def test_saved_model_folder(dirname,feats,output,filt=False):
"""
Test a saved model by loading it and applying to features.
Output is the output file, used with print_write()
RETURN
avgerage dist
nPatterns
nIters
totaltime
"""
print_write('*** MODEL SAVED IN: '+dirname+' ***',output)
if filt:
print_write('model uses FILTERING',output)
# load model
model = ANALYZE.unpickle(os.path.join(dirname,'model.p'))
print_write('model loaded',output)
# find nIters (#tracks), nPatterns, totaltime
nIters, nPatterns, totalTime = ANALYZE.traceback_stats(dirname)
if filt:
nPatterns = model._nPatternUsed
print_write('nIters (=nTracks): '+str(nIters),output)
print_write('nPatterns: '+str(nPatterns),output)
print_write('total time ran: '+str(totalTime),output)
# predict
best_code_per_p, dists = model.predicts(feats)
print_write('prediction done, avg. dist: '+str(np.average(dists)),output)
# return
return np.average(dists),nPatterns,nIters,totalTime
def test_saved_model_folder(dirname, feats, output):
"""
Test a saved model by loading it and applying to features.
Output is the output file, used with print_write()
RETURN
avgerage dist
nPatterns
nIters
totaltime
"""
print_write('*** MODEL SAVED IN: ' + dirname + ' ***', output)
# load model
model = ANALYZE.unpickle(os.path.join(dirname, 'model.p'))
print_write('model loaded', output)
# find nIters (#tracks), nPatterns, totaltime
nIters, nPatterns, totalTime = ANALYZE.traceback_stats(dirname)
print_write('nIters (=nTracks): ' + str(nIters), output)
print_write('nPatterns: ' + str(nPatterns), output)
print_write('total time ran: ' + str(totalTime), output)
# predict
best_code_per_p, dists = model.predicts(feats)
print_write('prediction done, avg. dist: ' + str(np.average(dists)),
output)
# return
return np.average(dists), nPatterns, nIters, totalTime
def train(savedmodel, expdir='', pSize=8, usebars=2, keyInv=True,
songKeyInv=False, positive=True, do_resample=True, partialbar=0,
lrate=1e-5, nThreads=4, oracle='EN', artistsdb='', matdir='',
nIterations=1e7, useModel='VQ', autobar=False, randoffset=False):
"""
Performs training
Grab track data from oracle
Pass them to model that updates itself
INPUT
savedmodel - previously saved model directory, to restart it
or matlab file, to start from a codebook
expdir - experiment directory, where to save experiments
pSize - pattern size
usebars - how many bars per pattern
keyInv - perform 'key invariance' on patterns
positive - replace negative values by 0
do_resample - if false, crop
partialbar - actual size, divides pSize
lrate - learning rate
nThreads - number of threads for the oracle, default=4
oracle - EN (EchoNest) or MAT (matfiles)
artistdb - SQLlite database containing artist names
matdir - matfiles directory, for oracle MAT
nIterations - maximum number of iterations
useModel - which model to use: 'VQ', 'VQFILT'
Saves everything when done.
"""
# create the StatLog object
statlog = StatLog()
# count iteration
# main - iteration since this instance started
# global - iteration adding those from the savedmodel
# last_printed - for verbose purposes
main_iterations = 0
global_iterations = 0
last_printed_iter = 1
# start from saved model
if os.path.isdir(savedmodel):
# load model
assert os.path.exists(os.path.join(savedmodel,'model.p')),'loading saved model, model.p does not exist? %s' % (os.path.join(savedmodel,'model.p'),)
f = open(os.path.join(savedmodel,'model.p'),'r')
model_unp = pickle.Unpickler(f)
model = model_unp.load()
f.close()
# load params, savedmodel will be modified
assert os.path.exists(os.path.join(savedmodel,'params.p')),'loading saved model, params.p does not exist? %s' % (os.path.join(savedmodel,'params.p'),)
f = open(os.path.join(savedmodel,'params.p'),'r')
param_unp = pickle.Unpickler(f)
oldparams = param_unp.load()
f.close()
for k in oldparams.keys():
if k == 'savedmodel': # special case
continue
exec_str = k + ' = oldparams["'+k+'"]'
exec( exec_str )
print 'from saved model,',k,'=',eval(k)
# get global_iterations
global_iterations,tmp1,tmp2 = ANALYZE.traceback_stats(savedmodel)
# initialized model from codebook
elif os.path.isfile(savedmodel):
codebook = load_codebook(savedmodel)
assert codebook != None,'Could not load codebook in: %s.'%savedmodel
if useModel == 'VQ':
model = MODEL.Model(codebook)
elif useModel == 'VQFILT':
model = MODEL.ModelFilter(codebook)
else:
assert False, 'wrong model codename: %s.'%useModel
statlog.startFromScratch()
# problem
else:
assert False,'saved model does not exist: %s.'%savedmodel
# creates a dictionary with all parameters
params = {'savedmodel':savedmodel, 'expdir':expdir,
'pSize':pSize, 'usebars':usebars,
'keyInv':keyInv, 'songKeyInv':songKeyInv,
'positive':positive, 'do_resample':do_resample,
'partialbar':partialbar, 'lrate':lrate,
'nThreads':nThreads, 'oracle':oracle,
'artistsdb':artistsdb, 'matdir':matdir,
'nIterations':nIterations, 'useModel':useModel,
'autobar':autobar,'randoffset':randoffset}
# creates the experiment folder
if not os.path.isdir(expdir):
print 'creating experiment directory:',expdir
os.mkdir(expdir)
# create oracle
if autobar:
assert oracle=='MAT','autobar implemented only for matfiles oracle'
if oracle == 'EN':
oracle = oracle_en.OracleEN(params,artistsdb)
elif oracle == 'MAT':
oracle = oracle_matfiles.OracleMatfiles(params,matdir)
else:
assert False, 'wrong oracle codename: %s.'%oracle
# starttime and save time
starttime = time.time()
last_save = starttime
# for estimate of distance
dist_estimate = deque()
dist_estimate_len = 2000
# main algorithm
try:
while True:
# increment iterations
main_iterations += 1
global_iterations += 1
if main_iterations == int(np.ceil(last_printed_iter * 1.1)) and len(dist_estimate) > 0:
print main_iterations,'/',global_iterations,'iterations (local/global), approx. avg dist:',np.average(dist_estimate)
last_printed_iter = main_iterations
statlog.iteration()
if global_iterations > nIterations:
raise StopIteration
# get features from the oracle
if not autobar:
feats = oracle.next_track()
else:
feats = oracle.next_track(auto_bar=model)
# check features, remove empty patterns
if feats == None:
continue
feats = feats[np.nonzero(np.sum(feats,axis=1))]
if feats.shape[0] == 0:
continue
assert not np.isnan(feats).any(),'features have NaN???'
# stats
statlog.patternsSeen(feats.shape[0])
# update model
avg_dist = model.update(feats,lrate=lrate)
assert not np.isnan(avg_dist)
# add to dist_estimate
dist_estimate.append(avg_dist)
if len(dist_estimate) > dist_estimate_len:
dist_estimate.popleft()
# save
if should_save(starttime,last_save):
savedir = save_experiment(expdir,model,starttime,statlog,params)
last_save = time.time()
# error, save and quit
except:
print ''
exc_type, exc_value, exc_traceback = sys.exc_info()
# normal stop or not? normal = StopIteration
exit_code = 1
if str(exc_type) == "exceptions.StopIteration" or str(exc_type) == "<type 'exceptions.StopIteration'>":
exit_code = 0 # normal exit
print "ERROR:", exc_type
if str(exc_type) != "<type 'exceptions.KeyboardInterrupt'>":
print '********** DEBUGGING INFO *******************'
formatted_lines = traceback.format_exc().splitlines()
if len(formatted_lines) > 2:
print formatted_lines[-3]
if len(formatted_lines) > 1:
print formatted_lines[-2]
print formatted_lines[-1]
print '*********************************************'
print 'Stoping after', main_iterations - 1, 'iterations.'
# EN oracle, try to stop/slow down threads
try:
oracle_en._en_queue_size = 0
except NameError:
pass # oracle_en not loaded on hog; I hate myself for doing that
# save
print 'saving...'
savedir = save_experiment(expdir,model,starttime,statlog,params,
crash=True)
print 'saved to: ',savedir
#quit
return exit_code