Nvidia 4060 TI 8GB在分类任务中比CPU慢。

huangapple
go评论105阅读模式
英文:

Nvidia 4060 TI 8GB slower than CPU in classification

问题

我正在使用这个教程：https://www.tensorflow.org/tutorials/images/classification

在测试中，CPU运行大约需要50秒，GPU大约需要7-8分钟。我猜我做错了什么。

我的CPU是一台拥有96GB内存的Intel i5第10代处理器。我希望GPU至少能运行快2倍。

我启用了混合精度（mixed precision），以确保使用了Tensor Cores。

  1. from tensorflow.keras import mixed_precision
  2. mixed_precision.set_global_policy('mixed_float16')

我错过了什么吗？使用分类算法时，RTX 4060 Ti 8GB VRAM为什么这么慢？

我有大约1000个类别，但这并不重要，因为CPU更快...

我使用的批次大小为512，VRAM占用6/8，CPU大部分时间占用50%左右。

我还使用了BatchNormalization。

  1. model = Sequential([
  2.   data_augmentation,
  3.   layers.Rescaling(1./255),
  4.   layers.Conv2D(16, 3, padding='same', activation='relu'),
  5.   BatchNormalization(),
  6.   layers.MaxPooling2D(),
  7.   layers.Conv2D(32, 3, padding='same', activation='relu'),
  8.   BatchNormalization(),
  9.   layers.MaxPooling2D(),
  10.   layers.Conv2D(64, 3, padding='same', activation='relu'),
  11.   BatchNormalization(),
  12.   layers.MaxPooling2D(),
  13.   layers.Dropout(0.2),
  14.   BatchNormalization(),
  15.   layers.Flatten(),
  16.   layers.Dense(128, activation='relu'),
  17.   layers.Dense(num_classes, name="outputs")
  18. ])

附注：我对AI还不熟悉。

我尝试使用不同的批次大小。

我尝试禁用GPU，只在CPU上运行测试。

我检查了RAM、磁盘和CPU是否存在瓶颈（没有一个达到100%）。当我在CPU上运行时，使用率为100%，而GPU的使用率为1%或更低。

这是我进行的批次测试：

  1. 批次大小 时间
  2. 4 377秒
  3. 8 304秒
  4. 16 317秒
  5. 32 335秒
  6. 64 446秒

这是模型的结构：

  1.  Layer (type)                Output Shape              Param #   
  2. =================================================================
  3.  sequential_1 (Sequential)   (None, 256, 256, 3)       0         
  4.                                                                  
  5.  rescaling_2 (Rescaling)     (None, 256, 256, 3)       0         
  6.                                                                  
  7.  conv2d_3 (Conv2D)           (None, 256, 256, 16)      448       
  8.                                                                  
  9.  batch_normalization (BatchN  (None, 256, 256, 16)     64        
  10.  ormalization)                                                   
  11.                                                                  
  12.  max_pooling2d_3 (MaxPooling  (None, 128, 128, 16)     0         
  13.  2D)                                                             
  14.                                                                  
  15.  conv2d_4 (Conv2D)           (None, 128, 128, 32)      4640      
  16.                                                                  
  17.  batch_normalization_1 (Batc  (None, 128, 128, 32)     128       
  18.  hNormalization)                                                 
  19.                                                                  
  20.  max_pooling2d_4 (MaxPooling  (None, 64, 64, 32)       0         
  21.  2D)                                                             
  22.                                                                  
  23.  conv2d_5 (Conv2D)           (None, 64, 64, 64)        18496     
  24.                                                                  
  25.  batch_normalization_2 (Batc  (None, 64, 64, 64)       256       
  26.  hNormalization)                                                 
  27.                                                                  
  28.  max_pooling2d_5 (MaxPooling  (None, 32, 32, 64)       0         
  29.  2D)                                                             
  30.                                                                  
  31.  dropout (Dropout)           (None, 32, 32, 64)        0         
  32.                                                                  
  33.  batch_normalization_3 (Batc  (None, 32, 32, 64)       256       
  34.  hNormalization)                                                 
  35.                                                                  
  36.  flatten_1 (Flatten)         (None, 65536)             0         
  37.                                                                  
  38.  dense_2 (Dense)             (None, 128)               8388736   
  39.                                                                  
  40.  outputs (Dense)             (None, 863)               111327    
  41.                                                                  
  42. =================================================================
  43. Total params: 8,524,351
  44. Trainable params: 8,523,999
  45. Non-trainable params: 352
  46. _________________________________________________________________

这是我加载数据的方式：

  1. folder = "some-folder"
  2. train_ds = tf.keras.utils.image_dataset_from_directory(
  3.     folder,
  4.     validation_split=0.2,
  5.     subset="training",
  6.     seed=1,
  7.     image_size=image_size,
  8.     batch_size=batch_size
  9. )
  11. val_ds = tf.keras.utils.image_dataset_from_directory(
  12.     folder,
  13.     validation_split=0.2,
  14.     subset="validation",
  15.     seed=1,
  16.     image_size=image_size,
  17.     batch_size=batch_size
  18. )

这是自动调整部分：

  1. AUTOTUNE = tf.data.AUTOTUNE
  3. train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
  4. val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)
  5. normalization_layer = layers.Rescaling(1. / 255)
  7. normalized_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))
  8. niter = iter(normalized_ds)
  9. image_batch, labels_batch = next(niter)
  10. first_image = image_batch[0]
  11. # 注意像素值现在在[0,1]之间。
  12. print(np.min(first_image), np.max(first_image))
英文:

I'm using this tutorial:
https://www.tensorflow.org/tutorials/images/classification

In tests, the cpu takes about 50seconds to run and GPU about 7-8 minutes. I'm guessing I'm doing something wrong.

My cpu is a intel i5 10th generation with 96 ram. I would expect gpu to run at least 2x faster

I enabled mixed precision so that I make sure it is using tensorcores

from tensorflow.keras import mixed_precision
mixed_precision.set_global_policy('mixed_float16')

What am I missing.... is rtx 4060 ti 8gb vram so slow when using classification algorithm?

I have about 1000 classes,but this is not relevant because cpu is alot faster ...

I'm using batches of 512, vram is at 6/8 and cpu most of the time is about 50%

I'm als doing BatchNormalization

  1. model = Sequential([
  2.   data_augmentation,
  3.   layers.Rescaling(1./255),
  4.   layers.Conv2D(16, 3, padding='same', activation='relu'),
  5.   BatchNormalization(),
  6.   layers.MaxPooling2D(),
  7.   layers.Conv2D(32, 3, padding='same', activation='relu'),
  8.   BatchNormalization(),
  9.   layers.MaxPooling2D(),
  10.   layers.Conv2D(64, 3, padding='same', activation='relu'),
  11.   BatchNormalization(),
  12.   layers.MaxPooling2D(),
  13.   layers.Dropout(0.2),
  14.   BatchNormalization(),
  15.   layers.Flatten(),
  16.   layers.Dense(128, activation='relu'),
  17.   layers.Dense(num_classes, name="outputs")
  18. ])

ps: i'm new to ai

I tried using different batch size

I tried disabling the gpu and running just on cpu to make the test

I check out the ram, disk and cpu for bottleneck(none are 100%). When I run on cpu the usage is 100% and gpu 1% or less

These are the batch test I've done

  1. Batch Time 
  2. 4 377s 
  3. 8  304s
  4. 16  317s 
  5. 32 335s
  6. 64 446s
  8. And this is the model:
  10.  Layer (type)                Output Shape              Param #   
  11. =================================================================
  12.  sequential_1 (Sequential)   (None, 256, 256, 3)       0         
  13.                                                                  
  14.  rescaling_2 (Rescaling)     (None, 256, 256, 3)       0         
  15.                                                                  
  16.  conv2d_3 (Conv2D)           (None, 256, 256, 16)      448       
  17.                                                                  
  18.  batch_normalization (BatchN  (None, 256, 256, 16)     64        
  19.  ormalization)                                                   
  20.                                                                  
  21.  max_pooling2d_3 (MaxPooling  (None, 128, 128, 16)     0         
  22.  2D)                                                             
  23.                                                                  
  24.  conv2d_4 (Conv2D)           (None, 128, 128, 32)      4640      
  25.                                                                  
  26.  batch_normalization_1 (Batc  (None, 128, 128, 32)     128       
  27.  hNormalization)                                                 
  28.                                                                  
  29.  max_pooling2d_4 (MaxPooling  (None, 64, 64, 32)       0         
  30.  2D)                                                             
  31.                                                                  
  32.  conv2d_5 (Conv2D)           (None, 64, 64, 64)        18496     
  33.                                                                  
  34.  batch_normalization_2 (Batc  (None, 64, 64, 64)       256       
  35.  hNormalization)                                                 
  36.                                                                  
  37.  max_pooling2d_5 (MaxPooling  (None, 32, 32, 64)       0         
  38.  2D)                                                             
  39.                                                                  
  40.  dropout (Dropout)           (None, 32, 32, 64)        0         
  41.                                                                  
  42.  batch_normalization_3 (Batc  (None, 32, 32, 64)       256       
  43.  hNormalization)                                                 
  44.                                                                  
  45.  flatten_1 (Flatten)         (None, 65536)             0         
  46.                                                                  
  47.  dense_2 (Dense)             (None, 128)               8388736   
  48.                                                                  
  49.  outputs (Dense)             (None, 863)               111327    
  50.                                                                  
  51. =================================================================
  52. Total params: 8,524,351
  53. Trainable params: 8,523,999
  54. Non-trainable params: 352
  55. _________________________________________________________________

And this is how I load the data:

  1. folder = "some-folder"
  2. train_ds = tf.keras.utils.image_dataset_from_directory(
  4.     folder,
  5.  validation_split=0.2,
  6.   subset="training",
  7.   seed=1,
  8.   image_size=image_size,
  9.   batch_size=batch_size)
  11. val_ds = tf.keras.utils.image_dataset_from_directory(
  12.     folder,
  13.     validation_split=0.2,
  14.   subset="validation",
  15.   seed=1,
  16.   image_size=image_size,
  17.   batch_size=batch_size)

And this is the autotune part

  1. AUTOTUNE = tf.data.AUTOTUNE
  3. train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
  4. val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)
  5. normalization_layer = layers.Rescaling(1. / 255)
  7. normalized_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))
  8. niter = iter(normalized_ds)
  9. image_batch, labels_batch = next(niter)
  10. first_image = image_batch[0]
  11. # Notice the pixel values are now in `[0,1]`.
  12. print(np.min(first_image), np.max(first_image))

答案1

得分: 0

RTX 4060系列不兼容CUDA工具包。

英文:

RTX 4060 SERIES is not CUDA toolkit compatible.

huangapple
  • 本文由 发表于 2023年7月27日 14:55:18
  • 转载请务必保留本文链接：https://go.coder-hub.com/76777171.html
  • nvidia
  • tensorflow
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定