preProbSquared = x * x
l1loss = torch.sum(0.01 * torch.abs(preProbSquared))
window_size = (int(x.size()[2]), int(x.size()[3]))
x = torch.nn.functional.avg_pool2d(x, kernel_size=window_size)
x = torch.squeeze(x)
probs = self.softmax(x)
logprobs = torch.log(probs)
nll = nllloss(logprobs, labels)
loss = l1loss + nll
return loss
# Set arbitrary seed so model weights are initialized to the
# same values each time the test is run
torch.manual_seed(1)
torchWriter = torchwriter.PytorchNetWriter(
inNames=inNames,
outNames=outNames,
optimizer=popart.ConstSGD(0.001),
inputShapeInfo=inputShapeInfo,
dataFlow=dataFlow,
### Torch specific:
module=Module0(),
samplesPerBatch=batchSize)
c10driver.run(torchWriter, willowOptPatterns, args.outputdir, cifarInIndices,
args.device, args.hw_id)
x = torch.nn.functional.avg_pool2d(x,
kernel_size=window_size,
stride=window_size)
preprobs = torch.squeeze(x)
probs = self.softmax(preprobs)
logprobs = torch.log(probs)
loss = nllloss(logprobs, inputs[1])
return loss
# Set arbitrary seed so model weights are initialized to the
# same values each time the test is run
torch.manual_seed(1)
torchWriter = torchwriter.PytorchNetWriter(
inNames=inNames,
outNames=outNames,
# large weight_decay term to test that it is definitely working
optimizer=popart.SGD({
"defaultLearningRate": (0.001, False),
"defaultWeightDecay": (10.0, False)
}),
inputShapeInfo=inputShapeInfo,
dataFlow=dataFlow,
### Torch specific:
module=Module0(),
samplesPerBatch=batchSize)
c10driver.run(torchWriter, None, args.outputdir, cifarInIndices, args.device,
args.hw_id)
def forward(self, inputs):
"""out = relu(conv(in))"""
image0 = inputs[0]
x = self.conv1(image0)
x = self.relu(x)
return x
# Set arbitrary seed so model weights are initialized to the
# same values each time the test is run
torch.manual_seed(1)
torchWriter = torchwriter.PytorchNetWriter(
inNames=inNames,
outNames=outNames,
optimizer=optimizer,
inputShapeInfo=inputShapeInfo,
dataFlow=dataFlow,
### Torch specific:
module=Module0(),
samplesPerBatch=samplesPerBatch)
# Passes if torch and popart models match
c10driver.run(torchWriter=torchWriter,
patterns=None,
outputdir=args.outputdir,
cifarInIndices=cifarInIndices,
device=args.device,
device_hw_id=args.hw_id,
mode="infer")
x = self.sin(image0)
x = self.conv1(x)
x = self.sin(x)
x = self.conv2(x)
window_size = (int(x.size()[2]), int(x.size()[3]))
x = torch.nn.functional.avg_pool2d(x, kernel_size=window_size)
x = torch.squeeze(x)
# This is the where the GEMM happens:
x = self.linear(x)
out = torch.sum(0.1 * torch.abs(x))
return out
# Set arbitrary seed so model weights are initialized to the
# same values each time the test is run
torch.manual_seed(1)
torchWriter = torchwriter.PytorchNetWriter(
inNames=inNames,
outNames=outNames,
# large weight_decay term to test that it is definitely working
optimizer=popart.ConstSGD(learning_rate=0.001, weight_decay=10),
inputShapeInfo=inputShapeInfo,
dataFlow=dataFlow,
### Torch specific:
module=Module0(),
samplesPerBatch=batchSize)
c10driver.run(torchWriter, None, args.outputdir, cifarInIndices, args.device,
args.hw_id)
]
class Module0(torch.nn.Module):
def __init__(self):
torch.nn.Module.__init__(self)
self.a = torch.nn.Parameter(torch.rand(2, 1, 4, 5, 6))
self.b = torch.nn.Parameter(torch.rand(3, 1, 6, 7))
self.matmul = torch.matmul
def forward(self, inputs):
return self.matmul(self.a, self.b)
# Set arbitrary seed so model weights are initialized to the
# same values each time the test is run
torch.manual_seed(1)
torchWriter = torchwriter.PytorchNetWriter(
inNames=inNames,
outNames=outNames,
losses=losses,
optimizer=popart_core.ConstSGD(0.001),
inputShapeInfo=inputShapeInfo,
dataFlow=dataFlow,
### Torch specific:
module=Module0())
c10driver.run(torchWriter, willowOptPasses, args.outputdir, cifarInIndices,
args.device, args.hw_id)
return preProbSquared, probs
# Set arbitrary seed so model weights are initialized to the
# same values each time the test is run
torch.manual_seed(1)
# Test with ConstSGD
anchors_1 = c10driver.run(
torchwriter.PytorchNetWriter(
inNames=inNames,
outNames=outNames,
losses=losses,
# default loss scaling (1.0f)
optimizer=popart.SGD({"defaultLearningRate": (0.001, True)}),
inputShapeInfo=inputShapeInfo,
dataFlow=dataFlow,
### Torch specific:
module=Module0(),
samplesPerBatch=batchSize),
willowOptPasses,
args.outputdir,
cifarInIndices,
args.device,
args.hw_id)
# Test with ConstSGD
anchors_2 = c10driver.run(
torchwriter.PytorchNetWriter(
inNames=inNames,
