hub.load导入"retinanet_res50_1x_800size"报错

[davidishere]

环境

1.系统环境：Ubuntu 18.04
2.MegEngine版本：0.4.0
3.python版本：3.6.8
4.模型名称：retinanet_res50_1x_800size

复现步骤

请提供关键的代码片段便于追查问题

from megengine import hub
model = hub.load(
"megengine/models",
"retinanet_res50_1x_800size",
pretrained=True,
)

请提供完整的日志及报错信息

Traceback (most recent call last):
File "", line 4, in
File "/usr/local/lib/python3.6/dist-packages/megengine/hub/hub.py", line 202, in load
raise RuntimeError("Cannot find callable {} in hubconf.py".format(entry))
RuntimeError: Cannot find callable retinanet_res50_1x_800size in hubconf.py

[FateScript]

环境

1.系统环境：Ubuntu 18.04
2.MegEngine版本：0.4.0
3.python版本：3.6.8
4.模型名称：retinanet_res50_1x_800size

复现步骤

请提供关键的代码片段便于追查问题

from megengine import hub
model = hub.load(
"megengine/models",
"retinanet_res50_1x_800size",
pretrained=True,
)

请提供完整的日志及报错信息

Traceback (most recent call last):
File "", line 4, in
File "/usr/local/lib/python3.6/dist-packages/megengine/hub/hub.py", line 202, in load
raise RuntimeError("Cannot find callable {} in hubconf.py".format(entry))
RuntimeError: Cannot find callable retinanet_res50_1x_800size in hubconf.py

看起来是cache没更新的问题，可以使用

model = hub.load(
    "megengine/models", "retinanet_res50_1x_800size",
    pretrained=True, use_cache=False
)

试一下，应该会重新download一份新的retinanet

[davidishere]

修改后报下面的错误：

Traceback (most recent call last):
File "", line 1, in
File "/usr/local/lib/python3.6/dist-packages/megengine/hub/hub.py", line 206, in load
module = getattr(hubmodule, entry)(*args, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/megengine/hub/hub.py", line 317, in pretrained_model_func
model = func(**kwargs)
File "/root/.cache/megengine/hub/e093aa62cdf8f313/official/vision/detection/retinanet_res50_coco_1x_800size.py", line 23, in retinanet_res50_coco_1x_800size
return models.RetinaNet(models.RetinaNetConfig(), batch_size=batch_size, **kwargs)
File "/root/.cache/megengine/hub/e093aa62cdf8f313/official/vision/detection/models/retinanet.py", line 39, in init
bottom_up = resnet50(norm=layers.get_norm(self.cfg.resnet_norm))
File "/root/.cache/megengine/hub/e093aa62cdf8f313/official/vision/detection/layers/basic/norm.py", line 71, in get_norm
"SyncBN": M.SyncBatchNorm,
AttributeError: module 'megengine.module' has no attribute 'SyncBatchNorm'

[FateScript]

修改后报下面的错误：

Traceback (most recent call last):
File "", line 1, in
File "/usr/local/lib/python3.6/dist-packages/megengine/hub/hub.py", line 206, in load
module = getattr(hubmodule, entry)(*args, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/megengine/hub/hub.py", line 317, in pretrained_model_func
model = func(**kwargs)
File "/root/.cache/megengine/hub/e093aa62cdf8f313/official/vision/detection/retinanet_res50_coco_1x_800size.py", line 23, in retinanet_res50_coco_1x_800size
return models.RetinaNet(models.RetinaNetConfig(), batch_size=batch_size, **kwargs)
File "/root/.cache/megengine/hub/e093aa62cdf8f313/official/vision/detection/models/retinanet.py", line 39, in init
bottom_up = resnet50(norm=layers.get_norm(self.cfg.resnet_norm))
File "/root/.cache/megengine/hub/e093aa62cdf8f313/official/vision/detection/layers/basic/norm.py", line 71, in get_norm
"SyncBN": M.SyncBatchNorm,
AttributeError: module 'megengine.module' has no attribute 'SyncBatchNorm'

哦哦，这个是0.5.0里面即将发布的Sync bn，目前0.4.0还不支持，可以把这一行注释掉

[davidishere]

您好，我看了retinanet的配置，并没有用Sync bn
class RetinaNetConfig:
def init(self):
self.resnet_norm = "FrozenBN"
self.backbone_freeze_at = 2

retinanet_res50_coco_1x_800size.py

from megengine import hub

from official.vision.detection import models

@hub.pretrained(
"https://data.megengine.org.cn/models/weights/"
"retinanet_d3f58dce_res50_1x_800size_36dot0.pkl"
)
def retinanet_res50_coco_1x_800size(batch_size=1, **kwargs):
r"""
RetinaNet trained from COCO dataset.
"RetinaNet" <https://arxiv.org/abs/1708.02002>_
"""
return models.RetinaNet(models.RetinaNetConfig(), batch_size=batch_size, **kwargs)

Net = models.RetinaNet
Cfg = models.RetinaNetConfig

并没有调用retinanet_res50_coco_1x_800size_syncbn.py

[FateScript]

您好，我看了retinanet的配置，并没有用Sync bn
class RetinaNetConfig:
def init(self):
self.resnet_norm = "FrozenBN"
self.backbone_freeze_at = 2

retinanet_res50_coco_1x_800size.py

from megengine import hub

from official.vision.detection import models

@hub.pretrained(
"https://data.megengine.org.cn/models/weights/"
"retinanet_d3f58dce_res50_1x_800size_36dot0.pkl"
)
def retinanet_res50_coco_1x_800size(batch_size=1, **kwargs):
r"""
RetinaNet trained from COCO dataset.
"RetinaNet" <https://arxiv.org/abs/1708.02002>_
"""
return models.RetinaNet(models.RetinaNetConfig(), batch_size=batch_size, **kwargs)

Net = models.RetinaNet
Cfg = models.RetinaNetConfig

并没有调用retinanet_res50_coco_1x_800size_syncbn.py

没有使用sync bn的原因是相关的模型还在训练中，另外模型提交到hub也需要一定时间

[davidishere]

那是不是现在retinanet模型的预训练权重还无法使用？需要自己重新训练

[FateScript]

那是不是现在retinanet模型的预训练权重还无法使用？需要自己重新训练

可以使用啊，只不过是sync bn的功能不支持而已，正常RetinaNet使用的是Frozen BN

[davidishere]

您好，现在的代码中就是使用的Frozen BN，不知道为什么说用了sync bn，您可以试一下看看是不是有问题？

[FateScript]

您好，现在的代码中就是使用的Frozen BN，不知道为什么说用了sync bn，您可以试一下看看是不是有问题？

您好，这个地方多了一个SyncBn，而现在0.4.0版本的MegEngine并不支持，我的意思是注释掉这一行你就可以正常使用RetinaNet了，您可以试一下

[davidishere]

OK，谢谢！