gmalivenko / gluon2pytorch Goto Github PK

Implement:

• LSTM
• GRU

I loaded the model from .json and .params，
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix[3], epoch[3])
mod = mx.mod.Module(symbol=L106_Net96_v2("test"), context=mx.gpu(), data_names=['data'], label_names=['landmark_target'])
mod.bind(for_training=False, data_shapes=[('data', (batch_size[3], 3, 96, 96))])
mod.set_params(arg_params, aux_params)
When I call pytorch_model = gluon2pytorch(mod, [(batch_size[3], 3, 96, 96)], dst_dir=None, pytorch_module_name='Test')
Return: TypeError: 'Module' object is not callable

How can I solve this problem?How to convert module to be hybrid ?

I would like to convert a Gluoncv pretrained model 'gluoncv.model_zoo.DeepLabV3Plus' to Pytorch one, but SyntaxError appears during conversion:

self.17 = nn.ConstantPad2d(padding=(1, 1, 1, 1), value=0)
          ^
SyntaxError: invalid syntax

The executable code:

net = gluoncv.model_zoo.DeepLabV3Plus(nclass=4, crop_size=224)
net.hybridize()
net.collect_params().initialize()
model = gluon2pytorch(net, [(1, 3, 224, 224)], dst_dir=None, pytorch_module_name='DeepLabv3')

Thank you in advance for your attention.

Nice work! I wonder why you change the class attributes name from the original ones like self.conv1, self.bn1, etc to self.x345, self.x347? That would potentially pose problems when we need to use the class attributes after conversion.

Could you share your thoughts on how to change the script to keep the original class attributes? Thanks.

Auto-generated layers name is not a good decision, It's very complicated to debug model with names like "x122" and "x1823".