From b7c6fc52b47f943b656376e1734a64562c7bd411 Mon Sep 17 00:00:00 2001 From: MPolaris <540492239@qq.com> Date: Wed, 12 Oct 2022 09:33:53 +0800 Subject: [PATCH 1/3] fix suffix typos --- utils/builder.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/utils/builder.py b/utils/builder.py index 0e31239..c5187ad 100644 --- a/utils/builder.py +++ b/utils/builder.py @@ -24,7 +24,7 @@ def representative_dataset_gen(img_root, img_size, mean=[0.485, 0.456, 0.406], s else: VALID_FORMAT = ['jpg', 'png', 'jpeg'] for i, fn in enumerate(os.listdir(img_root)): - if fn.split(".")[-1] not in VALID_FORMAT: + if fn.split(".")[-1].lower() not in VALID_FORMAT: continue _input = cv2.imread(os.path.join(img_root, fn)) _input = cv2.resize(_input, (img_size[1], img_size[0]))[:, :, ::-1] From 0169fb719219dccd21aefef5eb390d978fb9de01 Mon Sep 17 00:00:00 2001 From: MPolaris <540492239@qq.com> Date: Wed, 12 Oct 2022 09:44:52 +0800 Subject: [PATCH 2/3] reorganize code --- layers/__init__.py | 2 +- layers/activations_layers.py | 4 ++-- layers/calculations_layers.py | 16 ++++++++-------- layers/deformation_layers.py | 22 +++++++++++----------- layers/dimension_utils.py | 33 +++++++++++++++++++++++++++++++++ layers/shape_axis_utils.py | 22 ---------------------- 6 files changed, 55 insertions(+), 44 deletions(-) create mode 100644 layers/dimension_utils.py delete mode 100644 layers/shape_axis_utils.py diff --git a/layers/__init__.py b/layers/__init__.py index 43cb666..60fda4e 100644 --- a/layers/__init__.py +++ b/layers/__init__.py @@ -1,6 +1,6 @@ from utils import OPERATOR from .conv_layers import * -from .shape_axis_utils import * +from .dimension_utils import * from .common_layers import * from .activations_layers import * from .calculations_layers import * diff --git a/layers/activations_layers.py b/layers/activations_layers.py index 09469f0..7c7e5be 100644 --- a/layers/activations_layers.py +++ b/layers/activations_layers.py @@ -1,7 +1,7 @@ import tensorflow as tf from tensorflow import keras -from .shape_axis_utils import Torch2TFAxis +from .dimension_utils import channel_to_last_dimension from . import OPERATOR @OPERATOR.register_operator("Relu") @@ -122,7 +122,7 @@ def __call__(self, inputs): class TFSoftmax(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: super().__init__() - self.axis = Torch2TFAxis(node_attribute.get('axis', -1)) + self.axis = channel_to_last_dimension(node_attribute.get('axis', -1)) def __call__(self, inputs): return keras.activations.softmax(inputs, axis=self.axis) diff --git a/layers/calculations_layers.py b/layers/calculations_layers.py index 2f1fd20..3fd7dbf 100644 --- a/layers/calculations_layers.py +++ b/layers/calculations_layers.py @@ -3,7 +3,7 @@ import tensorflow as tf from . import OPERATOR -from . import shape_axis_utils +from . import dimension_utils LOG = logging.getLogger("calculations_layers :") @@ -31,7 +31,7 @@ def get_number(tensor_grap, node_weights, node_inputs): for _ in range(len(first_operand.shape) - 2): second_operand = second_operand[..., np.newaxis] else: - second_operand = shape_axis_utils.TorchWeights2TF(second_operand) + second_operand = dimension_utils.tensor_NCD_to_NDC_format(second_operand) elif (not first_operand_flg) and second_operand_flg: # 当second_operand为计算得出的，first_operand来自weight时 if len(first_operand.shape) == 1: @@ -40,7 +40,7 @@ def get_number(tensor_grap, node_weights, node_inputs): for _ in range(len(second_operand.shape) - 2): first_operand = first_operand[..., np.newaxis] else: - first_operand = shape_axis_utils.TorchWeights2TF(first_operand) + first_operand = dimension_utils.tensor_NCD_to_NDC_format(first_operand) return first_operand, second_operand @@ -148,7 +148,7 @@ def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args super().__init__() self.keep_dims = node_attribute.get("keepdims", 1) == 1 input_shape_len = len(tensor_grap[node_inputs[0]].shape) - self.axes = [shape_axis_utils.Torch2TFAxis(i) if i >=0 else shape_axis_utils.Torch2TFAxis(input_shape_len + i) for i in node_attribute.get("axes", [-1])] + self.axes = [dimension_utils.channel_to_last_dimension(i) if i >=0 else dimension_utils.channel_to_last_dimension(input_shape_len + i) for i in node_attribute.get("axes", [-1])] def __call__(self, inputs, *args, **kwargs): return tf.math.reduce_mean(inputs, axis=self.axes, keepdims=self.keep_dims) @@ -159,7 +159,7 @@ def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args super().__init__() self.keep_dims = node_attribute.get("keepdims", 1) == 1 input_shape_len = len(tensor_grap[node_inputs[0]].shape) - self.axes = [shape_axis_utils.Torch2TFAxis(i) if i >=0 else shape_axis_utils.Torch2TFAxis(input_shape_len + i) for i in node_attribute.get("axes", [-1])] + self.axes = [dimension_utils.channel_to_last_dimension(i) if i >=0 else dimension_utils.channel_to_last_dimension(input_shape_len + i) for i in node_attribute.get("axes", [-1])] def __call__(self, inputs, *args, **kwargs): return tf.math.reduce_max(inputs, axis=self.axes, keepdims=self.keep_dims) @@ -170,7 +170,7 @@ def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args super().__init__() self.keep_dims = node_attribute.get("keepdims", 1) == 1 input_shape_len = len(tensor_grap[node_inputs[0]].shape) - self.axes = [shape_axis_utils.Torch2TFAxis(i) if i >=0 else input_shape_len + i for i in node_attribute.get("axes", [-1])] + self.axes = [dimension_utils.channel_to_last_dimension(i) if i >=0 else input_shape_len + i for i in node_attribute.get("axes", [-1])] def __call__(self, inputs, *args, **kwargs): return tf.math.reduce_min(inputs, axis=self.axes, keepdims=self.keep_dims) @@ -179,7 +179,7 @@ def __call__(self, inputs, *args, **kwargs): class TFArgMax(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): super().__init__() - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute.get('axis', 0)) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute.get('axis', 0)) self.keepdims = node_attribute.get("keepdims", 1) == 1 def __call__(self, inputs, *args, **kwargs): @@ -192,7 +192,7 @@ def __call__(self, inputs, *args, **kwargs): class TFArgMin(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): super().__init__() - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute.get('axis', 0)) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute.get('axis', 0)) self.keepdims = node_attribute.get("keepdims", 1) == 1 def __call__(self, inputs, *args, **kwargs): diff --git a/layers/deformation_layers.py b/layers/deformation_layers.py index 7289b0f..273eb80 100644 --- a/layers/deformation_layers.py +++ b/layers/deformation_layers.py @@ -2,7 +2,7 @@ import tensorflow as tf from . import OPERATOR -from . import shape_axis_utils +from . import dimension_utils LOG = logging.getLogger("deformation_layers :") @@ -16,11 +16,11 @@ def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args elif len(node_attribute['perm']) > 4: self.perm_list = [] for axis in node_attribute['perm']: - new_axis = shape_axis_utils.Torch2TFAxis(axis) + new_axis = dimension_utils.channel_to_last_dimension(axis) if new_axis == -1: new_axis = max(node_attribute['perm']) self.perm_list.append(new_axis) - self.perm_list = shape_axis_utils.TorchShape2TF(self.perm_list) + self.perm_list = dimension_utils.shape_NCD_to_NDC_format(self.perm_list) else: self.perm_list = [i for i in node_attribute['perm']] LOG.info("Transpose will process tensor after change back to NCHW format.") @@ -44,12 +44,12 @@ def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args if len(node_inputs) == 1: self.starts = node_attribute['starts'][0] self.ends = node_attribute['ends'][0] - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute['axes'][0]) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute['axes'][0]) self.steps = 1 else: self.starts = node_weights[node_inputs[1]][0] if node_inputs[1] in node_weights else tensor_grap[node_inputs[1]][0] self.axis = node_weights[node_inputs[3]][0] if node_inputs[3] in node_weights else tensor_grap[node_inputs[3]][0] - self.axis = shape_axis_utils.Torch2TFAxis(self.axis) + self.axis = dimension_utils.channel_to_last_dimension(self.axis) self.ends = node_weights[node_inputs[2]][0] if node_inputs[2] in node_weights else tensor_grap[node_inputs[2]][0] self.ends = min(self.ends, tensor_grap[node_inputs[0]].shape[self.axis]) if len(node_inputs) < 5: @@ -65,7 +65,7 @@ def __call__(self, inputs): class TFGather(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: super().__init__() - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute.get('axis', 0)) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute.get('axis', 0)) self.indices = tensor_grap[node_inputs[1]] if node_inputs[1] in tensor_grap else node_weights[node_inputs[1]] def __call__(self, inputs): @@ -75,7 +75,7 @@ def __call__(self, inputs): class TFConcat(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): super().__init__() - _axis = shape_axis_utils.Torch2TFAxis(node_attribute['axis']) + _axis = dimension_utils.channel_to_last_dimension(node_attribute['axis']) _gather = [tensor_grap[x] for x in node_inputs] self.out = tf.concat(_gather, axis=_axis) @@ -129,7 +129,7 @@ def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args start += int(node_attribute['split'][i]) end = start + node_attribute['split'][index] self.indices = tf.keras.backend.arange(start, end, 1) - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute.get("axis", 0)) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute.get("axis", 0)) def __call__(self, inputs): return tf.gather(inputs, indices=self.indices, axis=self.axis) @@ -138,7 +138,7 @@ def __call__(self, inputs): class TFExpand(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs)->None: super().__init__() - self.shape = shape_axis_utils.TorchShape2TF(node_weights[node_inputs[1]]) + self.shape = dimension_utils.shape_NCD_to_NDC_format(node_weights[node_inputs[1]]) def __call__(self, inputs): for i in range(len(self.shape)): @@ -152,7 +152,7 @@ def __call__(self, inputs): class TFUnsqueeze(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs)->None: super().__init__() - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute['axes'][0]) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute['axes'][0]) def __call__(self, inputs): return tf.expand_dims(inputs, self.axis) @@ -161,7 +161,7 @@ def __call__(self, inputs): class TFSqueeze(): def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs)->None: super().__init__() - self.axis = shape_axis_utils.Torch2TFAxis(node_attribute['axes'][0]) + self.axis = dimension_utils.channel_to_last_dimension(node_attribute['axes'][0]) def __call__(self, inputs): return tf.squeeze(inputs, self.axis) \ No newline at end of file diff --git a/layers/dimension_utils.py b/layers/dimension_utils.py new file mode 100644 index 0000000..b2d70ba --- /dev/null +++ b/layers/dimension_utils.py @@ -0,0 +1,33 @@ +''' + shape and axis transform utils func. +''' +def channel_to_last_dimension(axis): + ''' + make channel first to channel last + ''' + if axis == 0: + axis = 0 + elif axis == 1: + axis = -1 + else: + axis -= 1 + return axis + +def shape_NCD_to_NDC_format(shape:list or tuple): + ''' + make shape format from channel first to channel last + ''' + if len(shape) <= 2: + return tuple(shape) + new_shape = [shape[0], *shape[2:], shape[1]] + return tuple(new_shape) + +def tensor_NCD_to_NDC_format(tensor): + ''' + make tensor format from channel first to channel last + ''' + if(len(tensor.shape) > 2): + shape = [i for i in range(len(tensor.shape))] + shape = shape_NCD_to_NDC_format(shape) + tensor = tensor.transpose(*shape) + return tensor \ No newline at end of file diff --git a/layers/shape_axis_utils.py b/layers/shape_axis_utils.py deleted file mode 100644 index b7150e5..0000000 --- a/layers/shape_axis_utils.py +++ /dev/null @@ -1,22 +0,0 @@ - -def Torch2TFAxis(axis): - if axis == 0: - axis = 0 - elif axis == 1: - axis = -1 - else: - axis -= 1 - return axis - -def TorchShape2TF(shape:list or tuple): - if len(shape) <= 2: - return tuple(shape) - new_shape = [shape[0], *shape[2:], shape[1]] - return tuple(new_shape) - -def TorchWeights2TF(weights): - if(len(weights.shape) > 2): - shape = [i for i in range(len(weights.shape))] - shape = TorchShape2TF(shape) - weights = weights.transpose(*shape) - return weights \ No newline at end of file From d3c0c881d5f2911ddf796888fff89d4d7a94c20a Mon Sep 17 00:00:00 2001 From: MPolaris <540492239@qq.com> Date: Wed, 12 Oct 2022 09:55:25 +0800 Subject: [PATCH 3/3] update quantity strategy --- converter.py | 2 +- readme.md | 13 ++++++++----- utils/builder.py | 5 ++--- 3 files changed, 11 insertions(+), 9 deletions(-) diff --git a/converter.py b/converter.py index 016f853..eb6ac09 100644 --- a/converter.py +++ b/converter.py @@ -9,7 +9,7 @@ def onnx_converter(onnx_model_path:str, output_path:str=None, input_node_names:list=None, output_node_names:list=None, need_simplify:bool=True, target_formats:list = ['keras', 'tflite'], weight_quant:bool=False, int8_model:bool=False, image_root:str=None, - int8_mean:list or float = [0.485, 0.456, 0.406], int8_std:list or float = [0.229, 0.224, 0.225]): + int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375]): if not isinstance(target_formats, list) and 'keras' not in target_formats and 'tflite' not in target_formats: raise KeyError("'keras' or 'tflite' should in list") diff --git a/readme.md b/readme.md index 61c4090..605c358 100644 --- a/readme.md +++ b/readme.md @@ -1,6 +1,9 @@ # ONNX->Keras and ONNX->TFLite tools ## How to use +```cmd +pip install -r requirements.txt +``` ```python # base python converter.py --weights "./your_model.onnx" @@ -17,12 +20,12 @@ python converter.py --weights "./your_model.onnx" --outpath "./save_path" --form # cutoff model, redefine inputs and outputs, support middle layers python converter.py --weights "./your_model.onnx" --outpath "./save_path" --formats "tflite" --input-node-names "layer_name" --output-node-names "layer_name1" "layer_name2" -# quantitative model weight, only weight +# quantify model weight, only weight python converter.py --weights "./your_model.onnx" --formats "tflite" --weigthquant -# quantitative model weight, include input and output +# quantify model weight, include input and output ## recommend -python converter.py --weights "./your_model.onnx" --formats "tflite" --int8 --imgroot "./dataset_path" --int8mean 0 0 0 --int8std 1 1 1 +python converter.py --weights "./your_model.onnx" --formats "tflite" --int8 --imgroot "./dataset_path" --int8mean 0 0 0 --int8std 255 255 255 ## generate random data, instead of read from image file python converter.py --weights "./your_model.onnx" --formats "tflite" --int8 ``` @@ -95,8 +98,8 @@ onnx_converter( target_formats = ['tflite'], #or ['keras'], ['keras', 'tflite'] weight_quant = False, int8_model = True, # do quantification - int8_mean = [0.485, 0.456, 0.406], # give mean of image preprocessing - int8_std = [0.229, 0.224, 0.225], # give std of image preprocessing + int8_mean = [123.675, 116.28, 103.53], # give mean of image preprocessing + int8_std = [58.395, 57.12, 57.375], # give std of image preprocessing image_root = "./dataset/train" # give image folder of train ) ``` diff --git a/utils/builder.py b/utils/builder.py index c5187ad..df73830 100644 --- a/utils/builder.py +++ b/utils/builder.py @@ -6,7 +6,7 @@ from onnx import numpy_helper from .op_registry import OPERATOR -def representative_dataset_gen(img_root, img_size, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]): +def representative_dataset_gen(img_root, img_size, mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]): if isinstance(mean, list): mean = np.array(mean, dtype=np.float32) if isinstance(std, list): @@ -28,7 +28,6 @@ def representative_dataset_gen(img_root, img_size, mean=[0.485, 0.456, 0.406], s continue _input = cv2.imread(os.path.join(img_root, fn)) _input = cv2.resize(_input, (img_size[1], img_size[0]))[:, :, ::-1] - _input = _input/255 if mean is not None: _input = (_input - mean) if std is not None: @@ -115,7 +114,7 @@ def keras_builder(onnx_model, new_input_nodes:list=None, new_output_nodes:list=N return keras_model def tflite_builder(keras_model, weight_quant:bool=False, int8_model:bool=False, image_root:str=None, - int8_mean:list or float = [0.485, 0.456, 0.406], int8_std:list or float = [0.229, 0.224, 0.225]): + int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375]): converter = tf.lite.TFLiteConverter.from_keras_model(keras_model) converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS] if weight_quant or int8_model: