problem with converting custom Pytorch model to TensorFlow

[nestyme]

Platform (like ubuntu 16.04/win10): Ubuntu 16.04.5 LTS

Python version: Python 3.7

Source framework with version (like Tensorflow 1.4.1 with GPU): PyTorch 1.6.0

Destination framework with version (like CNTK 2.3 with GPU): TensorFlow 2.2.0

Pre-trained model path (webpath or webdisk path): A Structured Self-attentive Sentence Embedding model

class SentenceEmbeddingsModel(nn.Module):
    def __init__(self,
                 vocab_size,
                 embedding_dim,
                 max_length=40,
                 use_pretrained_word_vectors=True,
                 word_vectors=None,
                 device=device,
                 C=0.001,
                 d_a=10,
                 r_a=4,
                 hidden_size=100,
                 need_masking=True,
                 need_attention=True,
                 need_lstm=False):

        super(SentenceEmbeddingsModel, self).__init__()

        self.embeddings = nn.Embedding(vocab_size, embedding_dim)

        self.d_a = d_a
        self.C = C
        self.r_a = r_a
        self.rnn_hidden_size = hidden_size

        if use_pretrained_word_vectors:
            w =torch.FloatTensor(word_vectors)
            self.embeddings = self.embeddings.from_pretrained(w)
            self.embeddings.weight.requires_grad = False

        self.need_masking = need_masking
        self.need_attention = need_attention
        self.need_lstm = need_lstm

        if self.need_lstm:
            ws_d = 2 * self.rnn_hidden_size
        else:
            ws_d = embedding_dim

        self.ws1 = nn.Parameter(torch.FloatTensor(1, self.d_a, ws_d))
        nn.init.xavier_uniform_(self.ws1)
        self.ws1.requires_grad = True

        self.ws2 = nn.Parameter(torch.FloatTensor(1, self.r_a, self.d_a))
        nn.init.xavier_uniform_(self.ws2)
        self.ws2.requires_grad = True

        self.dropout1 = nn.Dropout(0.1)

        self.need_masking = need_masking
        self.need_attention = need_attention
        self.need_lstm = need_lstm

        self.device = device

        self.dense = nn.Sequential(
            nn.Linear(ws_d, 20, bias=True),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(20, num_classes, bias=True),
        )

        self.linear = nn.Linear(ws_d * self.r_a, ws_d)

    def forward(self, inputs):
        e = self.embeddings(inputs)
        if self.need_masking:
            mask = (inputs != 0)[:, :, None].float().to(self.device)
            masked = e.mul(mask)
            r = self.dropout1(masked)
        else:
            r = self.dropout1(e)

        z = r

        if self.need_attention:
            a1 =torch.tanh(self.ws1.matmul(z.transpose(dim0=1, dim1=2)))

            attention = F.softmax(self.ws2.matmul(a1), dim=2)  # n_batch - r_a - max_lentgh
            m = attention.matmul(z)  # n_batch - r_a - ws_d

            # here we get r_a * ws_d embedding matrix per sentence
            flatten = m.view(z.shape[0], -1, 1)[:, :, 0]

            m = self.linear(flatten)

        else:
            m = z.sum(1)
            attention = None

        out = torch.sigmoid(self.dense(m))
        return out

    model = SentenceEmbeddingsModel(vocab_len,
                             vocab_dim,
                             50,
                             use_pretrained_word_vectors=True,
                             word_vectors=word_vectors,
                             device=device,
                             hidden_size=100,
                             need_masking=True,
                             need_attention=True,
                             need_lstm=False)

I'm saving my model with torch.save(model, model path) then I'm trying to convert Pytorch model to TF via mmconvert -sf pytorch -in model.pth -df tensorflow -om model.ckpt I get Traceback (most recent call last): File "/usr/local/bin/mmconvert", line 10, in <module> sys.exit(_main()) File "/usr/local/lib/python3.6/dist-packages/mmdnn/conversion/_script/convert.py", line 102, in _main ret = convertToIR._convert(ir_args) File "/usr/local/lib/python3.6/dist-packages/mmdnn/conversion/_script/convertToIR.py", line 97, in _convert parser = PytorchParser151(model, inputshape[0]) File "/usr/local/lib/python3.6/dist-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 526, in __init__ super(PytorchParser151, self).__init__(model_file_name, input_shape) File "/usr/local/lib/python3.6/dist-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 83, in __init__ model = torch.load(model_file_name, map_location='cpu') File "/usr/local/lib/python3.6/dist-packages/torch/serialization.py", line 584, in load return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args) File "/usr/local/lib/python3.6/dist-packages/torch/serialization.py", line 842, in _load result = unpickler.load() AttributeError: Can't get attribute 'SentenceEmbeddingsModel' on <module '__main__' from '/usr/local/bin/mmconvert'>

Could anyone help me, please?)

[linmajia]

@nestyme , thank you very much for the feedback. You need to keep the model module in the same folder when loading the serialized model. Please refer to https://github.com/pytorch/pytorch/issues/18325 and https://github.com/pytorch/pytorch/issues/18325 for solution.