251 lines
9.4 KiB
C++
251 lines
9.4 KiB
C++
#include <algorithm>
|
|
#include <numeric>
|
|
#include <caffe/util/math_functions.hpp>
|
|
#include "visualisations.hpp"
|
|
#include "DummyLayerVisualisation.hpp"
|
|
#include "MultiImageVisualisation.hpp"
|
|
#include "FlatLayerVisualisation.hpp"
|
|
#include "Range.hpp"
|
|
#include "ActivityAnimation.hpp"
|
|
#include "InputLayerVisualisation.hpp"
|
|
#include "PoolingLayerAnimation.hpp"
|
|
#include "ImageInteractionAnimation.hpp"
|
|
|
|
using namespace fmri;
|
|
using namespace std;
|
|
|
|
// Maximum number of interactions shown
|
|
static constexpr size_t INTERACTION_LIMIT = 10000;
|
|
|
|
typedef vector<pair<float, pair<size_t, size_t>>> EntryList;
|
|
|
|
/**
|
|
* Normalizer for node positions.
|
|
*
|
|
* Since not every neuron in a layer may get a node in the visualisation,
|
|
* this function maps those neurons back to a node number that does.
|
|
*
|
|
* Usage: node / getNodeNormalizer(layer).
|
|
*
|
|
* @param layer Layer to compute normalization for
|
|
* @return Number to divide node numbers by.
|
|
*/
|
|
static inline int getNodeNormalizer(const LayerData& layer) {
|
|
const auto& shape = layer.shape();
|
|
switch(shape.size()) {
|
|
case 2:
|
|
return 1;
|
|
|
|
case 4:
|
|
return shape[2] * shape[3];
|
|
|
|
default:
|
|
CHECK(false) << "Unsupported shape " << shape.size() << endl;
|
|
exit(EINVAL);
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Deduplicate interaction entries.
|
|
*
|
|
* For duplicate interactions, the interaction strengths are summed.
|
|
*
|
|
* @param entries
|
|
* @return the deduplicated entries.
|
|
*/
|
|
static EntryList deduplicate(const EntryList& entries)
|
|
{
|
|
map<pair<size_t, size_t>, float> combiner;
|
|
for (auto entry : entries) {
|
|
combiner[entry.second] += entry.first;
|
|
}
|
|
|
|
EntryList result;
|
|
transform(combiner.begin(), combiner.end(), back_inserter(result), [](const auto& item) {
|
|
return make_pair(item.second, item.first);
|
|
});
|
|
|
|
return result;
|
|
}
|
|
|
|
fmri::LayerVisualisation *fmri::getVisualisationForLayer(const fmri::LayerData &data, const fmri::LayerInfo &info)
|
|
{
|
|
switch (info.type()) {
|
|
case LayerInfo::Type::Input:
|
|
if (data.shape().size() == 4) {
|
|
return new InputLayerVisualisation(data);
|
|
} else {
|
|
return new FlatLayerVisualisation(data, FlatLayerVisualisation::Ordering::SQUARE);
|
|
}
|
|
|
|
default:
|
|
switch (data.shape().size()) {
|
|
case 2:
|
|
return new FlatLayerVisualisation(data, FlatLayerVisualisation::Ordering::SQUARE);
|
|
|
|
case 4:
|
|
return new MultiImageVisualisation(data);
|
|
|
|
default:
|
|
return new DummyLayerVisualisation();
|
|
}
|
|
}
|
|
}
|
|
|
|
static Animation *getFullyConnectedAnimation(const fmri::LayerData &prevState, const fmri::LayerInfo &layer,
|
|
const vector<float> &prevPositions, const vector<float> &curPositions)
|
|
{
|
|
LOG(INFO) << "Computing top interactions for " << layer.name() << endl;
|
|
|
|
auto data = prevState.data();
|
|
|
|
CHECK_GE(layer.parameters().size(), 1) << "Layer should have correct parameters";
|
|
|
|
const auto shape = layer.parameters()[0]->shape();
|
|
auto weights = layer.parameters()[0]->cpu_data();
|
|
const auto numEntries = accumulate(shape.begin(), shape.end(), static_cast<size_t>(1), multiplies<void>());
|
|
|
|
vector<float> interactions(numEntries);
|
|
const auto stepSize = shape[1];
|
|
|
|
for (auto i : Range(numEntries / stepSize)) {
|
|
caffe::caffe_mul(shape[1], &weights[i * stepSize], data, &interactions[i * stepSize]);
|
|
}
|
|
|
|
const auto desiredSize = min(INTERACTION_LIMIT, numEntries);
|
|
auto idx = arg_partial_sort(interactions.begin(), interactions.begin() + desiredSize, interactions.end(),
|
|
[](auto a, auto b) {
|
|
return abs(a) > abs(b);
|
|
});
|
|
|
|
EntryList result;
|
|
result.reserve(desiredSize);
|
|
const auto normalizer = getNodeNormalizer(prevState);
|
|
for (auto i : idx) {
|
|
if (abs(interactions[i]) < EPSILON){
|
|
break;
|
|
}
|
|
result.emplace_back(interactions[i], make_pair((i % shape[1]) / normalizer, i / shape[1]));
|
|
}
|
|
|
|
return new ActivityAnimation(result, prevPositions.data(), curPositions.data());
|
|
}
|
|
|
|
static Animation *getDropOutAnimation(const fmri::LayerData &prevState,
|
|
const fmri::LayerData &curState,
|
|
const vector<float> &prevPositions,
|
|
const vector<float> &curPositions) {
|
|
const auto sourceNormalize = getNodeNormalizer(prevState);
|
|
const auto sinkNormalize = getNodeNormalizer(curState);
|
|
|
|
auto data = curState.data();
|
|
EntryList results;
|
|
results.reserve(curState.numEntries());
|
|
for (auto i : Range(curState.numEntries())) {
|
|
if (data[i] != 0) {
|
|
results.emplace_back(data[i], make_pair(i / sourceNormalize, i / sinkNormalize));
|
|
}
|
|
}
|
|
|
|
results = deduplicate(results);
|
|
|
|
return new ActivityAnimation(results, prevPositions.data(), curPositions.data());
|
|
}
|
|
|
|
static Animation *getReLUAnimation(const fmri::LayerData &prevState,
|
|
const fmri::LayerData &curState,
|
|
const vector<float> &prevPositions,
|
|
const vector<float> &curPositions) {
|
|
CHECK_EQ(curState.numEntries(), prevState.numEntries()) << "Layers should be of same size!";
|
|
|
|
std::vector<float> changes(prevState.numEntries());
|
|
caffe::caffe_sub(prevState.numEntries(), curState.data(), prevState.data(), changes.data());
|
|
|
|
if (curState.shape().size() == 2) {
|
|
EntryList results;
|
|
for (auto i : Range(curState.numEntries())) {
|
|
results.emplace_back(changes[i], make_pair(i, i));
|
|
}
|
|
|
|
const auto maxValue = max_element(results.begin(), results.end())->first;
|
|
|
|
return new ActivityAnimation(results, prevPositions.data(), curPositions.data(),
|
|
[=](float i) -> Color {
|
|
if (maxValue == 0) {
|
|
return {1, 1, 1, 1};
|
|
} else {
|
|
return {1 - i / maxValue, 1 - i / maxValue, 1, 1};
|
|
}
|
|
});
|
|
} else {
|
|
return new ImageInteractionAnimation(changes.data(), prevState.shape(), prevPositions, curPositions);
|
|
}
|
|
}
|
|
|
|
static Animation *getNormalizingAnimation(const fmri::LayerData &prevState, const LayerData &curState,
|
|
const vector<float> &prevPositions,
|
|
const vector<float> &curPositions) {
|
|
CHECK(prevState.shape() == curState.shape()) << "Shapes should be of equal size" << endl;
|
|
vector<DType> scaling(std::accumulate(prevState.shape().begin(), prevState.shape().end(), 1u, multiplies<void>()));
|
|
caffe::caffe_div(scaling.size(), prevState.data(), curState.data(), scaling.data());
|
|
|
|
// Fix divisions by zero. For those cases, pick 1 since it doesn't matter anyway.
|
|
normalize(scaling.begin(), scaling.end());
|
|
|
|
if (prevState.shape().size() == 2) {
|
|
EntryList entries;
|
|
entries.reserve(scaling.size());
|
|
for (auto i : Range(scaling.size())) {
|
|
entries.emplace_back(scaling[i], make_pair(i, i));
|
|
}
|
|
|
|
auto max_val = *max_element(scaling.begin(), scaling.end());
|
|
|
|
return new ActivityAnimation(entries, prevPositions.data(), curPositions.data(),
|
|
[=](float i) -> Color {
|
|
auto intensity = clamp((i - 1) / (max_val - 1), 0.f, 1.f);
|
|
return {
|
|
1 - intensity,
|
|
1,
|
|
1,
|
|
1
|
|
};
|
|
});
|
|
} else {
|
|
transform(scaling.begin(), scaling.end(), scaling.begin(), [](float x) { return log(x); });
|
|
return new ImageInteractionAnimation(scaling.data(), prevState.shape(), prevPositions, curPositions);
|
|
}
|
|
}
|
|
|
|
Animation * fmri::getActivityAnimation(const fmri::LayerData &prevState, const fmri::LayerData &curState,
|
|
const fmri::LayerInfo &layer, const vector<float> &prevPositions,
|
|
const vector<float> &curPositions)
|
|
{
|
|
if (prevPositions.empty() || curPositions.empty()) {
|
|
// Not all positions known, no visualisation possible.
|
|
return nullptr;
|
|
}
|
|
|
|
|
|
switch (layer.type()) {
|
|
case LayerInfo::Type::InnerProduct:
|
|
return getFullyConnectedAnimation(prevState, layer,
|
|
prevPositions, curPositions);
|
|
|
|
case LayerInfo::Type::DropOut:
|
|
return getDropOutAnimation(prevState, curState, prevPositions, curPositions);
|
|
|
|
case LayerInfo::Type::ReLU:
|
|
return getReLUAnimation(prevState, curState, prevPositions, curPositions);
|
|
|
|
case LayerInfo::Type::Pooling:
|
|
return new PoolingLayerAnimation(prevState, curState, prevPositions, curPositions);
|
|
|
|
case LayerInfo::Type::LRN:
|
|
return getNormalizingAnimation(prevState, curState, prevPositions, curPositions);
|
|
|
|
default:
|
|
return nullptr;
|
|
}
|
|
}
|