Vadim Mazalov
GitHub
Коммит
c6d6570d91
|
|
@ -333,10 +333,10 @@ namespace Microsoft.MSR.CNTK.Extensibility.Managed.CSEvalClient
|
|||
// Specifies the number of times to iterate through the test file (epochs)
|
||||
int numRounds = 1;
|
||||
|
||||
// Counts the number of evaluations accross all models
|
||||
// Counts the number of evaluations across all models
|
||||
int count = 0;
|
||||
|
||||
// Counts the number of failed evaluations (output != expected) accross all models
|
||||
// Counts the number of failed evaluations (output != expected) across all models
|
||||
int errorCount = 0;
|
||||
|
||||
// The examples assume the executable is running from the data folder
|
||||
|
|
|
|||
|
|
@ -10,7 +10,7 @@ See License.md in the root level folder of the CNTK repository for full license
|
|||
|
||||
## Overview
|
||||
|
||||
|Data |The Penn Treebank Project (https://www.cis.upenn.edu/~treebank/) annotates naturally-occuring text for linguistic structure .
|
||||
|Data |The Penn Treebank Project (https://www.cis.upenn.edu/~treebank/) annotates naturally-occurring text for linguistic structure .
|
||||
|:---------|:---|
|
||||
|Purpose |Showcase how to train a recurrent network for text data.
|
||||
|Network |SimpleNetworkBuilder for recurrent network with two hidden layers.
|
||||
|
|
|
|||
|
|
@ -473,7 +473,7 @@
|
|||
"In order to report progress, please provide an instance of the [ProgressWriter](https://www.cntk.ai/pythondocs/cntk.logging.progress_print.html#module-cntk.logging.progress_print). It has its own set of parameters to control how often to print the loss value. If you need to have a custom logic for retrieving current status, please consider implementing your own ProgressWriter.\n",
|
||||
"\n",
|
||||
"### Checkpointing\n",
|
||||
"[Checkpoint configuration](https://www.cntk.ai/pythondocs/cntk.train.training_session.html#cntk.train.training_session.CheckpointConfig) specifies how often to save a checkpoint to the given file. The checkpointing frequency is specified in samples. When given, the method takes care of saving/restoring the state accross the trainer/learners/minibatch source and propagating this information among distributed workers. If you need to preserve all checkpoints that were taken during training, please set `preserveAll` to true. \n",
|
||||
"[Checkpoint configuration](https://www.cntk.ai/pythondocs/cntk.train.training_session.html#cntk.train.training_session.CheckpointConfig) specifies how often to save a checkpoint to the given file. The checkpointing frequency is specified in samples. When given, the method takes care of saving/restoring the state across the trainer/learners/minibatch source and propagating this information among distributed workers. If you need to preserve all checkpoints that were taken during training, please set `preserveAll` to true. \n",
|
||||
"\n",
|
||||
"### Validation\n",
|
||||
"When [cross validation](https://www.cntk.ai/pythondocs/cntk.train.training_session.html#cntk.train.training_session.CrossValidationConfig) config is given, the training session runs the validation on the specified minibatch source with the specified frequency and reports average metric error. The user can also provide a cross validation callback, that will be called with the specified frequency. It is up to the user to perform validation in the callback and return back `True` if the training should be continued, or `False` otherwise. \n",
|
||||
|
|
|
|||
|
|
@ -920,7 +920,7 @@ static wstring FormatConfigValue(ConfigValuePtr arg, const wstring &how);
|
|||
// StringFunction implements
|
||||
// - Format
|
||||
// - Chr(c) -- gives a string of one character with Unicode value 'c'
|
||||
// - Replace(s,what,withwhat) -- replace all occurences of 'what' with 'withwhat'
|
||||
// - Replace(s,what,withwhat) -- replace all occurrences of 'what' with 'withwhat'
|
||||
// - Substr(s,begin,num) -- get a substring
|
||||
// TODO: RegexReplace()
|
||||
class StringFunction : public String
|
||||
|
|
|
|||
|
|
@ -35,7 +35,7 @@ namespace CNTK
|
|||
{
|
||||
///
|
||||
/// Checked mode enables additional runtime verification such as:
|
||||
/// - Tracking NaN occurences in sequence gaps.
|
||||
/// - Tracking NaN occurrences in sequence gaps.
|
||||
/// - Function graph verification after binding of free static axes to actual values at runtime
|
||||
///
|
||||
/// Enabling checked mode incurs additional runtime costs and is meant to be used as a debugging aid.
|
||||
|
|
|
|||
|
|
@ -75,7 +75,7 @@ namespace CNTK
|
|||
|
||||
// This is used to generate a default seed value for random parameter initializer and also
|
||||
// for stateful nodes (dropout, and both flavors of random sample). The 'perWorkerLocalValue' flag
|
||||
// indicates if the generated value should be identical accross individual workers in distributed
|
||||
// indicates if the generated value should be identical across individual workers in distributed
|
||||
// setting or if each worker should get a different seed value.
|
||||
size_t GenerateRandomSeed(bool perWorkerLocalValue /*= false*/)
|
||||
{
|
||||
|
|
|
|||
|
|
@ -27,7 +27,7 @@ static inline size_t rand(const size_t begin, const size_t end)
|
|||
|
||||
// Rand based on Mersenne Twister.
|
||||
// We use our own distribution in order to match baselines between different operating systems,
|
||||
// because uniform_distribution is not guranteed to provide the same numbers on different platforms.
|
||||
// because uniform_distribution is not guaranteed to provide the same numbers on different platforms.
|
||||
// TODO: Switching to Boost would eliminate this problem.
|
||||
static inline size_t RandMT(const size_t begin, const size_t end, std::mt19937_64& rng)
|
||||
{
|
||||
|
|
|
|||
|
|
@ -247,7 +247,7 @@ double RandomSampleInclusionFrequencyNode<ElemType>::EstimateNumberOfTries()
|
|||
return totalTries / (double)numExperiments;
|
||||
}
|
||||
|
||||
// Estimates the expected number of occurences of each class in the sampled set.
|
||||
// Estimates the expected number of occurrences of each class in the sampled set.
|
||||
// For sampling without replacement we use estimate using average number of tries. (Inspired by TensorFlow)
|
||||
// BUGBUG: Consider to reimplement using a less biased estimate as proposed by Nikos.
|
||||
template<class ElemType>
|
||||
|
|
@ -340,4 +340,4 @@ template class DropoutNode<double>;
|
|||
template class BatchNormalizationNode<float>;
|
||||
template class BatchNormalizationNode<double>;
|
||||
|
||||
}}}
|
||||
}}}
|
||||
|
|
|
|||
|
|
@ -1447,7 +1447,7 @@ public:
|
|||
virtual void /*ComputationNode::*/ ForwardPropNonLooping() override;
|
||||
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override;
|
||||
private:
|
||||
// Approximates the expected number of occurences of a class in the sampled set.
|
||||
// Approximates the expected number of occurrences of a class in the sampled set.
|
||||
// Assuming (falsely) that the number of tries to get a sampled set with the requested number of distinct values is always estimatedNumTries
|
||||
// the probability that a specific class in the sampled set is (1 - (1-p)^estimatedNumTries), where p is the probablity to pick the clas in one draw.
|
||||
// The estimate can be quite a bit off but should be better than nothing. Better alternatives?
|
||||
|
|
|
|||
|
|
@ -1283,7 +1283,7 @@ size_t SGD<ElemType>::TrainOneEpoch(ComputationNetworkPtr net,
|
|||
// independent of their actual content (which is considered outdated).
|
||||
|
||||
// Sum of actualMBSize across all nodes when using parallel training
|
||||
// 'aggregate' here means accross-worker aggregate for this one minibatch.
|
||||
// 'aggregate' here means across-worker aggregate for this one minibatch.
|
||||
size_t aggregateNumSamples = actualMBSize; // (0 for empty MB)
|
||||
size_t aggregateNumSamplesWithLabel = CriterionAccumulator<ElemType>::GetNumSamples(criterionNodes[0], numSamplesWithLabelOfNetwork); // (0 for empty MB)
|
||||
|
||||
|
|
|
|||
|
|
@ -412,7 +412,7 @@ Test module "MathTests" has passed with:
|
|||
Test case "GPUMatrixSuite/GPUBlasInnerProduct" has passed with:
|
||||
2 assertions out of 2 passed
|
||||
|
||||
Test case "GPUMatrixSuite/MatrixCopyAssignAccrossDevices" has passed
|
||||
Test case "GPUMatrixSuite/MatrixCopyAssignAcrossDevices" has passed
|
||||
|
||||
Test case "GPUMatrixSuite/GPUMatrixConstructorNoFlag" has passed with:
|
||||
6 assertions out of 6 passed
|
||||
|
|
|
|||
|
|
@ -15,7 +15,7 @@ namespace Microsoft { namespace MSR { namespace CNTK { namespace Test {
|
|||
|
||||
BOOST_AUTO_TEST_SUITE(GPUMatrixSuite)
|
||||
|
||||
BOOST_FIXTURE_TEST_CASE(MatrixCopyAssignAccrossDevices, RandomSeedFixture)
|
||||
BOOST_FIXTURE_TEST_CASE(MatrixCopyAssignAcrossDevices, RandomSeedFixture)
|
||||
{
|
||||
bool hasTwoGpus = false;
|
||||
#ifndef CPUONLY
|
||||
|
|
|
|||
|
|
@ -107,7 +107,7 @@ class _DebugState(object):
|
|||
def set_checked_mode(enable):
|
||||
'''
|
||||
Checked mode enables additional runtime verification such as:
|
||||
- Tracking NaN occurences in sequence gaps.
|
||||
- Tracking NaN occurrences in sequence gaps.
|
||||
- Function graph verification after binding of free static axes to actual values at runtime
|
||||
|
||||
Enabling checked mode incurs additional runtime costs and is meant to be used as a debugging aid.
|
||||
|
|
|
|||
|
|
@ -2467,7 +2467,7 @@ def random_sample_inclusion_frequency(
|
|||
name=''):
|
||||
'''
|
||||
For weighted sampling with the specifed sample size (`num_samples`)
|
||||
this operation computes the expected number of occurences of each class
|
||||
this operation computes the expected number of occurrences of each class
|
||||
in the sampled set. In case of sampling without replacement
|
||||
the result is only an estimate which might be quite rough in the
|
||||
case of small sample sizes.
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче