action_net_v1图像动作识别——Action Recognition

哈喽，事实说明那两个服务器都是垃圾。我用第一个服务器就可以跑。但是调用非常慢！！！！运行起来很慢。

Epoch 00096: val_acc did not improve from 0.55416
Epoch 97/100
15285/15285 [==============================] - 14s 925us/step - loss: 0.1484 - acc: 0.9999 - val_loss: 2.1899 - val_acc: 0.5518

Epoch 00097: val_acc did not improve from 0.55416
Epoch 98/100
15285/15285 [==============================] - 14s 919us/step - loss: 0.1483 - acc: 0.9998 - val_loss: 2.1897 - val_acc: 0.5521

Epoch 00098: ReduceLROnPlateau reducing learning rate to 2.4414063659605745e-07.

Epoch 00098: val_acc did not improve from 0.55416
Epoch 99/100
15285/15285 [==============================] - 14s 918us/step - loss: 0.1484 - acc: 0.9999 - val_loss: 2.1908 - val_acc: 0.5521

Epoch 00099: val_acc did not improve from 0.55416
Epoch 100/100
15285/15285 [==============================] - 14s 919us/step - loss: 0.1483 - acc: 0.9999 - val_loss: 2.1902 - val_acc: 0.5515

Epoch 00100: val_acc did not improve from 0.55416

跑个ResNet就到55%验证集的acc，感觉有点Low。

既然我有能跑的服务器为何还要折腾？？？？

因为这个是tf1.12.0的版本，不能升级，升级需要cudnn及cuda的相应升级，否则不行。

因为我需要tf1.14.0的版本，这是落地必须的，不行的话没法玩。

因为项目多，不只跑一个，这个只有两个一般卡，虽说目前也没有更多卡的。

我想要256个卡，或者给我100个能用GPU的服务器。不能用GPU的服务器一分钱不值，没一点屁用。

下面试试减少add。[2,1],我用的是RGB的，灰度的话应该对动作影响不是太大。

只有53%，有点低了。。。。微笑。

Epoch 00097: val_acc did not improve from 0.53218
Epoch 98/100
15285/15285 [==============================] - 14s 887us/step - loss: 0.1340 - acc: 0.9999 - val_loss: 2.4703 - val_acc: 0.5288

Epoch 00098: val_acc did not improve from 0.53218
Epoch 99/100
15285/15285 [==============================] - 13s 881us/step - loss: 0.1340 - acc: 0.9998 - val_loss: 2.4732 - val_acc: 0.5293

Epoch 00099: val_acc did not improve from 0.53218
Epoch 100/100
15285/15285 [==============================] - 14s 885us/step - loss: 0.1338 - acc: 0.9998 - val_loss: 2.4735 - val_acc: 0.5301

Epoch 00100: val_acc did not improve from 0.53218

要不试试ResNet18吧，尽管不太想巨无霸的模型。

Epoch 00096: val_acc did not improve from 0.48325
Epoch 97/100
15285/15285 [==============================] - 20s 1ms/step - loss: 0.2240 - acc: 1.0000 - val_loss: 3.4173 - val_acc: 0.4686

Epoch 00097: val_acc did not improve from 0.48325
Epoch 98/100
15285/15285 [==============================] - 20s 1ms/step - loss: 0.2239 - acc: 0.9998 - val_loss: 3.4182 - val_acc: 0.4689

Epoch 00098: ReduceLROnPlateau reducing learning rate to 2.4414063659605745e-07.

Epoch 00098: val_acc did not improve from 0.48325
Epoch 99/100
15285/15285 [==============================] - 20s 1ms/step - loss: 0.2237 - acc: 0.9999 - val_loss: 3.4192 - val_acc: 0.4689

Epoch 00099: val_acc did not improve from 0.48325
Epoch 100/100
15285/15285 [==============================] - 20s 1ms/step - loss: 0.2237 - acc: 0.9998 - val_loss: 3.4194 - val_acc: 0.4686

Epoch 00100: val_acc did not improve from 0.48325

结果惨不忍睹啊，这就是18层，这是18层地狱吧

下面试试34层。可能会有奇迹吧，想想加勒比海盗中，海底即天堂。

我能说效果不行吗？也没进展。看来要用bottleneck了啊。

Epoch 00097: val_acc did not improve from 0.51622
Epoch 98/100
15285/15285 [==============================] - 30s 2ms/step - loss: 0.2098 - acc: 0.9999 - val_loss: 3.2552 - val_acc: 0.5118

Epoch 00098: ReduceLROnPlateau reducing learning rate to 2.4414063659605745e-07.

Epoch 00098: val_acc did not improve from 0.51622
Epoch 99/100
15285/15285 [==============================] - 30s 2ms/step - loss: 0.2097 - acc: 1.0000 - val_loss: 3.2569 - val_acc: 0.5123

Epoch 00099: val_acc did not improve from 0.51622
Epoch 100/100
15285/15285 [==============================] - 30s 2ms/step - loss: 0.2096 - acc: 0.9999 - val_loss: 3.2612 - val_acc: 0.5113

Epoch 00100: val_acc did not improve from 0.51622

上50层的bottleneck。不行啊，服务器直接死了。

只能改小，bottleneck[2323]，效果很差劲。垃圾玩意。验证集36%acc

改成[2,3]，只有50%。哎[2,2]也是差不多。

改到[1,1]也不到50%，失败，纯ResNet根本不行啊，我看有个达到70%多。

 

也试试VGG，参考了一个keras版本的，发现卷积池化不一起用，而是卷积BN，卷积BN池化，这样参数很多，跑不起来。只好将每层都加池化。用了个最简单的，200多万参数的。验证集55%acc，比ResNet要好，可能我用的ResNet也是假的。卧槽

再次修改了VGG，可以到60%了，看图，难道我的epoch还不够？？200次？？

 

use tf-org's model ,pls see the blog https://blog.youkuaiyun.com/SPESEG/article/details/103079746

另外有相关问题可以加入QQ群讨论，不设微信群

QQ群：868373192 

语音深度学习群