2G大小的GPU对深度学习的加速效果如何？

训练数据情况

总共42776张224*224*3张图片
Found 42776 files belonging to 9 classes.
Using 12833 files for training.

模型参数情况

Total params: 10,917,385
Trainable params: 10,860,745
Non-trainable params: 56,640

batch-size:12

GPU信息

    NVIDIA GeForce GT 730

    驱动程序版本:    27.21.14.6133
    驱动程序日期:    2021/1/19
    DirectX 版本:    12 (FL 11.0)
    物理位置：    PCI 总线 1、设备 0、功能 0

    利用率    11%
    专用 GPU 内存    0.3/2.0 GB
    共享 GPU 内存    0.0/31.9 GB
    GPU 内存    0.4/33.9 GB

训练情况分析

完全使用CPU进行训练的时候，每次训练大约需要2750s。

Epoch 1/65
2496/2496 [==============================] - 2937s 1s/step - loss: 0.4254 - accuracy: 0.8403 - val_loss: 0.3192 - val_accuracy: 0.8867
Epoch 2/65
2496/2496 [==============================] - 2756s 1s/step - loss: 0.2890 - accuracy: 0.8973 - val_loss: 0.4358 - val_accuracy: 0.8520
Epoch 3/65
2496/2496 [==============================] - 2737s 1s/step - loss: 0.2464 - accuracy: 0.9102 - val_loss: 0.2689 - val_accuracy: 0.9020

使用GPU加速进行训练的时候，每次训练的时间从2750s缩短到2100s左右，每次训练大约节省了650秒，效果也是比较明显的。

Epoch 1/65
2023-10-04 10:38:26.686146: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cudnn64_8.dll
2023-10-04 10:38:27.343524: I tensorflow/stream_executor/cuda/cuda_dnn.cc:359] Loaded cuDNN version 8101
2023-10-04 10:38:28.439803: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cublas64_11.dll
2023-10-04 10:38:29.088670: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cublasLt64_11.dll
2023-10-04 10:38:31.502277: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 606.50MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:31.805129: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.16GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:32.084683: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 599.75MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:32.129001: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 620.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:32.738828: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 620.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:32.801711: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 592.75MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:33.034554: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 592.19MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:33.056645: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 599.75MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:33.099135: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.14GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 10:38:33.124441: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.17GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
1070/1070 [==============================] - 2120s 2s/step - loss: 0.5235 - accuracy: 0.8034 - val_loss: 0.5122 - val_accuracy: 0.8171
Epoch 2/65
1070/1070 [==============================] - 2060s 2s/step - loss: 0.3620 - accuracy: 0.8668 - val_loss: 2.9616 - val_accuracy: 0.4629
Epoch 3/65124/1070 [==>...........................] - ETA: 18:02 - loss: 0.3194 - accuracy: 0.8844
Process finished with exit code -1

使用更小的数据集效果分析

数据集

10249
Found 10249 files belonging to 16 classes.
Using 3075 files for training.

参数

Total params: 6,143,760
Trainable params: 6,113,168
Non-trainable params: 30,592

只使用CPU

Epoch 1/65
684/684 [==============================] - 758s 1s/step - loss: 1.1408 - accuracy: 0.5963 - val_loss: 3.0769 - val_accuracy: 0.2738
Epoch 2/65
684/684 [==============================] - 744s 1s/step - loss: 0.7745 - accuracy: 0.7173 - val_loss: 1.0438 - val_accuracy: 0.6369
Epoch 3/65
684/684 [==============================] - 769s 1s/step - loss: 0.6504 - accuracy: 0.7602 - val_loss: 0.8624 - val_accuracy: 0.6964

使用GPU

小数据集速度节省了接近50%

Epoch 1/65
2023-10-04 16:58:19.928226: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cudnn64_8.dll
2023-10-04 16:58:20.236817: I tensorflow/stream_executor/cuda/cuda_dnn.cc:359] Loaded cuDNN version 8101
2023-10-04 16:58:20.791072: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cublas64_11.dll
2023-10-04 16:58:21.096985: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cublasLt64_11.dll
2023-10-04 16:58:23.704576: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.16GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 16:58:24.962633: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.14GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 16:58:24.987354: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.17GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
256/257 [============================>.] - ETA: 0s - loss: 1.3775 - accuracy: 0.51432023-10-04 17:01:47.489983: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.15GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 17:01:48.530265: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.14GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 17:01:48.550469: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.15GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
257/257 [==============================] - ETA: 0s - loss: 1.3776 - accuracy: 0.51452023-10-04 17:04:21.899587: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 626.56MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 17:04:23.391704: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.15GiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2023-10-04 17:04:23.801583: W tensorflow/core/common_runtime/bfc_allocator.cc:271] Allocator (GPU_0_bfc) ran out of memory trying to allocate 615.50MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
257/257 [==============================] - 368s 1s/step - loss: 1.3776 - accuracy: 0.5145 - val_loss: 5.9301 - val_accuracy: 0.2432
Epoch 2/65
257/257 [==============================] - 376s 1s/step - loss: 1.0042 - accuracy: 0.6237 - val_loss: 1.0432 - val_accuracy: 0.6183
Epoch 3/65