def benchmark(args):
print('Batch size: {}'.format(args.batch_size))
mf = ModelDownloader()
init_net, pred_net, value_info = mf.get_c2_model(args.model)
input_shapes = {
k: [args.batch_size] + v[-1][1:]
for (k, v) in value_info.items()
}
print("input info: {}".format(input_shapes))
external_inputs = {}
for k, v in input_shapes.items():
external_inputs[k] = np.random.randn(*v).astype(np.float32)
if args.device == 'CPU':
device_option = core.DeviceOption(caffe2_pb2.CPU)
elif args.device == 'MKL':
device_option = core.DeviceOption(caffe2_pb2.MKLDNN)
elif args.device == 'IDEEP':
device_option = core.DeviceOption(caffe2_pb2.IDEEP)
else:
raise Exception("Unknown device: {}".format(args.device))
print("Device option: {}, {}".format(args.device, device_option))
pred_net.device_option.CopyFrom(device_option)
for op in pred_net.op:
op.device_option.CopyFrom(device_option)
# Hack to initialized weights into MKL/IDEEP context
workspace.RunNetOnce(init_net)
bb = workspace.Blobs()
weights = {}
for b in bb:
weights[b] = workspace.FetchBlob(b)
for k, v in external_inputs.items():
weights[k] = v
workspace.ResetWorkspace()
with core.DeviceScope(device_option):
for name, blob in weights.items():
#print("{}".format(name))
workspace.FeedBlob(name, blob, device_option)
workspace.CreateNet(pred_net)
start = time.time()
res = workspace.BenchmarkNet(pred_net.name, args.warmup_iterations,
args.iterations,
args.layer_wise_benchmark)
print("FPS: {:.2f}".format(1 / res[0] * 1000 * args.batch_size))
def benchmark(args):
print('Batch size: {}'.format(args.batch_size))
mf = ModelDownloader()
init_net, pred_net, value_info = mf.get_c2_model(args.model)
input_shapes = {k : [args.batch_size] + v[-1][1:] for (k, v) in value_info.items()}
print("input info: {}".format(input_shapes))
external_inputs = {}
for k, v in input_shapes.items():
external_inputs[k] = np.random.randn(*v).astype(np.float32)
if args.device == 'CPU':
device_option = core.DeviceOption(caffe2_pb2.CPU)
elif args.device == 'MKL':
device_option = core.DeviceOption(caffe2_pb2.MKLDNN)
elif args.device == 'IDEEP':
device_option = core.DeviceOption(caffe2_pb2.IDEEP)
else:
raise Exception("Unknown device: {}".format(args.device))
print("Device option: {}, {}".format(args.device, device_option))
pred_net.device_option.CopyFrom(device_option)
for op in pred_net.op:
op.device_option.CopyFrom(device_option)
# Hack to initialized weights into MKL/IDEEP context
workspace.RunNetOnce(init_net)
bb = workspace.Blobs()
weights = {}
for b in bb:
weights[b] = workspace.FetchBlob(b)
for k, v in external_inputs.items():
weights[k] = v
workspace.ResetWorkspace()
with core.DeviceScope(device_option):
for name, blob in weights.items():
#print("{}".format(name))
workspace.FeedBlob(name, blob, device_option)
workspace.CreateNet(pred_net)
start = time.time()
res = workspace.BenchmarkNet(pred_net.name,
args.warmup_iterations,
args.iterations,
args.layer_wise_benchmark)
print("FPS: {:.2f}".format(1/res[0]*1000*args.batch_size))
# for cpu
conda install pytorch-nightly-cpu -c pytorch
# for gpu with CUDA 8
conda install pytorch-nightly cuda80 -c pytorch
or please refer to official site
https://caffe2.ai/docs/getting-started.html
"""
######################################################################
# Load pretrained Caffe2 model
# ----------------------------
# We load a pretrained resnet50 classification model provided by Caffe2.
from caffe2.python.models.download import ModelDownloader
mf = ModelDownloader()
class Model:
def __init__(self, model_name):
self.init_net, self.predict_net, self.value_info = mf.get_c2_model(
model_name)
resnet50 = Model("resnet50")
######################################################################
# Load a test image
# ------------------
# A single cat dominates the examples!
from tvm.contrib.download import download_testdata
def setUp(self):
self.model_downloader = ModelDownloader()
class TensorRTTransformTest(TestCase):
def setUp(self):
self.model_downloader = ModelDownloader()
def _add_head_tail(self, pred_net, new_head, new_tail):
orig_head = pred_net.external_input[0]
orig_tail = pred_net.external_output[0]
# Add head
head = caffe2_pb2.OperatorDef()
head.type = "Copy"
head.input.append(new_head)
head.output.append(orig_head)
dummy = caffe2_pb2.NetDef()
dummy.op.extend(pred_net.op)
del pred_net.op[:]
pred_net.op.extend([head])
pred_net.op.extend(dummy.op)
pred_net.external_input[0] = new_head
# Add tail
tail = caffe2_pb2.OperatorDef()
tail.type = "Copy"
tail.input.append(orig_tail)
tail.output.append(new_tail)
pred_net.op.extend([tail])
pred_net.external_output[0] = new_tail
@unittest.skipIf(not workspace.C.use_trt, "No TensortRT support")
def test_resnet50_core(self):
N = 2
warmup = 20
repeat = 100
print("Batch size: {}, repeat inference {} times, warmup {} times".format(N, repeat, warmup))
init_net, pred_net, _ = self.model_downloader.get_c2_model('resnet50')
self._add_head_tail(pred_net, 'real_data', 'real_softmax')
input_blob_dims = (N, 3, 224, 224)
input_name = "real_data"
device_option = core.DeviceOption(caffe2_pb2.CUDA, 0)
init_net.device_option.CopyFrom(device_option)
pred_net.device_option.CopyFrom(device_option)
for op in pred_net.op:
op.device_option.CopyFrom(device_option)
op.engine = 'CUDNN'
net_outputs = pred_net.external_output
Y_c2 = None
data = np.random.randn(*input_blob_dims).astype(np.float32)
c2_time = 1
workspace.SwitchWorkspace("gpu_test", True)
with core.DeviceScope(device_option):
workspace.FeedBlob(input_name, data)
workspace.RunNetOnce(init_net)
workspace.CreateNet(pred_net)
for _ in range(warmup):
workspace.RunNet(pred_net.name)
start = time.time()
for _ in range(repeat):
workspace.RunNet(pred_net.name)
end = time.time()
c2_time = end - start
output_values = [workspace.FetchBlob(name) for name in net_outputs]
Y_c2 = namedtupledict('Outputs', net_outputs)(*output_values)
workspace.ResetWorkspace()
# Fill the workspace with the weights
with core.DeviceScope(device_option):
workspace.RunNetOnce(init_net)
# Cut the graph
start = time.time()
pred_net_cut = transform_caffe2_net(pred_net,
{input_name: input_blob_dims},
build_serializable_op=False)
del init_net, pred_net
pred_net_cut.device_option.CopyFrom(device_option)
for op in pred_net_cut.op:
op.device_option.CopyFrom(device_option)
#_print_net(pred_net_cut)
Y_trt = None
input_name = pred_net_cut.external_input[0]
print("C2 runtime: {}s".format(c2_time))
with core.DeviceScope(device_option):
workspace.FeedBlob(input_name, data)
workspace.CreateNet(pred_net_cut)
end = time.time()
print("Conversion time: {:.2f}s".format(end -start))
for _ in range(warmup):
workspace.RunNet(pred_net_cut.name)
start = time.time()
for _ in range(repeat):
workspace.RunNet(pred_net_cut.name)
end = time.time()
trt_time = end - start
print("TRT runtime: {}s, improvement: {}%".format(trt_time, (c2_time-trt_time)/c2_time*100))
output_values = [workspace.FetchBlob(name) for name in net_outputs]
Y_trt = namedtupledict('Outputs', net_outputs)(*output_values)
np.testing.assert_allclose(Y_c2, Y_trt, rtol=1e-3)
def setUp(self):
self.model_downloader = ModelDownloader('ONNX_MODELS')
class TestCaffe2End2End(TestCase):
def setUp(self):
self.model_downloader = ModelDownloader('ONNX_MODELS')
def _test_net(self, net_name, input_blob_dims=(1, 3, 224, 224), decimal=7):
np.random.seed(seed=0)
try:
c2_init_net, c2_predict_net, value_info, debug_str = self.model_downloader.get_c2_model_dbg(
net_name)
except (OSError, IOError) as e:
# catch IOError/OSError that is caused by FileNotFoundError and PermissionError
# This is helpful because sometimes we get errors due to gfs not available
print("\n_test_net exception: ", e)
self.skipTest(str(e))
# start to run the model and compare outputs
n, c, h, w = input_blob_dims
data = np.random.randn(n, c, h, w).astype(np.float32)
inputs = [data]
_, c2_outputs = c2_native_run_net(c2_init_net, c2_predict_net, inputs,
debug_str)
del _
model = c2_onnx.caffe2_net_to_onnx_model(
predict_net=c2_predict_net,
init_net=c2_init_net,
value_info=value_info,
)
c2_ir = c2.prepare(model)
onnx_outputs = c2_ir.run(inputs)
self.assertSameOutputs(c2_outputs, onnx_outputs, decimal=decimal)
def test_alexnet(self):
self._test_net('bvlc_alexnet', decimal=4)
def test_resnet50(self):
self._test_net('resnet50')
@unittest.skipIf(os.environ.get('JENKINS_URL'),
'Taking too long to download!')
def test_vgg16(self):
self._test_net('vgg16')
@unittest.skipIf(os.environ.get('JENKINS_URL'),
'Taking too long to download!')
def test_zfnet(self):
self._test_net('zfnet')
def test_inception_v1(self):
self._test_net('inception_v1', decimal=2)
def test_inception_v2(self):
self._test_net('inception_v2')
def test_squeezenet(self):
self._test_net('squeezenet')
def test_densenet121(self):
self._test_net('densenet121')
def test_bvlc_googlenet(self):
self._test_net('bvlc_googlenet')
def test_bvlc_reference_caffenet(self):
self._test_net('bvlc_reference_caffenet')
def test_bvlc_reference_rcnn_ilsvrc13(self):
self._test_net('bvlc_reference_rcnn_ilsvrc13')
