Pytorch如何转tflite-创新互联

这篇文章主要为大家展示了Pytorch如何转tflite，内容简而易懂，希望大家可以学习一下，学习完之后肯定会有收获的，下面让小编带大家一起来看看吧。

创新互联是一家集网站建设,昌江企业网站建设,昌江品牌网站建设,网站定制,昌江网站建设报价,网络营销,网络优化,昌江网站推广为一体的创新建站企业，帮助传统企业提升企业形象加强企业竞争力。可充分满足这一群体相比中小企业更为丰富、高端、多元的互联网需求。同时我们时刻保持专业、时尚、前沿，时刻以成就客户成长自我，坚持不断学习、思考、沉淀、净化自己，让我们为更多的企业打造出实用型网站。

目标是想把在服务器上用pytorch训练好的模型转换为可以在移动端运行的tflite模型。

最直接的思路是想把pytorch模型转换为tensorflow的模型，然后转换为tflite。但是这个转换目前没有发现比较靠谱的方法。

经过调研发现最新的tflite已经支持直接从keras模型的转换，所以可以采用keras作为中间转换的桥梁，这样就能充分利用keras高层API的便利性。

转换的基本思想就是用pytorch中的各层网络的权重取出来后直接赋值给keras网络中的对应layer层的权重。

转换为Keras模型后，再通过tf.contrib.lite.TocoConverter把模型直接转为tflite.

下面是一个例子，假设转换的是一个两层的CNN网络。

import tensorflow as tf
from tensorflow import keras
import numpy as np

import torch
from torchvision import models
import torch.nn as nn
# import torch.nn.functional as F
from torch.autograd import Variable

class PytorchNet(nn.Module):
 def __init__(self):
 super(PytorchNet, self).__init__()
 conv1 = nn.Sequential(
  nn.Conv2d(3, 32, 3, 2),
  nn.BatchNorm2d(32),
  nn.ReLU(inplace=True),
  nn.MaxPool2d(2, 2))
 conv2 = nn.Sequential(
  nn.Conv2d(32, 64, 3, 1, groups=1),
  nn.BatchNorm2d(64),
  nn.ReLU(inplace=True),
  nn.MaxPool2d(2, 2))
 self.feature = nn.Sequential(conv1, conv2)
 self.init_weights()

 def forward(self, x):
 return self.feature(x)

 def init_weights(self):
 for m in self.modules():
  if isinstance(m, nn.Conv2d):
  nn.init.kaiming_normal_(
   m.weight.data, mode='fan_out', nonlinearity='relu')
  if m.bias is not None:
   m.bias.data.zero_()
  if isinstance(m, nn.BatchNorm2d):
  m.weight.data.fill_(1)
  m.bias.data.zero_()

def KerasNet(input_shape=(224, 224, 3)):
 image_input = keras.layers.Input(shape=input_shape)
 # conv1
 network = keras.layers.Conv2D(
 32, (3, 3), strides=(2, 2), padding="valid")(image_input)
 network = keras.layers.BatchNormalization(
 trainable=False, fused=False)(network)
 network = keras.layers.Activation("relu")(network)
 network = keras.layers.MaxPool2D(pool_size=(2, 2), strides=(2, 2))(network)

 # conv2
 network = keras.layers.Conv2D(
 64, (3, 3), strides=(1, 1), padding="valid")(network)
 network = keras.layers.BatchNormalization(
 trainable=False, fused=True)(network)
 network = keras.layers.Activation("relu")(network)
 network = keras.layers.MaxPool2D(pool_size=(2, 2), strides=(2, 2))(network)

 model = keras.Model(inputs=image_input, outputs=network)

 return model

class PytorchToKeras(object):
 def __init__(self, pModel, kModel):
 super(PytorchToKeras, self)
 self.__source_layers = []
 self.__target_layers = []
 self.pModel = pModel
 self.kModel = kModel
 tf.keras.backend.set_learning_phase(0)

 def __retrieve_k_layers(self):
 for i, layer in enumerate(self.kModel.layers):
  if len(layer.weights) > 0:
  self.__target_layers.append(i)

 def __retrieve_p_layers(self, input_size):

 input = torch.randn(input_size)
 input = Variable(input.unsqueeze(0))
 hooks = []

 def add_hooks(module):

  def hook(module, input, output):
  if hasattr(module, "weight"):
   # print(module)
   self.__source_layers.append(module)

  if not isinstance(module, nn.ModuleList) and not isinstance(module, nn.Sequential) and module != self.pModel:
  hooks.append(module.register_forward_hook(hook))

 self.pModel.apply(add_hooks)

 self.pModel(input)
 for hook in hooks:
  hook.remove()

 def convert(self, input_size):
 self.__retrieve_k_layers()
 self.__retrieve_p_layers(input_size)

 for i, (source_layer, target_layer) in enumerate(zip(self.__source_layers, self.__target_layers)):
  print(source_layer)
  weight_size = len(source_layer.weight.data.size())
  transpose_dims = []
  for i in range(weight_size):
  transpose_dims.append(weight_size - i - 1)
  if isinstance(source_layer, nn.Conv2d):
  transpose_dims = [2,3,1,0]
  self.kModel.layers[target_layer].set_weights([source_layer.weight.data.numpy(
  ).transpose(transpose_dims), source_layer.bias.data.numpy()])
  elif isinstance(source_layer, nn.BatchNorm2d):
  self.kModel.layers[target_layer].set_weights([source_layer.weight.data.numpy(), source_layer.bias.data.numpy(),
        source_layer.running_mean.data.numpy(), source_layer.running_var.data.numpy()])
 def save_model(self, output_file):
 self.kModel.save(output_file)

 def save_weights(self, output_file):
 self.kModel.save_weights(output_file, save_format='h6')

pytorch_model = PytorchNet()
keras_model = KerasNet(input_shape=(224, 224, 3))

torch.save(pytorch_model, 'test.pth')

#Load the pretrained model
pytorch_model = torch.load('test.pth')

# #Time to transfer weights
converter = PytorchToKeras(pytorch_model, keras_model)
converter.convert((3, 224, 224))

# #Save the converted keras model for later use
# converter.save_weights("keras.h6")
converter.save_model("keras_model.h6")

# convert keras model to tflite model
converter = tf.contrib.lite.TocoConverter.from_keras_model_file(
 "keras_model.h6")
tflite_model = converter.convert()
open("convert_model.tflite", "wb").write(tflite_model)

当前名称：Pytorch如何转tflite-创新互联
当前地址：http://cxhlcq.com/article/dcjgip.html