Day #229 - Running Pytorch Model in OpenVino

Step 1 - ONNX Pre-Requisites Install

Step 2 - Save Pytorch Model in ONNX Format

	from torchreid import models
	print(models.show_avai_models())
	from torch.autograd import Variable
	import torch.onnx
	import torchvision
	#H-256
	#W-128
	dummy_input = torch.randn(1,3,256,128)
	#dummy_input = torch.randn(1,3,128,256)
	model = models.build_model('resnet50',751,loss='softmax')
	#model = models.resnet50mid(pretrained=True)
	torch.onnx.export(model,dummy_input,"resnet50mid.onnx")

view raw pytorch_onnx.py hosted with ❤ by GitHub

#Customize and Run this model

Step 3 - Goto Model Optimizer Directory

sudo python3 mo.py --input_model /home/ubuntu/code/resnet51mid.onnx 

<code>python3 mo.py --input_model <INPUT_MODEL>.onnx</code>

This will generate the required xml to be run with OpenVino Model

Step 4 - Custom Model Training

	import torchreid
	datamanager = torchreid.data.ImageDataManager(
	root='reid-data',
	sources='market1501',
	height=160,
	width=64,
	batch_size=32,
	market1501_500k=False)
	model = torchreid.models.build_model(
	name = 'resnet50',
	num_classes = datamanager.num_train_pids,
	loss='softmax',
	pretrained=True)
	optimizer = torchreid.optim.build_optimizer(
	model,
	optim='adam',
	lr=0.0003)
	scheduler = torchreid.optim.build_lr_scheduler(
	optimizer,
	lr_scheduler='single_step',
	stepsize=20)
	engine = torchreid.engine.ImageSoftmaxEngine(datamanager,
	model,
	optimizer=optimizer,
	scheduler=scheduler,
	label_smooth=True)
	engine.run(save_dir='log/resnet50',
	max_epoch=3,
	eval_freq=10,
	print_freq=10,
	test_only=False)

view raw Reid_Train.py hosted with ❤ by GitHub

Step 5 - Custom Model Export

	from torchreid import models
	print(models.show_avai_models())
	from torch.autograd import Variable
	import torch.onnx
	import torchvision
	import torchreid
	datamanager = torchreid.data.ImageDataManager(
	root='reid-data',
	sources='market1501',
	height=160,
	width=64,
	batch_size=32,
	market1501_500k=False)
	#import mymodel
	#H - 164
	#W - 64
	dummy_input = torch.randn(1,3,164,64)
	model = torchreid.models.build_model(
	name = 'resnet50',
	num_classes = datamanager.num_train_pids,
	loss='softmax',
	pretrained=False)
	state_dict = torch.load('mymodel.pth')
	model.load_state_dict(state_dict,strict=False)
	torch.onnx.export(model,dummy_input,"resnet51mid.onnx")

view raw Re_idonnx.py hosted with ❤ by GitHub

./pedestrian_tracker_demo -i /home/ubuntu/code/smarthub_915am_cut.mp4.mp4 -m_det /opt/intel/computer_vision_sdk/deployment_tools/intel_models/person-detection-retail-0013/FP32/person-detection-retail-0013.xml -m_reid /opt/intel/computer_vision_sdk/deployment_tools/model_optimizer/resnet51mid.xml -d_det CPU

Happy Mastering DL!!!