I am trying to run pretrained TabFlex and MotherNet models using the current TICL repository. The package installs correctly and both model classes import successfully, but their checkpoints cannot be downloaded because the configured Azure Blob Storage hostname no longer resolves.
The affected hostname is:
amuellermothernet.blob.core.windows.net
Both my notebook server and a separate web browser return a DNS NXDOMAIN error.
Steps to reproduce
from ticl.prediction.tabflex import TabFlex
model = TabFlex()
This produces:
urllib.error.URLError:
<urlopen error [Errno -2] Name or service not known>
The DNS failure can also be reproduced directly:
getent hosts amuellermothernet.blob.core.windows.net
curl -I --connect-timeout 15
https://amuellermothernet.blob.core.windows.net/
Result:
curl: (6) Could not resolve host:
amuellermothernet.blob.core.windows.net
Environment
OS: Linux
Python: 3.10
PyTorch: 2.0.0
CUDA build: 11.8
CUDA available: True
GPU: NVIDIA RTX A6000
TICL installation: editable installation from the current GitHub repository
The following imports work correctly:
from ticl.prediction.tabflex import TabFlex
from ticl.prediction import MotherNetClassifier, EnsembleMeta
Required checkpoints
TabFlex appears to require these checkpoints:
-
ssm_tabpfn_b4_maxnumclasses100_modellinear_attention_numfeatures1000_n1024_validdatanew_warm_08_23_2024_19_25_40_epoch_3140.cpkt
-
ssm_tabpfn_b4_largedatasetTrue_modellinear_attention_nsamples50000_08_01_2024_22_05_50_epoch_110.cpkt
-
ssm_tabpfn_modellinear_attention_08_28_2024_19_00_44_epoch_3110.cpkt
The default MotherNet classifier appears to require:
- mn_Dclass_average_03_25_2024_17_14_32_epoch_3970.cpkt
The microsoft/mothernet Hugging Face repository currently contains only tabpfn_07_24_2023_epoch1650.cpkt, which appears to be an older TabPFN reference checkpoint rather than the MotherNet checkpoint required by MotherNetClassifier.
The TabFlex Hugging Face paper page also does not provide model files.
Request
Could the maintainers please:
- Restore the Azure checkpoint host, or
- Upload the required TabFlex and MotherNet checkpoints to Hugging Face or another persistent public location, and
- Update fetch_model() to use the new checkpoint URLs?
It would also be helpful to publish checksums for the checkpoint files so downloaded or mirrored copies can be verified.
This currently prevents reproduction of both the MotherNet and TabFlex inference examples, even when TICL and its CUDA dependencies are installed successfully.
I am trying to run pretrained TabFlex and MotherNet models using the current TICL repository. The package installs correctly and both model classes import successfully, but their checkpoints cannot be downloaded because the configured Azure Blob Storage hostname no longer resolves.
The affected hostname is:
amuellermothernet.blob.core.windows.net
Both my notebook server and a separate web browser return a DNS NXDOMAIN error.
Steps to reproduce
from ticl.prediction.tabflex import TabFlex
model = TabFlex()
This produces:
urllib.error.URLError:
<urlopen error [Errno -2] Name or service not known>
The DNS failure can also be reproduced directly:
getent hosts amuellermothernet.blob.core.windows.net
curl -I --connect-timeout 15
https://amuellermothernet.blob.core.windows.net/
Result:
curl: (6) Could not resolve host:
amuellermothernet.blob.core.windows.net
Environment
OS: Linux
Python: 3.10
PyTorch: 2.0.0
CUDA build: 11.8
CUDA available: True
GPU: NVIDIA RTX A6000
TICL installation: editable installation from the current GitHub repository
The following imports work correctly:
from ticl.prediction.tabflex import TabFlex
from ticl.prediction import MotherNetClassifier, EnsembleMeta
Required checkpoints
TabFlex appears to require these checkpoints:
ssm_tabpfn_b4_maxnumclasses100_modellinear_attention_numfeatures1000_n1024_validdatanew_warm_08_23_2024_19_25_40_epoch_3140.cpkt
ssm_tabpfn_b4_largedatasetTrue_modellinear_attention_nsamples50000_08_01_2024_22_05_50_epoch_110.cpkt
ssm_tabpfn_modellinear_attention_08_28_2024_19_00_44_epoch_3110.cpkt
The default MotherNet classifier appears to require:
The microsoft/mothernet Hugging Face repository currently contains only tabpfn_07_24_2023_epoch1650.cpkt, which appears to be an older TabPFN reference checkpoint rather than the MotherNet checkpoint required by MotherNetClassifier.
The TabFlex Hugging Face paper page also does not provide model files.
Request
Could the maintainers please:
It would also be helpful to publish checksums for the checkpoint files so downloaded or mirrored copies can be verified.
This currently prevents reproduction of both the MotherNet and TabFlex inference examples, even when TICL and its CUDA dependencies are installed successfully.