Pipelines are the default API for deploying a model with DeepSparse.
Similar to Hugging Face Pipelines, DeepSparse Pipelines wrap inference with task-specific pre- and post-processing, enabling you to pass raw data and receive the predictions.
Let us try a quick example of the Pipeline API. All we have to do is pass a task and model to the
the Pipeline.create function, and then we can run inference on raw data using DeepSparse!
This example creates a sentiment analysis Pipeline with a 90% pruned-quantized verion of BERT from the SparseZoo.
from deepsparse import Pipeline
# download and compile onnx, create pipeline
zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
sentiment_analysis_pipeline = Pipeline.create(
task="sentiment-analysis", # name of the task
model_path=zoo_stub, # zoo stub or path to local onnx file
)
# run inference
print(sentiment_analysis_pipeline("I love using DeepSparse Pipelines"))
# >>> labels=['positive'] scores=[0.9954759478569031]In this case we passed a SparseZoo stub as the model, which instructs DeepSparse to download the
relevant ONNX file from the SparseZoo. To deploy your own model, pass a path to a model.onnx file or to a
folder containing model.onnx and supporting files (e.g., the Hugging Face tokenizer.json and config.json).
Pipelines support many CV and NLP use cases out of the box. Check out the Use Cases page for details on task-specific APIs.
Beyond officially supported tasks, Pipelines can be extended to additional tasks via the CustomTaskPipeline.
CustomTaskPipelines are passed the following arguments:
model_path- a SparseZoo stub or path to a local ONNX fileprocess_inputs_fn- an optional function that handles pre-processing of input into a list of numpy arrays that can be passed directly to the inference forward passprocess_outputs_fn- an optional function that handles post-processing of the list of numpy arrays that are the output of the engine forward pass
To replicate the functionality of the image classification pipeline as a custom Pipeline, an example is provided.
Download an image and ONNX file (a 95% pruned-quantized ResNet-50) for the demo:
wget https://github.com/raw/neuralmagic/docs/main/files-for-examples/use-cases/embedding-extraction/goldfish.jpg
sparsezoo.download zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned95_quant-none --save-dir ./resnet-50-pruned-quant
We can create a custom image classification Pipeline which returns the raw logits and class probabilities for the 1000 ImageNet classes with the following:
For the purposes of this quick example, make sure you have torch torchvision and Pillow installed.
from deepsparse.pipelines.custom_pipeline import CustomTaskPipeline
from torchvision import transforms
from PIL import Image
import numpy as np
import torch
IMAGENET_RGB_MEANS = [0.485, 0.456, 0.406]
IMAGENET_RGB_STDS = [0.229, 0.224, 0.225]
preprocess_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=IMAGENET_RGB_MEANS, std=IMAGENET_RGB_STDS),
])
def preprocess(img_file):
with open(img_file, "rb") as img_file:
img = Image.open(img_file)
img = img.convert("RGB")
img = preprocess_transforms(img)
batch = torch.stack([img])
return [batch.numpy()]
custom_pipeline = CustomTaskPipeline(
model_path="./resnet-50-pruned-quant/model.onnx",
process_inputs_fn=preprocess,
)
scores, probs = custom_pipeline("goldfish.jpg")
print(scores.shape)
print(probs.shape)
print(np.sum(probs))
print(np.argmax(probs))
# >> (1,1000)
# >> (1,1000)
# >> ~1.00000
# >> 1 << index of the goldfish class in ImageNetBeyond supporting pre- and post-processing, Pipelines also offer additional utilities that simplify the deployment process.
The batch_size argument configures the batch size of the Pipeline, modifying the underlying ONNX graph for you.
The default is batch size 1, and but we can override to batch size 3 with the following:
from deepsparse import Pipeline
zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
batch_size = 3
sentiment_analysis_pipeline = Pipeline.create(
task="sentiment-analysis", # name of the task
model_path=zoo_stub, # zoo stub or path to local onnx file
batch_size=batch_size # default is batch 1
)
sentences = [
"I love DeepSparse Pipelines",
"I hated changing the batch size with my prior Deep Learning framework",
"DeepSparse makes it very easy to adjust the batch size"
]
output = sentiment_analysis_pipeline(sentences)
print(output)
# >>> labels=['positive', 'negative', 'positive'] scores=[0.9969560503959656, 0.9964107871055603, 0.7127435207366943]The num_cores argument configures the number of physical cores used by DeepSparse. The default is None, which
instructs DeepSparse to use all physical cores available on the system. We can override to use only
one core with the following:
from deepsparse import Pipeline
zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
sentiment_analysis_pipeline = Pipeline.create(
task="sentiment-analysis", # name of the task
model_path=zoo_stub, # zoo stub or path to local onnx file
num_cores=1
)
sentences = "I love how DeepSparse makes it easy to configure the number of cores used"
output = sentiment_analysis_pipeline(sentences)
print(output)
# >> labels=['positive'] scores=[0.9951152801513672] << but runs slower than if using all coresWe can utilize an the multi-stream capabilites of DeepSparse to make requests with dynamic batch sizes.
Let us create an example with a single sentiment analysis Pipeline with dynamic batch sizes by
setting the batch_size argument to None. Under the hood, the pipeline will split the batch into
multiple asynchronous requests using the multi-stream scheduler.
from deepsparse import Pipeline
zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
sentiment_analysis_pipeline = Pipeline.create(
task="sentiment_analysis",
model_path=zoo_stub,
batch_size=None, # setting to None enables dynamic batch
)
b1_request = ["This multi model concept is great!"]
b4_request = b1_request * 4
output_b1 = sentiment_analysis_pipeline(b1_request)
output_b4 = sentiment_analysis_pipeline(b4_request)
print(output_b1)
# >> labels=['positive'] scores=[0.9995297789573669]
print(output_b4)
# >> labels=['positive', 'positive', 'positive', 'positive'] scores=[0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669]Some deployment scenarios will require running multiple instances of DeepSparse on a single machine. DeepSparse includes a concept called Context. Contexts can be used to run multiple models with the same scheduler, enabling DeepSparse to manage the resources of the system effectively, keeping engines that are running different models from fighting over resources.
To create an example with multiple sentiment analysis Pipelines, one with batch size 1 (for maximum latency) and one with batch size 32 (for maximum throughput):
from concurrent.futures import ThreadPoolExecutor
from deepsparse.engine import Context
from deepsparse import Pipeline
context = Context()
executor = ThreadPoolExecutor(max_workers=context.num_streams)
zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
sentiment_analysis_pipeline_b1 = Pipeline.create(
task="sentiment_analysis",
model_path=zoo_stub,
batch_size=1,
context=context,
executor=executor
)
sentiment_analysis_pipeline_b32 = Pipeline.create(
task="sentiment_analysis",
model_path=zoo_stub,
batch_size=32,
context=context,
executor=executor
)
b1_request = ["This multi model concept is great!"]
b32_request = b1_request * 32
output_b1 = sentiment_analysis_pipeline_b1(b1_request)
output_b32 = sentiment_analysis_pipeline_b32(b32_request)
print(output_b1)
print(output_b32)
# >> labels=['positive'] scores=[0.9995297789573669]
# >> labels=['positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive', 'positive'] scores=[0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669, 0.9995297789573669]If you are deploying multiple models on a same system, you may want to answer multiple
requests concurrently. We can enable this but setting the num_streams argument in the Context argument.
from concurrent.futures import ThreadPoolExecutor
from deepsparse.engine import Context
from deepsparse.pipeline import Pipeline
import threading
class ExecutorThread(threading.Thread):
def __init__(self, pipeline, input, iters=1):
super(ExecutorThread, self).__init__()
self.pipeline = pipeline
self.input = input
self.iters = iters
def run(self):
for _ in range(self.iters):
output = self.pipeline(self.input)
print(output)
num_concurrent_requests = 2
context = Context(num_streams=num_concurrent_requests)
executor = ThreadPoolExecutor(max_workers=context.num_streams)
zoo_stub = "zoo:nlp/sentiment_analysis/obert-base/pytorch/huggingface/sst2/pruned90_quant-none"
sentiment_analysis_pipeline_b1 = Pipeline.create(
task="sentiment_analysis",
model_path=zoo_stub,
batch_size=1,
context=context,
executor=executor
)
sentiment_analysis_pipeline_b32 = Pipeline.create(
task="sentiment_analysis",
model_path=zoo_stub,
batch_size=32,
context=context,
executor=executor
)
b1_request = ["This multi model concept is great!"]
b64_request = b1_request * 32
threads = [
ExecutorThread(sentiment_analysis_pipeline_b1, input=b1_request, iters=64),
ExecutorThread(sentiment_analysis_pipeline_b32, input=b64_request, iters=1),
]
for thread in threads:
thread.start()
for thread in threads:
thread.join()
# mutiple b=1 queries print results before the b=32 query returnsNote that requests will be execute in a FIFO manner, with a maximum of num_concurrent_requests running at once.
As a result, high traffic on one of your Pipelines can impact performance on the other Pipeline. If you prefer to
isolate your Pipelines, we recommend using an orchestration framework such as Docker and Kubernetes with
one DeepSparse Pipeline running in each container for proper process isolation.
Stay tuned for documentation on enabling multi-stream scheduling with DeepSparse Pipelines.
Stay tuned for documentation on enabling bucketing with DeepSparse Pipelines.
Stay tuned for documentation on enabling logging with DeepSparse Pipelines.