转换 Pytorch -> Onnx -> Tensorflow 的张量格式问题

Question

我对从 Pytorch -> Onnx -> Tensorflow 转换而来的 Tensorflow 模型有疑问。问题是转换后的 Tensorflow 模型需要 Pytorch 格式的输入（批量大小、通道数、高度、宽度），而不是 Tensorflow 格式的输入（批量大小、高度、宽度、通道数）。因此，我无法使用该模型通过 Vitis AI 进行进一步处理。

所以我想问一下，有什么方法可以使用 Onnx、Tensorflow 1 或其他工具将这种 Pytorch 输入格式转换为 Tensorflow 格式吗？

我的代码如下：

Pytorch -> Onnx

from hardnet import hardnet
import torch
import onnx

ckpt = torch.load('../hardnet.pth')
model_state_dict = ckpt['model_state_dict']
optimizer_state_dict = ckpt['optimizer_state_dict']

model = hardnet(11)
model.load_state_dict(model_state_dict)
model.eval()     

dummy_input = torch.randn(1, 3, 1080, 1920)
input_names = ['input0']
output_names = ['output0']

output_file = 'hardnet.onnx'
torch.onnx.export(model, dummy_input, output_file, verbose=True,
    input_names=input_names, output_names=output_names,
    opset_version=11, keep_initializers_as_inputs=True)

onnx_model = onnx.load(output_file)
onnx.checker.check_model(onnx_model)
print('Passed Onnx')

Onnx -> Tensorflow 1（使用 Tensorflow 1.15）

import cv2
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import onnx
from onnx_tf.backend import prepare

output_file = 'hardnet.onnx'
onnx_model = onnx.load(output_file)
output = prepare(onnx_model)
output.export_graph('hardnet.pb')
tf.compat.v1.disable_eager_execution()

def load_pb(path_to_pb: str):
    """From: 
    """
    with tf.gfile.GFile(path_to_pb, "rb") as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
    with tf.Graph().as_default() as graph:
        tf.import_graph_def(graph_def, name='')
        return graph


graph = load_pb('hardnet.pb')
input = graph.get_tensor_by_name('input0:0')
output = graph.get_tensor_by_name('output0:0')
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
img = cv2.imread('train_0.jpg', cv2.IMREAD_COLOR)
img = cv2.resize(img, (1920,  1080))

img = img/255
img = img - mean
img = img/std
img = np.expand_dims(img, -1)
# To Pytorch format.
img = np.transpose(img, (3, 2, 0, 1))
img = img

with tf.Session(graph=graph) as sess:
    pred = sess.run(output, {input: img})

Answer 1

您可以将您的 Pytorch 模型包装到另一个可以执行您希望在 TensorFlow 中进行转置的模型中。请参阅以下示例：

假设您有以下玩具 NN：

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.rnn = nn.LSTM(10, 20, 2)

    def forward(self, x):
        h0 = torch.zeros(2, 3, 20)
        c0 = torch.zeros(2, 3, 20)
        return self.rnn(x, (h0, c0))

典型的 pytorch/tensorflow 输入形状为：

>> pytorch_input  = torch.randn(5, 3, 10)
>> tf_input  = torch.transpose(pytorch_input, 1, 2)

>> print("PyTorch input shape: ", pytorch_input.shape)
>> print("TensorFlow input shape: ", tf_input.shape)

PyTorch input shape:  torch.Size([5, 3, 10])
TensorFlow input shape:  torch.Size([5, 10, 3])

现在，首先转置输入然后将转置输入传递给某个模型的包装器：

class NetTensorFlowWrapper(nn.Module):
    def __init__(self, main_module: nn.Module):
        super(NetTensorFlowWrapper, self).__init__()
        self.main_module = main_module
        
    def forward(self, x):
        x = torch.transpose(x, 1, 2)
        return self.main_module(x)

那么，这是可能的：

net = Net()
net_wrapper = NetTensorFlowWrapper(net)

net(pytorch_input)
net_wrapper(tf_input)

然后，当您最终像之前那样通过 torch.onnx.export 保存您的模型并通过 onnx 包（而不是 torch.onnx）读取它们的图表时，您将拥有...

for Net- 输入 5x3x10 and no transpose layer

graph torch-jit-export (
  %input0[FLOAT, 5x3x10]

 {
  %76 = Shape(%input0)
  %77 = Constant[value = <Scalar Tensor []>]()

for NetTensorFlowWrapper- 输入 5x10x3 和转置层

graph torch-jit-export (
  %input0[FLOAT, 5x10x3]

{
  %9 = Transpose[perm = [0, 2, 1]](%input0)
  %77 = Shape(%9)
  %78 = Constant[value = <Scalar Tensor []>]()
...

转换 Pytorch -> Onnx -> Tensorflow 的张量格式问题

Tensor format issue from converting Pytorch -> Onnx -> Tensorflow

tensorflow

pytorch

onnx

vitis-ai