Implement C/C++ PairwiseANN w/ Python API #268
Conversation
pecos/ann/pairwise/model.py
Outdated
| def train(cls, X, Y, train_params=None, pred_params=None): | ||
| """train and return the PairwiseANN indexer | ||
| Args: | ||
| X (numpy.array or smat.csr_matrix): database matrix to be indexed. (num_item x feat_dim). |
There was a problem hiding this comment.
Missing Y in docstring.
There was a problem hiding this comment.
Fixed. See Revision 2.
pecos/ann/pairwise/model.py
Outdated
| """constructor of PairwiseANN class | ||
| Args: | ||
| model_ptr (c_void_p): pointer to C instance pecos::ann:PairwiseANN. | ||
| num_item (int): number of item being indexed |
There was a problem hiding this comment.
Fixed. See Revision 2.
pecos/core/ann/pairwise.hpp
Outdated
|
|
||
| struct Searcher { | ||
| typedef PairwiseANN<feat_vec_t, mat_t> pairwise_ann_t; | ||
| typedef pecos::ann::heap_t<pair_t, std::less<pair_t>> max_heap_t; |
There was a problem hiding this comment.
Don't need to define this again.
There was a problem hiding this comment.
Fixed. See Revision 2.
pecos/core/ann/pairwise.hpp
Outdated
| typedef pecos::ann::KeyValPair<index_type, value_type> pair_t; | ||
| typedef pecos::ann::heap_t<pair_t, std::less<pair_t>> max_heap_t; | ||
|
|
||
| struct Searcher { |
There was a problem hiding this comment.
Fixed. See Revision 2.
pecos/core/ann/pairwise.hpp
Outdated
| } | ||
| } | ||
|
|
||
| void train(const mat_t &X_trn, const pecos::csr_t &Y_trn) { |
There was a problem hiding this comment.
should we just assume Y_trn to be CSC here? We can support CSR/CSC at python level.
There was a problem hiding this comment.
Fixed. See Revision 2.
pecos/core/libpecos.cpp
Outdated
| uint32_t* ret_Mmat, \ | ||
| float* ret_Dmat, \ | ||
| float* ret_Vmat, \ | ||
| const bool is_sample_input) { \ |
There was a problem hiding this comment.
Should this be is_same_input?
There was a problem hiding this comment.
Fixed. See Revision 2.
| """ | ||
| return copy.deepcopy(self.pred_params) | ||
|
|
||
| def predict(self, input_feat, label_keys, searchers, pred_params=None, is_same_input=False): |
There was a problem hiding this comment.
I wonder if there's a more extendable design for the predict API. Now we are basically supporting two cases (1) is_samp_input=True, a csr_vector and (2) is_same_input=False, a coo matrix.
bdab871
to
6a3feba
Compare
Issue #, if available:
Description of changes:
Implement C/C++ PairwiseANN function with Python API.
Given a (input, label) pair, PairwiseANN finds top-K nearest queries in the indexed input-to-label graph.
The searcher return four arrays of size
K:Imat: top-K input indicesMmat: the masking indicators. 1 indicates the kNN index is presented; otherwise 0.Dmat: the kNN distances between the test input and top-K indexed input.Vmat: the corresponding values stored in the input-to-label graph.Indexing usage
where
X_trnis the dense or sparse input feature matrixY_csris the sparse input-to-label graphSave/load Usage
Searchers Usage
where
max_batch_sizedefines the maximum number of rows in the returned arraymax_only_topkdefines the maximum number of cols in the returned arrayPrediction Usage
With same input (suitable for real-time inference):
query_vecis a single feature vector of shape(1, feat_dim)label_keysis a np.array of shape(batch_size, )With different inputs (suitable for batch prediction):
Query_matis a batch feature matrix of shape(batch_size, feat_dim)label_keysis a np.array of shape(batch_size, )By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.