I tried to run this script
tutorials/frontend/deploy_prequantized.py, however, an unexpected error appeared.
AssertionError: parameter classifier.1._packed_params._packed_params not found in state dict
and I also found a post proposed by this author. it said that after jitting, parameters are packed in a different way.
So he added some code in tvm to address this. But I also tested it by using colab.
import numpy as np import torch from torchvision.models.quantization import mobilenet as qmobilenet qmodel = qmobilenet.mobilenet_v2(pretrained=True).eval() input_size = (1, 3, 224, 224) inp = np.random.randn(*input_size).astype("float32") trace = torch.jit.trace(qmodel, torch.from_numpy(inp)) state_dict = trace.state_dict() for (k, v) in state_dict.items(): print(k, v.size())
the output is
...... features.18.1.running_var torch.Size() features.18.1.num_batches_tracked torch.Size() classifier.1.weight torch.Size([1000, 1280]) classifier.1.bias torch.Size()
I guess that this error probably caused by different versions of PyTorch. Then I degraded my PyTorch to 1.4, the error is disappeared.@masahi