mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-06 12:20:52 +01:00
ad07f5f05d
Summary: Moved code for global norm-based gradient clipping from fb specific workflows (seq2seq) to the open-source caffe2 optimizer library Reviewed By: jhcross Differential Revision: D5637453 fbshipit-source-id: 7e73c9a1c97c28a152c188467b27a6449f79242e
213 lines
8.1 KiB
Python
213 lines
8.1 KiB
Python
## @package optimizer_test_util
|
|
# Module caffe2.python.optimizer_test_util
|
|
from __future__ import absolute_import
|
|
from __future__ import division
|
|
from __future__ import print_function
|
|
from __future__ import unicode_literals
|
|
|
|
import unittest
|
|
import numpy as np
|
|
from caffe2.python import brew, core, workspace, cnn
|
|
from caffe2.proto import caffe2_pb2
|
|
from caffe2.python.modeling.initializers import (
|
|
Initializer, pFP16Initializer)
|
|
|
|
from caffe2.python.model_helper import ModelHelper
|
|
|
|
|
|
class OptimizerTestBase(object):
|
|
"""
|
|
This is an abstract base class.
|
|
Don't inherit from unittest.TestCase, and don't name it 'Test*'.
|
|
Do, however, do these things in classes which inherit from this.
|
|
"""
|
|
|
|
def _createDense(self, dtype=core.DataType.FLOAT):
|
|
perfect_model = np.array([2, 6, 5, 0, 1]).astype(np.float32)
|
|
np.random.seed(123) # make test deterministic
|
|
numpy_dtype = np.float32 if dtype == core.DataType.FLOAT else np.float16
|
|
initializer = Initializer if dtype == core.DataType.FLOAT else pFP16Initializer
|
|
data = np.random.randint(
|
|
2,
|
|
size=(20, perfect_model.size)).astype(numpy_dtype)
|
|
label = np.dot(data, perfect_model)[:, np.newaxis]
|
|
|
|
model = ModelHelper(name="test", arg_scope={'order':'NCHW'})
|
|
out = brew.fc(
|
|
model,
|
|
'data', 'fc', perfect_model.size, 1, ('ConstantFill', {}),
|
|
('ConstantFill', {}), axis=0,
|
|
WeightInitializer=initializer, BiasInitializer=initializer
|
|
)
|
|
if dtype == core.DataType.FLOAT16:
|
|
out = model.HalfToFloat(out, out + "_fp32")
|
|
sq = model.SquaredL2Distance([out, 'label'])
|
|
loss = model.AveragedLoss(sq, "avg_loss")
|
|
grad_map = model.AddGradientOperators([loss])
|
|
self.assertIsInstance(grad_map['fc_w'], core.BlobReference)
|
|
return (model, perfect_model, data, label)
|
|
|
|
def testDense(self):
|
|
model, perfect_model, data, label = self._createDense()
|
|
optimizer = self.build_optimizer(model)
|
|
|
|
workspace.FeedBlob('data', data[0])
|
|
workspace.FeedBlob('label', label[0])
|
|
workspace.RunNetOnce(model.param_init_net)
|
|
workspace.CreateNet(model.net, True)
|
|
for _ in range(2000):
|
|
idx = np.random.randint(data.shape[0])
|
|
workspace.FeedBlob('data', data[idx])
|
|
workspace.FeedBlob('label', label[idx])
|
|
workspace.RunNet(model.net.Proto().name)
|
|
|
|
np.testing.assert_allclose(
|
|
perfect_model[np.newaxis, :],
|
|
workspace.FetchBlob('fc_w'),
|
|
atol=1e-2
|
|
)
|
|
self.check_optimizer(optimizer)
|
|
|
|
@unittest.skipIf(not workspace.has_gpu_support, "No gpu support")
|
|
def testGPUDense(self, dtype=core.DataType.FLOAT):
|
|
device_opt = core.DeviceOption(caffe2_pb2.CUDA, 0)
|
|
with core.DeviceScope(device_opt):
|
|
model, _perfect_model, data, label = self._createDense(dtype)
|
|
if dtype == core.DataType.FLOAT16:
|
|
fc_fp32_for_host = model.HalfToFloat('fc', 'fc_fp32_for_host')
|
|
model.CopyGPUToCPU(fc_fp32_for_host, 'fc_cpu')
|
|
else:
|
|
model.CopyGPUToCPU('fc', 'fc_cpu')
|
|
workspace.FeedBlob('data', data[0])
|
|
workspace.FeedBlob('label', label[0])
|
|
|
|
# Add some CPU ops
|
|
brew.fc(model, 'fc_cpu', 'fc2', dim_in=1, dim_out=10, axis=0)
|
|
|
|
# Create optimizer in default device scope
|
|
self.build_optimizer(model)
|
|
|
|
if self._skip_gpu:
|
|
return
|
|
|
|
# Run net to see it does not crash
|
|
workspace.RunNetOnce(model.param_init_net)
|
|
workspace.CreateNet(model.net, True)
|
|
workspace.RunNet(model.net.Proto().name)
|
|
|
|
|
|
def testSparse(self):
|
|
# to test duplicated indices we assign two indices to each weight and
|
|
# thus each weight might count once or twice
|
|
DUPLICATION = 2
|
|
perfect_model = np.array([2, 6, 5, 0, 1]).astype(np.float32)
|
|
np.random.seed(123) # make test deterministic
|
|
data = np.random.randint(
|
|
2,
|
|
size=(20, perfect_model.size * DUPLICATION)).astype(np.float32)
|
|
label = np.dot(data, np.repeat(perfect_model, DUPLICATION))
|
|
|
|
model = cnn.CNNModelHelper("NCHW", name="test")
|
|
# imitate what model wrapper does
|
|
w = model.param_init_net.ConstantFill(
|
|
[], 'w', shape=[perfect_model.size], value=0.0)
|
|
model.params.append(w)
|
|
picked = model.net.Gather([w, 'indices'], 'gather')
|
|
out = model.ReduceFrontSum(picked, 'sum')
|
|
|
|
sq = model.SquaredL2Distance([out, 'label'])
|
|
loss = model.AveragedLoss(sq, "avg_loss")
|
|
grad_map = model.AddGradientOperators([loss])
|
|
self.assertIsInstance(grad_map['w'], core.GradientSlice)
|
|
optimizer = self.build_optimizer(model)
|
|
|
|
workspace.CreateBlob('indices')
|
|
workspace.CreateBlob('label')
|
|
|
|
for indices_type in [np.int32, np.int64]:
|
|
workspace.RunNetOnce(model.param_init_net)
|
|
workspace.CreateNet(model.net, True)
|
|
for _ in range(2000):
|
|
idx = np.random.randint(data.shape[0])
|
|
# transform into indices of binary features
|
|
indices = np.repeat(np.arange(perfect_model.size),
|
|
DUPLICATION)[data[idx] == 1]
|
|
if indices.size == 0:
|
|
continue
|
|
workspace.FeedBlob(
|
|
'indices',
|
|
indices.reshape((indices.size,)).astype(indices_type)
|
|
)
|
|
workspace.FeedBlob('label',
|
|
np.array(label[idx]).astype(np.float32))
|
|
workspace.RunNet(model.net.Proto().name)
|
|
|
|
np.testing.assert_allclose(
|
|
perfect_model,
|
|
workspace.FetchBlob('w'),
|
|
atol=1e-2
|
|
)
|
|
self.check_optimizer(optimizer)
|
|
|
|
|
|
class GradientClippingTestBase(object):
|
|
"""
|
|
This is an abstract base class.
|
|
Don't inherit from unittest.TestCase, and don't name it 'Test*'.
|
|
Do, however, do these things in classes which inherit from this.
|
|
"""
|
|
|
|
def _gradient_ratio_reference(self, model, params, max_gradient_norm):
|
|
from caffe2.python import core
|
|
sum_squared_norms = 0.0
|
|
for param in params:
|
|
grad = (
|
|
model.param_to_grad[param]
|
|
if not isinstance(
|
|
model.param_to_grad[param],
|
|
core.GradientSlice,
|
|
) else model.param_to_grad[param].values
|
|
)
|
|
val = workspace.FetchBlob(grad)
|
|
sum_squared_norms += np.power(np.linalg.norm(val), 2.0)
|
|
global_norm = np.sqrt(sum_squared_norms)
|
|
clip_norm = max_gradient_norm
|
|
norm_ratio = clip_norm / np.maximum(clip_norm, global_norm)
|
|
return norm_ratio
|
|
|
|
def test_global_norm_based_gradient_clipping(self):
|
|
max_gradient_norm = 1
|
|
model, perfect_model, data, label = self._createDense()
|
|
self.build_optimizer(model, max_gradient_norm=max_gradient_norm)
|
|
|
|
params = []
|
|
for param in model.GetParams(top_scope=True):
|
|
if param in model.param_to_grad:
|
|
if not isinstance(
|
|
model.param_to_grad[param],
|
|
core.GradientSlice,
|
|
):
|
|
params.append(param)
|
|
|
|
workspace.FeedBlob('data', data[0])
|
|
workspace.FeedBlob('label', label[0])
|
|
workspace.RunNetOnce(model.param_init_net)
|
|
workspace.CreateNet(model.net, True)
|
|
for _ in range(1):
|
|
idx = np.random.randint(data.shape[0])
|
|
workspace.FeedBlob('data', data[idx])
|
|
workspace.FeedBlob('label', label[idx])
|
|
workspace.RunNet(model.net.Proto().name)
|
|
|
|
reference = self._gradient_ratio_reference(
|
|
model,
|
|
params,
|
|
max_gradient_norm,
|
|
)
|
|
norm_ratio = workspace.FetchBlob(
|
|
'norm_clipped_grad_update/norm_ratio')
|
|
np.testing.assert_almost_equal(norm_ratio, reference)
|
|
self.assertTrue(
|
|
norm_ratio < 1.0, "Bad test, gradient not being scaled."
|
|
)
|