[WIP] Sparse Tensor (apache#5800)

* squash merge with 38f7c55 compiles on GPU update check alloc: Checkpoint. Pass elem-sum gpu test bug fix for copyfromto. sparse sgd test pass on gpu inefficient implementation for csr copy update submodule fix lint Simple bind with infer storage type (apache#32) * Symbol binding for sparse tensor development. (apache#31) * Initial checkin * Add init functions for simple bind in graph_executor * Add simple_bind c_api * Add simple bind c-api * Assign zeros to in_args, arg_grads, and aux_states * Add simple_bind2 python interface * Fix python interface bugs * Interface changes * Fix * Fix core dump * Add bind_ith_exec c_api * Change simple_bind2 * Fix seg fault * Finish simple_bind * Change _bind_ith_exec * Refactor simple_bind initialization flow for bind * Consolidate bind and simple_bind graph init flow * Fix bug * Clean up * Add comments * Clean up * Clean up * Minor correction * Rename APIs in graph executor * Refactor * Rebase * Delete deprecated functions * Move more front-end work to backend * Bug fix * Fix failed tests * Minor fix * Fix lint * Fix lint * Revert unnecessary changes * Revert * Revert * Clean up * Fix lint Conflicts: python/mxnet/symbol.py src/executor/graph_executor.cc * Add inferstorage to graph executor * re-enable tests for sparse embedding with simple_bind * type switch fix in sparse embedding" ; change `default` to `default_storage` for cast storage op (apache#33) * change default to default_storage * disable cpp test build temporarily attempt to fix windows build error, and fix lint (apache#34) update nnvm submodule (apache#37) Scipy build (apache#38) * update nnvm submodule * add scipy pip install for dockerfile Python3 unit tests (apache#39) * change xrange to range for python3 compatiblity" * remove more xrange from tests replace long with int for python3 (apache#40) fix the rest of TShape constructor errors (apache#41) fix lint (apache#42) fix wrong usage of mshadow::Shape1" (apache#43) implementation for Csr slice on cpu (apache#36) * CPU implementation for CSR remove seg_len from csr slice add some docs for slice csr change indptr, values, etc to be private member bug fix in sparse embedding update nnvm submoduel fix lint update unit test for sparse nd" * add const for SliceCsrIndPtr kernel Fix sparse dot according to the new RSP definition (apache#35) * Fix csr dot dns * Fix sparse dot * Add fallback and test cases for dot(csr, dns)=dns * Add int type switch * Fix * Fix * Fix update mshadow submodule (apache#44) Fix dns to rsp (apache#46) fix lint (apache#47) add runtime storage fallback detection" (apache#48) * add runtime storage fallback detection" * replace cast storage ex with cast storage impl Fm example (apache#45) * update csr slice logic to avoid confusion. add more exmaples. * add hint to module.update * more testcases(fallback) for sparse_nd * add to_csr() and to_rsp() method. More unit test (fallback now) * add fm test. fix lint * register sparse sgd under Optim.SGD * update dmlc-core submoduel * change indptr to _indptr temporarily. add const ref to fname fix lint fix lint; (apache#51) Guard gpu cast storage (apache#50) * Clean up * Fix typo Rearrange unit test files (apache#52) fix lint. add scipy for python_test. fix scipy.sparse import error. fix truediv for python3 fix travis test (apache#54) * remove pyc files * add verbose for travis nosetests cleanup some testing code and enums (apache#57) * update Makefile * refactor test_sparse_operator * change `default_storage` back to `default` * remove unused cpp tests port libsvm parser to mxnet as libsvm iter (apache#55) * copied csv iter to libsvm iter test libsvm iter draft handle round batch == false for csr batch loader code refactoring add get stype, shape interface to iiter separate class for sparse iter add missing file fix mem corruption' rename variables add comments also read label from libsvm add test. update docs. update submodule Conflicts: python/mxnet/sparse_ndarray.py * update submodule * fix lint * update test * revert naming change add benchmark scritp for dot (apache#59) * add benchmark scritp for dot add gpu option for bench add get_data funciton for benchmark print t_sparse, too; add comment change nnz to dnesity add backward * add comment update fm test (apache#62) introduce CSRNDarray and rowsparseNDarray to python frontend api (apache#58) * introduce CSRNDarray and rowsparseNDarray to python frontend api * temporarily disable fm_module test fix lint (apache#64) fix typo. disable libsvm io test (apache#65) Improve dot (apache#61) * Init checkin * Fix * Adjust dot parallelization methods * Set num_omp_threads for benchmark from command line * Fix omp thread number * Clean up * Add scipy as dot baseline * Fix format sparse_retain op (apache#66) * Initial checkin * Fix bugs * Add unit test for sparse_retain * Add example and modify test add storage cast for outputs that have non-default storage (apache#67) fix gpu build (apache#69) Fix test_sparse_retain python3 issue (apache#68) revert nnvm version * draft for sgd rsp rsp (apache#75) support sgd(rsp, rsp) support dot(csr, rsp) when rsp is full add ref to const ndarray params support sparse embedding with rsp weight' fix lint modify embedding backward to produce dense grad remove invalid_rid for rsp->dns remove previous embedding op changes pass sparse embedding test add STORAGE_TYPE_ASSIGN_CHECK remove backward storage infer * fix lint (apache#78) * fix lint (apache#79) * serial elemwise sum impl (apache#80) update module kvstore interface add other missing params and functions revert some interface changes revert some more changes reomve explicit casting for gradients on kvstore update Comm interface update fm example Conflicts: python/mxnet/model.py python/mxnet/ndarray.py * bug fix for initializing module with row_sparse weight (apache#81) * bug fix for initializing module with row_sparse weight * update log message * Sparse ndarray serialization and deserialization (apache#77) * Initial checkin * Add unit tests * Fix lint * Fix lint (apache#84) * Sgd with row_sparse weight, dns gradient (apache#83) * sgd rsp dns draft * support sgd_mom(rsp, dns, rsp) * update doc * remove cast storage for kv updater * code refactoring * update mshadow version (apache#88) * csr slice bug fix (apache#90) * benchmark dot code refactor (apache#87) * q^x6x add some code in benchmark * refactor * minor fixes * fix * lint fix * Add unit test (apache#91) * add unittest * minor fix * remove commented lines * change test func name * add test rsp * kvstore push row sparse (apache#93) * Add multi-thread cpu elemwise sum for rsps * Minor fix * Add flag to switch between serial and multi-thread kvstore push * Fix lint in sparse_ndarray.py * Revert "Fix lint in sparse_ndarray.py" This reverts commit d7225ec. * Fix ndarray init in copy(ctx) * Add env var to control the flow of serial/parallel reduce * Refactor * Fix copy ndarray bug * Fix lint * Refactor * Fix windows openmp build failure (apache#94) * update mshadow submoduel (apache#95) * Revert "update mshadow submoduel (apache#95)" (apache#96) This reverts commit 1a129e4. * Refactor sparse tensor code (apache#99) * Initial checkin test_sparse_ndarray passes * Fix test failure * Clean up * Clean up * Move init backend op to ndarray_utils * Fix lint * Eliminate circular dependency on headers * More refactor * Fix gpu build and consolidate Slice for dense and sparse * Clean up * More refactor * Clean up * Fix gpu build * Fix comment * fix pylint (apache#100) * Fix refactor sparse gpu test (apache#104) * Fix gpu build * Fix * Fix gpu test failure * change idx types from int32 to int64 (apache#101) Conflicts: python/mxnet/test_utils.py tests/python/unittest/test_sparse_operator.py update mshadow submodule fix extra quotes in test script change indptr type to int64 better err message for rsp" * revert LOG(DEBUG) change (apache#105) * fix undefined zeros in optimizer.py (apache#106) * move init dns zeros to init_op.h for kvstore to use (apache#107) * Refactor cast storage (apache#109) * Refactor cast_storage * Add cast_storage cc and cu files * Remove redundant comments * Replace std::accumulate with ParallelAccumulate * Clean up * Fix windows build * Rowsparse kv (apache#111) * update kvstore unit test Conflicts: tests/python/unittest/test_kvstore.py update model/module.py Conflicts: python/mxnet/model.py python/mxnet/module/module.py fix lint resolve conflict remove int keys in kvstore update cast to str function * fix failed dist_sync_kv test * bug fix in comm to ensure merged gradient is of the right type bug fix in comm * row sparse dist kvstore draft (push only) row_sparse pull * add ndarray row sparse shared mem constructor * code refactoring * add test for row_sparse weight bug fix for kv server slicing add async support rsolve race condition in kvstore * resolve error after reb ase * fix lint (apache#113) * rename some python funciton (apache#114) * _to_rsp * _to_csr. raise NotImplementedError * todense * fix lint (apache#115) enable libsvm uniit test (apache#6839) remove shared mem slice for csr add csr ndarray iter test make osx nose test verbose disable libsvm iter test Move InferAttr to mxnet from nnvm (apache#6830) * Move InferAttr to mxnet from nnvm Replace nnvm infer attr functions in c_api Initial checkin Clean up Remove nnvm namespace for FInferShape, FInferType, and FInferStorageType Add new interface for InferStorageType Revert "Remove nnvm namespace for FInferShape, FInferType, and FInferStorageType" This reverts commit 8aedf05. Fix and clean up Fix lint Add nnvm changes Change infer function interface to accept only rvalue reference of graph Clean up Flush commits to show up in PR Add error handling for storage type inference failure Update nnvm * Fix pylint Change idx type switch for aux data (apache#6860) * Change idx type switch for aux data * Add mshadow commit Sparse dot enhancement (apache#6842) * Initial checkin Initial checkin Fix sparse dot test Fix unitest and add fallback for sparse dot * Add benchmark code * Revert "Add benchmark code" This reverts commit be009fe. * Fix bug * Fix storage shape * Remove unnecessary test code * Use idx type switch Implement dot(csr, rsp)=dns and dot(csr.T, rsp)=rsp and refactor (apache#6902) * Initial checkin Add dot(csr.T, rsp)=rsp2 Add infer storage for dot(csr, rsp)=dns and dot(csr.T, rsp)=rsp2 * Fix comments * Replace std::lower_bound with own impl for gpu use too * Add time profiling * Revert "Add time profiling" This reverts commit 8f5bb98. * Move dot and batch_dot to a single file * Move dot gpu impl to a .cuh file * More refactor * Fix include error LibsvmIter fix (apache#6898) * fix bug in libsvm iter which causes mem corruption * add test for news dataset * fix wrong path in test * fix import error for urllib * update url * replace bz command with bz module Optimized gpu dot kernels (apache#6937) * pulled update to mshadow * mshadow update * added optimized gpu kernels for dot(csr,dns)=dns and dot(csr.T,dns)=dns, and unit test * added __syncwarp to vector kernel and reduced number of writes to shared memory Refactor sparse tensor code (apache#6955) * Save stype in frontend to avoid c-api call for stype * Change storage_type to stype * Revert "Change storage_type to stype" This reverts commit 90db7d1. * Revert "Revert "Change storage_type to stype"" This reverts commit 0932838. Move ndarray.py, sparse_ndarray.py, ndarray_utils.py, and _ndarray_internal to ndarrary folder More refactor Move elementwise sum for rsp to ndarray_function.cc Remove unnecessary import in ndarray module Fix pylint Remove redundant code Remove _stype from slots Fix cpp-package build error caused by the change to imperative invoke interface Use relative import Remove print line Rename _ndarray_internal.py to _internal.py * Relaunch test... minor bug fix in warp synchronous code (apache#7029)
cjolivier01 · Jul 25, 2017 · c5f6648 · c5f6648
1 parent 4fef6fd
commit c5f6648
Show file tree

Hide file tree

Showing 105 changed files with 9,154 additions and 1,294 deletions.
diff --git a/benchmark/python/sparse_op.py b/benchmark/python/sparse_op.py
@@ -0,0 +1,228 @@
+import ctypes
+
+from mxnet.test_utils import *
+import scipy.sparse as sp
+import os
+import time
+import argparse
+
+from mxnet.base import check_call, _LIB
+from util import get_data, estimate_density
+
+parser = argparse.ArgumentParser(description="Benchmark sparse operators",
+                                 formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+parser.add_argument('--num-omp-threads', type=int, default=1, help='number of omp threads to set in MXNet')
+args = parser.parse_args()
+
+# some data information
+kdda = {
+    'data_mini': 'kdda.t.mini',
+    'data_name': 'kdda.t',
+    'data_origin_name': 'kdda.t.bz2',
+    'url': "https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary/kdda.t.bz2",
+    'feature_dim': 20216830,
+    'm': 200,
+    'batch_size': [64]
+}
+
+avazu = {
+    'data_mini': 'avazu-app.t.mini',
+    'data_name': 'avazu-app.t',
+    'data_origin_name': 'avazu-app.t.bz2',
+    'url': "https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary/avazu-app.t.bz2",
+    'feature_dim': 1000000,
+    'm': 500,
+    'batch_size': [64, 128]
+}
+
+
+def measure_cost(repeat, f, *args, **kwargs):
+    # start bench
+    start = time.time()
+    results = []
+    for i in range(repeat):
+        results.append(f(*args, **kwargs))
+    for result in results:
+        result.wait_to_read()
+    end = time.time()
+    diff = end - start
+    return diff / repeat
+
+
+def test_dot_real(data_dict):
+    def get_iter(path, data_shape, batch_size):
+        data_train = mx.io.LibSVMIter(data_libsvm=path,
+                                      data_shape=data_shape,
+                                      batch_size=batch_size)
+        data_iter = iter(data_train)
+        return data_iter
+
+    data_dir = os.path.join(os.getcwd(), 'data')
+
+    path = os.path.join(data_dir, data_dict['data_name'])
+    if not os.path.exists(path):
+        get_data(
+            data_dir,
+            data_dict['data_name'],
+            data_dict['url'],
+            data_dict['data_origin_name']
+        )
+        assert os.path.exists(path)
+
+    k = data_dict['feature_dim']
+    m = data_dict['m']
+    density = estimate_density(path, data_dict['feature_dim'])
+
+    mini_path = os.path.join(data_dir, data_dict['data_mini'])
+    if not os.path.exists(mini_path):
+        os.system("head -n 2000 %r > %r" % (path, mini_path))
+        assert os.path.exists(mini_path)
+
+    print "Running Benchmarking on %r data" % data_dict['data_mini']
+    for batch_size in data_dict['batch_size']:  # iterator through different batch size of choice
+        print "batch_size is %d" % batch_size
+        # model
+        data_shape = (k, )
+        train_iter = get_iter(mini_path, data_shape, batch_size)
+        weight = mx.nd.random_uniform(low=0, high=1, shape=(k, m))
+
+        csr_data = []
+        dns_data = []
+        num_batch = 0
+        for batch in train_iter:
+            data = train_iter.getdata()
+            csr_data.append(data)
+            dns_data.append(data.todense())
+            num_batch += 1
+        bag_of_data = [csr_data, dns_data]
+        num_repeat = 5
+        costs = []
+        for d in bag_of_data:
+            weight.wait_to_read()
+            cost = 0.
+            count = 0
+            for d_batch in d:
+                d_batch.wait_to_read()
+                cost += measure_cost(num_repeat, mx.nd.dot, d_batch, weight)
+                count += 1
+            costs.append(cost/count)
+        t_sparse = costs[0]
+        t_dense = costs[1]
+        ratio = t_dense / t_sparse
+        print('density(%)\tn\tm\tk\tt_dense/t_sparse\tt_dense\tt_sparse')
+        fmt = "%0.4f\t\t%d\t%d\t%d\t%0.2f\t\t\t%0.4f\t%0.6f"
+        print(fmt % (density * 100, batch_size, m, k, ratio, t_dense, t_sparse))
+
+
+def test_dot_synthetic():
+    """benchmark mx.nd.dot(sparse_ndarray, dense_ndarray) with given density.
+    `t_sparse` is the time cost of dot(csr, dns), while `t_dense` is the time cost
+    of dot(dns, dns), with the same matrix except that it is in default storage type.
+    """
+    def measure_cost_forward_baseline(repeat, dot, lhs, rhs):
+        start = time.time()
+        for i in range(repeat):
+            dot(lhs, rhs)
+        end = time.time()
+        diff = end - start
+        return diff / repeat
+
+    def measure_cost_backward_baseline(repeat, dot, transpose, lhs, rhs):
+        start = time.time()
+        for i in range(repeat):
+            dot(transpose(lhs), rhs)
+        end = time.time()
+        diff = end - start
+        return diff / repeat
+
+    def bench_dot_forward(m, k, n, density, ctx, repeat):
+        set_default_context(ctx)
+        dns = mx.nd.random_uniform(shape=(k, n)).copyto(ctx)
+        data_shape = (m, k)
+        csr_data = rand_ndarray(data_shape, 'csr', density)
+        dns_data = csr_data.todense()
+        rhs_dns_np = dns.asnumpy()
+        lhs_csr_sp = sp.csr_matrix(dns_data.asnumpy())  # csr in scipy
+        lhs_dns_np = lhs_csr_sp.todense()
+
+        data = [dns_data, csr_data]
+        costs = []
+        for d in data:
+            dns.wait_to_read()
+            d.wait_to_read()
+            cost = measure_cost(repeat, mx.nd.dot, d, dns)
+            costs.append(cost)
+        ratio = costs[0] / costs[1]
+
+        costs_baseline = []
+        cost = measure_cost_forward_baseline(repeat, np.dot, lhs_dns_np, rhs_dns_np)
+        costs_baseline.append(cost)
+        cost = measure_cost_forward_baseline(repeat, sp.spmatrix.dot, lhs_csr_sp, rhs_dns_np)
+        costs_baseline.append(cost)
+        ratio_baseline = costs_baseline[0] / costs_baseline[1]
+        fmt = "%0.1f\t\t%s\t%d\t%d\t%d\t%0.2f\t\t\t%0.2f\t%0.5f\t\t%0.2f\t\t\t\t%0.6f\t%0.5f"
+        print(fmt % (density * 100, str(ctx), n, m, k, ratio, costs[0], costs[1],
+                     ratio_baseline, costs_baseline[0], costs_baseline[1]))
+
+    def bench_dot_backward(m, k, n, density, ctx, repeat):
+        set_default_context(ctx)
+        dns = mx.nd.random_uniform(shape=(m, n)).copyto(ctx)
+        data_shape = (m, k)
+        csr_data = rand_ndarray(data_shape, 'csr', density)
+        dns_data = csr_data.todense()
+        rhs_dns_np = dns.asnumpy()
+        lhs_csr_sp = sp.csr_matrix(dns_data.asnumpy())
+        lhs_dns_np = lhs_csr_sp.todense()
+
+        data = [dns_data, csr_data]
+        costs = []
+        for d in data:
+            dns.wait_to_read()
+            d.wait_to_read()
+            cost = measure_cost(repeat, mx.nd.dot, d, dns, transpose_a=True)
+            costs.append(cost)
+        ratio = costs[0] / costs[1]
+
+        costs_baseline = []
+        cost = measure_cost_backward_baseline(repeat, np.dot, np.transpose, lhs_dns_np, rhs_dns_np)
+        costs_baseline.append(cost)
+        cost = measure_cost_backward_baseline(repeat, sp.spmatrix.dot, sp.spmatrix.transpose, lhs_csr_sp, rhs_dns_np)
+        costs_baseline.append(cost)
+        ratio_baseline = costs_baseline[0] / costs_baseline[1]
+        fmt = "%0.1f\t\t%s\t%d\t%d\t%d\t%0.2f\t\t\t%0.2f\t%0.5f\t\t%0.2f\t\t\t\t%0.6f\t%0.5f"
+        print(fmt % (density * 100, str(ctx), n, m, k, ratio, costs[0], costs[1],
+                     ratio_baseline, costs_baseline[0], costs_baseline[1]))
+
+    print("A = sparse NDArray of shape(m, k)")
+    print("B = dense NDArray of shape(k, n)")
+    print("dot_forward\tdot(csr, dns)")
+    print('density(%)\tcontext\tn\tm\tk\tt_dense/t_sparse\tt_dense\tt_sparse'
+          '\tt_scipy_dense/t_scipy_sparse\tt_scipy_dense\tt_scipy_sparse')
+
+    check_call(_LIB.MXSetNumOMPThreads(ctypes.c_int(args.num_omp_threads)))
+    # TODO(haibin) make these runtime options
+    m = 512
+    k = [50000, 100000]
+    n = [64, 128]
+    density = [1.00, 0.90, 0.70, 0.50, 0.30, 0.20, 0.10, 0.07, 0.05, 0.02, 0.01, 0.005, 0.001]
+    num_repeat = 10
+    # contexts = [mx.cpu(), mx.gpu(0)]
+    contexts = [mx.cpu()]
+    for i in range(2):
+        for ctx in contexts:
+            for den in density:
+                bench_dot_forward(m, k[i], n[i], den, ctx, num_repeat)
+
+    print("dot_backward\tdot(csr.T, dns)")
+    print('density(%)\tcontext\tn\tm\tk\tt_dense/t_sparse\tt_dense\tt_sparse'
+          '\tt_scipy_dense/t_scipy_sparse\tt_scipy_dense\tt_scipy_sparse')
+    for i in range(2):
+        for ctx in contexts:
+            for den in density:
+                bench_dot_backward(m, k[i], n[i], den, ctx, num_repeat)
+
+
+if __name__ == "__main__":
+    test_dot_real(avazu)
+    test_dot_real(kdda)
+    test_dot_synthetic()
diff --git a/benchmark/python/util.py b/benchmark/python/util.py
@@ -0,0 +1,33 @@
+import os
+import random
+
+
+def get_data(data_dir, data_name, url, data_origin_name):
+    if not os.path.isdir(data_dir):
+        os.system("mkdir " + data_dir)
+    os.chdir(data_dir)
+    if (not os.path.exists(data_name)):
+        import urllib
+        zippath = os.path.join(data_dir, data_origin_name)
+        urllib.urlretrieve(url, zippath)
+        os.system("bzip2 -d %r" % data_origin_name)
+    os.chdir("..")
+
+
+def estimate_density(DATA_PATH, feature_size):
+    """sample 10 times of a size of 1000 for estimating the density of the sparse dataset"""
+    if not os.path.exists(DATA_PATH):
+        raise Exception("Data is not there!")
+    density = []
+    P = 0.01
+    for _ in xrange(10):
+        num_non_zero = 0
+        num_sample = 0
+        with open(DATA_PATH) as f:
+            for line in f:
+                if (random.random() < P):
+                    num_non_zero += len(line.split(" ")) - 1
+                    num_sample += 1
+        density.append(num_non_zero * 1.0 / (feature_size * num_sample))
+    return sum(density) / len(density)
+