diff --git a/tools/README.md b/tools/README.md index fb4fe8f85e0b9483cc72a35dc37efcb177cbff9f..1a507890bc18ea85906812dec28ad5aea4fec373 100644 --- a/tools/README.md +++ b/tools/README.md @@ -6,7 +6,6 @@ python tools/check_for_lite.py paddle_model/inference_model/__model__ ``` > 附：check_for_lite工具并不能完全判断模型是否被支持，PaddleLite详细支持的算子请参考[PaddleLite支持算子集](https://github.com/PaddlePaddle/Paddle-Lite/blob/develop/docs/introduction/support_operation_list.md) - ### 二、模型参数合并 X2Paddle转换后产出的路径下包括两个目录， 1. `model_with_code`: 包含保存的参数文件和模型python代码文件，供用户debug diff --git a/x2paddle/convert.py b/x2paddle/convert.py index 365e1144dc6a049c6c465499e4a868ac73c89cef..7e7f439bfe743663c06972b8dc8af1c0803361b0 100644 --- a/x2paddle/convert.py +++ b/x2paddle/convert.py @@ -117,6 +117,7 @@ def tf2paddle(model_path, from x2paddle.optimizer.tensorflow.bias import BiasOpt from x2paddle.optimizer.tensorflow.transpose import TransposeOpt from x2paddle.optimizer.tensorflow.batch_norm import BatchNormOpt + from x2paddle.optimizer.tensorflow.prelu import PReLUOpt print("Now translating model from tensorflow to paddle.") model = TFDecoder(model_path, define_input_shape=define_input_shape) @@ -125,8 +126,10 @@ def tf2paddle(model_path, bias_opt = BiasOpt() transpose_opt = TransposeOpt() batch_norm_opt = BatchNormOpt() + prelu_opt = PReLUOpt() bias_opt.run(program) batch_norm_opt.run(program) + prelu_opt.run(program) transpose_opt.run(program) program.gen_model(save_dir) diff --git a/x2paddle/op_mapper/tf_op_mapper.py b/x2paddle/op_mapper/tf_op_mapper.py index d2ed2944fd257df2ef0e5cfc7e5451abde6245aa..82e2694f1745874c9256921be66f22300e0236e1 100644 --- a/x2paddle/op_mapper/tf_op_mapper.py +++ b/x2paddle/op_mapper/tf_op_mapper.py @@ -882,7 +882,7 @@ class TFOpMapper(OpMapper): begin = self.graph.get_node(node.layer.input[1]) size = self.graph.get_node(node.layer.input[2]) - inputs = {"x": input.name} + inputs = {"input": input.name} attrs = {} if begin.layer_type == "Const": begin = begin.value.tolist() @@ -901,20 +901,30 @@ class TFOpMapper(OpMapper): if size.layer_type == "Const": size = size.value.tolist() attrs['shape'] = size + shape = size else: shape = size.out_shapes[0] - reshape_name = gen_name("slice", "reshape") - program.add_layer( - kernel="fluid.layers.reshape", - inputs={"x": size.name}, - outputs=[reshape_name], - shape=shape) - inputs['shape'] = reshape_name +# reshape_name = gen_name("slice", "reshape") +# program.add_layer( +# kernel="fluid.layers.reshape", +# inputs={"x": size.name}, +# outputs=[reshape_name], +# shape=shape) +# inputs['shape'] = reshape_name + +# inputs.pop('shape') program.add_layer( - kernel="fluid.layers.crop_tensor", + kernel="fluid.layers.slice", inputs=inputs, outputs=[node.name], - **attrs) + axes=list(range(len(attrs['offsets']))), + starts=attrs['offsets'], + ends=[attrs['offsets'][i] + shape[i] for i in range(len(shape))]) +# program.add_layer( +# kernel="fluid.layers.crop_tensor", +# inputs=inputs, +# outputs=[node.name], +# **attrs) def ResizeNearestNeighbor(self, node): input = self.graph.get_node(node.layer.input[0]) diff --git a/x2paddle/optimizer/tensorflow/batch_norm.py b/x2paddle/optimizer/tensorflow/batch_norm.py index 3e3d81360598c2b66a00735cd23059998f3bef2d..f2d1db83bf64eb1c886ad1a4def3a7e2861b1be4 100644 --- a/x2paddle/optimizer/tensorflow/batch_norm.py +++ b/x2paddle/optimizer/tensorflow/batch_norm.py @@ -20,10 +20,12 @@ class BatchNormOpt: input_ids0 = graph.edges_in[layer_id] mul_layer0 = graph.layers[input_ids0[0]] sub_layer0 = graph.layers[input_ids0[1]] + if mul_layer0.kernel != "fluid.layers.elementwise_mul": continue if sub_layer0.kernel != "fluid.layers.elementwise_sub": continue + axis = mul_layer0.attrs.get('axis', -1) if axis != -1 and axis != 3: continue @@ -116,7 +118,7 @@ class BatchNormOpt: other = graph.layers[input_ids6[1]] if variance.kernel != "fluid.layers.create_parameter": continue - if other.kernel != "fluid.layers.create_parameter": + if other.kernel != "fluid.layers.fill_constant": continue if len(graph.edges_out.get(input_ids6[0], [])) != 1: continue @@ -127,10 +129,6 @@ class BatchNormOpt: variance_shape = graph.parameters[variance.outputs[0]].shape if variance_shape != beta_shape: continue - if other.outputs[0] not in graph.parameters: - continue - if graph.parameters[other.outputs[0]].size != 1: - continue ids = set([ layer_id, mul_layer0.id, sub_layer0.id, mul_layer1.id, beta.id, @@ -163,7 +161,7 @@ class BatchNormOpt: kernel="fluid.layers.batch_norm", inputs={"input": "transpose_for_bn"}, outputs=layer.outputs, - epsilon=graph.parameters[other.outputs[0]], + epsilon=other.attrs["value"], param_attr="'{}'".format(gamma.outputs[0]), bias_attr="'{}'".format(beta.outputs[0]), moving_mean_name="'{}'".format(mean.outputs[0]), diff --git a/x2paddle/optimizer/tensorflow/prelu.py b/x2paddle/optimizer/tensorflow/prelu.py new file mode 100644 index 0000000000000000000000000000000000000000..cec49f4e62dd0ecdf6c2c0a1ee723a596deda371 --- /dev/null +++ b/x2paddle/optimizer/tensorflow/prelu.py @@ -0,0 +1,123 @@ +import copy +import numpy as np +from collections import OrderedDict +from x2paddle.core.program import PaddleLayer +from x2paddle.core.util import * + + +class PReLUOpt: + def __init__(self): + pass + + def run(self, graph): + print("Optimize: PReLUOpt...") + layers = copy.deepcopy(graph.layers) + for layer_id, layer in layers.items(): + if layer.kernel != "fluid.layers.elementwise_add": + continue + axis = layer.attrs.get('axis', -1) + if axis != -1 and axis != 3: + continue + + input_ids0 = graph.edges_in[layer_id] + relu_layer0 = graph.layers[input_ids0[0]] + mul_layer0 = graph.layers[input_ids0[1]] + + if relu_layer0.kernel != "fluid.layers.relu": + continue + if mul_layer0.kernel != "fluid.layers.elementwise_mul": + continue + + axis = mul_layer0.attrs.get('axis', -1) + if axis != -1 and axis != 3: + continue + if len(graph.edges_out.get(input_ids0[0], [])) != 1: + continue + if len(graph.edges_out.get(input_ids0[1], [])) != 1: + continue + + input_ids1_0 = graph.edges_in[input_ids0[0]] + input_ids1_1 = graph.edges_in[input_ids0[1]] + fill_layer = graph.layers[input_ids1_1[1]] + mul_layer1 = graph.layers[input_ids1_1[0]] + if fill_layer.kernel != "fluid.layers.fill_constant": + continue + if mul_layer1.kernel != "fluid.layers.elementwise_mul": + continue + axis = mul_layer1.attrs.get('axis', -1) + if axis != -1 and axis != 0: + continue + if len(graph.edges_out.get(input_ids1_1[1], [])) != 1: + continue + if len(graph.edges_out.get(input_ids1_0[0], [])) != 3: + continue + + input_ids2 = graph.edges_in[input_ids1_1[0]] + alpha = graph.layers[input_ids2[0]] + sub_layer = graph.layers[input_ids2[1]] + if alpha.kernel != "fluid.layers.create_parameter": + continue + if sub_layer.kernel != "fluid.layers.elementwise_sub": + continue + axis = sub_layer.attrs.get('axis', -1) + if axis != -1 and axis != 3: + continue + if len(graph.edges_out.get(input_ids2[0], [])) != 1: + continue + if len(graph.edges_out.get(input_ids2[1], [])) != 1: + continue + if alpha.outputs[0] not in graph.parameters: + continue + + input_ids3 = graph.edges_in[input_ids2[1]] + add_layer = graph.layers[input_ids3[0]] + abs_layer = graph.layers[input_ids3[1]] + if abs_layer.kernel != "fluid.layers.abs": + continue + if len(graph.edges_out.get(input_ids3[1], [])) != 1: + continue + + + ids = set([ + layer.id, relu_layer0.id, mul_layer0.id, fill_layer.id, mul_layer1.id, alpha.id, + sub_layer.id, abs_layer.id]) + + for id in ids: + del graph.layers[id] + if id in graph.edges_in: + del graph.edges_in[id] + if id in graph.edges_out: + del graph.edges_out[id] + + copy_layers = copy.deepcopy(graph.layers) + graph.layers = OrderedDict() + for k, v in copy_layers.items(): + if k != add_layer.id: + graph.layers[k] = v + continue + graph.layers[k] = v + transpose0 = PaddleLayer( + id='{}_1'.format(k), + kernel="fluid.layers.transpose", + inputs={"x": v.outputs[0]}, + outputs=["transpose_for_prelu"], + perm=[0, 3, 1, 2]) + prelu = PaddleLayer( + id='{}_2'.format(k), + kernel="fluid.layers.prelu", + inputs={"x": "transpose_for_prelu"}, + outputs=layer.outputs, + mode=string("channel"), + param_attr="'{}'".format(alpha.outputs[0])) + transpose1 = PaddleLayer( + id=layer_id, + kernel="fluid.layers.transpose", + inputs={"x": layer.outputs[0]}, + outputs=layer.outputs, + perm=[0, 2, 3, 1]) + graph.layers[transpose0.id] = transpose0 + graph.layers[prelu.id] = prelu + graph.layers[transpose1.id] = transpose1 + graph.parameters[alpha.outputs[0]] = np.expand_dims(graph.parameters[alpha.outputs[0]], 0) + graph.build() +