diff --git a/README.md b/README.md index 0d1caeba1da7f478c5475677c800f1b036f924b8..acc08fa3bb815e33d39b5121d172830211d5ec95 100644 --- a/README.md +++ b/README.md @@ -50,11 +50,11 @@ X2Paddle是飞桨生态下的模型转换工具，致力于帮助其它深度学 ### 环境依赖 - python >= 3.5 -- paddlepaddle >= 2.0.0 +- paddlepaddle >= 2.2.2 - tensorflow == 1.14 (如需转换TensorFlow模型) - onnx >= 1.6.0 (如需转换ONNX模型) - torch >= 1.5.0 (如需转换PyTorch模型) -- paddlelite == 2.9.0 (如需一键转换成Paddle-Lite支持格式) +- paddlelite >= 2.9.0 (如需一键转换成Paddle-Lite支持格式,推荐最新版本) ### pip安装(推荐） diff --git a/x2paddle/convert.py b/x2paddle/convert.py index d4f12f9ea9767091f5db98157cdcf1a7db829ce8..53a65a9ef0895a84cd6f886d864502599da9235c 100644 --- a/x2paddle/convert.py +++ b/x2paddle/convert.py @@ -141,30 +141,35 @@ def tf2paddle(model_path, version = tf.__version__ if version >= '2.0.0' or version < '1.0.0': logging.info( - "[ERROR] 1.0.0<=tensorflow<2.0.0 is required, and v1.14.0 is recommended" + "[ERROR] 1.0.0<=TensorFlow<2.0.0 is required, and v1.14.0 is recommended" ) return except: logging.info( - "[ERROR] Tensorflow is not installed, use \"pip install tensorflow\"." + "[ERROR] TensorFlow is not installed, use \"pip install TensorFlow\"." ) return from x2paddle.decoder.tf_decoder import TFDecoder from x2paddle.op_mapper.tf2paddle.tf_op_mapper import TFOpMapper - logging.info("Now translating model from tensorflow to paddle.") + logging.info("Now translating model from TensorFlow to Paddle.") model = TFDecoder(model_path, define_input_shape=define_input_shape) mapper = TFOpMapper(model) mapper.paddle_graph.build() + logging.info("Model optimizing ...") from x2paddle.optimizer.optimizer import GraphOptimizer graph_opt = GraphOptimizer(source_frame="tf") graph_opt.optimize(mapper.paddle_graph) + logging.info("Model optimized!") mapper.paddle_graph.gen_model(save_dir) + logging.info("Successfully exported Paddle static graph model!") ConverterCheck(task="TensorFlow", convert_state="Success").start() if convert_to_lite: + logging.info("Now translating model from Paddle to Paddle Lite ...") ConverterCheck(task="TensorFlow", lite_state="Start").start() convert2lite(save_dir, lite_valid_places, lite_model_type) + logging.info("Successfully exported Paddle Lite support model!") ConverterCheck(task="TensorFlow", lite_state="Success").start() @@ -193,12 +198,15 @@ def caffe2paddle(proto_file, from x2paddle.optimizer.optimizer import GraphOptimizer graph_opt = GraphOptimizer(source_frame="caffe") graph_opt.optimize(mapper.paddle_graph) - logging.info("Model optimized.") + logging.info("Model optimized!") mapper.paddle_graph.gen_model(save_dir) + logging.info("Successfully exported Paddle static graph model!") ConverterCheck(task="Caffe", convert_state="Success").start() if convert_to_lite: + logging.info("Now translating model from Paddle to Paddle Lite ...") ConverterCheck(task="Caffe", lite_state="Start").start() convert2lite(save_dir, lite_valid_places, lite_model_type) + logging.info("Successfully exported Paddle Lite support model!") ConverterCheck(task="Caffe", lite_state="Success").start() @@ -234,10 +242,13 @@ def onnx2paddle(model_path, graph_opt.optimize(mapper.paddle_graph) logging.info("Model optimized.") mapper.paddle_graph.gen_model(save_dir) + logging.info("Successfully exported Paddle static graph model!") ConverterCheck(task="ONNX", convert_state="Success").start() if convert_to_lite: + logging.info("Now translating model from Paddle to Paddle Lite ...") ConverterCheck(task="ONNX", lite_state="Start").start() convert2lite(save_dir, lite_valid_places, lite_model_type) + logging.info("Successfully exported Paddle Lite support model!") ConverterCheck(task="ONNX", lite_state="Success").start() @@ -261,17 +272,17 @@ def pytorch2paddle(module, version_sum = int(v0) * 100 + int(v1) * 10 + int(v2) if version_sum < 150: logging.info( - "[ERROR] pytorch>=1.5.0 is required, 1.6.0 is the most recommended" + "[ERROR] PyTorch>=1.5.0 is required, 1.6.0 is the most recommended" ) return if version_sum > 160: - logging.info("[WARNING] pytorch==1.6.0 is recommended") + logging.info("[WARNING] PyTorch==1.6.0 is recommended") except: logging.info( - "[ERROR] Pytorch is not installed, use \"pip install torch==1.6.0 torchvision\"." + "[ERROR] PyTorch is not installed, use \"pip install torch==1.6.0 torchvision\"." ) return - logging.info("Now translating model from pytorch to paddle.") + logging.info("Now translating model from PyTorch to Paddle.") from x2paddle.decoder.pytorch_decoder import ScriptDecoder, TraceDecoder from x2paddle.op_mapper.pytorch2paddle.pytorch_op_mapper import PyTorchOpMapper @@ -286,13 +297,16 @@ def pytorch2paddle(module, from x2paddle.optimizer.optimizer import GraphOptimizer graph_opt = GraphOptimizer(source_frame="pytorch", jit_type=jit_type) graph_opt.optimize(mapper.paddle_graph) - logging.info("Model optimized.") + logging.info("Model optimized!") mapper.paddle_graph.gen_model( save_dir, jit_type=jit_type, enable_code_optim=enable_code_optim) + logging.info("Successfully exported Paddle static graph model!") ConverterCheck(task="PyTorch", convert_state="Success").start() if convert_to_lite: + logging.info("Now translating model from Paddle to Paddle Lite ...") ConverterCheck(task="PyTorch", lite_state="Start").start() convert2lite(save_dir, lite_valid_places, lite_model_type) + logging.info("Successfully exported Paddle Lite support model!") ConverterCheck(task="PyTorch", lite_state="Success").start() diff --git a/x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/nms.py b/x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/nms.py index 2815cda9860fe5a87edf6194ce50b182c386ef9d..17d573dbffaf9f0d37a7e1d1b366946cd3ae624f 100644 --- a/x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/nms.py +++ b/x2paddle/op_mapper/onnx2paddle/onnx_custom_layer/nms.py @@ -106,7 +106,8 @@ class NMS(object): if bboxes.shape[0] == 1: batch = paddle.zeros_like(clas, dtype="int64") else: - bboxes_count = bboxes.shape[1] + bboxes_count = paddle.shape(bboxes)[1] + bboxes_count = paddle.cast(bboxes_count, dtype="int64") batch = paddle.divide(index, bboxes_count) index = paddle.mod(index, bboxes_count) res = paddle.concat([batch, clas, index], axis=1) diff --git a/x2paddle/op_mapper/onnx2paddle/opset9/opset.py b/x2paddle/op_mapper/onnx2paddle/opset9/opset.py index 381b31cf04f6cc1c396eebe65507b11f7ed673c9..faf681df1298814c91fea1072b7f84480e4e16ad 100755 --- a/x2paddle/op_mapper/onnx2paddle/opset9/opset.py +++ b/x2paddle/op_mapper/onnx2paddle/opset9/opset.py @@ -620,15 +620,23 @@ class OpSet9(): pads) # NCHW if assume_pad: paddle_op = 'paddle.nn.Pad2D' + # x1_begin,x2_begin,x3_begin,x4_begin,x1_end,x2_end,x3_end,x4_end->x1_begin,x1_end,x2_begin,x2_end,x3_begin,x3_end,x4_begin,x4_end paddings = np.array(pads).reshape( (2, -1)).transpose().astype("int32") - paddings = np.flip(paddings, axis=0).flatten().tolist() - if sum(paddings[:4]) == 0: - paddings = paddings[4:] + if mode == 'constant': + paddings = paddings.flatten().tolist() layer_attrs['padding'] = paddings else: - layer_attrs["pad"] = paddings - paddle_op = "custom_layer:PadAllDim4WithOneInput" + paddings = np.flip(paddings, axis=0).flatten().tolist() + if sum(paddings[:4]) == 0: + paddings = paddings[4:] + layer_attrs['padding'] = paddings + else: + layer_attrs["pad"] = paddings + paddle_op = "custom_layer:PadAllDim4WithOneInput" + else: + paddle_op = 'paddle.nn.functional.pad' + layer_attrs["pad"] = np.array(pads).tolist() else: pad_data_temp = pads[0::2] pad_data_all = [] @@ -1464,11 +1472,18 @@ class OpSet9(): outputs_list.append("{}_p{}".format(node.layer_name, i)) else: outputs_list.append(node.name) - self.paddle_graph.add_layer( - 'paddle.split', - inputs={"x": val_x.name}, - outputs=outputs_list, - **layer_attrs) + if len(split) > 1: + self.paddle_graph.add_layer( + 'paddle.split', + inputs={"x": val_x.name}, + outputs=outputs_list, + **layer_attrs) + else: + self.paddle_graph.add_layer( + "paddle.cast", + inputs={"x": val_x.name}, + outputs=outputs_list, + dtype=string(val_x.dtype)) @print_mapping_info def Reshape(self, node): @@ -2698,28 +2713,36 @@ class OpSet9(): layer_outputs = [nn_op_name, output_name] boxes = self.graph.get_input_node(node, idx=0, copy=True) scores = self.graph.get_input_node(node, idx=1, copy=True) - num_classes = scores.out_shapes[0][1] inputs_len = len(node.layer.input) layer_attrs = dict() + layer_attrs["keep_top_k"] = -1 + layer_attrs["nms_threshold"] = 0.0 + layer_attrs["score_threshold"] = 0.0 if inputs_len > 2: max_output_boxes_per_class = self.graph.get_input_node( node, idx=2, copy=True) - layer_attrs["keep_top_k"] = _const_weight_or_none( - max_output_boxes_per_class).tolist()[0] * num_classes - else: - layer_attrs["keep_top_k"] = 0 + max_output_boxes_per_class = _const_weight_or_none( + max_output_boxes_per_class) + if len(scores.out_shapes[0]) != 0: + num_classes = scores.out_shapes[0][1] + else: + num_classes = 1 + if max_output_boxes_per_class is not None: + max_output_boxes_per_class = max_output_boxes_per_class.tolist() + if isinstance(max_output_boxes_per_class, int): + layer_attrs[ + "keep_top_k"] = max_output_boxes_per_class * num_classes + else: + layer_attrs["keep_top_k"] = max_output_boxes_per_class[ + 0] * num_classes if inputs_len > 3: iou_threshold = self.graph.get_input_node(node, idx=3, copy=True) layer_attrs["nms_threshold"] = _const_weight_or_none( iou_threshold).tolist()[0] - else: - layer_attrs["nms_threshold"] = 0.0 if inputs_len > 4: score_threshold = self.graph.get_input_node(node, idx=4, copy=True) layer_attrs["score_threshold"] = _const_weight_or_none( score_threshold).tolist()[0] - else: - layer_attrs["score_threshold"] = 0.0 self.paddle_graph.add_layer( "custom_layer:NMS", inputs={"bboxes": boxes.name,