def createTestSet(ark_file, sDataFileTemplate, sDataDir, sDataFilePattern,
sDataProtoFile):
dbInfo = dl.DatabaseInfo()
dbInfo.name = 'dataset_test'
dbInfo.data_handler = 'deeplearn'
dbInfo.path_prefix = sDataDir
datasetInfo = dbInfo.data.add()
datasetInfo.data_format = dl.DatasetInfo.PBM
datasetInfo.size = 0
datasetInfo.sparse_label = True
datasetInfo.type = dl.DatasetInfo.TEST_SET
datasetInfo.file_pattern = sDataFilePattern
uttIDBatch = []
uttIDLength = []
featMatBatch = None
batchSz = -1
iFileIdx = 0
kaldiIn = KaldiReadIn(ark_file)
kaldiIn.open()
uttID, featMat = kaldiIn.next()
while featMat is not None:
if batchSz < 0:
batchSz = 400 * 1024 * 1024 / (4 * featMat.shape[1])
datasetInfo.dimensions = featMat.shape[1] + 1
datasetInfo.label_start_index = datasetInfo.dimensions - 1
if featMatBatch is None:
featMatBatch = featMat
else:
featMatBatch = np.vstack([featMatBatch, featMat])
uttIDBatch.append(uttID)
uttIDLength.append(featMat.shape[0])
if featMatBatch.shape[0] >= batchSz:
util.WriteProto(
sDataFileTemplate % iFileIdx,
util.npy2ProtoMat(
np.hstack(
[featMatBatch,
np.zeros((featMatBatch.shape[0], 1))])))
iFileIdx += 1
datasetInfo.size += featMatBatch.shape[0]
featMatBatch = None
uttID, featMat = kaldiIn.next()
kaldiIn.close()
# last batch
if featMatBatch is not None:
util.WriteProto(
sDataFileTemplate % iFileIdx,
util.npy2ProtoMat(
np.hstack([featMatBatch,
np.zeros((featMatBatch.shape[0], 1))])))
iFileIdx += 1
datasetInfo.size += featMatBatch.shape[0]
util.WriteProto(sDataProtoFile, dbInfo)
return (uttIDBatch, uttIDLength)
def pfile2Proto(pfilePath, filePrefix, pbmDir):
pfile = read_dataset(pfilePath, {'partition': 1024 * 1024 * 400})
dsInfo = dl.DatasetInfo()
dsInfo.data_format = dl.DatasetInfo.PBM
dsInfo.sparse_label = True
dsInfo.file_pattern = '%s[0-9]*.pbm' % filePrefix
dim = None
sz = 0
for i, (data, label) in enumerate(zip(pfile.feat_mats, pfile.label_vecs)):
dataset = util.npy2ProtoMat(np.hstack([data, label[:, None]]))
util.WriteProto(os.path.join(pbmDir, '%s%05d.pbm' % (filePrefix, i)),
dataset)
if dim is None:
dim = data.shape[1] + 1
if dim != data.shape[1] + 1:
print dim, sz, data, shape, label.shape
assert dim == data.shape[1] + 1
sz += data.shape[0]
dsInfo.size = sz
dsInfo.dimensions = dim
dsInfo.label_start_index = dim - 1
return dsInfo
def createPbmDataset(pfiles, pbmDir, protoFilePath, gpuMem):
assert gpuMem > 0.1
dbInfo = dl.DatabaseInfo()
for (name, sPath) in pfiles:
assert name == 'train' or name == 'valid' or name == 'test'
dsInfo = pfile2Proto(sPath, name + '_part', pbmDir)
if name == 'train':
dsInfo.type = dl.DatasetInfo.TRAIN_SET
elif name == 'valid':
dsInfo.type = dl.DatasetInfo.EVAL_SET
else:
dsInfo.type = dl.DatasetInfo.TEST_SET
dbInfo.data.extend([dsInfo])
dbInfo.name = 'dataset'
dbInfo.data_handler = 'deeplearn'
dbInfo.main_memory = 6.0
dbInfo.gpu_memory = float(gpuMem)
dbInfo.path_prefix = pbmDir
util.WriteProto(protoFilePath, dbInfo)
evalOp.randomize = False
evalOp.get_last_piece = True
evalOp.verbose = False
# test on all models
for i in xrange(0, config.PHONES):
for j in xrange(i + 1, config.PHONES):
sResultFile = os.path.join(sResultDir, '%d_%d.csv' % (i, j))
if os.path.exists(sResultFile):
continue
print 'Testing for %d-%d, writing results to %s' % (i, j,
sResultFile)
evalOp.result_file = sResultFile
util.WriteProto(sEvalOpFile, evalOp)
sModelFile = sModelFiles % (i, j)
args = [
sDeeplearnPath, 'eval', sModelFile,
'--eval-op=%s' % sEvalOpFile
]
pr = subprocess.Popen(args, stderr=subprocess.STDOUT)
pr.wait()
if pr.returncode != 0:
print 'Failed to test %d-%d' % (i, j)
exit(1)
# run majorityVote, compute "probabilities"
sHardVoteFile = os.path.join(sMajorVoteDir, 'hard.csv')
sSoftVoteFile = os.path.join(sMajorVoteDir, 'soft.csv')
if not os.path.exists(sOutputModelFile):
# modify architecture...
sModelDir = os.path.join(wdir, 'model/')
createDir(sModelDir)
sCurrentDir = os.path.split(os.path.realpath(
os.path.abspath(__file__)))[0]
model = util.ReadProto(
os.path.join(sCurrentDir, 'prototype/conv_timit.pbtxt'),
dl.ModelData())
model.name = 'spn_conv'
for n in model.nodes:
if n.name == 'output':
n.dimension = num_outputs
break
sModelFile = os.path.join(sModelDir, 'spn_conv.pbtxt')
util.WriteProto(sModelFile, model)
trainOp = util.ReadProto(
os.path.join(sCurrentDir, 'prototype/train.pbtxt'), dl.Operation())
sCheckpointDir = os.path.join(sModelDir, 'cp')
trainOp.name = 'train'
trainOp.data_proto = sDataProtoFile
trainOp.checkpoint_directory = sCheckpointDir
trainOp.verbose = False
sTrainOpFile = os.path.join(sModelDir, 'train.pbtxt')
util.WriteProto(sTrainOpFile, trainOp)
evalOp = util.ReadProto(
os.path.join(sCurrentDir, 'prototype/eval.pbtxt'), dl.Operation())
evalOp.data_proto = sDataProtoFile
evalOp.verbose = False
def extractRepresentation(data, wdir, sDeeplearnPath, sModelFile):
# append label column
data = np.hstack([data, np.zeros((data.shape[0], 1))])
npyData = util.npy2ProtoMat(data)
sDataFile = os.path.join(wdir, 'input.pbm')
util.WriteProto(sDataFile, npyData)
sDataProtoFile = os.path.join(wdir, 'data.pbtxt')
dbInfo = dl.DatabaseInfo()
dbInfo.name = 'dataset_extract'
dbInfo.data_handler = 'deeplearn'
dbInfo.path_prefix = wdir
datasetInfo = dbInfo.data.add()
datasetInfo.data_format = dl.DatasetInfo.PBM
datasetInfo.size = data.shape[0]
datasetInfo.dimensions = data.shape[1]
datasetInfo.label_start_index = datasetInfo.dimensions - 1
datasetInfo.sparse_label = True
datasetInfo.type = dl.DatasetInfo.TEST_SET
datasetInfo.file_pattern = 'input.pbm'
util.WriteProto(sDataProtoFile, dbInfo)
sEvalOpFile = os.path.join(wdir, 'eval.pbtxt')
sExtractedActs = os.path.join(wdir, 'acts')
sLayerName = 'conv2'
if not os.path.exists(sExtractedActs):
os.mkdir(sExtractedActs)
evalOp = dl.Operation()
evalOp.name = 'extract'
evalOp.stop_condition.all_processed = True
evalOp.operation_type = dl.Operation.TEST
evalOp.data_proto = sDataProtoFile
evalOp.randomize = False
evalOp.get_last_piece = True
evalOp.verbose = False
evalOp.extracted_layers.append(sLayerName)
evalOp.extracted_output_dir = sExtractedActs
evalOp.extracted_data_format = dl.DatasetInfo.PBM
evalOp.extracted_data_sets.append(dl.DatasetInfo.TEST_SET)
util.WriteProto(sEvalOpFile, evalOp)
sOutFileTemplate = os.path.join(sExtractedActs, sLayerName, '*.pbm')
for s in sorted(glob.glob(sOutFileTemplate)):
try:
os.remove(s)
except Exception:
pass
# run the network...
args = [
sDeeplearnPath, 'extract', sModelFile,
'--eval-op=%s' % sEvalOpFile
]
pr = subprocess.Popen(args, stderr=subprocess.STDOUT)
pr.wait()
if pr.returncode != 0:
print 'Failed to extract representations'
exit(1)
# read Dataset
mOutput = None
for s in sorted(glob.glob(sOutFileTemplate)):
m = util.proto2Npy(util.ReadProto(s, dl.Matrix()))
if mOutput is None:
mOutput = m
else:
mOutput = np.vstack([mOutput, m])
if mOutput.shape[0] != data.shape[0]:
print 'Invalid results'
exit(1)
return mOutput