permutations (%d:%02d left)" % (idx, len(all_features), remaining // 60, remaining % 60))
print("")
return sorted(results, key=results.get, reverse=True)
# trim to only key and tempo for each song
data_trimmed = extract_params(['key', 'tempo'], data_raw)
data_genres = extract_params(['class'], data_raw)
# select training set using boolean mask of a random subset
training_indices = random.sample(range(data_trimmed.shape[0]), data_trimmed.shape[0]//2)
mask = np.zeros(data_trimmed.shape[0], dtype=bool)
mask[training_indices] = True
training_params = data_trimmed[mask,:]
training_values = data_genres[mask]
# build model
model = build_model(training_params, training_values)
# ** TEST **
# only predict genres for those not used by mask
guessed_genres = model.predict(data_trimmed[~mask])
correct_genres = data_genres[~mask]
guessed_genres = [int_as_genre(int(genre)) for genre in guessed_genres]
correct_genres = [int_as_genre(int(genre)) for genre in correct_genres]
print(data_raw[0]['keystrength'][0][0])
# evaluate and graph
vis.present_results(guessed_genres, correct_genres)
# all params availible
params = ['eng', 'chroma', 'keystrength', 'zerocross', 'tempo', 'mfc', 'brightness', 'roughness', 'inharmonic']
param_dict = dict()
for param in params:
if param in ['inharmonic']:
param_dict[param] = (True, True, False)
elif param in ['chroma', 'keystrength']:
param_dict[param] = (True, False, True)
else:
param_dict[param] = (True, True, True)
start = time.clock()
p, r = leave_one_out(param_dict, glob, 1, "K Neighbors (Opt)")
t = time.clock() - start
vis.present_results(p, r, "K Neighbors (Opt)", t, print_results=True, show_results=True)
start = time.clock()
p, r = leave_one_out(param_dict, glob, 2, "Voting")
t = time.clock() - start
vis.present_results(p, r, "Voting", t, print_results=True, show_results=True)
start = time.clock()
p, r = leave_one_out(param_dict, glob, 3, GaussianNB.__name__)
t = time.clock() - start
vis.present_results(p, r, GaussianNB.__name__, t, print_results=True, show_results=True)
start = time.clock()
p, r = leave_one_out(param_dict, glob, 4, "Stochastic Gradient Descent (Opt)")
t = time.clock() - start
vis.present_results(p, r, "Stochastic Gradient Descent (Opt)", t, print_results=True, show_results=True)