mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-06 12:20:52 +01:00
812bc1dde6
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/62058 This is the second diff in this stack. This diff includes the changes to DPER3; the first diff includes the changes to Caffe2. We want to decay learning parameters properly. Previously this was not done when a parameter is absent from a minibatch. We fix this by keeping track of missed minibatches and making decay catch up accordingly. The exponential moving averages (EMA) for the first and second moments used in Adam are updated only for parameters seen in a minibatch. Actually, for these parameters, 0 should be added to the EMAs and the EMAs should then be decayed by multiplying by beta1 and beta2 respectively. To avoid the computational overhead of touching every parameter for every minibatch, we: * keep track of the last time a parameter is seen * instead of decaying the EMAs by multiplying by beta1 and beta2, we multiply by beta1^k and beta2^k, where k is the number of minibatches since the parameter was last seen. We hope this will significantly improve the inconsistent learning parameter issue we have seen with Adam. Differential Revision: D29638897 fbshipit-source-id: 18d8e227d72c2e23010ca81e0f6eeb78872c8d3c
758 lines
30 KiB
Python
758 lines
30 KiB
Python
|
|
|
|
|
|
from caffe2.proto import caffe2_pb2
|
|
import caffe2.python.optimizer as optimizer
|
|
from caffe2.python.optimizer import (
|
|
build_sgd, build_multi_precision_sgd, build_ftrl, build_gftrl, build_wngrad,
|
|
build_adagrad, build_adadelta, build_adam, build_yellowfin, build_rms_prop,
|
|
build_storm, build_decay_adagrad, add_weight_decay, SgdOptimizer)
|
|
from caffe2.python.optimizer_context import UseOptimizer
|
|
from caffe2.python.optimizer_test_util import (
|
|
OptimizerTestBase, LRModificationTestBase
|
|
)
|
|
from caffe2.python import core, workspace
|
|
from caffe2.python.test_util import TestCase
|
|
import numpy as np
|
|
from numpy.testing import assert_allclose, assert_equal
|
|
import math
|
|
import unittest
|
|
|
|
|
|
class TestLars(OptimizerTestBase, TestCase):
|
|
def testSparse(self):
|
|
raise unittest.SkipTest("no sparse support")
|
|
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_sgd(model, base_learning_rate=0.1, lars=0.5, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
tensor = workspace.FetchBlob(param)
|
|
np.testing.assert_allclose(np.array([1.0]), tensor, atol=1e-5)
|
|
|
|
|
|
class TestMomentumSgd(OptimizerTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_sgd(model, base_learning_rate=0.1, momentum=0.1, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
tensor = workspace.FetchBlob(param)
|
|
np.testing.assert_allclose(np.array([1.0]), tensor, atol=1e-5)
|
|
|
|
|
|
class TestSgd(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_sgd(model, base_learning_rate=0.1, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
tensor = workspace.FetchBlob(param)
|
|
np.testing.assert_allclose(np.array([1.0]), tensor, atol=1e-5)
|
|
|
|
|
|
class TestMultiPrecisionSgd(
|
|
OptimizerTestBase, LRModificationTestBase, TestCase
|
|
):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_multi_precision_sgd(
|
|
model, base_learning_rate=0.1, **kwargs
|
|
)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
tensor = workspace.FetchBlob(param)
|
|
np.testing.assert_allclose(np.array([1.0]), tensor, atol=1e-5)
|
|
|
|
@unittest.skipIf(not workspace.has_gpu_support, "No GPU support")
|
|
def testGPUDense(self):
|
|
super(TestMultiPrecisionSgd, self).testGPUDense(core.DataType.FLOAT16)
|
|
|
|
|
|
class TestFtrl(OptimizerTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_ftrl(
|
|
model,
|
|
engine=None,
|
|
alpha=1.0,
|
|
beta=0.1,
|
|
lambda1=0.0,
|
|
lambda2=0.0,
|
|
**kwargs
|
|
)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestGFtrl(OptimizerTestBase, TestCase):
|
|
def testSparse(self):
|
|
raise unittest.SkipTest("no sparse support")
|
|
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_gftrl(
|
|
model,
|
|
engine=None,
|
|
alpha=1.0,
|
|
beta=0.1,
|
|
lambda1=0.0,
|
|
lambda2=0.0,
|
|
**kwargs
|
|
)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestAdagrad(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_adagrad(model, base_learning_rate=1.0, lars=0.5, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestRowWiseAdagrad(OptimizerTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_adagrad(
|
|
model, base_learning_rate=1.0, lars=0.5, rowWise=True, **kwargs
|
|
)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
def testDense(self):
|
|
raise unittest.SkipTest("no dense support")
|
|
|
|
def testGPUDense(self):
|
|
raise unittest.SkipTest("no dense support")
|
|
|
|
class TestRowWiseAdagradWithCounter(OptimizerTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_adagrad(
|
|
model,
|
|
base_learning_rate=1.0,
|
|
lars=0.5,
|
|
rowWise=True,
|
|
counter_halflife=5,
|
|
**kwargs
|
|
)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
self.assertTrue(workspace.HasBlob("optimizer_iteration"))
|
|
iteration_tensor = workspace.FetchBlob("optimizer_iteration")
|
|
np.testing.assert_allclose(np.array([2000]),
|
|
iteration_tensor,
|
|
atol=1e-5)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
workspace.FetchBlob(param)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
def testDense(self):
|
|
raise unittest.SkipTest("no dense support")
|
|
|
|
def testGPUDense(self):
|
|
raise unittest.SkipTest("no dense support")
|
|
|
|
class TestWngrad(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_wngrad(model, base_learning_rate=25.0, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestStorm(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_storm(model, base_learning_rate=2.0, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestAdadelta(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_adadelta(model, base_learning_rate=1.0, decay=0.995, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestAdam(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_adam(model, base_learning_rate=0.1, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
self.assertTrue(workspace.HasBlob("optimizer_iteration"))
|
|
iteration_tensor = workspace.FetchBlob("optimizer_iteration")
|
|
np.testing.assert_allclose(np.array([2000]),
|
|
iteration_tensor,
|
|
atol=1e-5)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
workspace.FetchBlob(param)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
class TestSmartDecayAdam(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
kwargs['beta1'] = 0.0
|
|
return build_adam(model, base_learning_rate=0.1, use_smart_decay=True, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
self.assertTrue(workspace.HasBlob("optimizer_iteration"))
|
|
blob_names = workspace.Blobs()
|
|
self.assertTrue(any((bn.endswith('_last_seen') for bn in blob_names)))
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
workspace.FetchBlob(param)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
class TestDecayAdagrad(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_decay_adagrad(model, base_learning_rate=1.0, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
self.assertTrue(workspace.HasBlob("optimizer_iteration"))
|
|
iteration_tensor = workspace.FetchBlob("optimizer_iteration")
|
|
np.testing.assert_allclose(np.array([2000]),
|
|
iteration_tensor,
|
|
atol=1e-5)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
workspace.FetchBlob(param)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
def testSparse(self):
|
|
raise unittest.SkipTest("no sparse support")
|
|
|
|
class TestSparseRAdam(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = True
|
|
return build_adam(model, base_learning_rate=0.1, enableRAdam=True, **kwargs)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
self.assertTrue(workspace.HasBlob("optimizer_iteration"))
|
|
iteration_tensor = workspace.FetchBlob("optimizer_iteration")
|
|
np.testing.assert_allclose(np.array([2000]),
|
|
iteration_tensor,
|
|
atol=1e-5)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
workspace.FetchBlob(param)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
|
|
class TestYellowFin(OptimizerTestBase, TestCase):
|
|
# YellowFin: An automatic tuner for momentum SGD
|
|
# (https://arxiv.org/abs/1706.03471)
|
|
def build_optimizer(self, model):
|
|
self._skip_gpu = False
|
|
return build_yellowfin(model, base_learning_rate=0.1)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
self.assertTrue(workspace.HasBlob("optimizer_iteration"))
|
|
iteration_tensor = workspace.FetchBlob("optimizer_iteration")
|
|
np.testing.assert_allclose(np.array([2000]),
|
|
iteration_tensor,
|
|
atol=1e-5)
|
|
for param in optimizer.get_auxiliary_parameters().shared:
|
|
workspace.FetchBlob(param)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
def testSparse(self):
|
|
raise unittest.SkipTest("no sparse support")
|
|
|
|
def deb(self, val, beta, i, zero_debias):
|
|
if zero_debias:
|
|
return val / (1.0 - beta ** i)
|
|
else:
|
|
return val
|
|
|
|
def get_lr_mu(self, distance, grad_var, h_min, h_max):
|
|
# First tune based on dynamic range
|
|
if grad_var == 0:
|
|
dr = h_max / h_min
|
|
mu = ((np.sqrt(dr) - 1) / (np.sqrt(dr) + 1)) ** 2
|
|
lr_min = (1 + np.sqrt(mu)) ** 2 / h_max
|
|
return lr_min, mu
|
|
|
|
p = distance ** 2 * h_min ** 2 / 2 / grad_var
|
|
w3 = (-math.sqrt(p * p + 4.0 / 27.0 * p * p * p) - p) / 2.0
|
|
w = (1.0 if w3 > 0.0 else -1.0) * math.pow(math.fabs(w3), 1.0 / 3.0)
|
|
y = w - p / 3.0 / w
|
|
root = y + 1
|
|
root = min(root, 1.0 - 1e-6)
|
|
dr = h_max / h_min
|
|
mu = max(((np.sqrt(dr) - 1) / (np.sqrt(dr) + 1)) ** 2, root**2)
|
|
lr_min = (1 - np.sqrt(mu)) ** 2 / h_min
|
|
return lr_min, mu
|
|
|
|
def caffe2_yellowfin(self, zero_debias, grad_coef, n_dim, n_iter, gpu):
|
|
caffe2_res = {}
|
|
|
|
alpha = 1.0
|
|
mu = 0.0
|
|
beta = 0.999
|
|
curv_win_width = 20
|
|
epsilon = 1e-6
|
|
|
|
net = core.Net("net")
|
|
param_init_net = core.Net("param_init_net")
|
|
workspace.ResetWorkspace()
|
|
|
|
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU)):
|
|
iteration = param_init_net.ConstantFill(
|
|
[],
|
|
"iteration",
|
|
shape=[1],
|
|
value=0,
|
|
dtype=core.DataType.INT64)
|
|
iter_mutex = param_init_net.CreateMutex([], ["iteration_mutex"])
|
|
net.AtomicIter([iter_mutex, iteration], [iteration])
|
|
pre_grad = param_init_net.ConstantFill(
|
|
[],
|
|
"pre_grad",
|
|
shape=[n_dim],
|
|
value=grad_coef
|
|
)
|
|
if gpu:
|
|
iteration = net.CopyCPUToGPU(
|
|
[iteration],
|
|
"iteration_cpu"
|
|
)
|
|
iteration_float = net.Cast([iteration], "iteration_float")
|
|
grad = net.Mul([pre_grad, iteration_float], "grad", broadcast=True)
|
|
w = param_init_net.ConstantFill([], "w", shape=[n_dim], value=0.0)
|
|
|
|
# a hack to create an object with __dict__
|
|
param_info = lambda: None
|
|
param_info.blob = w
|
|
param_info.grad = grad
|
|
|
|
optimizer.YellowFinOptimizer(
|
|
alpha=alpha,
|
|
mu=mu,
|
|
beta=beta,
|
|
curv_win_width=curv_win_width,
|
|
epsilon=epsilon,
|
|
zero_debias=zero_debias
|
|
)._run(
|
|
net,
|
|
param_init_net,
|
|
param_info
|
|
)
|
|
|
|
workspace.RunNetOnce(param_init_net)
|
|
workspace.CreateNet(net, overwrite=True)
|
|
for i in range(n_iter):
|
|
workspace.RunNet(net)
|
|
scalars_memory_blob = workspace.FetchBlob("w_scalars_memory")
|
|
g_norm2_avg = scalars_memory_blob[1]
|
|
g_norm2_min_avg = scalars_memory_blob[2]
|
|
g_norm2_max_avg = scalars_memory_blob[3]
|
|
distance_avg = scalars_memory_blob[4]
|
|
g_avg_blob = workspace.FetchBlob("w_g_avg")
|
|
res_lr = workspace.FetchBlob("w_lr_avg")[0]
|
|
res_mu = workspace.FetchBlob("w_mu_avg")[0]
|
|
g_deb = self.deb(g_avg_blob, beta, i + 1, zero_debias)
|
|
variance = max(
|
|
self.deb(g_norm2_avg, beta, i + 1, zero_debias) -
|
|
g_deb.dot(g_deb),
|
|
epsilon
|
|
)
|
|
if i > 0:
|
|
caffe2_res[i] = {
|
|
'h_max': np.exp(self.deb(g_norm2_max_avg,
|
|
beta,
|
|
i + 1,
|
|
zero_debias)),
|
|
'h_min': np.exp(self.deb(g_norm2_min_avg,
|
|
beta,
|
|
i + 1,
|
|
zero_debias)),
|
|
'var': variance,
|
|
'dist': self.deb(distance_avg, beta, i + 1, zero_debias),
|
|
'lr': res_lr,
|
|
'mu': res_mu
|
|
}
|
|
return caffe2_res
|
|
|
|
def numpy_yellowfin(self, zero_debias, grad_coef, n_dim, n_iter, gpu):
|
|
numpy_res = {}
|
|
|
|
target_h_max = 0.0
|
|
target_h_min = 0.0
|
|
target_g_norm_squared_avg = 0.0
|
|
target_g_norm_avg = 0.0
|
|
target_g_avg = 0.0
|
|
target_dist_avg = 0.0
|
|
target_lr = 1.0
|
|
target_mu = 0.0
|
|
|
|
for i in range(n_iter):
|
|
grad_val = (i + 1) * grad_coef
|
|
target_g_norm_squared_avg = 0.999 * target_g_norm_squared_avg + \
|
|
0.001 * np.sum((grad_val * np.ones([n_dim, ])) ** 2)
|
|
target_g_norm_avg = 0.999 * target_g_norm_avg + \
|
|
0.001 * np.linalg.norm(grad_val * np.ones([n_dim, ]))
|
|
target_g_avg = 0.999 * target_g_avg + 0.001 * grad_val
|
|
|
|
target_h_max = 0.999 * target_h_max + \
|
|
0.001 * np.log(grad_val ** 2 * n_dim)
|
|
target_h_min = 0.999 * target_h_min + \
|
|
0.001 * np.log((max(1, i + 2 - 20) * grad_coef) ** 2 * n_dim)
|
|
if zero_debias:
|
|
target_var = target_g_norm_squared_avg / \
|
|
(1 - 0.999 ** (i + 1)) - \
|
|
target_g_avg ** 2 * n_dim / (1 - 0.999 ** (i + 1)) ** 2
|
|
else:
|
|
target_var = target_g_norm_squared_avg - \
|
|
target_g_avg ** 2 * n_dim
|
|
target_dist_avg = 0.999 * target_dist_avg + \
|
|
0.001 * target_g_norm_avg / target_g_norm_squared_avg
|
|
|
|
if i > 0:
|
|
if zero_debias:
|
|
lr, mu = self.get_lr_mu(
|
|
target_dist_avg / (1.0 - 0.999 ** (i + 1)),
|
|
target_var,
|
|
np.exp(target_h_min / (1.0 - 0.999 ** (i + 1))),
|
|
np.exp(target_h_max / (1.0 - 0.999 ** (i + 1))))
|
|
target_lr = 0.999 * target_lr + 0.001 * lr
|
|
target_mu = 0.999 * target_mu + 0.001 * mu
|
|
numpy_res[i] = {
|
|
'h_max': np.exp(target_h_max / (1 - 0.999 ** (i + 1))),
|
|
'h_min': np.exp(target_h_min / (1 - 0.999 ** (i + 1))),
|
|
'var': target_var,
|
|
'dist': target_dist_avg / (1 - 0.999 ** (i + 1)),
|
|
'lr': target_lr,
|
|
'mu': target_mu
|
|
}
|
|
else:
|
|
lr, mu = self.get_lr_mu(
|
|
target_dist_avg,
|
|
target_var,
|
|
np.exp(target_h_min),
|
|
np.exp(target_h_max))
|
|
target_lr = 0.999 * target_lr + 0.001 * lr
|
|
target_mu = 0.999 * target_mu + 0.001 * mu
|
|
numpy_res[i] = {
|
|
'h_max': np.exp(target_h_max),
|
|
'h_min': np.exp(target_h_min),
|
|
'var': target_var,
|
|
'dist': target_dist_avg,
|
|
'lr': target_lr,
|
|
'mu': target_mu
|
|
}
|
|
return numpy_res
|
|
|
|
def compare_yellowfin_models(self,
|
|
model0,
|
|
model1,
|
|
zero_debias,
|
|
grad_coef,
|
|
n_dim,
|
|
n_iter,
|
|
gpu):
|
|
model0_res = model0(zero_debias, grad_coef, n_dim, n_iter, gpu)
|
|
model1_res = model1(zero_debias, grad_coef, n_dim, n_iter, gpu)
|
|
assert_equal(len(model0_res), len(model1_res))
|
|
for i in range(1, len(model0_res)):
|
|
assert_equal(model0_res[i].keys(), model1_res[i].keys())
|
|
for feat in model0_res[i].keys():
|
|
err_msg = \
|
|
'i=' + str(i) + ',\n' + \
|
|
'feat=' + feat + ',\n' + \
|
|
'grad_coef=' + str(grad_coef) + ',\n' + \
|
|
'zero_debias=' + str(zero_debias)
|
|
assert_allclose(model0_res[i][feat],
|
|
model1_res[i][feat],
|
|
rtol=1e-2,
|
|
err_msg=err_msg)
|
|
|
|
@unittest.skip("Results might vary too much. Only for individual use.")
|
|
def test_caffe2_cpu_vs_numpy(self):
|
|
n_dim = 1000000
|
|
n_iter = 50
|
|
cpu_device_opt = core.DeviceOption(caffe2_pb2.CPU)
|
|
with core.DeviceScope(cpu_device_opt):
|
|
for zero_debias, grad_coef in [
|
|
(False, 1.0),
|
|
(False, 0.1),
|
|
(False, 0.01),
|
|
(True, 1.0)
|
|
]:
|
|
self.compare_yellowfin_models(
|
|
self.caffe2_yellowfin,
|
|
self.numpy_yellowfin,
|
|
zero_debias,
|
|
grad_coef,
|
|
n_dim,
|
|
n_iter,
|
|
gpu=False
|
|
)
|
|
|
|
@unittest.skip("Results might vary too much. Only for individual use.")
|
|
@unittest.skipIf(not workspace.has_gpu_support, "No gpu support")
|
|
def test_caffe2_gpu_vs_numpy(self):
|
|
n_dim = 1000000
|
|
n_iter = 50
|
|
gpu_device_opt = core.DeviceOption(workspace.GpuDeviceType, 0)
|
|
with core.DeviceScope(gpu_device_opt):
|
|
for zero_debias in [False, True]:
|
|
for grad_coef in [1.0, 0.1, 0.01]:
|
|
self.compare_yellowfin_models(
|
|
self.caffe2_yellowfin,
|
|
self.numpy_yellowfin,
|
|
zero_debias,
|
|
grad_coef,
|
|
n_dim,
|
|
n_iter,
|
|
gpu=True
|
|
)
|
|
|
|
|
|
class TestRmsProp(OptimizerTestBase, LRModificationTestBase, TestCase):
|
|
def build_optimizer(self, model, **kwargs):
|
|
self._skip_gpu = False
|
|
return build_rms_prop(
|
|
model, base_learning_rate=0.1, epsilon=0.1, **kwargs
|
|
)
|
|
|
|
def check_optimizer(self, optimizer):
|
|
self.assertFalse(optimizer.get_auxiliary_parameters().shared)
|
|
self.assertTrue(optimizer.get_auxiliary_parameters().local)
|
|
for param in optimizer.get_auxiliary_parameters().local:
|
|
workspace.FetchBlob(param)
|
|
|
|
def testSparse(self):
|
|
raise unittest.SkipTest("no sparse support")
|
|
|
|
|
|
class TestMultiOptimizers(TestCase):
|
|
def test_multiple_optimizers(self):
|
|
from caffe2.python import brew, core, optimizer
|
|
from caffe2.python.model_helper import ModelHelper
|
|
|
|
model = ModelHelper(name="test")
|
|
fc1 = brew.fc(model, 'data', 'fc1', 100, 50)
|
|
fc2 = brew.fc(model, fc1, 'fc2', 50, 25)
|
|
pred = brew.fc(model, fc2, 'fc3', 25, 10)
|
|
(softmax, loss) = model.SoftmaxWithLoss(
|
|
[pred, 'label'],
|
|
['softmax', 'loss'],
|
|
)
|
|
model.AddGradientOperators([loss])
|
|
|
|
param_to_device = optimizer._get_param_to_device(model)
|
|
|
|
def infer_blob_device(blob_name):
|
|
return optimizer.get_param_device(
|
|
blob_name, "{}_grad".format(blob_name), param_to_device
|
|
)
|
|
|
|
sgd_1 = optimizer.SgdOptimizer(base_learning_rate=0.1)
|
|
sgd_2 = optimizer.SgdOptimizer(base_learning_rate=0.2)
|
|
adagrad = optimizer.AdagradOptimizer()
|
|
|
|
# Check same optimizer share the same learning rate.
|
|
with core.DeviceScope(infer_blob_device("fc1_w")):
|
|
sgd_1(model.net, model.param_init_net, "fc1_w", "fc1_w_grad")
|
|
with core.DeviceScope(infer_blob_device("fc1_b")):
|
|
sgd_1(model.net, model.param_init_net, "fc1_b", "fc1_b_grad")
|
|
fc1_lr_blobs = []
|
|
for op in model.net.Proto().op:
|
|
if op.type == 'WeightedSum' and op.input[0] == 'fc1_w' or \
|
|
op.input[0] == 'fc1_b':
|
|
fc1_lr_blobs.append(op.input[3])
|
|
self.assertEqual(fc1_lr_blobs[0], fc1_lr_blobs[1])
|
|
|
|
# Check different instance of the same optimizer has a different lr.
|
|
with core.DeviceScope(infer_blob_device("fc2_w")):
|
|
sgd_2(model.net, model.param_init_net, "fc2_w", "fc2_w_grad")
|
|
with core.DeviceScope(infer_blob_device("fc2_b")):
|
|
sgd_2(model.net, model.param_init_net, "fc2_b", "fc2_b_grad")
|
|
fc2_lr_blobs = []
|
|
for op in model.net.Proto().op:
|
|
if op.type == 'WeightedSum' and op.input[0] == 'fc2_w' or \
|
|
op.input[0] == 'fc2_b':
|
|
self.assertTrue(op.input[3] not in fc1_lr_blobs)
|
|
fc2_lr_blobs.append(op.input[3])
|
|
self.assertEqual(fc2_lr_blobs[0], fc2_lr_blobs[1])
|
|
|
|
# Check different optimizer type case
|
|
with core.DeviceScope(infer_blob_device("fc3_w")):
|
|
adagrad(model.net, model.param_init_net, "fc3_w", "fc3_w_grad")
|
|
with core.DeviceScope(infer_blob_device("fc3_b")):
|
|
adagrad(model.net, model.param_init_net, "fc3_b", "fc3_b_grad")
|
|
fc3_lr_blobs = []
|
|
for op in model.net.Proto().op:
|
|
if op.type == 'Adagrad' and op.input[0] == 'fc3_w' or \
|
|
op.input[0] == 'fc3_b':
|
|
self.assertTrue(op.input[3] not in fc2_lr_blobs)
|
|
self.assertTrue(op.input[3] not in fc1_lr_blobs)
|
|
fc3_lr_blobs.append(op.input[3])
|
|
self.assertEqual(fc3_lr_blobs[0], fc3_lr_blobs[1])
|
|
|
|
|
|
class TestWeightDecay(TestCase):
|
|
|
|
def test_weight_decay(self):
|
|
from caffe2.python import brew
|
|
from caffe2.python.model_helper import ModelHelper
|
|
|
|
model = ModelHelper(name="test", arg_scope={'order': 'NCHW'})
|
|
cnv = brew.conv(model, 'data', 'cnv', 32, 32, 4)
|
|
a = brew.fc(model, cnv, 'a', 100, 200)
|
|
pred = brew.fc(model, a, 'b', 200, 5)
|
|
(softmax, loss) = model.SoftmaxWithLoss(
|
|
[pred, 'label'],
|
|
['softmax', 'loss'],
|
|
)
|
|
model.AddGradientOperators([loss])
|
|
|
|
add_weight_decay(model, weight_decay=1e-4)
|
|
build_sgd(model, 0.11)
|
|
|
|
expected_weight_grad = {'b_w_grad', 'a_w_grad', 'cnv_w_grad'}
|
|
|
|
# Check the proto that all weights are decayed and not non-weights
|
|
# are decayed.
|
|
for op in model.net.Proto().op:
|
|
if op.type == 'WeightedSum' and 'wd_0_0' in op.input:
|
|
if op.output[0] not in expected_weight_grad:
|
|
print(
|
|
"Unexpected param for weight_decay: {}".
|
|
format(op.output[0])
|
|
)
|
|
self.assertTrue(op.output[0] in expected_weight_grad)
|
|
expected_weight_grad.remove(op.output[0])
|
|
|
|
self.assertEqual(
|
|
expected_weight_grad,
|
|
set(),
|
|
"Not all weights were decayed: {}".format(expected_weight_grad)
|
|
)
|
|
|
|
|
|
class TestOptimizerContext(TestCase):
|
|
|
|
def test_optimizer_context(self):
|
|
from caffe2.python import brew, optimizer
|
|
from caffe2.python.model_helper import ModelHelper
|
|
|
|
model = ModelHelper(name="test", arg_scope={'order': 'NCHW'})
|
|
count = optimizer._optimizer_instance_count['SgdOptimizer']
|
|
cnv_optim = SgdOptimizer(0.15)
|
|
weight_optim = SgdOptimizer(0.2)
|
|
bias_optim = SgdOptimizer(0.1)
|
|
|
|
with UseOptimizer(cnv_optim):
|
|
cnv = brew.conv(model, 'data', 'cnv', 32, 32, 4)
|
|
with UseOptimizer({'WEIGHT': weight_optim, 'BIAS': bias_optim}):
|
|
a = brew.fc(model, cnv, 'a', 100, 200)
|
|
pred = brew.fc(model, a, 'b', 200, 5)
|
|
(softmax, loss) = model.SoftmaxWithLoss(
|
|
[pred, 'label'],
|
|
['softmax', 'loss'],
|
|
)
|
|
model.AddGradientOperators([loss])
|
|
|
|
add_weight_decay(model, weight_decay=1e-4)
|
|
# use the following optimizer if none specified in param_info
|
|
build_sgd(model, 0.11)
|
|
expected_weight_grad = {'b_w_grad', 'a_w_grad', 'cnv_w_grad'}
|
|
expected_learning_rate = {
|
|
"SgdOptimizer_{}_lr_cpu".format(count): -0.15,
|
|
"SgdOptimizer_{}_lr_cpu".format(count + 1): -0.2,
|
|
"SgdOptimizer_{}_lr_cpu".format(count + 2): -0.1,
|
|
"SgdOptimizer_{}_lr_cpu".format(count + 3): -0.11
|
|
}
|
|
|
|
for op in model.net.Proto().op:
|
|
# Check the proto that all weights are decayed and not non-weights
|
|
# are decayed.
|
|
if op.type == 'WeightedSum' and 'wd_0_0' in op.input:
|
|
if op.output[0] not in expected_weight_grad:
|
|
print(
|
|
"Unexpected param for weight_decay: {}".
|
|
format(op.output[0])
|
|
)
|
|
self.assertTrue(op.output[0] in expected_weight_grad)
|
|
expected_weight_grad.remove(op.output[0])
|
|
# Check the learning rate for each parameter
|
|
if op.type == 'LearningRate':
|
|
val = 0
|
|
for arg in op.arg:
|
|
if arg.name == 'base_lr':
|
|
val = arg.f
|
|
self.assertAlmostEqual(
|
|
val,
|
|
expected_learning_rate[op.output[0]]
|
|
)
|
|
|
|
self.assertEqual(
|
|
expected_weight_grad,
|
|
set(),
|
|
"Not all weights were decayed: {}".format(expected_weight_grad)
|
|
)
|