這篇文章主要為大家展示了如何實現Pytorch轉keras，內容簡而易懂，希望大家可以學習一下，學習完之后肯定會有收獲的，下面讓小編帶大家一起來看看吧。

創新互聯專注于企業成都營銷網站建設、網站重做改版、萊山網站定制設計、自適應品牌網站建設、H5開發、購物商城網站建設、集團公司官網建設、成都外貿網站建設、高端網站制作、響應式網頁設計等建站業務，價格優惠性價比高，為萊山等各大城市提供網站開發制作服務。

Pytorch憑借動態圖機制，獲得了廣泛的使用，大有超越tensorflow的趨勢，不過在工程應用上，TF仍然占據優勢。有的時候我們會遇到這種情況，需要把模型應用到工業中，運用到實際項目上，TF支持的PB文件和TF的C++接口就成為了有效的工具。今天就給大家講解一下Pytorch轉成Keras的方法，進而我們也可以獲得Pb文件，因為Keras是支持tensorflow的，我將會在下一篇博客講解獲得Pb文件，并使用Pb文件的方法。

Pytorch To Keras

首先，我們必須有清楚的認識，網上以及github上一些所謂的pytorch轉換Keras或者Keras轉換成Pytorch的工具代碼幾乎不能運行或者有使用的局限性（比如僅僅能轉換某一些模型），但是我們是可以用這些轉換代碼中看出一些端倪來，比如二者的參數的尺寸（shape）的形式、channel的排序（first or last）是否一樣，掌握到差異性，就能根據這些差異自己編寫轉換代碼，沒錯，自己編寫轉換代碼，是最穩妥的辦法。整個過程也就分為兩個部分。筆者將會以Nvidia開源的FlowNet為例，將開源的Pytorch代碼轉化為Keras模型。

按照Pytorch中模型的結構，編寫對應的Keras代碼，用keras的函數式API，構建起來會非常方便。

把Pytorch的模型參數，按照層的名稱依次賦值給Keras的模型

以上兩步雖然看上去簡單，但實際我也走了不少彎路。這里一個關鍵的地方，就是參數的shape在兩個框架中是否統一，那當然是不統一的。下面我以FlowNet為例。

Pytorch中的FlowNet代碼

我們僅僅展示層名稱和層參數，就不把整個結構貼出來了，否則會占很多的空間，形成水文。

先看用Keras搭建的flowNet模型，直接用model.summary()輸出模型信息

__________________________________________________________________________________________________
Layer (type)   Output Shape  Param # Connected to   
==================================================================================================
input_1 (InputLayer)  (None, 6, 512, 512) 0      
__________________________________________________________________________________________________
conv0 (Conv2D)   (None, 64, 512, 512) 3520 input_1[0][0]   
__________________________________________________________________________________________________
leaky_re_lu_1 (LeakyReLU) (None, 64, 512, 512) 0  conv0[0][0]   
__________________________________________________________________________________________________
zero_padding2d_1 (ZeroPadding2D (None, 64, 514, 514) 0  leaky_re_lu_1[0][0]  
__________________________________________________________________________________________________
conv1 (Conv2D)   (None, 64, 256, 256) 36928 zero_padding2d_1[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_2 (LeakyReLU) (None, 64, 256, 256) 0  conv1[0][0]   
__________________________________________________________________________________________________
conv1_1 (Conv2D)  (None, 128, 256, 256 73856 leaky_re_lu_2[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_3 (LeakyReLU) (None, 128, 256, 256 0  conv1_1[0][0]   
__________________________________________________________________________________________________
zero_padding2d_2 (ZeroPadding2D (None, 128, 258, 258 0  leaky_re_lu_3[0][0]  
__________________________________________________________________________________________________
conv2 (Conv2D)   (None, 128, 128, 128 147584 zero_padding2d_2[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_4 (LeakyReLU) (None, 128, 128, 128 0  conv2[0][0]   
__________________________________________________________________________________________________
conv2_1 (Conv2D)  (None, 128, 128, 128 147584 leaky_re_lu_4[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_5 (LeakyReLU) (None, 128, 128, 128 0  conv2_1[0][0]   
__________________________________________________________________________________________________
zero_padding2d_3 (ZeroPadding2D (None, 128, 130, 130 0  leaky_re_lu_5[0][0]  
__________________________________________________________________________________________________
conv3 (Conv2D)   (None, 256, 64, 64) 295168 zero_padding2d_3[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_6 (LeakyReLU) (None, 256, 64, 64) 0  conv3[0][0]   
__________________________________________________________________________________________________
conv3_1 (Conv2D)  (None, 256, 64, 64) 590080 leaky_re_lu_6[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_7 (LeakyReLU) (None, 256, 64, 64) 0  conv3_1[0][0]   
__________________________________________________________________________________________________
zero_padding2d_4 (ZeroPadding2D (None, 256, 66, 66) 0  leaky_re_lu_7[0][0]  
__________________________________________________________________________________________________
conv4 (Conv2D)   (None, 512, 32, 32) 1180160 zero_padding2d_4[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_8 (LeakyReLU) (None, 512, 32, 32) 0  conv4[0][0]   
__________________________________________________________________________________________________
conv4_1 (Conv2D)  (None, 512, 32, 32) 2359808 leaky_re_lu_8[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_9 (LeakyReLU) (None, 512, 32, 32) 0  conv4_1[0][0]   
__________________________________________________________________________________________________
zero_padding2d_5 (ZeroPadding2D (None, 512, 34, 34) 0  leaky_re_lu_9[0][0]  
__________________________________________________________________________________________________
conv5 (Conv2D)   (None, 512, 16, 16) 2359808 zero_padding2d_5[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_10 (LeakyReLU) (None, 512, 16, 16) 0  conv5[0][0]   
__________________________________________________________________________________________________
conv5_1 (Conv2D)  (None, 512, 16, 16) 2359808 leaky_re_lu_10[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_11 (LeakyReLU) (None, 512, 16, 16) 0  conv5_1[0][0]   
__________________________________________________________________________________________________
zero_padding2d_6 (ZeroPadding2D (None, 512, 18, 18) 0  leaky_re_lu_11[0][0]  
__________________________________________________________________________________________________
conv6 (Conv2D)   (None, 1024, 8, 8) 4719616 zero_padding2d_6[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_12 (LeakyReLU) (None, 1024, 8, 8) 0  conv6[0][0]   
__________________________________________________________________________________________________
conv6_1 (Conv2D)  (None, 1024, 8, 8) 9438208 leaky_re_lu_12[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_13 (LeakyReLU) (None, 1024, 8, 8) 0  conv6_1[0][0]   
__________________________________________________________________________________________________
deconv5 (Conv2DTranspose) (None, 512, 16, 16) 8389120 leaky_re_lu_13[0][0]  
__________________________________________________________________________________________________
predict_flow6 (Conv2D)  (None, 2, 8, 8) 18434 leaky_re_lu_13[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_14 (LeakyReLU) (None, 512, 16, 16) 0  deconv5[0][0]   
__________________________________________________________________________________________________
upsampled_flow6_to_5 (Conv2DTra (None, 2, 16, 16) 66  predict_flow6[0][0]  
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 1026, 16, 16) 0  leaky_re_lu_11[0][0]  
         leaky_re_lu_14[0][0]  
         upsampled_flow6_to_5[0][0] 
__________________________________________________________________________________________________
inter_conv5 (Conv2D)  (None, 512, 16, 16) 4728320 concatenate_1[0][0]  
__________________________________________________________________________________________________
deconv4 (Conv2DTranspose) (None, 256, 32, 32) 4202752 concatenate_1[0][0]  
__________________________________________________________________________________________________
predict_flow5 (Conv2D)  (None, 2, 16, 16) 9218 inter_conv5[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_15 (LeakyReLU) (None, 256, 32, 32) 0  deconv4[0][0]   
__________________________________________________________________________________________________
upsampled_flow5_to4 (Conv2DTran (None, 2, 32, 32) 66  predict_flow5[0][0]  
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 770, 32, 32) 0  leaky_re_lu_9[0][0]  
         leaky_re_lu_15[0][0]  
         upsampled_flow5_to4[0][0] 
__________________________________________________________________________________________________
inter_conv4 (Conv2D)  (None, 256, 32, 32) 1774336 concatenate_2[0][0]  
__________________________________________________________________________________________________
deconv3 (Conv2DTranspose) (None, 128, 64, 64) 1577088 concatenate_2[0][0]  
__________________________________________________________________________________________________
predict_flow4 (Conv2D)  (None, 2, 32, 32) 4610 inter_conv4[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_16 (LeakyReLU) (None, 128, 64, 64) 0  deconv3[0][0]   
__________________________________________________________________________________________________
upsampled_flow4_to3 (Conv2DTran (None, 2, 64, 64) 66  predict_flow4[0][0]  
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 386, 64, 64) 0  leaky_re_lu_7[0][0]  
         leaky_re_lu_16[0][0]  
         upsampled_flow4_to3[0][0] 
__________________________________________________________________________________________________
inter_conv3 (Conv2D)  (None, 128, 64, 64) 444800 concatenate_3[0][0]  
__________________________________________________________________________________________________
deconv2 (Conv2DTranspose) (None, 64, 128, 128) 395328 concatenate_3[0][0]  
__________________________________________________________________________________________________
predict_flow3 (Conv2D)  (None, 2, 64, 64) 2306 inter_conv3[0][0]  
__________________________________________________________________________________________________
leaky_re_lu_17 (LeakyReLU) (None, 64, 128, 128) 0  deconv2[0][0]   
__________________________________________________________________________________________________
upsampled_flow3_to2 (Conv2DTran (None, 2, 128, 128) 66  predict_flow3[0][0]  
__________________________________________________________________________________________________
concatenate_4 (Concatenate) (None, 194, 128, 128 0  leaky_re_lu_5[0][0]  
         leaky_re_lu_17[0][0]  
         upsampled_flow3_to2[0][0] 
__________________________________________________________________________________________________
inter_conv2 (Conv2D)  (None, 64, 128, 128) 111808 concatenate_4[0][0]  
__________________________________________________________________________________________________
predict_flow2 (Conv2D)  (None, 2, 128, 128) 1154 inter_conv2[0][0]  
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D) (None, 2, 512, 512) 0  predict_flow2[0][0] 

另外有需要云服務器可以了解下創新互聯scvps.cn，海內外云服務器15元起步，三天無理由+7*72小時售后在線，公司持有idc許可證，提供“云服務器、裸金屬服務器、高防服務器、香港服務器、美國服務器、虛擬主機、免備案服務器”等云主機租用服務以及企業上云的綜合解決方案，具有“安全穩定、簡單易用、服務可用性高、性價比高”等特點與優勢，專為企業上云打造定制，能夠滿足用戶豐富、多元化的應用場景需求。

分享名稱：如何實現Pytorch轉keras-創新互聯
文章起源：http://vcdvsql.cn/article32/ccsosc.html

成都網站建設公司_創新互聯，為您提供云服務器、手機網站建設、網站制作、App設計、網站設計公司、網站收錄

廣告

聲明：本網站發布的內容（圖片、視頻和文字）以用戶投稿、用戶轉載內容為主，如果涉及侵權請盡快告知，我們將會在第一時間刪除。文章觀點不代表本網站立場，如需處理請聯系客服。電話：028-86922220；郵箱：631063699@qq.com。內容未經允許不得轉載，或轉載時需注明來源： 創新互聯