Deep Learning Inferencing on IBM Cloud with NVIDIA TensorRT Khoa - - PowerPoint PPT Presentation
Watson Cloud Platform Strategic Customer Success Deep Learning Inferencing on IBM Cloud with NVIDIA TensorRT Khoa Huynh Senior Technical Staff Member (STSM), IBM Larry Brown Senior Software Engineer, IBM Watson Cloud Platform Strategic
Watson Cloud Platform Strategic Customer Success
Watson Cloud Platform Strategic Customer Success
§ Introduction § Inferencing with PyCaffe § TensorRT Overview § TensorRT Implementation § Performance Results § Conclusions § Q & A
2
Watson Cloud Platform Strategic Customer Success
§ AI especially deep learning has seen rapid advancement in recent years § Initial focus on image processing, expanding to natural language, different neural network models, recurrent networks, and DL frameworks § Much attention to developing networks and training models § Very compute intensive, GPUs nearly a necessity § As DL becomes mainstream focus is shifting to inferencing (use of the trained network) § Inferencing cloud service could handle requests from multiple users – One request at a time or collect a batch of requests to inference at once – But only wait a short time to fill a batch or latency is affected – Or one user might submit a larger number of images to classify at once § Inferencing cloud service needs – quick response - latency seen by the user – to handle large volume – overall throughput
3
Watson Cloud Platform Strategic Customer Success
§ Bare Metal Servers – Nvidia M60 GPUs (monthly & hourly) – Nvidia K80 GPU PCIe cards (monthly & hourly) – Nvidia P100 GPUs (monthly) – Nvidia V100 GPUs (monthly) § Virtual Servers – Nvidia P100 GPUs (monthly & hourly) – Nvidia V100 (coming soon – monthly & hourly) § Deep Learning as a Service (DLaaS) – Part of Watson Machine Learning (WML) – Focused on deep learning training – Allows user to run training jobs on a cluster of GPU-enabled machines using various frameworks § PowerAI – Available in 2Q2018 with PowerAI R5 – Delivered through IBM Cloud Catalog & supported by IBM Trusted Partner Nimbix – On-demand cloud provisioning – Containerized – Native Distributed Deep Learning (DDL) and Large Model Support (LMS)
4
Watson Cloud Platform Strategic Customer Success
§ Given a trained model want to use it to classify images. § Study performance of various GPUs, FPGAs, etc. § Can use C++, but more familiar with Python so that was the language of choice. § Unlike training, a single GPU is used. Use multiple threads, processes, or services if more volume needs to take advantage of more GPUs. root@V100:~/infer_caffe# python infer_caffe.py -h usage: infer_caffe.py [-h] -m MODEL -w WEIGHTS -l LMDB [-b BATCH] [-i ITERATIONS] [-c CAFFEROOT] [--blobName BLOBNAME] [--labels LABELS] [--meanImage MEANIMAGE] [--debug] [--gpu] [--csvFile CSVFILE] [--quiet] Use a trained Caffe model to classify images from a LMDB database.
5
Watson Cloud Platform Strategic Customer Success
root@V100:~/infer_caffe# python infer_caffe.py -m ~/model_zoo/caffe/vgg16/pretrained/VGG_ILSVRC_16_layers_deploy.prototxt -w ~/model_zoo/caffe/vgg16/pretrained/VGG_ILSVRC_16_layers.caffemodel -l /datasets/x86_LMDB/LMDB/ilsvrc12_val_lmdb/ -c /opt/nvidia/caffe-0.16/caffe -b 1 -i 5 --gpu --csvFile ./pycaffe.csv --quiet Final Stats (times in seconds)
Time: 09:05:54 Host: V100 Iterations: 5 Batch size: 1 Data type: NA Total run time: 3.3347 Stats for all iterations
Top 1 accuracy: 100.00% Top 5 accuracy: 100.00% Inference time -- Total: 0.1645 Mean: 0.0329 Min: 0.0076 Max: 0.0726 Range: 0.0649 STD: 0.0308 Median: 0.0079 Inference time/prediction: 0.0329 Images/sec: 30.40
6
Watson Cloud Platform Strategic Customer Success
Parse command line … # Create the neural network. net = caffe.Net(model_def, # defines the structure of the model model_weights, # contains the trained weights caffe.TEST) # use test mode (e.g., don't perform dropout) … for each iteration read a batch of images from LMDB # Call the network.
…
7
Watson Cloud Platform Strategic Customer Success
§ Speeds up inferencing by – Merging layers and tensors to reduce size of network and execute in a single kernel. – Selects the best specialized kernel for the target hardware based on layer parameters and measured performance. § Stages – Build: optimize the network (layers, weights, labels) to produce a runtime plan or engine.
– Deploy: run the engine with given input data to get the resulting predictions. § Supports Python and C++. § TensorRT Lite is a simplified interface for Python (not used here). § Can create the TRT network yourself or use TRT utility to import and convert framework model into TRT form. – Caffe and UFF (Universal Framework Format) compatible frameworks such as TensorFlow.
8
Watson Cloud Platform Strategic Customer Success
§ Model trained in FLOAT (FP32). § TensorRT inferencing can use FLOAT, HALF, or INT8 (as supported by the GPU). § Increase speed with no or little loss of accuracy. – HALF could be some reduction of accuracy, not noticeable in general. – INT8 a small reduction of accuracy. § INT8 requires calibration files. – Uses sample runs of data through the net to determine the range of FLOAT values encountered. – Maps that range to INT8’s smaller range. – Caffe patch available to easily generate these calibration files during a short training run when environment variable TENSORRT_INT8_BATCH_DIRECTORY is set. – NVIDIA suggests “For ImageNet networks, around 500 calibration images is adequate”.
9
Watson Cloud Platform Strategic Customer Success
§ Re-implementation of caffe_infer.py using TRT instead of pycaffe. § Shares code for getting images, collecting stats, overall flow.
root@V100:~/infer_caffe# python infer_caffe_trt.py -h usage: infer_caffe_trt.py [-h] -m MODEL -w WEIGHTS -l LMDB [-b BATCH] [-i ITERATIONS] [-c CAFFEROOT] [--imageShape IMAGESHAPE] [--max_batch MAX_BATCH] [--outputLayer OUTPUTLAYER] [--outputSize OUTPUTSIZE] [--dtype {FLOAT,HALF,INT8}] [--labels LABELS] [--meanImage MEANIMAGE] [--csvFile CSVFILE] [--calBatchDir CALBATCHDIR] [--firstCalBatch FIRSTCALBATCH] [--numCalBatches NUMCALBATCHES] [--debug] [--quiet] Uses NVidia TensorRT to optimize and run inference on a trained Caffe model performing an image recognition task and prints performance and accuracy results.
10
Watson Cloud Platform Strategic Customer Success
# Create the engine using the TRT utilities for Caffe. # Use the caffe model converter utility in tensorrt.utils. # We provide it a logger, a path to the model prototxt, the model file, the max batch size, # the max workspace size, the output layer(s) and the data type of the weights. engine_dtype = trt.infer.DataType[dtype] calibrator = None if engine_dtype == trt.infer.DataType.INT8: calibrator = infer_utils.Calibrator.Calibrator(cal_batch_dir, first_cal_batch, num_cal_batches, debug) engine = trt.utils.caffe_to_trt_engine(trt_logger, model, weights, max_batch, 1 << 25, [output_layer], engine_dtype, calibrator=calibrator) … # Allocate memory on the GPU with PyCUDA and register it with the engine. # The size of the allocations is the size of the input and expected output * the batch size d_input = cuda.mem_alloc(batch_size * image_shape[0] * image_shape[1] * 3 * np.dtype(np.float32).itemsize) d_output = cuda.mem_alloc(batch_size * output.size * output.dtype.itemsize) # The engine needs bindings provided as pointers to the GPU memory. # PyCUDA lets us do this for memory allocations by casting those allocations to ints bindings = [int(d_input), int(d_output)] # Create a cuda stream to run inference in. stream = cuda.Stream()
11
Watson Cloud Platform Strategic Customer Success
# Time moving the data to the GPU, running the network, and getting the results back to the host as part of the # inference operation for this iteration. stats.begin_iteration() cuda.memcpy_htod_async(d_input, batchin, stream) # execute model context.enqueue(batch_size, bindings, stream.handle, None) # transfer predictions back cuda.memcpy_dtoh_async(output, d_output, stream) # syncronize threads stream.synchronize()
12
Watson Cloud Platform Strategic Customer Success
13
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 1 5 10 25 50 75 100 125 150 175
Average Latency/Batch (sec) Batch Size
V100 Caffe VGGNet16 TensorRT 3.01 Inference Latency
PyCaffe FLOAT HALF INT8
Watson Cloud Platform Strategic Customer Success
14
500 1000 1500 2000 2500 1 5 10 25 50 75 100 125 150 175
Images/Second Batch Size
V100 Caffe VGGNet16 TensorRT 3.01 Inference Throughput
PyCaffe FLOAT HALF INT8
Watson Cloud Platform Strategic Customer Success
15
64.65 85.77 64.65 85.77 64.65 85.77 64.37 85.49 10 20 30 40 50 60 70 80 90 100 Top 1 Top 5
Accuracy Per Cent
V100 Caffe VGGNet16 TensorRT 3.0.1 Inference Accuracy
PyCaffe FLOAT HALF INT8
Watson Cloud Platform Strategic Customer Success
16
10 20 30 40 50 60 70 80 90 1 10 25 50 100 150 200 300 400 500 600 700 800 900 1000
Accuracy Per Cent Number of Calibration Images
V100 Caffe VGGNet16 TensortRT 3.0.1 INT8 Inference Accuracy
Top 1 Top 5
Watson Cloud Platform Strategic Customer Success
17 50.12 128.74 289.28 373.29 45.61 51.7 52.41 440.82 424.22 842.51 2154.29 1519 500 1000 1500 2000 2500
1 x Nvidia K80 GPU 1 x Nvidia P4 GPU 1 x Nvidia P100 GPU 1 x Nvidia V100 GPU 1 x Intel DLIA 1 x Nvidia K80 GPU SL Bare-Metal (Dual Xeon E5-2690v4) Intel Bare-Metal (Dual Xeon E5-2650v3) SL Bare-Metal (Dual Xeon E5-2690v4) SL Bare-Metal (Dual Xeon E5-2690v4) Intel Bare-Metal (Dual Xeon E5-2690v4) AWS p2.16xlarge (Dual Xeon E5-2686v4)
Number of Images Processed Per Second (Inferening)
Image Classification with VGG-16 on Caffe (Single Precision)
PyCaffe TensorRT TensorRT (HALF2) TensorRT (INT8)
Higher is better
Notes:
Watson Cloud Platform Strategic Customer Success
1000 2000 3000 4000 5000 6000 7000 8000 9000 1 50 167 238 404 572 1135
Latency Per Iteration (ms)
Batch Size
Deep Learning Model Inferencing - Image Classification
VGG-16 Neural Net on Caffe Framework with TensorRT
(Single Precision Except Where Noted Otherwise)
K80 GPU P100 GPU P100 GPU w/ HALF2 V100 GPU V100 GPU w/ HALF2
0.5 1 1.5 2 2.5
Intel DLIA Nvidia P4 GPU
Classification Power Efficiency (Images/Second/Watt)
Deep-Learning Model Inferencing (Image Classification) Nvidia P4 GPU vs. FPGA for DL
VGG-16 Neural Net on Caffe Framework (Single Precision)
Lower is better Higher is better
Notes:
the K80 GPU
than previous-generation K80 GPU
consumption of a P100 GPU – making the P4 a very cost-effective DL inferencing engine for a cloud platform
Watson Cloud Platform Strategic Customer Success
§ Python interface was very important § Utility to convert Caffe model to TRT network saved much work – Building the network with native TRT calls is much more advanced – Allows flexibility and customization for those who need it § INT8 Calibration was challenging – Doc was not complete – Found a good blog that referenced classes not in the doc
– Need to write a Calibrator implementation by extending a class
§ Overall for Caffe our experience was good. Other frameworks might require more use of lower TRT calls.
19
Watson Cloud Platform Strategic Customer Success
§ V100 GPU is faster than anything else § Nvidia TensorRT with half-precision support improves DL inferencing performance by 5X on a V100 GPU § TensorRT is much faster than pycaffe § HALF and INT8 are significantly faster than FLOAT with no or little loss of accuracy § INT8 slightly better than HALF for batch size 1 – Latency 1.8 vs 2.0 msec – Throughput 558 vs 505 imagers/sec § HALF has better throughput and latency at larger batch sizes § Surprising that INT8 accuracy when calibration used only 1 image performed so well – Suggest to follow guidance of using more images as calibration is not that slow and need
§ Experiment with your own network and data as your results could vary
20
Watson Cloud Platform Strategic Customer Success
– Senior Technical Staff Member (STSM), IBM – khoa@us.ibm.com
– Senior Software Engineer, IBM – ltbrown@us.ibm.com
21