Merge pull request #16817 from dkurt:dnn_onnx_lstm

modules/dnn/src/layers/recurrent_layers.cpp

@@ -93,6 +93,7 @@ class LSTMLayerImpl CV_FINAL : public LSTMLayer
     float forgetBias, cellClip;
     bool useCellClip, usePeephole;
     bool reverse;   // If true, go in negative direction along the time axis
+    bool bidirectional;  // If true, produces both forward and reversed directions along time axis
 
 public:
 
@@ -101,6 +102,7 @@ public:
     {
         setParamsFrom(params);
 
+        bidirectional = params.get<bool>("bidirectional", false);
         if (!blobs.empty())
         {
             CV_Assert(blobs.size() >= 3);
@@ -110,10 +112,11 @@ public:
             const Mat& Wh = blobs[0];
             const Mat& Wx = blobs[1];
             const Mat& bias = blobs[2];
-            CV_Assert(Wh.dims == 2 && Wx.dims == 2);
-            CV_Assert(Wh.rows == Wx.rows);
-            CV_Assert(Wh.rows == 4*Wh.cols);
-            CV_Assert(Wh.rows == (int)bias.total());
+            CV_CheckEQ(Wh.dims, 2, "");
+            CV_CheckEQ(Wx.dims, 2, "");
+            CV_CheckEQ(Wh.rows, Wx.rows, "");
+            CV_CheckEQ(Wh.rows, (1 + static_cast<int>(bidirectional))*4*Wh.cols, "");
+            CV_CheckEQ(Wh.rows, (int)bias.total(), "");
             CV_Assert(Wh.type() == Wx.type() && Wx.type() == bias.type());
 
             // Peephole weights.
@@ -135,6 +138,7 @@ public:
         useCellClip = params.get<bool>("use_cell_clip", false);
         usePeephole = params.get<bool>("use_peephole", false);
         reverse = params.get<bool>("reverse", false);
+        CV_Assert(!reverse || !bidirectional);
 
         allocated = false;
         outTailShape.clear();
@@ -206,6 +210,7 @@ public:
 
         outResShape.push_back(_numSamples);
         outResShape.insert(outResShape.end(), outTailShape_.begin(), outTailShape_.end());
+        outResShape.back() *= (1 + static_cast<int>(bidirectional));
 
         size_t noutputs = produceCellOutput ? 2 : 1;
         outputs.assign(noutputs, outResShape);
@@ -252,6 +257,7 @@ public:
         outTsShape.clear();
         outTsShape.push_back(numSamples);
         outTsShape.insert(outTsShape.end(), outTailShape.begin(), outTailShape.end());
+        outTsShape.back() *= (1 + static_cast<int>(bidirectional));
 
         allocated = true;
     }
@@ -272,9 +278,12 @@ public:
         outputs_arr.getMatVector(output);
         internals_arr.getMatVector(internals);
 
-        const Mat &Wh = blobs[0];
-        const Mat &Wx = blobs[1];
-        const Mat &bias = blobs[2];
+        const int numDirs = 1 + static_cast<int>(bidirectional);
+        for (int i = 0; i < numDirs; ++i)
+        {
+            const Mat &Wh = blobs[0].rowRange(i * blobs[0].rows / numDirs, (i + 1) * blobs[0].rows / numDirs);
+            const Mat &Wx = blobs[1].rowRange(i * blobs[1].rows / numDirs, (i + 1) * blobs[1].rows / numDirs);
+            const Mat &bias = blobs[2].colRange(i * blobs[2].cols / numDirs, (i + 1) * blobs[2].cols / numDirs);
 
             int numOut = Wh.size[1];
 
@@ -288,10 +297,11 @@ public:
             Mat xTs = input[0].reshape(1, numSamplesTotal);
 
             Mat hOutTs = output[0].reshape(1, numSamplesTotal);
+            hOutTs = hOutTs.colRange(i * hOutTs.cols / numDirs, (i + 1) * hOutTs.cols / numDirs);
             Mat cOutTs = produceCellOutput ? output[1].reshape(1, numSamplesTotal) : Mat();
 
             int tsStart, tsEnd, tsInc;
-        if (reverse) {
+            if (reverse || i == 1) {
                 tsStart = numTimeStamps - 1;
                 tsEnd = -1;
                 tsInc = -1;
@@ -359,6 +369,7 @@ public:
                     cInternal.copyTo(cOutTs.rowRange(curRowRange));
             }
         }
+    }
 };
 
 Ptr<LSTMLayer> LSTMLayer::create(const LayerParams& params)

modules/dnn/test/test_onnx_importer.cpp

@@ -405,6 +405,8 @@ TEST_P(Test_ONNX_layers, Reshape)
 
 TEST_P(Test_ONNX_layers, Squeeze)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
     testONNXModels("squeeze");
 }
 
@@ -451,6 +453,16 @@ TEST_P(Test_ONNX_layers, Split_EltwiseMax)
     testONNXModels("split_max");
 }
 
+TEST_P(Test_ONNX_layers, LSTM)
+{
+    testONNXModels("lstm", npy, 0, 0, false, false);
+}
+
+TEST_P(Test_ONNX_layers, LSTM_bidirectional)
+{
+    testONNXModels("lstm_bidirectional", npy, 0, 0, false, false);
+}
+
 INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_ONNX_layers, dnnBackendsAndTargets());
 
 class Test_ONNX_nets : public Test_ONNX_layers