Modified code using LoDTensor

paddle/framework/lod_tensor.cc

@@ -103,25 +103,19 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
   lod_ = new_lod;
 }
 
-Vector<size_t> expand_lod(Vector<size_t> level, Vector<size_t> starts,
+Vector<size_t> expand_lod(Vector<size_t> level, Vector<size_t> indexes,
                           Vector<size_t> scales, bool repeat) {
   Vector<size_t> result;
   result.push_back(level[0]);
-  size_t p = 0, start = 0, end = 0;
+  size_t start = 0, end = 0;
   if (!repeat) {
     for (size_t i = 0; i < scales.size(); ++i) {
       result.push_back(result.back() + scales[i] * (level[i + 1] - level[i]));
     }
   } else {
     for (size_t i = 0; i < scales.size(); ++i) {
-      while (starts[i] != level[p] && p < level.size()) {
-        ++p;
-      }
-      start = p;
-      while (starts[i + 1] != level[p] && p < level.size()) {
-        ++p;
-      }
-      end = p + 1;
+      start = indexes[i];
+      end = indexes[i + 1];
       for (size_t j = 0; j < scales[i]; ++j) {
         for (size_t index = start; index < end - 1; ++index) {
           result.push_back(result.back() + level[index + 1] - level[index]);

paddle/framework/lod_tensor.h

@@ -123,7 +123,7 @@ class LoDTensor : public Tensor {
   LoD lod_;
 };
 
-Vector<size_t> expand_lod(Vector<size_t> level, Vector<size_t> starts,
+Vector<size_t> expand_lod(Vector<size_t> level, Vector<size_t> indexes,
                           Vector<size_t> scales, bool repeat);
 
 }  // namespace framework

paddle/operators/seq_expand_op.cc

@@ -77,15 +77,15 @@ by lod of input(Y) or 'repeat' attribute.
 Case 1:
 
 Given a 2-level LoDTensor X:
-    X.data = [1, 2 , 3, 4]
+    X.data = [a, b , c, d]
     X.lod = [[0, 3, 4], [0, 1, 3, 4]]
 and
     repeat = 2
 then we get 3-level LoDTensor
-    Out.data = [1, 2, 3, 1, 2, 3, 4, 4]
     Out.lod = [[0,                6,    8],
                [0,       3,       6, 7, 8],
                [0, 1,    3, 4,    6, 7, 8]]
+    Out.data = [a, b, c, a, b, c, d, d]
 
 Case 2:
 

paddle/operators/seq_expand_op.h

@@ -33,15 +33,12 @@ class SeqExpandKernel : public framework::OpKernel<T> {
     auto x_dims = x->dims();
     auto x_lod = x->lod();
 
-    if (x_lod.size() == 0) {
     framework::Vector<size_t> level;
-      for (int i = 0; i < x->dims()[0] + 1; ++i) {
+    size_t num = (x_lod.size() == 0) ? (x->dims()[0] + 1) : x_lod[0].size();
+    for (int i = 0; i < num; ++i) {
       level.push_back(i);
     }
     x_lod.push_back(level);
-    } else {
-      x_lod.insert(x_lod.begin(), x_lod[0]);
-    }
 
     size_t repeat = static_cast<size_t>(context.Attr<int>("repeat"));
     framework::Vector<size_t> scales;
@@ -56,19 +53,27 @@ class SeqExpandKernel : public framework::OpKernel<T> {
     } else {
       auto* y = context.Input<LoDTensor>("Y");
       auto y_lod = y->lod();
-      for (int i = 0; i < y_lod[0].size() - 1; ++i) {
-        scales.push_back((y_lod[0][i + 1] - y_lod[0][i]) /
-                         (x_lod[0][i + 1] - x_lod[0][i]));
+      auto y_abs_lod = y_lod.ToAbsOffset();
+      auto x_abs_lod = x_lod.ToAbsOffset();
+      for (int i = 0; i < y_abs_lod[0].size() - 1; ++i) {
+        scales.push_back((y_abs_lod[0][i + 1] - y_abs_lod[0][i]) /
+                         (x_abs_lod[0][i + 1] - x_abs_lod[0][i]));
       }
       out->Resize(y->dims());
     }
 
+    framework::Vector<size_t> indexes;
+    for (int size_t i = 0; i < x_lod[0]; ++i) {
+      indexes[i] = x_lod[0];
+    }
     framework::LoD out_lod;
-    auto level0 = framework::expand_lod(x_lod[0], x_lod[0], scales, false);
+    auto level0 = framework::expand_lod(indexes, x_lod[0], scales, false);
     out_lod.push_back(level0);
     for (int i = 1; i < x_lod.size(); ++i) {
-      out_lod.push_back(
-          framework::expand_lod(x_lod[i], x_lod[0], scales, true));
+      for (int j = 0; j < indexes.size(); ++j) {
+        indexes[j] = x_lod[i - 1][indexes[j]];
+      }
+      out_lod.push_back(framework::expand_lod(x_lod[i], indexes, scales, true));
     }
 
     size_t element_len = framework::product(x_dims) / x_dims[0];
@@ -80,7 +85,7 @@ class SeqExpandKernel : public framework::OpKernel<T> {
     if (platform::is_cpu_place(place)) {
       auto& cpu_place = boost::get<platform::CPUPlace>(place);
       for (size_t i = 0; i < scales.size(); ++i) {
-        count = element_len * (x_lod[0][i + 1] - x_lod[0][i]);
+        count = element_len * (x_abs_lod[0][i + 1] - x_abs_lod[0][i]);
         for (size_t j = 0; j < scales[i]; ++j) {
           memory::Copy(cpu_place, out_data, cpu_place, x_data,
                        sizeof(T) * count);
@@ -95,7 +100,7 @@ class SeqExpandKernel : public framework::OpKernel<T> {
                         context.device_context())
                         .stream();
       for (size_t i = 0; i < scales.size(); ++i) {
-        count = element_len * (x_lod[0][i + 1] - x_lod[0][i]);
+        count = element_len * (x_abs_lod[0][i + 1] - x_abs_lod[0][i]);
         for (size_t j = 0; j < scales[i]; ++j) {
           memory::Copy(gpu_place, out_data, gpu_place, x_data,
                        sizeof(T) * count, stream);
@@ -109,6 +114,11 @@ class SeqExpandKernel : public framework::OpKernel<T> {
     }
 
     out->set_lod(out_lod);
+    for (size_t i = 0; i < lod.size; i++) {
+      for (size_t j = 0; j < lod[i].size(); j++) {
+        LOG(INFO) << "lod[" << i << "][" << j "] = " << lod[i][j];
+      }
+    }
   }
 };
 
@@ -121,13 +131,14 @@ class SeqExpandGradKernel : public framework::OpKernel<T> {
     auto* out = context.Input<LoDTensor>("Out");
     auto* d_x = context.Output<LoDTensor>(framework::GradVarName("X"));
     auto out_lod = out->lod();
+    auto out_abs_lod = out_lod.ToAbsOffset();
     d_x->set_lod(x->lod());
     const T* d_out_data = d_out->data<T>();
     auto d_out_dims = d_out->dims();
     T* d_x_data = d_x->mutable_data<T>(context.GetPlace());
     size_t element_len = framework::product(d_out_dims) / d_out_dims[0];
     for (size_t i = 0; i < out->NumElements(); ++i) {
-      size_t ele_count = out_lod[0][i + 1] - out_lod[0][i];
+      size_t ele_count = out_abs_lod[0][i + 1] - out_abs_lod[0][i];
       size_t repeat = out->NumElements(0, i);
       Eigen::TensorMap<Eigen::Tensor<const T, 2>> d_out_t(
           d_out_data, static_cast<int>(repeat),

python/paddle/v2/framework/tests/op_test.py

@@ -246,6 +246,8 @@ class OpTest(unittest.TestCase):
             else:
                 actual = np.array(self.scope.find_var(out_name).get_tensor())
                 expect = self.outputs[out_name]
+                print "actual= %s" % actual
+                print "expect = %s" % expect
                 self.assertTrue(
                     np.allclose(
                         actual, expect, atol=atol),

python/paddle/v2/framework/tests/test_seq_expand.py

@@ -27,7 +27,15 @@ def repeat_array(array, starts, times):
     return newlist
 
 
+def toAbsOffset(lod):
+    for i in range(len(lod) - 2, -1, -1):
+        for j in range(len(lod[i])):
+            lod[i][j] = lod[i + 1][lod[i][j]]
+    return lod
+
+
 class TestSeqExpand(OpTest):
+    #class TestSeqExpand():
     def set_data(self):
         x_data = np.random.uniform(0.1, 1, [4, 1]).astype('float32')
         self.inputs = {'X': x_data}
@@ -35,23 +43,26 @@ class TestSeqExpand(OpTest):
 
     def compute(self):
         x = self.inputs['X']
+        print "x= %s" % x
         x_data, x_lod = x if type(x) == tuple else (x, None)
-        if not x_lod:
-            x_lod = [[i for i in range(1 + x_data.shape[0])]]
-        else:
-            x_lod = [x_lod[0]] + x_lod
+        n = 1 + x_data.shape[0] if not x_lod else len(x_lod[0])
+        x_lod = [[i for i in range(n)]] + x_lod
+        x_abs_lod = toAbsOffset(x_lod)
         if self.repeat:
+            print "repeat= %s" % self.repeat
             self.attrs = {'repeat': self.repeat}
             repeats = (len(x_lod[0]) - 1) * [self.repeat]
         else:
             y_data, y_lod = self.inputs['Y']
-            repeats = [((y_lod[0][i + 1] - y_lod[0][i]) /
-                        (x_lod[0][i + 1] - x_lod[0][i]))
-                       for i in range(len(y_lod[0]) - 1)]
-        out_lod = [repeat(x_lod[0], x_lod[0], repeats, True)] + [
-            repeat(lod, x_lod[0], repeats, False) for lod in x_lod[1:]
-        ]
-        out = repeat_array(x_data.tolist(), x_lod[0], repeats)
+            print "y_lod: %s" % y_lod
+            y_abs_lod = toAbsOffset(y_lod)
+            repeats = [((y_abs_lod[0][i + 1] - y_abs_lod[0][i]) /
+                        (x_abs_lod[0][i + 1] - x_abs_lod[0][i]))
+                       for i in range(len(y_abs_lod[0]) - 1)]
+        #out_lod = [repeat(x_lod[0], x_lod[0], repeats, True)] + [
+        #    repeat(lod, x_lod[0], repeats, False) for lod in x_lod[1:]
+        #]
+        out = repeat_array(x_data.tolist(), x_abs_lod[0], repeats)
         self.outputs = {'Out': out}
 
     def setUp(self):
@@ -69,7 +80,7 @@ class TestSeqExpand(OpTest):
 class TestSeqExpandCase1(TestSeqExpand):
     def set_data(self):
         x_data = np.random.uniform(0.1, 1, [7, 1]).astype('float32')
-        x_lod = [[0, 5, 7], [0, 2, 5, 7]]
+        x_lod = [[0, 2, 3], [0, 2, 5, 7]]
         self.inputs = {'X': (x_data, x_lod)}
         self.repeat = 2
 
@@ -95,10 +106,11 @@ class TestSeqExpandCase4(TestSeqExpand):
         x_data = np.random.uniform(0.1, 1, [5, 1]).astype('float32')
         x_lod = [[0, 2, 5]]
         y_data = np.random.uniform(0.1, 1, [13, 1]).astype('float32')
-        y_lod = [[0, 4, 13], [0, 2, 4, 7, 10, 13]]
+        y_lod = [[0, 2, 5], [0, 2, 4, 7, 10, 13]]
         self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
         self.repeat = None
 
 
 if __name__ == '__main__':
     unittest.main()
+#    TestSeqExpandCase4().setUp()