After logging in on your machine, you can download the checkpoints:
from huggingface_hub import hf_hub_download
REPO_ID = "micromind/ImageNet"
FILENAME = "v5/state_dict.pth.tar"
model_path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)
followed by:
model = PhiNet(
input_shape=(3, 224, 224),
alpha=...,
num_layers=...,
beta=...,
t_zero=...,
include_top=True,
num_classes=1000,
compatibility=False,
divisor=8,
downsampling_layers=[4,5,7]
)
model.load_state_dict(torch.load(model_path))
Note for v1, when initializing the network, use:
downsampling_layers=[5,7]
Performance:
| Model name | Acc@1 | Acc@5 |
|---|---|---|
| v1 | 71.18% | 89.65% |
| v2 | 65.21% | 85.82% |
| v3 | 64.69% | 86.15% |
| v5 | 67.99% | 87.53% |
| v6 | 61.86% | 83.44% |
| v7 | 53.66% | 77.13% |