1. Which of the following AI framework development models can be directly trained on the Ascend AI processor without adaptation?
A. PyTorch
B. Caffe
C. Mindspore
D. Tensorflow
C
2. The general steps for script training and development using Estimator are (A)
A. Data preprocessing --> Model building --> Run configuration --> Execute training
B. Run configuration --> Model building --> Data preprocessing --> Execute Training
C. Model Construction --> Data Preprocessing --> Run Configuration --> Execute Training
D. Data Preprocessing --> Execute Training --> Run Configuration --> Model Construction
3. Regarding the use of Pycharm to train in Modelarts, the following statement is wrong (C)
A.
The data sets and projects required for training will be uploaded to OBS
B.
The device side logs generated by the training will be saved to OBS
C.
Closing Pycharm training during training will stop
D.
During training, you can view the usage of NPU in Modelarts
5. For the automatic migration of TensorFlow 1.15 training scripts, which of the following statements is correct (D)
A. The Ascend platform provides the Tensorflow 1.15 network migration tool, which is suitable for native Tensorflow training script migration scenarios
B. After the CANN software is installed, the migration tool is in the "tfplugin installation directory/tfplugin/latest/python/site-packages/npu_bridge/convert_tf2npu/" directory
C. The original script is required to use only TensorFlow 1.15 official API and Horovod official API
D. All of the above are correct
6. Which of the following does not belong to the Ascend computing service layer (B)
A.
Ascend operator library
B.
Ascend tensor compiler
C.
Shengteng tuning engine
D.
Frame Adapter
7. The advantages of AscendCL include (D)
A.
High abstraction: APIs for operator compilation, loading, and execution are unified. Compared with one API for each operator, AscendCL greatly reduces the number of APIs and reduces complexity
B.
Backward Compatibility: AscendCL has backward compatibility, ensuring that after the software is upgraded, programs compiled based on the old version can still run on the new version
C.
Zero perception chip: a set of AscendCL interface can realize the unification of application code, and there is no difference between multiple Ascend processors
D.
All of the above
8. The basic process of developing applications using AscendCL, which of the following is correct? D.
A.
Operation management resource application –> model loading –> model execution –> model unloading –> operation management resource release
B.
AscendCL initialization –> operation management resource application –> model loading –> model execution –> operation management resource release –> AscendCL deinitialization
C.
AscendCL initialization –> operation management resource application –> model execution –> model unloading –> operation management resource release –> AscendCL deinitialization D. AscendCL initialization –> operation management resource
application
–> model loading –> model execution –>Model unloading –>running management resource release –>AscendCL deinitialization
9. Which of the following statements about the ATC tool is correct (D)
A. The ATC tool can convert the network model of the open source framework (such as Caffe, TensorFlow, etc.) into a model dedicated to the Davinci architecture
B. The ATC tool can convert the single-operator description file defined by Ascend IR into an offline model supported by the Ascend AI processor
C. The ATC tool will perform specific operations such as operator scheduling optimization, weight data rearrangement, and memory usage optimization during the conversion process
D. All of the above are correct
10. What is the name of the model conversion tool? ©
A. NO
B. Profiling
C. ATC
D. AMCT
11. What preparatory work should be done before model migration? (AB)
A. Repeat training on GPU or CPU, and run through the training
B. Training recurrence on GPU or CPU, Loss convergence
C. Repeat the training on the NPU, and the training runs through
D. Training recurrence on NPU, Loss convergence
12. Which of the following steps are required to run the TensorFlow network model on the Ascend AI processor? (ABCD)
A. Precision tuning
B. Model Migration
C. Model training
D. Performance tuning
13. Regarding automatic migration, which of the following statements are correct (ABCD)
A. Automatic migration refers to scanning the AST syntax tree of the original training script through tools.
B. During automatic migration, it can automatically analyze the support degree of the native Tensorflow API on the Ascend AI processor.
C. Automatic migration can automatically migrate the original TensorFlow training script. Scripts supported by Chengsheng AI processor
D. For APIs that cannot be automatically migrated, you can refer to the migration report output by the tool to adapt and modify the training script accordingly
15. When using Pycharm to configure training tasks on Modelarts, what parameters need to be filled in? (ABCD)
A.
Start training script path
B.
Path to store training logs on OBS
C.
The path of the training data set on OBS
D. AI engine used for training
16. How can I ask for help when encountering problems during training? (ABCD)
A. You can raise an issue on the Modelzoo warehouse
B. You can check experience posts on the Shengteng community blog
C. If it is a scene of centralized organization of activities, you can ask questions to the contact person
D. You can post questions on the Shengteng community forum
17. AscendCL can perform ()
A.
Model loading and execution
B.
Runtime Resource Management
C.
Operator loading and execution
D.
Image/Video Processing
ABCD
18. Which of the following statements about the logical architecture of the Ascend heterogeneous computing architecture CANN is correct (ABCD)
A.
The Ascend computing service layer provides a high-performance operator library AOL and the Ascend optimization engine AOE, which is the basis for the Ascend AI processor to support a variety of artificial intelligence algorithms and optimize execution performance
B.
The Ascend Computing Compilation Layer, which compiles the calculation graph of the user input intermediate representation (IR) into the Ascend hardware executable model through the graph compiler
C.
Ascension computing execution layer, responsible for the execution of models and operators, providing runtime library (Runtim), graph executor (Graph Executor), digital vision pre-processing (Digital Vision Pre-Processing, DVPP), artificial intelligence pre-processing (Artificial Intelligence Pre-Processing, AIPP) and other functional units
D.
Ascend computing base layer, which mainly provides basic services for the layers above it, such as shared virtual memory (Shared Virtual Memory, SVM), device virtualization (Virtual Machine, VM), host-device communication (Host Device Communication, HDC )wait
19. What information is recorded in the operation log? (ABCD)
A.
The name of the module that generated the log
B.
Process name and ID
C.
Log printing time
D.
Specific log content of each module
