# Examples

## 1. Compile a Pytorch Resnet Model

```python
import torchvision
from torchvision import datasets, models
import torchvision.transforms as transforms
import torch

from onnc.bench import login, Project

# Acquire the pretrained Resnet18 Model from TorchVision model zoo
# To fit the cifar10 dataset, we need to modify the fullly connected layer
model = models.resnet18(pretrained=False)
numft = model.fc.in_features
model.fc = torch.nn.Sequential(torch.nn.Linear(numft, 1000),
                               torch.nn.ReLU(),
                               torch.nn.Linear(1000, 10))


# Define preprocess steps
transform = transforms.Compose([
    transforms.Pad(4),
    transforms.RandomHorizontalFlip(),
    transforms.RandomCrop(32),
    transforms.ToTensor()])

# Acquire ImageNet dataset from TorchVision Datasets
test_dataset = torchvision.datasets.CIFAR10(root='./cifar10',
                                            train=False,
                                            download=True,
                                            transform=transform)
# onnc-bench also accept a DataLoader object
data_loader = torch.utils.data.DataLoader(test_dataset)

# Code for training the model
# ...

# Set the model to evaluation mode
model.eval()

# Set up for compilation
login("your_account", "your_password")

project = Project(name='experiment')
project.add_model(model=model,  # feed a Pytorch model object
                  samples=test_dataset,  # feed a Pytorch Dataset object
                  #samples=data_loader  # or you can feed a DataLoader instead
                  model_inputs=[["input", (1, 3, 32, 32), float]]
                  )
project.compile(target='NVDLA-NV-SMALL-DEFAULT')

deployment = project.save('./output')
print(deployment.compiled_files)
```
**Note: Due to the dynamic graph nature of Pytorch, input signatures is required.**


## 2. Compile a Keras Resnet model

```python
import keras
import tensorflow as tf
import numpy as np

from onnc.bench import login, Project

# (x_train, y_train), (x_test, y_test) = tf.keras.datasets.imagenet.load_data()
# Imagenet is not public available now.
# Please download and preprocess Imagenet dataset from:
# https://image-net.org/challenges/LSVRC/2012/

# Here we use random data to demostrate the procedure
x_test = np.random.rand(100, 224, 224, 3)

model = tf.keras.applications.resnet50.ResNet50(weights="imagenet")


# Set up for compilation
login("your_account", "your_password")

project = Project(name='experiment')
project.add_model(model=model,  # feed a Keras model object
                  samples=x_test,  # feed a Keras/Numpy Dataset object
                 )
project.compile(target='NVDLA-NV-SMALL-DEFAULT')

deployment = project.save('./output')
print(deployment.compiled_files)

```

## 3. Compile a ONNX Model Zoo model

Please download onnx model and dataset from [ONNX Model Zoo](https://github.com/onnx/models#image_classification)

Currently ONNC supports below models:

+ bvlc_alexnetbvlc_googlenet
+ bvlc_reference_caffenet
+ bvlc_reference_rcnn_ilsvrc13
+ densenet121
+ inception_v1
+ inception_v2
+ resnet50
+ shufflenet
+ squeezenet
+ vgg19
+ zfnet512

**Note: Please make sure the Opset version of the ONNX file you download is `8`**

To run below example, please download and untar Resnet50 model and calibration samples from [here](https://github.com/onnx/models/blob/master/vision/classification/resnet/model/resnet50-caffe2-v1-8.tar.gz).


```python

from onnc.bench import login, Project

# Set up for compilation
login("your_account", "your_password")

project = Project(name='experiment')
project.add_model(model='resnet50/model.onnx',  # feed a onnx file
                  samples='resnet50/test_data_set_0/input_0.pb',  # calibration samples
                 )
project.compile(target='NVDLA-NV-SMALL-DEFAULT')

deployment = project.save('./output')
print(deployment.compiled_files)

```