Skip to content

Save log prob #3208

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
luotao1 merged 3 commits into from
Aug 4, 2017
Merged

Save log prob #3208

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
82801f2
save log probabilty for every generated words.
lcy-seso Jul 10, 2017
86d184e
Merge branch 'develop' into save_log_prob
luotao1 Aug 3, 2017
a58a77e
Merge branch 'develop' into save_log_prob
luotao1 Aug 3, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 19 additions & 7 deletions paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -967,8 +967,9 @@ void RecurrentGradientMachine::generateSequence() {
size_t numSequences = getGenBatchSize();

resizeBootFrame(numSequences);
// We create only two sub-network in generation for alternate use.
// Thus, we can reduce total memory of output_ in layer forward.
// We create only two sub-network in generation, one stores states of all
// layers in previous time step and the other storing the states at current
// time step.
resizeOrCreateFrames(2);

// outFrameLines_.size() > 1UL
Expand Down Expand Up @@ -1001,17 +1002,21 @@ void RecurrentGradientMachine::generateSequence() {

// init outArg
size_t resultNum = generator_.config.num_results_per_sample();
IVector::resizeOrCreate(
generator_.outArg.ids,
generator_.config.max_num_frames() * numSequences * resultNum,
false);
size_t maxGenWordCount =
generator_.config.max_num_frames() * numSequences * resultNum;
IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false);
if (resultNum > 1) {
CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
Matrix::resizeOrCreate(generator_.outArg.in,
/* height */ numSequences,
/* width */ resultNum,
false,
/* useGpu */ false);
Matrix::resizeOrCreate(generator_.outArg.value,
/* height */ maxGenWordCount,
/* width */ 1,
false,
/* useGpu */ false);
}
ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
numSequences + 1,
Expand Down Expand Up @@ -1313,13 +1318,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
starts[0] = 0;
if (numResults > 1) {
real* probs = generator_.outArg.in->getData();
real* idsProb = generator_.outArg.value->getData();
size_t curPos = 0;
for (size_t i = 0; i < finalPaths_.size(); ++i) {
for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
Path& path = finalPaths_[i][j];
generator_.ids.push_back(path.ids.size()); // sequence size
size_t genLen = path.ids.size();
generator_.ids.push_back(genLen); // sequence size
generator_.ids.insert(
generator_.ids.end(), path.ids.begin(), path.ids.end());
generator_.ids.push_back(-1); // end of sequence

memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen);
curPos += genLen;
idsProb[curPos++] = -1.0;
probs[i * numResults + j] = path.logProb;

if (!j && dataArgsSize_) {
Expand Down
12 changes: 10 additions & 2 deletions paddle/gserver/gradientmachines/RecurrentGradientMachine.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,6 +189,11 @@ class RecurrentGradientMachine : public NeuralNetwork {
*/
std::vector<int> ids;

/**
* @brief idsProb, log probability of each generated words.
*/
std::vector<real> idsProb;

/**
* @brief logProb, current probability of path.
*/
Expand Down Expand Up @@ -228,11 +233,13 @@ class RecurrentGradientMachine : public NeuralNetwork {
*/
Path(Path& old, int newId, real logProb, int machineId, int topIndex)
: ids(old.ids),
idsProb(old.idsProb),
logProb(old.logProb + logProb),
machineId(machineId),
topIndex(topIndex),
seqId(old.seqId) {
ids.push_back(newId);
idsProb.push_back(logProb);
if (!old.probHistory.empty()) {
this->probHistory = old.probHistory;
// probHistory store current prob, not sum
Expand Down Expand Up @@ -411,8 +418,9 @@ class RecurrentGradientMachine : public NeuralNetwork {

struct Generator {
GeneratorConfig config;
std::vector<int> ids; // store generated sequences
Argument outArg; // final output argument
std::vector<int> ids; // store generated sequences
std::vector<real> idsProb; // log probability of each generated word
Argument outArg; // final output argument
};
bool generating_;
Generator generator_;
Expand Down