non_max_suppression结果size为0

[linklist2]

请问我使用n系列的pt文件转为weights文件后，进行训练后，训练的过程是正常的，但是在进行val的时候，non_max_suppression后pds的size为0，后面的代码就报错了。

但是如果使用您提供的yolov5s.weights训练的话，就不会出现问题。

我转换的代码参考您提供的代码，如下所示：

from pathlib import Path

from models.common import DetectMultiBackend
import numpy as np
import torch
from utils.torch_utils import select_device

weights = 'yolov5s.weights'

pt_weights = 'yolov5s.pt'

device = select_device("cuda:0")
model = DetectMultiBackend(pt_weights, device=device, dnn=False, data='data/coco128.yaml', fp16=False)

# parms "model" is DetectMultiBackend's instance
ddd = model.model.state_dict()
i = 0
if Path(weights).exists():
    f = open(weights, 'rb')
    for k, v in ddd.items():
        if 'weight' in k or 'bias' in k or 'running_mean' in k or 'running_var' in k:
            # t = v.cpu().numpy()
            # if 'weight' in k or 'bias' in k:
            #     print(k, i)
            #     i += 1

            nb = v.element_size() * v.numel()
            data_ = f.read(nb)
            y = np.frombuffer(data_, np.float32).reshape(v.shape)
            y = torch.from_numpy(y).to(v.device)
            v[...] = torch.zeros(v.shape, device=v.device)[...]
            v[...] = y[...]
            print(i, k, v.shape, nb)
            i += 1
            # nb2 = t.nbytes
            # print(k, t.shape, t.dtype)
            # f.write(t.tobytes())
        # v[...] = torch.zeros(v.shape,device=v.device)[...]

    f.close()

else:
    f = open(weights, 'wb')
    for k, v in ddd.items():
        if 'weight' in k or 'bias' in k or 'running_mean' in k or 'running_var' in k:
            t = v.cpu().numpy()
            print(k, t.shape, t.dtype)
            f.write(t.tobytes())
        # v[...] = torch.zeros(v.shape,device=v.device)[...]

    f.close()

[wangzhenlin123]

你用的是哪个libtorch版本呢