def score(data, hmm_depth=3, cache=None, obs=1000, smooth=True, check_len=True):
song = data if isinstance(data, list) else utils.parse_song(data)
song = utils.trim_song(song, length=2500)
song_len = len(song[0])
# I don't know of any reliable way to normalize the log-likelyhood
# for different length probabilities, so we're just going to limit
# things to a fixed number of observations to normalize the lenght
# of observations per file, intestead of normalizing the probabilities
# of different length songs
if check_len and song_len < obs:
print " !! %s is too short (%d)" % (data, song_len)
return None
scores = []
for x in range(0, obs):
frame = utils.sized_observation_from_index(song, start=x, length=hmm_depth)
frame_obs = frame.split(".")
top_frame = frame_obs[-1].split("|")
song_chunk = [[], [], []]
for a_frame in frame_obs[:-1]:
note_1, note_2, note_3 = a_frame.split("|")
song_chunk[0].append(note_1)
song_chunk[1].append(note_2)
song_chunk[2].append(note_3)
scores.append(score_transition(song_chunk, top_frame, smooth, cache))
return sum(scores)
def train_on_files(files, max_hmm_order=8):
for a_file in files:
if has_file_been_recorded(a_file):
print "Not recalculating counts for %s" % (a_file,)
continue
else:
print "Beginning to calculate counts for %s" % (a_file,)
record_file(a_file)
song = parse_song(a_file)
song = trim_song(song, length=2500)
song_len = len(song[0])
if song_len < 10:
print "Song is too short for consideration. May be a sound effect or something trivial. Ignoring."
continue
record_obs('S|S|S')
for x in range(0, song_len):
for y in range(0, max_hmm_order + 1):
if y > 0:
frame = sized_observation_from_index(song, start=x, length=y)
record_obs(frame)
else:
frame = serialize_observation(song, x)
commit()
print "finished calculating counts from %s" % (a_file,)
def train_on_files(files):
for a_file in files:
if has_file_been_recorded(a_file):
print "Not recalculating counts for %s" % (a_file,)
continue
else:
print "Beginning to calculate counts for %s" % (a_file,)
record_file(a_file)
song = parse_song(a_file)
song = trim_song(song, length=2500)
song_len = len(song[0])
if song_len < 100:
print "Song is too short for consideration. May be a sound effect or something trivial. Ignoring."
continue
for x in range(0, song_len):
frame = sized_observation_from_index(song, start=x, length=2)
frame_1, frame_2 = frame.split(".")
frame_1_chan_1, frame_1_chan_2, frame_1_chan_3 = frame_1.split("|")
frame_2_chan_1, frame_2_chan_2, frame_2_chan_3 = frame_2.split("|")
# We treat channel one as the most important one, so first
# record the counts of all channels in frame 1 transitioning
# into channel 1 in frame 2
record_cross_obs_count(frame_1_chan_1, 1, frame_2_chan_1, 1)
record_cross_obs_count(frame_1_chan_2, 2, frame_2_chan_1, 1)
record_cross_obs_count(frame_1_chan_3, 3, frame_2_chan_1, 1)
# Now also record the remaining channels trainsitioning to
# themselves in the next channel
record_cross_obs_count(frame_1_chan_2, 2, frame_2_chan_2, 2)
record_cross_obs_count(frame_1_chan_3, 3, frame_2_chan_3, 3)
# Last, record the transitioning from the assumed "melody" in the
# new frame to the assumed "background" channels in the new
# frame
record_inner_obs_count(frame_2_chan_1, frame_2_chan_2, 2)
record_inner_obs_count(frame_2_chan_1, frame_2_chan_3, 3)
commit()
print "finished calculating counts from %s" % (a_file,)
def score(data, cache=None, obs=1000, smooth=False, check_len=True):
song = data if isinstance(data, list) else utils.parse_song(data)
song = utils.trim_song(song, length=2500)
song_len = len(song[0])
# I don't know of any reliable way to normalize the log-likelyhood
# for different length probabilities, so we're just going to limit
# things to a fixed number of observations to normalize the lenght
# of observations per file, intestead of normalizing the probabilities
# of different length songs
if check_len and song_len < obs:
print " !! %s is too short (%d)" % (data, song_len)
return None
numerator_smooth, denom_smooth = (1, 128) if smooth else (0, 0)
scores = []
for x in range(0, obs):
frame = utils.sized_observation_from_index(song, start=x, length=2)
frame_1, frame_2 = frame.split(".")
scores.append(score_transition(frame_1.split("|"), frame_2.split("|"), smooth))
return sum(scores)
def get_begin_of_song(relative_file_path, depth):
song = utils.parse_song(relative_file_path)
song = utils.trim_song(song, length=2500)
begin = [channel[:depth] for channel in song]
return begin
import os
import midi
import utils
from utils import trim_song, parse_song, sized_observation_from_index
data_dir = os.path.join("data", "training_songs")
training_files = []
max_pitch = 0
min_pitch = 100
for root, dirs, files in os.walk(data_dir):
for name in [a_file for a_file in files if a_file[-4:] == ".mid"]:
relative_path = os.path.join(root, name)
print relative_path
song = parse_song(relative_path)
song = trim_song(song, length=2500)
song_len = len(song[0])
for x in range(0, song_len):
frame = sized_observation_from_index(song, start=x, length=1)
current_observation = [str(chan_data[x]) for chan_data in song.values()]
for i in current_observation:
print i
if i > max_pitch:
max_pitch = i
if i < min_pitch:
print i
min_pitch = i
print max_pitch
print min_pitch