Here is a note for converting PyTorch to PaddlePaddle:
In most cases, one just replace torch by paddle:
| PyTorch | PaddlePaddle |
|---|---|
import torch |
import paddle |
import torch.nn as nn |
import paddle.nn as nn |
import torch.nn.functional as F |
import paddle.nn.functional as F |
Some cases one must notice (inp and out are torch.Tensor or paddle.Tensor):
| Description | PyTorch | PaddlePaddle |
|---|---|---|
| paddle's shape parameter must be a tuple/list | out = inp.view(b, c, h, w) |
out = inp.reshape((b, c, h, w)) |
| different transpose | out = inp.transpose(1, 2) |
out = inp.transpose((0, 1, 2)) |
| pytorch's max returns a tuple, while paddle returns the max values | out = inp.max(0)[0] |
out = inp.max(0) |
| pytorch's min returns a tuple, while paddle returns the min values | out = inp.min(0)[0] |
out = inp.min(0) |
| paddle has no inplace operations | out.unsqueeze_(0) |
out = inp.unsqueeze(0) |
| different api names | optimizer.zero_grad() |
optimizer.clear_grad() |
| different input shape | cross_entropy(pred, mask) |
cross_entropy(pred.transpose((0, 2, 3, 1)), mask) (pred has shape [B, C, H, W]) |
| paddle's linear weights has transposed dimensions | nn.Linear(2, 6).weight.shape == torch.Size([6, 2]) |
nn.Linear(2, 6).weight.shape == [2, 6] |
| paddle's conv weights has the same dimensions | nn.Conv2d(2, 6, 3).weight.shape == torch.Size([6, 2, 3, 3]) |
nn.Conv2D(2, 6, 3).weight.shape == [6, 2, 3, 3] |