通过NCNN部署YOLOv_Deploy YOLOv6 by NCNN.zip

# YOLOv6-NCNN Deploy YOLOv6 by NCNN ## 1. Change Detect module Change the forward function of Detect module (In effidehead*.py) as follow: ``` def forward(self, x): if self.training: ...(Do not change) else: cls_score_list = [] reg_lrtb_list = [] for i in range(self.nl): b, _, h, w = x[i].shape l = h * w x[i] = self.stems[i](x[i]) cls_x = x[i] reg_x = x[i] cls_feat = self.cls_convs[i](cls_x) cls_output = self.cls_preds[i](cls_feat) reg_feat = self.reg_convs[i](reg_x) reg_output_lrtb = self.reg_preds_lrtb[i](reg_feat) cls_output = torch.sigmoid(cls_output) cls_score_list.append(cls_output.reshape([b, self.nc, l])) reg_lrtb_list.append(reg_output_lrtb.reshape([b, 4, l])) cls_score_list = torch.cat(cls_score_list, axis=-1).permute(0, 2, 1) reg_lrtb_list = torch.cat(reg_lrtb_list, axis=-1).permute(0, 2, 1) # Disable the part of parsing bbox EXPORT_NCNN = True if EXPORT_NCNN: return torch.cat([ reg_lrtb_list, torch.ones((b, reg_lrtb_list.shape[1], 1), device=reg_lrtb_list.device, dtype=reg_lrtb_list.dtype), cls_score_list], axis=-1 ) anchor_points, stride_tensor = generate_anchors( x, self.stride, self.grid_cell_size, self.grid_cell_offset, device=x[0].device, is_eval=True, mode='af') pred_bboxes = dist2bbox(reg_lrtb_list, anchor_points, box_format='xywh') pred_bboxes *= stride_tensor return torch.cat( [ pred_bboxes, torch.ones((b, pred_bboxes.shape[1], 1), device=pred_bboxes.device, dtype=pred_bboxes.dtype), cls_score_list ], axis=-1) ``` Set the **EXPORT_NCNN** to **True** before export. ## 2. Export to ONNX Use following commands to export Pytorch model to ONNX model: ``` cd YOLOv6 python deploy/ONNX/export_onnx.py --weights path-to-pt-model --img-size 640 640 --batch-size 1 --simplify --device 0 ``` **--simplify** must be required, the othe parameters can be changed according to the actual situation. ## 3. Convert ONNX to NCNN Use the executable file "onnx2ncnn.exe" to convert ONNX model to NCNN model: ``` path-to-onnx2ncnn.exe path-to-yolov6.onnx path-to-yolov6.param path-to-yolov6.bin ``` Then, modify the first dimension of **Reshape** layer in .param file: ``` # Before ... Reshape Reshape_143 1 1 290 303 0=6400 1=17 2=4 ... Reshape Reshape_153 1 1 307 316 0=6400 1=80 Reshape Reshape_159 1 1 306 325 0=6400 1=4 ... Reshape Reshape_184 1 1 346 359 0=1600 1=17 2=4 ... Reshape Reshape_194 1 1 363 372 0=1600 1=80 Reshape Reshape_200 1 1 362 381 0=1600 1=4 ... Reshape Reshape_225 1 1 402 415 0=400 1=17 2=4 ... Reshape Reshape_235 1 1 419 428 0=400 1=80 Reshape Reshape_241 1 1 418 437 0=400 1=4 ... # After ... Reshape Reshape_143 1 1 290 303 0=-1 1=17 2=4 ... Reshape Reshape_153 1 1 307 316 0=-1 1=80 Reshape Reshape_159 1 1 306 325 0=-1 1=4 ... Reshape Reshape_184 1 1 346 359 0=-1 1=17 2=4 ... Reshape Reshape_194 1 1 363 372 0=-1 1=80 Reshape Reshape_200 1 1 362 381 0=-1 1=4 ... Reshape Reshape_225 1 1 402 415 0=-1 1=17 2=4 ... Reshape Reshape_235 1 1 419 428 0=-1 1=80 Reshape Reshape_241 1 1 418 437 0=-1 1=4 ... ``` Otherwise, you can use fp16 to save and inference model by this command: ``` path-to-ncnnoptimize.exe path-to-yolov6.param path-to-yolov6.bin path-to-yolov6-opt.param path-to-yolov6-opt.bin 65536 ``` ## 4. Deploy NCNN model Load and inference NCNN models as the way in **yolov6.cpp**. Here is an example of detection result:  See more details at: https://zhuanlan.zhihu.com/p/609567390