Quick Start Tutorial for Compiling Deep Learning Models

Author: Yao Wang, Truman Tian

This example shows how to build a neural network with Relay python frontend and generates a runtime library for Nvidia GPU with TVM. Notice that you need to build TVM with cuda and llvm enabled.

Overview for Supported Hardware Backend of TVM

The image below shows hardware backend currently supported by TVM:

https://github.com/dmlc/web-data/raw/master/tvm/tutorial/tvm_support_list.png

In this tutorial, we’ll choose cuda and llvm as target backends. To begin with, let’s import Relay and TVM.

import numpy as np

from tvm import relay
from tvm.relay import testing
import tvm
from tvm import te
from tvm.contrib import graph_runtime

Define Neural Network in Relay

First, let’s define a neural network with relay python frontend. For simplicity, we’ll use pre-defined resnet-18 network in Relay. Parameters are initialized with Xavier initializer. Relay also supports other model formats such as MXNet, CoreML, ONNX and Tensorflow.

In this tutorial, we assume we will do inference on our device and the batch size is set to be 1. Input images are RGB color images of size 224 * 224. We can call the tvm.relay.TupleWrapper.astext() to show the network structure.

batch_size = 1
num_class = 1000
image_shape = (3, 224, 224)
data_shape = (batch_size,) + image_shape
out_shape = (batch_size, num_class)

mod, params = relay.testing.resnet.get_workload(
    num_layers=18, batch_size=batch_size, image_shape=image_shape)

# set show_meta_data=True if you want to show meta data
print(mod.astext(show_meta_data=False))

Out:

v0.0.4
def @main(%data: Tensor[(1, 3, 224, 224), float32], %bn_data_gamma: Tensor[(3), float32], %bn_data_beta: Tensor[(3), float32], %bn_data_moving_mean: Tensor[(3), float32], %bn_data_moving_var: Tensor[(3), float32], %conv0_weight: Tensor[(64, 3, 7, 7), float32], %bn0_gamma: Tensor[(64), float32], %bn0_beta: Tensor[(64), float32], %bn0_moving_mean: Tensor[(64), float32], %bn0_moving_var: Tensor[(64), float32], %stage1_unit1_bn1_gamma: Tensor[(64), float32], %stage1_unit1_bn1_beta: Tensor[(64), float32], %stage1_unit1_bn1_moving_mean: Tensor[(64), float32], %stage1_unit1_bn1_moving_var: Tensor[(64), float32], %stage1_unit1_conv1_weight: Tensor[(64, 64, 3, 3), float32], %stage1_unit1_bn2_gamma: Tensor[(64), float32], %stage1_unit1_bn2_beta: Tensor[(64), float32], %stage1_unit1_bn2_moving_mean: Tensor[(64), float32], %stage1_unit1_bn2_moving_var: Tensor[(64), float32], %stage1_unit1_conv2_weight: Tensor[(64, 64, 3, 3), float32], %stage1_unit1_sc_weight: Tensor[(64, 64, 1, 1), float32], %stage1_unit2_bn1_gamma: Tensor[(64), float32], %stage1_unit2_bn1_beta: Tensor[(64), float32], %stage1_unit2_bn1_moving_mean: Tensor[(64), float32], %stage1_unit2_bn1_moving_var: Tensor[(64), float32], %stage1_unit2_conv1_weight: Tensor[(64, 64, 3, 3), float32], %stage1_unit2_bn2_gamma: Tensor[(64), float32], %stage1_unit2_bn2_beta: Tensor[(64), float32], %stage1_unit2_bn2_moving_mean: Tensor[(64), float32], %stage1_unit2_bn2_moving_var: Tensor[(64), float32], %stage1_unit2_conv2_weight: Tensor[(64, 64, 3, 3), float32], %stage2_unit1_bn1_gamma: Tensor[(64), float32], %stage2_unit1_bn1_beta: Tensor[(64), float32], %stage2_unit1_bn1_moving_mean: Tensor[(64), float32], %stage2_unit1_bn1_moving_var: Tensor[(64), float32], %stage2_unit1_conv1_weight: Tensor[(128, 64, 3, 3), float32], %stage2_unit1_bn2_gamma: Tensor[(128), float32], %stage2_unit1_bn2_beta: Tensor[(128), float32], %stage2_unit1_bn2_moving_mean: Tensor[(128), float32], %stage2_unit1_bn2_moving_var: Tensor[(128), float32], %stage2_unit1_conv2_weight: Tensor[(128, 128, 3, 3), float32], %stage2_unit1_sc_weight: Tensor[(128, 64, 1, 1), float32], %stage2_unit2_bn1_gamma: Tensor[(128), float32], %stage2_unit2_bn1_beta: Tensor[(128), float32], %stage2_unit2_bn1_moving_mean: Tensor[(128), float32], %stage2_unit2_bn1_moving_var: Tensor[(128), float32], %stage2_unit2_conv1_weight: Tensor[(128, 128, 3, 3), float32], %stage2_unit2_bn2_gamma: Tensor[(128), float32], %stage2_unit2_bn2_beta: Tensor[(128), float32], %stage2_unit2_bn2_moving_mean: Tensor[(128), float32], %stage2_unit2_bn2_moving_var: Tensor[(128), float32], %stage2_unit2_conv2_weight: Tensor[(128, 128, 3, 3), float32], %stage3_unit1_bn1_gamma: Tensor[(128), float32], %stage3_unit1_bn1_beta: Tensor[(128), float32], %stage3_unit1_bn1_moving_mean: Tensor[(128), float32], %stage3_unit1_bn1_moving_var: Tensor[(128), float32], %stage3_unit1_conv1_weight: Tensor[(256, 128, 3, 3), float32], %stage3_unit1_bn2_gamma: Tensor[(256), float32], %stage3_unit1_bn2_beta: Tensor[(256), float32], %stage3_unit1_bn2_moving_mean: Tensor[(256), float32], %stage3_unit1_bn2_moving_var: Tensor[(256), float32], %stage3_unit1_conv2_weight: Tensor[(256, 256, 3, 3), float32], %stage3_unit1_sc_weight: Tensor[(256, 128, 1, 1), float32], %stage3_unit2_bn1_gamma: Tensor[(256), float32], %stage3_unit2_bn1_beta: Tensor[(256), float32], %stage3_unit2_bn1_moving_mean: Tensor[(256), float32], %stage3_unit2_bn1_moving_var: Tensor[(256), float32], %stage3_unit2_conv1_weight: Tensor[(256, 256, 3, 3), float32], %stage3_unit2_bn2_gamma: Tensor[(256), float32], %stage3_unit2_bn2_beta: Tensor[(256), float32], %stage3_unit2_bn2_moving_mean: Tensor[(256), float32], %stage3_unit2_bn2_moving_var: Tensor[(256), float32], %stage3_unit2_conv2_weight: Tensor[(256, 256, 3, 3), float32], %stage4_unit1_bn1_gamma: Tensor[(256), float32], %stage4_unit1_bn1_beta: Tensor[(256), float32], %stage4_unit1_bn1_moving_mean: Tensor[(256), float32], %stage4_unit1_bn1_moving_var: Tensor[(256), float32], %stage4_unit1_conv1_weight: Tensor[(512, 256, 3, 3), float32], %stage4_unit1_bn2_gamma: Tensor[(512), float32], %stage4_unit1_bn2_beta: Tensor[(512), float32], %stage4_unit1_bn2_moving_mean: Tensor[(512), float32], %stage4_unit1_bn2_moving_var: Tensor[(512), float32], %stage4_unit1_conv2_weight: Tensor[(512, 512, 3, 3), float32], %stage4_unit1_sc_weight: Tensor[(512, 256, 1, 1), float32], %stage4_unit2_bn1_gamma: Tensor[(512), float32], %stage4_unit2_bn1_beta: Tensor[(512), float32], %stage4_unit2_bn1_moving_mean: Tensor[(512), float32], %stage4_unit2_bn1_moving_var: Tensor[(512), float32], %stage4_unit2_conv1_weight: Tensor[(512, 512, 3, 3), float32], %stage4_unit2_bn2_gamma: Tensor[(512), float32], %stage4_unit2_bn2_beta: Tensor[(512), float32], %stage4_unit2_bn2_moving_mean: Tensor[(512), float32], %stage4_unit2_bn2_moving_var: Tensor[(512), float32], %stage4_unit2_conv2_weight: Tensor[(512, 512, 3, 3), float32], %bn1_gamma: Tensor[(512), float32], %bn1_beta: Tensor[(512), float32], %bn1_moving_mean: Tensor[(512), float32], %bn1_moving_var: Tensor[(512), float32], %fc1_weight: Tensor[(1000, 512), float32], %fc1_bias: Tensor[(1000), float32]) -> Tensor[(1, 1000), float32] {
  %0 = nn.batch_norm(%data, %bn_data_gamma, %bn_data_beta, %bn_data_moving_mean, %bn_data_moving_var, epsilon=2e-05f, scale=False) /* ty=(Tensor[(1, 3, 224, 224), float32], Tensor[(3), float32], Tensor[(3), float32]) */;
  %1 = %0.0;
  %2 = nn.conv2d(%1, %conv0_weight, strides=[2, 2], padding=[3, 3, 3, 3], channels=64, kernel_size=[7, 7]) /* ty=Tensor[(1, 64, 112, 112), float32] */;
  %3 = nn.batch_norm(%2, %bn0_gamma, %bn0_beta, %bn0_moving_mean, %bn0_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 112, 112), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %4 = %3.0;
  %5 = nn.relu(%4) /* ty=Tensor[(1, 64, 112, 112), float32] */;
  %6 = nn.max_pool2d(%5, pool_size=[3, 3], strides=[2, 2], padding=[1, 1, 1, 1]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %7 = nn.batch_norm(%6, %stage1_unit1_bn1_gamma, %stage1_unit1_bn1_beta, %stage1_unit1_bn1_moving_mean, %stage1_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %8 = %7.0;
  %9 = nn.relu(%8) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %10 = nn.conv2d(%9, %stage1_unit1_conv1_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %11 = nn.batch_norm(%10, %stage1_unit1_bn2_gamma, %stage1_unit1_bn2_beta, %stage1_unit1_bn2_moving_mean, %stage1_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %12 = %11.0;
  %13 = nn.relu(%12) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %14 = nn.conv2d(%13, %stage1_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %15 = nn.conv2d(%9, %stage1_unit1_sc_weight, padding=[0, 0, 0, 0], channels=64, kernel_size=[1, 1]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %16 = add(%14, %15) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %17 = nn.batch_norm(%16, %stage1_unit2_bn1_gamma, %stage1_unit2_bn1_beta, %stage1_unit2_bn1_moving_mean, %stage1_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %18 = %17.0;
  %19 = nn.relu(%18) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %20 = nn.conv2d(%19, %stage1_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %21 = nn.batch_norm(%20, %stage1_unit2_bn2_gamma, %stage1_unit2_bn2_beta, %stage1_unit2_bn2_moving_mean, %stage1_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %22 = %21.0;
  %23 = nn.relu(%22) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %24 = nn.conv2d(%23, %stage1_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %25 = add(%24, %16) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %26 = nn.batch_norm(%25, %stage2_unit1_bn1_gamma, %stage2_unit1_bn1_beta, %stage2_unit1_bn1_moving_mean, %stage2_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %27 = %26.0;
  %28 = nn.relu(%27) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %29 = nn.conv2d(%28, %stage2_unit1_conv1_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %30 = nn.batch_norm(%29, %stage2_unit1_bn2_gamma, %stage2_unit1_bn2_beta, %stage2_unit1_bn2_moving_mean, %stage2_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %31 = %30.0;
  %32 = nn.relu(%31) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %33 = nn.conv2d(%32, %stage2_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %34 = nn.conv2d(%28, %stage2_unit1_sc_weight, strides=[2, 2], padding=[0, 0, 0, 0], channels=128, kernel_size=[1, 1]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %35 = add(%33, %34) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %36 = nn.batch_norm(%35, %stage2_unit2_bn1_gamma, %stage2_unit2_bn1_beta, %stage2_unit2_bn1_moving_mean, %stage2_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %37 = %36.0;
  %38 = nn.relu(%37) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %39 = nn.conv2d(%38, %stage2_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %40 = nn.batch_norm(%39, %stage2_unit2_bn2_gamma, %stage2_unit2_bn2_beta, %stage2_unit2_bn2_moving_mean, %stage2_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %41 = %40.0;
  %42 = nn.relu(%41) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %43 = nn.conv2d(%42, %stage2_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %44 = add(%43, %35) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %45 = nn.batch_norm(%44, %stage3_unit1_bn1_gamma, %stage3_unit1_bn1_beta, %stage3_unit1_bn1_moving_mean, %stage3_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %46 = %45.0;
  %47 = nn.relu(%46) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %48 = nn.conv2d(%47, %stage3_unit1_conv1_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %49 = nn.batch_norm(%48, %stage3_unit1_bn2_gamma, %stage3_unit1_bn2_beta, %stage3_unit1_bn2_moving_mean, %stage3_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %50 = %49.0;
  %51 = nn.relu(%50) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %52 = nn.conv2d(%51, %stage3_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %53 = nn.conv2d(%47, %stage3_unit1_sc_weight, strides=[2, 2], padding=[0, 0, 0, 0], channels=256, kernel_size=[1, 1]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %54 = add(%52, %53) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %55 = nn.batch_norm(%54, %stage3_unit2_bn1_gamma, %stage3_unit2_bn1_beta, %stage3_unit2_bn1_moving_mean, %stage3_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %56 = %55.0;
  %57 = nn.relu(%56) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %58 = nn.conv2d(%57, %stage3_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %59 = nn.batch_norm(%58, %stage3_unit2_bn2_gamma, %stage3_unit2_bn2_beta, %stage3_unit2_bn2_moving_mean, %stage3_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %60 = %59.0;
  %61 = nn.relu(%60) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %62 = nn.conv2d(%61, %stage3_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %63 = add(%62, %54) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %64 = nn.batch_norm(%63, %stage4_unit1_bn1_gamma, %stage4_unit1_bn1_beta, %stage4_unit1_bn1_moving_mean, %stage4_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %65 = %64.0;
  %66 = nn.relu(%65) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %67 = nn.conv2d(%66, %stage4_unit1_conv1_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %68 = nn.batch_norm(%67, %stage4_unit1_bn2_gamma, %stage4_unit1_bn2_beta, %stage4_unit1_bn2_moving_mean, %stage4_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %69 = %68.0;
  %70 = nn.relu(%69) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %71 = nn.conv2d(%70, %stage4_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %72 = nn.conv2d(%66, %stage4_unit1_sc_weight, strides=[2, 2], padding=[0, 0, 0, 0], channels=512, kernel_size=[1, 1]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %73 = add(%71, %72) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %74 = nn.batch_norm(%73, %stage4_unit2_bn1_gamma, %stage4_unit2_bn1_beta, %stage4_unit2_bn1_moving_mean, %stage4_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %75 = %74.0;
  %76 = nn.relu(%75) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %77 = nn.conv2d(%76, %stage4_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %78 = nn.batch_norm(%77, %stage4_unit2_bn2_gamma, %stage4_unit2_bn2_beta, %stage4_unit2_bn2_moving_mean, %stage4_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %79 = %78.0;
  %80 = nn.relu(%79) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %81 = nn.conv2d(%80, %stage4_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %82 = add(%81, %73) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %83 = nn.batch_norm(%82, %bn1_gamma, %bn1_beta, %bn1_moving_mean, %bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %84 = %83.0;
  %85 = nn.relu(%84) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %86 = nn.global_avg_pool2d(%85) /* ty=Tensor[(1, 512, 1, 1), float32] */;
  %87 = nn.batch_flatten(%86) /* ty=Tensor[(1, 512), float32] */;
  %88 = nn.dense(%87, %fc1_weight, units=1000) /* ty=Tensor[(1, 1000), float32] */;
  %89 = nn.bias_add(%88, %fc1_bias, axis=-1) /* ty=Tensor[(1, 1000), float32] */;
  nn.softmax(%89) /* ty=Tensor[(1, 1000), float32] */
}

Compilation

Next step is to compile the model using the Relay/TVM pipeline. Users can specify the optimization level of the compilation. Currently this value can be 0 to 3. The optimization passes include operator fusion, pre-computation, layout transformation and so on.

relay.build() returns three components: the execution graph in json format, the TVM module library of compiled functions specifically for this graph on the target hardware, and the parameter blobs of the model. During the compilation, Relay does the graph-level optimization while TVM does the tensor-level optimization, resulting in an optimized runtime module for model serving.

We’ll first compile for Nvidia GPU. Behind the scene, relay.build() first does a number of graph-level optimizations, e.g. pruning, fusing, etc., then registers the operators (i.e. the nodes of the optimized graphs) to TVM implementations to generate a tvm.module. To generate the module library, TVM will first transfer the high level IR into the lower intrinsic IR of the specified target backend, which is CUDA in this example. Then the machine code will be generated as the module library.

opt_level = 3
target = tvm.target.cuda()
with tvm.transform.PassContext(opt_level=opt_level):
    graph, lib, params = relay.build(mod, target, params=params)

Out:

...1%, 0.01 MB, 37 KB/s, 0 seconds passed
...3%, 0.02 MB, 75 KB/s, 0 seconds passed
...5%, 0.02 MB, 113 KB/s, 0 seconds passed
...7%, 0.03 MB, 150 KB/s, 0 seconds passed
...9%, 0.04 MB, 187 KB/s, 0 seconds passed
...11%, 0.05 MB, 224 KB/s, 0 seconds passed
...13%, 0.05 MB, 261 KB/s, 0 seconds passed
...15%, 0.06 MB, 298 KB/s, 0 seconds passed
...17%, 0.07 MB, 335 KB/s, 0 seconds passed
...19%, 0.08 MB, 372 KB/s, 0 seconds passed
...21%, 0.09 MB, 409 KB/s, 0 seconds passed
...23%, 0.09 MB, 445 KB/s, 0 seconds passed
...25%, 0.10 MB, 481 KB/s, 0 seconds passed
...27%, 0.11 MB, 518 KB/s, 0 seconds passed
...29%, 0.12 MB, 554 KB/s, 0 seconds passed
...31%, 0.12 MB, 590 KB/s, 0 seconds passed
...33%, 0.13 MB, 626 KB/s, 0 seconds passed
...35%, 0.14 MB, 662 KB/s, 0 seconds passed
...37%, 0.15 MB, 698 KB/s, 0 seconds passed
...39%, 0.16 MB, 734 KB/s, 0 seconds passed
...41%, 0.16 MB, 770 KB/s, 0 seconds passed
...43%, 0.17 MB, 804 KB/s, 0 seconds passed
...45%, 0.18 MB, 840 KB/s, 0 seconds passed
...47%, 0.19 MB, 876 KB/s, 0 seconds passed
...49%, 0.20 MB, 912 KB/s, 0 seconds passed
...51%, 0.20 MB, 948 KB/s, 0 seconds passed
...53%, 0.21 MB, 981 KB/s, 0 seconds passed
...55%, 0.22 MB, 1017 KB/s, 0 seconds passed
...57%, 0.23 MB, 1052 KB/s, 0 seconds passed
...59%, 0.23 MB, 1087 KB/s, 0 seconds passed
...61%, 0.24 MB, 1121 KB/s, 0 seconds passed
...63%, 0.25 MB, 1157 KB/s, 0 seconds passed
...65%, 0.26 MB, 1192 KB/s, 0 seconds passed
...67%, 0.27 MB, 1228 KB/s, 0 seconds passed
...69%, 0.27 MB, 1262 KB/s, 0 seconds passed
...71%, 0.28 MB, 1297 KB/s, 0 seconds passed
...73%, 0.29 MB, 1332 KB/s, 0 seconds passed
...75%, 0.30 MB, 1367 KB/s, 0 seconds passed
...77%, 0.30 MB, 1399 KB/s, 0 seconds passed
...79%, 0.31 MB, 1434 KB/s, 0 seconds passed
...81%, 0.32 MB, 1469 KB/s, 0 seconds passed
...83%, 0.33 MB, 1504 KB/s, 0 seconds passed
...85%, 0.34 MB, 1539 KB/s, 0 seconds passed
...87%, 0.34 MB, 1575 KB/s, 0 seconds passed
...89%, 0.35 MB, 1606 KB/s, 0 seconds passed
...91%, 0.36 MB, 1641 KB/s, 0 seconds passed
...93%, 0.37 MB, 1676 KB/s, 0 seconds passed
...95%, 0.38 MB, 1711 KB/s, 0 seconds passed
...97%, 0.38 MB, 1745 KB/s, 0 seconds passed
...99%, 0.39 MB, 1780 KB/s, 0 seconds passed
...100%, 0.40 MB, 1814 KB/s, 0 seconds passed
Cannot find config for target=cuda -model=unknown, workload=('dense_small_batch.cuda', ('TENSOR', (1, 512), 'float32'), ('TENSOR', (1000, 512), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.

Run the generate library

Now we can create graph runtime and run the module on Nvidia GPU.

# create random input
ctx = tvm.gpu()
data = np.random.uniform(-1, 1, size=data_shape).astype("float32")
# create module
module = graph_runtime.create(graph, lib, ctx)
# set input and parameters
module.set_input("data", data)
module.set_input(**params)
# run
module.run()
# get output
out = module.get_output(0, tvm.nd.empty(out_shape)).asnumpy()

# Print first 10 elements of output
print(out.flatten()[0:10])

Out:

[0.00089283 0.00103331 0.0009094  0.00102275 0.00108751 0.00106737
 0.00106262 0.00095838 0.00110792 0.00113151]

Save and Load Compiled Module

We can also save the graph, lib and parameters into files and load them back in deploy environment.

# save the graph, lib and params into separate files
from tvm.contrib import util

temp = util.tempdir()
path_lib = temp.relpath("deploy_lib.tar")
lib.export_library(path_lib)
with open(temp.relpath("deploy_graph.json"), "w") as fo:
    fo.write(graph)
with open(temp.relpath("deploy_param.params"), "wb") as fo:
    fo.write(relay.save_param_dict(params))
print(temp.listdir())

Out:

['deploy_lib.tar', 'deploy_param.params', 'deploy_graph.json']
# load the module back.
loaded_json = open(temp.relpath("deploy_graph.json")).read()
loaded_lib = tvm.runtime.load_module(path_lib)
loaded_params = bytearray(open(temp.relpath("deploy_param.params"), "rb").read())
input_data = tvm.nd.array(np.random.uniform(size=data_shape).astype("float32"))

module = graph_runtime.create(loaded_json, loaded_lib, ctx)
module.load_params(loaded_params)
module.run(data=input_data)
out_deploy = module.get_output(0).asnumpy()

# Print first 10 elements of output
print(out_deploy.flatten()[0:10])

# check whether the output from deployed module is consistent with original one
tvm.testing.assert_allclose(out_deploy, out, atol=1e-3)

Out:

[0.00090713 0.00105705 0.00094459 0.00103146 0.00110017 0.00105846
 0.00104143 0.00095862 0.0010827  0.00111618]

Gallery generated by Sphinx-Gallery