can't load pre-trained weights for YOLO-NAS using the SuperGradients

[ApodiAlsurihi]

💡 Your Question

Hello,
I am experiencing an issue while trying to load pre-trained weights for YOLO-NAS using the SuperGradients library.
Here is the code snippet I am using:

python
Copy code
from super_gradients.training import models

model = models.get('yolo_nas_s',
num_classes=len(dataset_params['classes']),
pretrained_weights="coco"
)
When I run this code, I get the following error:

Downloading: "https://sghub.deci.ai/models/yolo_nas_s_coco.pth" to C:\Users\Owner/.cache\torch\hub\checkpoints\yolo_nas_s_coco.pth
URLError: <urlopen error [Errno 11001] getaddrinfo failed>
It seems there is an issue with accessing the URL to download the pre-trained weights. I suspect it might be network-related, but I'm unsure if there are any alternative links or settings I need to change.
Any help or suggestions would be greatly appreciated!
Thank you in advance!

Versions

No response

[TychoBomer]

Hi First check this issue where it has been discussed:

#2057

then follow exactly what tatsuya-fukuoka gave as fix.

You basically change the original bucket where weights where stored to the new bucket https://sg-hub-nv.s3.amazonaws.com/. In pretrained_models.py you update the full path to your weights and in checkpoint_utils.py you change the line with unique_filename.

Let me know if you still need help after!

[rabion1234]

@TychoBomer I am still getting error after updating .

error: 2024-11-30 15:14:07] WARNING - checkpoint_utils.py - :warning: The pre-trained models provided by SuperGradients may have their own licenses or terms and conditions derived from the dataset used for pre-training.
It is your responsibility to determine whether you have permission to use the models for your use case.
The model you have requested was pre-trained on the coco dataset, published under the following terms: https://cocodataset.org/#termsofuse
[2024-11-30 15:14:07] INFO - checkpoint_utils.py - License Notification: YOLO-NAS pre-trained weights are subjected to the specific license terms and conditions detailed in
https://github.com/Deci-AI/super-gradients/blob/master/LICENSE.YOLONAS.md
By downloading the pre-trained weight files you agree to comply with these terms.
Downloading: "https://sghub.deci.ai/models/yolo_nas_l_coco.pth" to /root/.cache/torch/hub/checkpoints/yolo_nas_l_coco.pth

gaierror Traceback (most recent call last)
/usr/lib/python3.10/urllib/request.py in do_open(self, http_class, req, **http_conn_args)
1347 try:
-> 1348 h.request(req.get_method(), req.selector, req.data, headers,
1349 encode_chunked=req.has_header('Transfer-encoding'))

21 frames
/usr/lib/python3.10/http/client.py in request(self, method, url, body, headers, encode_chunked)
1282 """Send a complete request to the server."""
-> 1283 self._send_request(method, url, body, headers, encode_chunked)
1284

/usr/lib/python3.10/http/client.py in _send_request(self, method, url, body, headers, encode_chunked)
1328 body = _encode(body, 'body')
-> 1329 self.endheaders(body, encode_chunked=encode_chunked)
1330

/usr/lib/python3.10/http/client.py in endheaders(self, message_body, encode_chunked)
1277 raise CannotSendHeader()
-> 1278 self._send_output(message_body, encode_chunked=encode_chunked)
1279

/usr/lib/python3.10/http/client.py in _send_output(self, message_body, encode_chunked)
1037 del self._buffer[:]
-> 1038 self.send(msg)
1039

/usr/lib/python3.10/http/client.py in send(self, data)
975 if self.auto_open:
--> 976 self.connect()
977 else:

/usr/lib/python3.10/http/client.py in connect(self)
1447
-> 1448 super().connect()
1449

/usr/lib/python3.10/http/client.py in connect(self)
941 sys.audit("http.client.connect", self, self.host, self.port)
--> 942 self.sock = self._create_connection(
943 (self.host,self.port), self.timeout, self.source_address)

/usr/lib/python3.10/socket.py in create_connection(address, timeout, source_address)
823 err = None
--> 824 for res in getaddrinfo(host, port, 0, SOCK_STREAM):
825 af, socktype, proto, canonname, sa = res

/usr/lib/python3.10/socket.py in getaddrinfo(host, port, family, type, proto, flags)
954 addrlist = []
--> 955 for res in _socket.getaddrinfo(host, port, family, type, proto, flags):
956 af, socktype, proto, canonname, sa = res

gaierror: [Errno -2] Name or service not known

During handling of the above exception, another exception occurred:

URLError Traceback (most recent call last)
in <cell line: 2>()
1 from super_gradients.training import models
----> 2 model = models.get('yolo_nas_l',
3 num_classes=len(dataset_params['classes']),
4 pretrained_weights="coco"
5 )

/usr/local/lib/python3.10/dist-packages/super_gradients/common/decorators/factory_decorator.py in wrapper(*args, **kwargs)
34 new_value = factory.get(args[index])
35 args = _assign_tuple(args, index, new_value)
---> 36 return func(*args, **kwargs)
37
38 return wrapper

/usr/local/lib/python3.10/dist-packages/super_gradients/training/models/model_factory.py in get(model_name, arch_params, num_classes, strict_load, checkpoint_path, pretrained_weights, load_backbone, download_required_code, checkpoint_num_classes, num_input_channels)
228
229 if checkpoint_num_classes:
--> 230 net = instantiate_model(model_name, arch_params, checkpoint_num_classes, pretrained_weights, download_required_code)
231 else:
232 net = instantiate_model(model_name, arch_params, num_classes, pretrained_weights, download_required_code)

/usr/local/lib/python3.10/dist-packages/super_gradients/training/models/model_factory.py in instantiate_model(model_name, arch_params, num_classes, pretrained_weights, download_required_code)
166 load_pretrained_weights_local(net, model_name, pretrained_weights_path)
167 else:
--> 168 load_pretrained_weights(net, model_name, pretrained_weights)
169
170 if pretrained_weights in PRETRAINED_NUM_CLASSES.keys() and num_classes_new_head != arch_params.num_classes:

/usr/local/lib/python3.10/dist-packages/super_gradients/training/utils/checkpoint_utils.py in load_pretrained_weights(model, architecture, pretrained_weights)
1593 map_location = torch.device("cpu")
1594 with wait_for_the_master(get_local_rank()):
-> 1595 pretrained_state_dict = load_state_dict_from_url(url=url, map_location=map_location, file_name=unique_filename)
1596
1597 _load_weights(architecture, model, pretrained_state_dict)

/usr/local/lib/python3.10/dist-packages/torch/hub.py in load_state_dict_from_url(url, model_dir, map_location, progress, check_hash, file_name, weights_only)
865 r = HASH_REGEX.search(filename) # r is Optional[Match[str]]
866 hash_prefix = r.group(1) if r else None
--> 867 download_url_to_file(url, cached_file, hash_prefix, progress=progress)
868
869 if _is_legacy_zip_format(cached_file):

/usr/local/lib/python3.10/dist-packages/torch/hub.py in download_url_to_file(url, dst, hash_prefix, progress)
706 file_size = None
707 req = Request(url, headers={"User-Agent": "torch.hub"})
--> 708 u = urlopen(req)
709 meta = u.info()
710 if hasattr(meta, "getheaders"):

/usr/lib/python3.10/urllib/request.py in urlopen(url, data, timeout, cafile, capath, cadefault, context)
214 else:
215 opener = _opener
--> 216 return opener.open(url, data, timeout)
217
218 def install_opener(opener):

/usr/lib/python3.10/urllib/request.py in open(self, fullurl, data, timeout)
517
518 sys.audit('urllib.Request', req.full_url, req.data, req.headers, req.get_method())
--> 519 response = self._open(req, data)
520
521 # post-process response

/usr/lib/python3.10/urllib/request.py in _open(self, req, data)
534
535 protocol = req.type
--> 536 result = self._call_chain(self.handle_open, protocol, protocol +
537 '_open', req)
538 if result:

/usr/lib/python3.10/urllib/request.py in _call_chain(self, chain, kind, meth_name, *args)
494 for handler in handlers:
495 func = getattr(handler, meth_name)
--> 496 result = func(*args)
497 if result is not None:
498 return result

/usr/lib/python3.10/urllib/request.py in https_open(self, req)
1389
1390 def https_open(self, req):
-> 1391 return self.do_open(http.client.HTTPSConnection, req,
1392 context=self._context, check_hostname=self._check_hostname)
1393

/usr/lib/python3.10/urllib/request.py in do_open(self, http_class, req, **http_conn_args)
1349 encode_chunked=req.has_header('Transfer-encoding'))
1350 except OSError as err: # timeout error
-> 1351 raise URLError(err)
1352 r = h.getresponse()
1353 except:

URLError: <urlopen error [Errno -2] Name or service not known>

I am working On colab

[TychoBomer]

In your error message it still points to the wrong pre-trained weights location. In your pasted error i read this:

[2024-11-30 15:14:07] INFO - checkpoint_utils.py - License Notification: YOLO-NAS pre-trained weights are subjected to the specific license terms and conditions detailed in
https://github.com/Deci-AI/super-gradients/blob/master/LICENSE.YOLONAS.md
By downloading the pre-trained weight files you agree to comply with these terms.
Downloading: "https://sghub.deci.ai/models/yolo_nas_l_coco.pth" to /root/.cache/torch/hub/checkpoints/yolo_nas_l_coco.pth

The link: "https://sghub.deci.ai/models/yolo_nas_l_coco.pth" is incorrect. As we stated in this issue above, you must manually reconfigure it in the supergradients package code itself to point at: "https://sg-hub-nv.s3.amazonaws.com/"

If you post the first link in the webbrowser you will find a unresponsive server error, which is perfectly in line with your error:
URLError: <urlopen error [Errno -2] Name or service not known>

I suppose you did not change everything discussed in #2057 or you changed it but imported the original.

Hope this helps!

[rabion1234]

I checked all but could not found any error but now i have downloaded manually my weights and putted in root directory so code run with out problem. thanks for help.

[erfan-rfmhr]

As a temporary solution, I modified the library code myself. Below is the change I implemented:

else:
    url = url.replace("https://sghub.deci.ai", "https://sg-hub-nv.s3.amazonaws.com")
    unique_filename = url.split("https://sg-hub-nv.s3.amazonaws.com/models/")[1].replace("/", "_").replace(" ", "_")

I modified the following section in the code: Link. After else statement you replace the broken URL with new one, you will also change the value of split function. It works now!
Updating the old domain throughout the code proved challenging, as it is hardcoded in multiple places. Let me know if there's a better approach to handle this more efficiently!

Update: I added more details after this reply.

[ppraneth]

As a temporary solution, I modified the library code myself. Below is the change I implemented:

url = url.replace("https://sghub.deci.ai", "https://sg-hub-nv.s3.amazonaws.com")

I inserted this line just before the following section in the code: Link and it worked! Updating the old domain throughout the code proved challenging, as it is hardcoded in multiple places. Let me know if there's a better approach to handle this more efficiently!
@erfan-rfmhr
where should i paste it above line 1592? if not can you please specify?

[erfan-rfmhr]

As a temporary solution, I modified the library code myself. Below is the change I implemented:
url = url.replace("https://sghub.deci.ai", "https://sg-hub-nv.s3.amazonaws.com")
I inserted this line just before the following section in the code: Link and it worked! Updating the old domain throughout the code proved challenging, as it is hardcoded in multiple places. Let me know if there's a better approach to handle this more efficiently!
@erfan-rfmhr
where should i paste it above line 1592? if not can you please specify?

Sorry for the bad explanation. After line 1591, the else statement, you should modify the code:

else:
    url = url.replace("https://sghub.deci.ai", "https://sg-hub-nv.s3.amazonaws.com")
    unique_filename = url.split("https://sg-hub-nv.s3.amazonaws.com/models/")[1].replace("/", "_").replace(" ", "_")

Do not forget to modify the value in split function in unique_filename variable, line 1592.

[matthias83100]

Hi, thank you, in fact this solution is working really well for the yolo_nas_s and yolo_nas_m but I actually want to use the yolo_nas_l model... Do you have an idea of why the import is not working for this model version? (I'm using super_gradient==3.2.0 version and python3.10)
Thanks for your help!
I have this weird error here:

`---------------------------------------------------------------------------
EOFError Traceback (most recent call last)
Cell In[4], line 5
2 from super_gradients.training import models
3 from super_gradients.common.object_names import Models
----> 5 model = models.get(Models.YOLO_NAS_L, pretrained_weights='coco')

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\super_gradients\training\models\model_factory.py:224, in get(model_name, arch_params, num_classes, strict_load, checkpoint_path, pretrained_weights, load_backbone, download_required_code, checkpoint_num_classes)
222 net = instantiate_model(model_name, arch_params, checkpoint_num_classes, pretrained_weights, download_required_code)
223 else:
--> 224 net = instantiate_model(model_name, arch_params, num_classes, pretrained_weights, download_required_code)
226 if load_backbone and not checkpoint_path:
227 raise ValueError("Please set checkpoint_path when load_backbone=True")

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\super_gradients\training\models\model_factory.py:159, in instantiate_model(model_name, arch_params, num_classes, pretrained_weights, download_required_code)
157 load_pretrained_weights_local(net, model_name, pretrained_weights_path)
158 else:
--> 159 load_pretrained_weights(net, model_name, pretrained_weights)
161 if pretrained_weights in PRETRAINED_NUM_CLASSES.keys() and num_classes_new_head != arch_params.num_classes:
162 net.replace_head(new_num_classes=num_classes_new_head)

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\super_gradients\training\utils\checkpoint_utils.py:1575, in load_pretrained_weights(model, architecture, pretrained_weights)
1573 map_location = torch.device("cpu")
1574 with wait_for_the_master(get_local_rank()):
-> 1575 pretrained_state_dict = load_state_dict_from_url(url=url, map_location=map_location, file_name=unique_filename)
1576 _load_weights(architecture, model, pretrained_state_dict)

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\torch\hub.py:871, in load_state_dict_from_url(url, model_dir, map_location, progress, check_hash, file_name, weights_only)
869 if _is_legacy_zip_format(cached_file):
870 return _legacy_zip_load(cached_file, model_dir, map_location, weights_only)
--> 871 return torch.load(cached_file, map_location=map_location, weights_only=weights_only)

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\torch\serialization.py:1495, in load(f, map_location, pickle_module, weights_only, mmap, **pickle_load_args)
1493 except pickle.UnpicklingError as e:
1494 raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
-> 1495 return _legacy_load(
1496 opened_file, map_location, pickle_module, **pickle_load_args
1497 )

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\torch\serialization.py:1744, in _legacy_load(f, map_location, pickle_module, **pickle_load_args)
1737 if not hasattr(f, "readinto") and (3, 8, 0) <= sys.version_info < (3, 8, 2):
1738 raise RuntimeError(
1739 "torch.load does not work with file-like objects that do not implement readinto on Python 3.8.0 and 3.8.1. "
1740 f'Received object of type "{type(f)}". Please update to Python 3.8.2 or newer to restore this '
1741 "functionality."
1742 )
-> 1744 magic_number = pickle_module.load(f, **pickle_load_args)
1745 if magic_number != MAGIC_NUMBER:
1746 raise RuntimeError("Invalid magic number; corrupt file?")

EOFError: Ran out of input`

[erfan-rfmhr]

This is weird! No idea. I use version 3.7.1, python 3.10, and model yolo_nas_s here. Maybe you should try downloading the model yourself🤷‍♂️

[THCNN]

Hi, thank you, in fact this solution is working really well for the yolo_nas_s and yolo_nas_m but I actually want to use the yolo_nas_l model... Do you have an idea of why the import is not working for this model version? (I'm using super_gradient==3.2.0 version and python3.10) Thanks for your help! I have this weird error here:

`--------------------------------------------------------------------------- EOFError Traceback (most recent call last) Cell In[4], line 5 2 from super_gradients.training import models 3 from super_gradients.common.object_names import Models ----> 5 model = models.get(Models.YOLO_NAS_L, pretrained_weights='coco')

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\super_gradients\training\models\model_factory.py:224, in get(model_name, arch_params, num_classes, strict_load, checkpoint_path, pretrained_weights, load_backbone, download_required_code, checkpoint_num_classes) 222 net = instantiate_model(model_name, arch_params, checkpoint_num_classes, pretrained_weights, download_required_code) 223 else: --> 224 net = instantiate_model(model_name, arch_params, num_classes, pretrained_weights, download_required_code) 226 if load_backbone and not checkpoint_path: 227 raise ValueError("Please set checkpoint_path when load_backbone=True")

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\super_gradients\training\models\model_factory.py:159, in instantiate_model(model_name, arch_params, num_classes, pretrained_weights, download_required_code) 157 load_pretrained_weights_local(net, model_name, pretrained_weights_path) 158 else: --> 159 load_pretrained_weights(net, model_name, pretrained_weights) 161 if pretrained_weights in PRETRAINED_NUM_CLASSES.keys() and num_classes_new_head != arch_params.num_classes: 162 net.replace_head(new_num_classes=num_classes_new_head)

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\super_gradients\training\utils\checkpoint_utils.py:1575, in load_pretrained_weights(model, architecture, pretrained_weights) 1573 map_location = torch.device("cpu") 1574 with wait_for_the_master(get_local_rank()): -> 1575 pretrained_state_dict = load_state_dict_from_url(url=url, map_location=map_location, file_name=unique_filename) 1576 _load_weights(architecture, model, pretrained_state_dict)

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\torch\hub.py:871, in load_state_dict_from_url(url, model_dir, map_location, progress, check_hash, file_name, weights_only) 869 if _is_legacy_zip_format(cached_file): 870 return _legacy_zip_load(cached_file, model_dir, map_location, weights_only) --> 871 return torch.load(cached_file, map_location=map_location, weights_only=weights_only)

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\torch\serialization.py:1495, in load(f, map_location, pickle_module, weights_only, mmap, **pickle_load_args) 1493 except pickle.UnpicklingError as e: 1494 raise pickle.UnpicklingError(_get_wo_message(str(e))) from None -> 1495 return _legacy_load( 1496 opened_file, map_location, pickle_module, **pickle_load_args 1497 )

File c:\Users\matthias.sto\Desktop\Training.venv\lib\site-packages\torch\serialization.py:1744, in _legacy_load(f, map_location, pickle_module, **pickle_load_args) 1737 if not hasattr(f, "readinto") and (3, 8, 0) <= sys.version_info < (3, 8, 2): 1738 raise RuntimeError( 1739 "torch.load does not work with file-like objects that do not implement readinto on Python 3.8.0 and 3.8.1. " 1740 f'Received object of type "{type(f)}". Please update to Python 3.8.2 or newer to restore this ' 1741 "functionality." 1742 ) -> 1744 magic_number = pickle_module.load(f, **pickle_load_args) 1745 if magic_number != MAGIC_NUMBER: 1746 raise RuntimeError("Invalid magic number; corrupt file?")

EOFError: Ran out of input`

Hi @matthias83100
I applied the solution here: #2057 (comment)

I worked with:
Python 3.9.21
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
Supergradients 3.7.1

Modifying line 49 in “pretrained_models.py” results in:
"yolo_nas_l_coco": "https://sg-hub-nv.s3.amazonaws.com/models/yolo_nas_l_coco.pth",

Modifying line 1592 in “checkpoint_utils.py” results in:
unique_filename = url.split("https://sg-hub-nv.s3.amazonaws.com/")[1].replace("/", "").replace(" ", "")

It works for me, I hope it works for you.
[2025-02-10 17:46:33] INFO - checkpoint_utils.py - Successfully loaded pretrained weights for architecture yolo_nas_l
[2025-02-10 17:46:34] INFO - checkpoint_utils.py - Successfully loaded model weights from ../ckpt_best.pth EMA checkpoint.

[matthias83100]

Hi @matthias83100 I applied the solution here: #2057 (comment)

I worked with: Python 3.9.21 conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia Supergradients 3.7.1

Modifying line 49 in “pretrained_models.py” results in: "yolo_nas_l_coco": "https://sg-hub-nv.s3.amazonaws.com/models/yolo_nas_l_coco.pth",

Modifying line 1592 in “checkpoint_utils.py” results in: unique_filename = url.split("https://sg-hub-nv.s3.amazonaws.com/")[1].replace("/%5B1%5D.replace(%22/)", "").replace(" ", "")

It works for me, I hope it works for you. [2025-02-10 17:46:33] INFO - checkpoint_utils.py - Successfully loaded pretrained weights for architecture yolo_nas_l [2025-02-10 17:46:34] INFO - checkpoint_utils.py - Successfully loaded model weights from ../ckpt_best.pth EMA checkpoint.

Hi thanks for your responses, I installed the .pth file directly in the checkpoint folder, because I already created a setup with my colleagues that requires the 3.2.0 version of super_gradients so I didn't try your solution.

[MRo47]

Got this too, following from here

from super_gradients.common.object_names import Models
from super_gradients.training import models

model = models.get(Models.YOLO_NAS_S, pretrained_weights="coco")

export_result = model.export("yolo_nas_s.onnx")

got: list index out of range

just the URL needs to change.

--- "https://sghub.deci.ai/models/yolo_nas_s_coco.pth"
+++ "https://sg-hub-nv.s3.amazonaws.com/models/"

Should be fixed in #2061