def PytorchEmit(original_framework, architecture_name, architecture_path, weight_path, image_path):
import torch
from mmdnn.conversion.pytorch.pytorch_emitter import PytorchEmitter
# IR to code
converted_file = original_framework + '_pytorch_' + architecture_name + "_converted"
converted_file = converted_file.replace('.', '_')
emitter = PytorchEmitter((architecture_path, weight_path))
emitter.run(converted_file + '.py', converted_file + '.npy', 'test')
del emitter
del PytorchEmitter
# import converted model
model_converted = __import__(converted_file).KitModel(converted_file + '.npy')
model_converted.eval()
func = TestKit.preprocess_func[original_framework][architecture_name]
img = func(image_path)
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, 0).copy()
input_data = torch.from_numpy(img)
input_data = torch.autograd.Variable(input_data, requires_grad = False)
predict = model_converted(input_data)
predict = predict.data.numpy()
converted_predict = np.squeeze(predict)
del model_converted
del sys.modules[converted_file]
del torch
os.remove(converted_file + '.py')
os.remove(converted_file + '.npy')
return converted_predict
def PytorchEmit(original_framework, architecture_name, architecture_path, weight_path, image_path):
import torch
# IR to code
converted_file = original_framework + '_pytorch_' + architecture_name + "_converted"
converted_file = converted_file.replace('.', '_')
emitter = PytorchEmitter((architecture_path, weight_path))
emitter.run(converted_file + '.py', converted_file + '.npy', 'test')
del emitter
# import converted model
model_converted = __import__(converted_file).KitModel(converted_file + '.npy')
model_converted.eval()
func = TestKit.preprocess_func[original_framework][architecture_name]
img = func(image_path)
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, 0).copy()
input_data = torch.from_numpy(img)
input_data = torch.autograd.Variable(input_data, requires_grad = False)
predict = model_converted(input_data)
predict = predict.data.numpy()
converted_predict = np.squeeze(predict)
del model_converted
del sys.modules[converted_file]
del torch
os.remove(converted_file + '.py')
os.remove(converted_file + '.npy')
return converted_predict
def _convert(args):
if args.dstFramework == 'caffe':
from mmdnn.conversion.caffe.caffe_emitter import CaffeEmitter
if args.IRWeightPath is None:
emitter = CaffeEmitter(args.IRModelPath)
else:
assert args.dstWeightPath
emitter = CaffeEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstFramework == 'keras':
from mmdnn.conversion.keras.keras2_emitter import Keras2Emitter
emitter = Keras2Emitter((args.IRModelPath, args.IRWeightPath))
elif args.dstFramework == 'tensorflow':
from mmdnn.conversion.tensorflow.tensorflow_emitter import TensorflowEmitter
if args.IRWeightPath is None:
# Convert network architecture only
emitter = TensorflowEmitter(args.IRModelPath)
else:
emitter = TensorflowEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstFramework == 'cntk':
from mmdnn.conversion.cntk.cntk_emitter import CntkEmitter
if args.IRWeightPath is None:
emitter = CntkEmitter(args.IRModelPath)
else:
emitter = CntkEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstFramework == 'coreml':
raise NotImplementedError("CoreML emitter is not finished yet.")
elif args.dstFramework == 'pytorch':
if not args.dstWeightPath or not args.IRWeightPath:
raise ValueError("Need to set a target weight filename.")
from mmdnn.conversion.pytorch.pytorch_emitter import PytorchEmitter
emitter = PytorchEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstFramework == 'mxnet':
from mmdnn.conversion.mxnet.mxnet_emitter import MXNetEmitter
if args.IRWeightPath is None:
emitter = MXNetEmitter(args.IRModelPath)
else:
if args.dstWeightPath is None:
raise ValueError(
"MXNet emitter needs argument [dstWeightPath(dw)], like -dw mxnet_converted-0000.param"
)
emitter = MXNetEmitter(
(args.IRModelPath, args.IRWeightPath, args.dstWeightPath))
elif args.dstFramework == 'onnx':
from mmdnn.conversion.onnx.onnx_emitter import OnnxEmitter
if args.IRWeightPath is None:
raise NotImplementedError("ONNX emitter needs IR weight file")
else:
emitter = OnnxEmitter(args.IRModelPath, args.IRWeightPath)
else:
assert False
emitter.run(args.dstModelPath, args.dstWeightPath, args.phase)
return 0
def pytorch_emit(original_framework, architecture_name, architecture_path,
weight_path, test_input_path):
from mmdnn.conversion.pytorch.pytorch_emitter import PytorchEmitter
# IR to code
converted_file = TestModels.tmpdir + original_framework + '_pytorch_' + architecture_name + "_converted"
converted_file = converted_file.replace('.', '_')
emitter = PytorchEmitter((architecture_path, weight_path))
emitter.run(converted_file + '.py', converted_file + '.npy', 'test')
del emitter
del PytorchEmitter
# import converted model
import torch
model_converted = imp.load_source(
'PytorchModel',
converted_file + '.py').KitModel(converted_file + '.npy')
model_converted.eval()
original_framework = checkfrozen(original_framework)
if 'rnn' not in architecture_name:
func = TestKit.preprocess_func[original_framework][
architecture_name]
img = func(test_input_path(architecture_name))
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, 0).copy()
input_data = torch.from_numpy(img)
input_data = torch.autograd.Variable(input_data,
requires_grad=False)
else:
sentence = np.load(test_input_path(architecture_name))
input_data = torch.from_numpy(sentence)
input_data = torch.autograd.Variable(input_data,
requires_grad=False)
predict = model_converted(input_data)
predict = predict.data.numpy()
converted_predict = np.squeeze(predict)
del model_converted
del sys.modules['PytorchModel']
del torch
os.remove(converted_file + '.py')
os.remove(converted_file + '.npy')
return converted_predict
def _convert(args):
if args.dstModelFormat == 'caffe':
raise NotImplementedError(
"Destination [Caffe] is not implemented yet.")
elif args.dstModelFormat == 'keras':
from mmdnn.conversion.keras.keras2_emitter import Keras2Emitter
emitter = Keras2Emitter(args.IRModelPath)
elif args.dstModelFormat == 'tensorflow':
from mmdnn.conversion.tensorflow.tensorflow_emitter import TensorflowEmitter
if args.IRWeightPath == None:
emitter = TensorflowEmitter(args.IRModelPath)
else:
emitter = TensorflowEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstModelFormat == 'cntk':
from mmdnn.conversion.cntk.cntk_emitter import CntkEmitter
if args.IRWeightPath == None:
emitter = CntkEmitter(args.IRModelPath)
else:
emitter = CntkEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstModelFormat == 'coreml':
raise NotImplementedError("CoreML emitter is not finished yet.")
assert args.IRWeightPath != None
from mmdnn.conversion.coreml.coreml_emitter import CoreMLEmitter
emitter = CoreMLEmitter((args.IRModelPath, args.IRWeightPath))
model = emitter.gen_model()
print("Saving the CoreML model [{}].".format(args.dstModelPath +
'.mlmodel'))
model.save(args.dstModelPath + '.mlmodel')
print("The converted CoreML model saved as [{}].".format(
args.dstModelPath + '.mlmodel'))
return 0
elif args.dstModelFormat == 'pytorch':
if not args.dstWeightPath or not args.IRWeightPath:
raise ValueError("Need to set a target weight filename.")
from mmdnn.conversion.pytorch.pytorch_emitter import PytorchEmitter
emitter = PytorchEmitter((args.IRModelPath, args.IRWeightPath))
elif args.dstModelFormat == 'mxnet':
from mmdnn.conversion.mxnet.mxnet_emitter import MXNetEmitter
if args.IRWeightPath == None:
emitter = MXNetEmitter(args.IRModelPath)
else:
emitter = MXNetEmitter((args.IRModelPath, args.IRWeightPath,
args.inputShape, args.dstWeightPath))
else:
assert False
emitter.run(args.dstModelPath, args.dstWeightPath, args.phase)
return 0
def _convert(args):
if args.framework == 'caffe':
raise NotImplementedError(
"Destination [Caffe] is not implemented yet.")
elif args.framework == 'keras':
raise NotImplementedError(
"Destination [Keras] is not implemented yet.")
elif args.framework == 'tensorflow':
raise NotImplementedError(
"Destination [Tensorflow] is not implemented yet.")
elif args.framework == 'cntk':
raise NotImplementedError(
"Destination [Tensorflow] is not implemented yet.")
elif args.framework == 'coreml':
from mmdnn.conversion.coreml.coreml_emitter import CoreMLEmitter
assert args.inputNetwork is not None
assert args.inputWeight is not None
emitter = CoreMLEmitter(args.inputNetwork, args.inputWeight)
model = emitter.gen_model(
args.inputNames,
args.outputNames,
image_input_names=set(args.imageInputNames)
if args.imageInputNames else None,
is_bgr=args.isBGR,
red_bias=args.redBias,
blue_bias=args.blueBias,
green_bias=args.greenBias,
gray_bias=args.grayBias,
image_scale=args.scale,
class_labels=args.classInputPath if args.classInputPath else None,
predicted_feature_name=args.predictedFeatureName)
"""
from google.protobuf import text_format
with open(args.output+'.txt', 'w') as f:
f.write(text_format.MessageToString(model))
"""
with open(args.output, 'wb') as f:
model = model.SerializeToString()
f.write(model)
return 0
elif args.framework == 'pytorch':
if not args.dstWeightPath or not args.IRWeightPath:
raise ValueError("Need to set a target weight filename.")
from mmdnn.conversion.pytorch.pytorch_emitter import PytorchEmitter
emitter = PytorchEmitter((args.IRModelPath, args.IRWeightPath))
elif args.framework == 'mxnet':
from mmdnn.conversion.mxnet.mxnet_emitter import MXNetEmitter
if args.IRWeightPath == None:
emitter = MXNetEmitter(args.IRModelPath)
else:
emitter = MXNetEmitter((args.IRModelPath, args.IRWeightPath,
args.inputShape, args.dstWeightPath))
else:
assert False
emitter.run(args.output)
return 0