Tensorflow to pytorch notes; tf. gather_ Nd (x, y) to pytorch

Recorded will tensorflow turn pytorch when , Some commonly used function conversions ：

Can't convert directly

1. tf.transpose(input,[1, 0, 2]) -> input.permute([1, 0, 2]) It cannot be directly replaced by torch.transpose, Because you can't operate multidimensional
2. tf.expand_dims(input), axis=1)->input.unsqueeze(1)
3. tf.concat([content1,content2], axis=1->torch.cat((content1,content2), dim=1) Remember to axis Switch to dim
4. tf.tile(input, [2, 1])-> input.repeat([2, 1])
5. tf.range(10)->torch.arange(0)
6. tf.reduce_sum(x, axis=1, keep_dims=True)-> torch.sum(x,dim=1,keepdim=True)
7. tf.clip_by_value(x, min, max)->torch.clamp(x, min, max)
8. tf.multinomial(logits=a, num_samples=1)->torch.multinomial(input=a, num_samples=1, replacement=False)
9. tf.equal(x, y)->torch.eq(x, y)
10. tf.nn.embedding_lookup(W_fe, Feature_input + 1)-> torch.index_select(W_fe, 0, Feature_input + 1)
11. tf.one_hot()->functional.one_hot()

tf.gather_nd(x,y) transformation

Reference article

    def gather_nd(self,params, indices):
        ''' 4D example params: tensor shaped [n_1, n_2, n_3, n_4] --> 4 dimensional indices: tensor shaped [m_1, m_2, m_3, m_4, 4] --> multidimensional list of 4D indices returns: tensor shaped [m_1, m_2, m_3, m_4] ND_example params: tensor shaped [n_1, ..., n_p] --> d-dimensional tensor indices: tensor shaped [m_1, ..., m_i, d] --> multidimensional list of d-dimensional indices returns: tensor shaped [m_1, ..., m_1] '''
        out_shape = indices.shape[:-1]
        indices = indices.unsqueeze(0).transpose(0, -1) # roll last axis to fring
        ndim = indices.shape[0]
        indices = indices.long()
        idx = torch.zeros_like(indices[0], device=indices.device).long()
        m = 1
        for i in range(ndim)[::-1]:
            idx += indices[i] * m
            m *= params.size(i)
        out = torch.take(params, idx)
        return out.view(out_shape)

Can convert directly

1. tf.reshape()->torch.reshape()
2. tf.log()
3. tf.squeeze