OneShotChannelPruner get error when reshape / view to 3 (or higher) dimension about tinyneuralnetwork

[dinghuanghao]

The problem you mentioned above has been fixed, but there is still a reshape problem, which is temporarily not supported.

        bn1 = self.bn1(fc1)
        reshape_2 = bn1.reshape(47, 16, 1024)
        lstm = self.lstm(reshape_2)

752 is reshaped to 47 and 16. If pruning dim0, then 16 adjacent neurons need to be deleted at one time. If pruning dim1, 47 equidistant neurons need to be deleted at one time. Pruning of this complex strategy is already under development, and we will provide a usable version as soon as possible.

[peterjc123]

@dinghuanghao

[dinghuanghao]

If possible, please generate model code for me to debug, methods as below:

import torch
import torchvision

from tinynn.graph.tracer import model_tracer, trace

with model_tracer():
    model = torchvision.models.alexnet()  # your model
    model.eval()

    dummy_input = torch.rand((1, 3, 224, 224))
    graph = trace(model, dummy_input)
    graph.generate_code('my_alexnet.py', 'my_alexnet.pth', 'Alexnet')

[juyenshun]

Hi,
I wrote a test model to duplicate this problem, and follow the instruction to generate model code.
However, it seems that the pth file is too large to upload, please run the prune_test.py and generate it. Thanks

prune_test20220418_1.zip

[dinghuanghao]

The latest version already supports reshape, permute, transpose and other operators