sandroci/redisai-py

Python client for RedisAI

RedisAI Python Client

Installation

1. Install Redis 5.0 or above

2. Install RedisAI

3. Install the python client

$ pip install redisai

4. Install serialization-deserialization utility (optional)

$ pip install ml2rt

RedisAI example repo shows few examples made using redisai-py under python_client section. Checkout ml2rt for convenient functions those might help in converting models (sparkml, sklearn, xgboost to ONNX), serializing models to disk, loading it back to redisai-py etc.

Documentation

APIs available in redisai-py

tensorset

Set the values of the tensor on the server using the provided tensor object

Parameters

• name: Key on which tensor is saved
• tensor: a Tensor object
• shape: Shape of the tensor
• dtype: redisai.DType object represents data type of the tensor. Required if input is a list/tuple

Example

from redisai import Client, DType
client = Client()
arr = np.array([2, 3])
client.tensorset('x', arr)
client.tensorget('y', [1, 2], dtype=DType.float)
client.tensorget('z', [3, 4, 5, 6], dtype=DType.float, shape=(1, 2, 2))

tensorget

Parameters

Retrieve the value of a tensor from the server. By default it returns the numpy array but it can be controlled using as_type and meta_only arguments

• name: Key from where the tensor is saved
• as_type: the resultant tensor type. Returns numpy array if None
• meta_only: if true, then the value is not retrieved, only the shape and the type

Example

from redisai import Tensor
x = client.tensorget('x')  # numpy array
y = client.tensorget('y', as_type=Tensor)  # A Tensor object
z = client.tensorget('z', meta_only=True)  # A Tensor object but without value

loadbackend

RedisAI by default won't load any backends. User can either explicitly load the backend by using this function or let RedisAI load the required backend from the default path on-demand.

Parameters

• identifier: String representing which backend. Allowed values - TF, TORCH & ONNX
• path: Path to the shared object of the backend

Example

client.loadbackend('TORCH', 'install-cpu/backends/redisai_torch/redisai_torch.so')

modelset

Store a model of Tensorflow/PyTorch/ONNX format in RedisAI

Parameters

• name: Key on which model should be saved
• backend: redisai.Backend object - tf, torch or onnx
• device: redisai.Device object - cpu or gpu
• data: model as a byte string. ml2rt.load_model(path) returns this.

Example

Tensorflow requires the input and output nodes of the graph while storing the model. For exporting a normal tensorflow session to pb file, you could use ml2rt package

import ml2rt
ml2rt.save_tensorflow(sess, 'path/to/graph.pb', output_nodes)
model = ml2rt.load_model('path/to/graph.pb')
client.modelset('m', Backend.tf,
                Device.cpu,
                input=['input_1', 'input_2'],
                output='output',
                data=model)

Torch doesn't need input and output node information. You could use ml2rt for exporting torch model as well but ml2rt needs torchscript model rather than normal torch model. Checkout the document to learn more.

import ml2rt
ml2rt.save_torch('optimized_graph', 'path/to/graph.pt')
model = ml2rt.load_model('path/to/graph.pt')
client.modelset('m', Backend.torch, Device.cpu, data=model)

modelget

Fetch the stored model from RedisAI

Parameters

name: the name of the model

Example

mod_det = client.modelget('m')
print(mod_det['backend'], mod_det['device'])
model_binary = mod_det['data']

modeldel

Delete a stored model from RedisAI

Parameters

• name: Key of the model

Example

client.modeldel('m')

modelrun

Execute a model. Required inputs must be present in RedisAI before calling modelrun

Parameters

• name: Key of the model
• inputs: Key of the input tensors. It can be a single key or a list of keys
• outputs: Key on which the output tensors will be saved. It can be a single key or list of keys

Example

client.tensorset('a', [2, 3], dtype=DType.float, shape=(2,))
client.tensorset('b', [12, 10], dtype=DType.float)
model = ml2rt.load_model('test/testdata/graph.pt')
client.modelset('m', Backend.torch,
                Device.cpu,
                input=['input_1', 'input_2'],
                output='output',
                data=model)
client.modelrun('m', ['a', 'b'], ['mul'])
out = client.tensorget('mul')

scriptset

Store a SCRIPT in RedisAI. SCRIPT is a subset of python language itself but will be executed on high performance C++ runtime. RedisAI uses TORCH runtime to execute SCRIPT and it must follow the format required by the doc.

Parameters

• name: Key on which SCRIPT should be saved
• device: redisai.Device object - cpu or gpu
• script: SCRIPT as defined in TorchScript documentation. SCRIPT must have functions defined in it (you can have multiple functions).

Example

script = """
def myfunc(a, b):
    return a + b
"""
client.scriptset('script', Device.cpu, script)

scriptget

Fetch a stored SCRIPT from RedisAI

Parameters

• name: Key from which SCRIPT can be retrieved

Example

script_details = client.scriptget('script')
device = script_details['device']
script = script_details['script']

scriptdel

Delete a stored SCRIPT from RedisAI

Parameters

• name: Key from which SCRIPT can be retrieved

Example

client.scriptdel('script')

scriptrun

Execute a SCRIPT. Required inputs must be present in RedisAI before calling modelrun

Parameters

• name: Key from which SCRIPT can be retrieved
• function: name of the function to call. The function that you call can call other functions in the same SCRIPT
• inputs: Key of the input tensors. It can be a single key or a list of keys
• outputs: Key on which the output tensors will be saved. It can be a single key or list of keys

Example

script = """
def myfunc(a, b):
    return a + b
"""
client.scriptset('script', Device.cpu, script)
client.tensorget('a', [1, 2], dtype=DType.float)
client.tensorget('b', [3, 4], dtype=DType.float)
client.scriptrun('script', 'mufunc', ['a', 'b'], 'out')
out = client.tensorget('out')  # => [4, 6]