[v1.x] Backport of Fix LSTM and GRU layers gradient calculations (#18203

)

* Fix input gradient calculation for bidirectional LSTM

For bidiractional LSTM with number of layers > 2 input gradient calculation was incorrect.
Reason of wrong calculations was overwriting y derivative (dy) tensor by
calculated x derivative (dx) tensor before right2left layer could use dy for own
gradient calculations.
Propsed fix uses additional space to avoid overwriting.

* Fix gradient calculation for GRU

For GRU with number of layers > 2 i2h_weight gradient for
layers in the middle (all except last and first) was incorrect.
Wrong caluculations were caused by assigning output pointer to
input instead of calculating new input pointer.

* Enable tests for GRU and LSTM gradients

* Fix comments

* Change loop iteration deduction

* Add more test cases for fused rnn layers

src/operator/rnn-inl.h

@@ -194,7 +194,9 @@ inline size_t GetRNNWorkspaceSize(index_t seq_length,
     case rnn_enum::kLstm:
       size = seq_length * batch_size * hidden_size * (4 + direction) +  // wx*x + inter-y
           batch_size * hidden_size * 6 +                                // wh*h + h + c
-          seq_length * hidden_size * 8;                    // Used in Backward, Δbx, Δbh
+          seq_length * hidden_size * 8 +                   // Used in Backward, Δbx, Δbh
+          // temporary dy in backward computation for bidirectional layers
+          seq_length * batch_size * hidden_size * (direction - 1 ? direction : 0);
       break;
     case rnn_enum::kGru:
       // Differs with Lstm, the outputs of three gates are also held in memory

src/operator/rnn_impl.h

@@ -559,6 +559,7 @@ void LstmBackward(DType* ws,
   const index_t w_size1 = (I + H) * H * 4;      // first layer
   const index_t w_size2 = (D * H + H) * H * 4;  // other layers
   const index_t cell_size = N * H;
+  const index_t y_size = T * N * H * D;
   DType* dy_tmp_ptr = ws2 + T * cell_size * 4 + cell_size * 3;
   for (int i = L - 1; i >= 0; --i) {
     const index_t input_size = i ? H * D : I;
@@ -589,6 +590,10 @@ void LstmBackward(DType* ws,
                                      x, hx[idx], cx[idx], y, dy, dx, dhx[idx], dcx[idx],
                                      dhy_cur_ptr, dcy_cur_ptr, w_cur_ptr, dw_cur_ptr, db_cur_ptr,
                                      req_data, req_params, req_state, req_statecell);
+
+      // Prevent overwritting dy while calculating dx in left2right layer
+      const int loop_iteration = (L - 1) - i;
+      dy_tmp_ptr = loop_iteration % 2 ? dy_tmp_ptr - y_size : dy_tmp_ptr + y_size;
     }
     if (dropout > 0.0f && i > 0 && req_data != kNullOp) {
       dropout_random = dropout_random - T * N * D * H;
@@ -1502,7 +1507,7 @@ void GruBackward(DType* ws,
       if (dhy_l)
         dhy_l = dhy_l - D * N * H;
       y_l = y_l - T * N * H * D;
-      y_tmp = y_l;
+      y_tmp = y_tmp - T * N * H * D;
       if (l == 1) {
         wx_l = wx_l - (inputsize + H) * H * 3 * D;
         wh_l = wx_l + inputsize * 3 * H;

tests/python/unittest/test_gluon_rnn.py

@@ -685,15 +685,10 @@ def check_rnn_consistency(fused_layer, stack_layer, loss, input_size, hidden_siz
     stack_input_grad = sx.grad.asnumpy()
 
     assert_allclose(fused_out.asnumpy(), stack_out.asnumpy(), rtol=rtol, atol=atol)
-    if mx.context.current_context().device_type == 'cpu' and \
-            not mx.runtime.Features().is_enabled('MKLDNN') and \
-            'rnn' not in fused_layer.prefix:
-        print("LSTM and GRU on native CPU give wrong gradients. "
-              "Tracking issue: https://github.com/apache/incubator-mxnet/issues/17898.")
-    else:
-        assert_allclose(fused_input_grad, stack_input_grad, rtol=rtol, atol=atol)
-        for key, value in fused_grads.items():
-            assert_allclose(value.asnumpy(), stack_grads[key].asnumpy(), rtol=rtol, atol=atol)
+    assert_allclose(fused_input_grad, stack_input_grad, rtol=rtol, atol=atol)
+    for key, value in fused_grads.items():
+        assert_allclose(value.asnumpy(), stack_grads[key].asnumpy(), rtol=rtol, atol=atol)
+
     num_layers = fused_begin_state[0].shape[0] // (2 if bidirectional else 1)
     check_rnn_states(fused_states, stack_states, num_layers, bidirectional, len(fused_begin_state) == 2)
 
@@ -719,61 +714,32 @@ def create_op_by_mode(mode):
     return fused_op, stack_op, recurrent_block_prefix
 
 
-def check_rnn_unidir_layer_gradients(mode, input_size, hidden_size, loss):
+def check_rnn_unidir_layer_gradients(mode, input_size, hidden_size, num_layers, loss):
     fused_op, stack_op, recurrent_block_prefix = create_op_by_mode(mode)
-    # ==== Single layer ====
-    fused_layer = fused_op(hidden_size, num_layers=1, layout='NTC', bidirectional=False, prefix=recurrent_block_prefix)
-    fused_layer.initialize()
-
-    stack_layer = mx.gluon.rnn.HybridSequentialRNNCell(prefix=recurrent_block_prefix)
-    with stack_layer.name_scope():
-        stack_layer.add(stack_op(hidden_size, prefix='l0_'))
-    stack_layer.initialize()
 
-    check_rnn_consistency(fused_layer, stack_layer, loss, input_size, hidden_size)
-
-    # ==== Multiple layer ====
-    fused_layer = fused_op(hidden_size, num_layers=3, layout='NTC', bidirectional=False, prefix=recurrent_block_prefix)
+    fused_layer = fused_op(hidden_size, num_layers=num_layers, layout='NTC', bidirectional=False, prefix=recurrent_block_prefix)
     fused_layer.initialize()
 
     stack_layer = mx.gluon.rnn.HybridSequentialRNNCell(prefix=recurrent_block_prefix)
     with stack_layer.name_scope():
-        stack_layer.add(stack_op(hidden_size, prefix='l0_'))
-        stack_layer.add(stack_op(hidden_size, prefix='l1_'))
-        stack_layer.add(stack_op(hidden_size, prefix='l2_'))
+        for n in range(num_layers):
+            stack_layer.add(stack_op(hidden_size, prefix="l{}_".format(n)))
     stack_layer.initialize()
-
     check_rnn_consistency(fused_layer, stack_layer, loss, input_size, hidden_size)
 
 
-def check_rnn_bidir_layer_gradients(mode, input_size, hidden_size, loss):
+def check_rnn_bidir_layer_gradients(mode, input_size, hidden_size, num_layers, loss):
     fused_op, stack_op, recurrent_block_prefix = create_op_by_mode(mode)
-    # ==== Single layer ====
-    fused_layer = fused_op(hidden_size, num_layers=1, layout='NTC', bidirectional=True, prefix=recurrent_block_prefix)
-    fused_layer.initialize()
-
-    stack_layer = mx.gluon.rnn.HybridSequentialRNNCell(prefix=recurrent_block_prefix)
-    with stack_layer.name_scope():
-        stack_layer.add(gluon.rnn.BidirectionalCell(stack_op(hidden_size, prefix='l0_'),
-                                                    stack_op(hidden_size, prefix='r0_')))
-    stack_layer.initialize()
 
-    check_rnn_consistency(fused_layer, stack_layer, loss, input_size, hidden_size, bidirectional=True)
-
-    # ==== Multiple layer ====
-    fused_layer = fused_op(hidden_size, num_layers=3, layout='NTC', bidirectional=True, prefix=recurrent_block_prefix)
+    fused_layer = fused_op(hidden_size, num_layers=num_layers, layout='NTC', bidirectional=True, prefix=recurrent_block_prefix)
     fused_layer.initialize()
 
     stack_layer = mx.gluon.rnn.HybridSequentialRNNCell(prefix=recurrent_block_prefix)
     with stack_layer.name_scope():
-        stack_layer.add(gluon.rnn.BidirectionalCell(stack_op(hidden_size, prefix='l0_'),
-                                                    stack_op(hidden_size, prefix='r0_')))
-        stack_layer.add(gluon.rnn.BidirectionalCell(stack_op(hidden_size, prefix='l1_'),
-                                                    stack_op(hidden_size, prefix='r1_')))
-        stack_layer.add(gluon.rnn.BidirectionalCell(stack_op(hidden_size, prefix='l2_'),
-                                                    stack_op(hidden_size, prefix='r2_')))
-    stack_layer.initialize()
-
+        for n in range(num_layers):
+            stack_layer.add(gluon.rnn.BidirectionalCell(stack_op(hidden_size, prefix="l{}_".format(n)),
+                                                stack_op(hidden_size, prefix="r{}_".format(n))))
+        stack_layer.initialize()
     check_rnn_consistency(fused_layer, stack_layer, loss, input_size, hidden_size, bidirectional=True)
 
 
@@ -782,43 +748,47 @@ def check_rnn_bidir_layer_gradients(mode, input_size, hidden_size, loss):
 def test_fused_lstm_layer():
     input_sizes = [8]
     hidden_sizes = [8, 16]
-    for input_size, hidden_size in product(input_sizes, hidden_sizes):
+    num_layers = [1, 2, 3, 4]
+    for input_size, hidden_size, num_layers in product(input_sizes, hidden_sizes, num_layers):
         loss = mx.gluon.loss.L2Loss()
-        check_rnn_unidir_layer_gradients('lstm', input_size, hidden_size, loss)
-        check_rnn_bidir_layer_gradients('lstm', input_size, hidden_size, loss)
+        check_rnn_unidir_layer_gradients('lstm', input_size, hidden_size, num_layers, loss)
+        check_rnn_bidir_layer_gradients('lstm', input_size, hidden_size, num_layers, loss)
 
 
 @with_seed()
 @assert_raises_cudnn_not_satisfied(min_version='5.1.10')
 def test_fused_gru_layer():
     input_sizes = [8]
     hidden_sizes = [8, 16]
-    for input_size, hidden_size in product(input_sizes, hidden_sizes):
+    num_layers = [1, 2, 3, 4]
+    for input_size, hidden_size, num_layers in product(input_sizes, hidden_sizes, num_layers):
         loss = mx.gluon.loss.L2Loss()
-        check_rnn_unidir_layer_gradients('gru', input_size, hidden_size, loss)
-        check_rnn_bidir_layer_gradients('gru', input_size, hidden_size, loss)
+        check_rnn_unidir_layer_gradients('gru', input_size, hidden_size, num_layers, loss)
+        check_rnn_bidir_layer_gradients('gru', input_size, hidden_size, num_layers, loss)
 
 
 @with_seed()
 @assert_raises_cudnn_not_satisfied(min_version='5.1.10')
 def test_fused_rnnrelu_layer():
     input_sizes = [8]
     hidden_sizes = [8, 16]
-    for input_size, hidden_size in product(input_sizes, hidden_sizes):
+    num_layers = [1, 2, 3, 4]
+    for input_size, hidden_size, num_layers in product(input_sizes, hidden_sizes, num_layers):
         loss = mx.gluon.loss.L2Loss()
-        check_rnn_unidir_layer_gradients('rnn_relu', input_size, hidden_size, loss)
-        check_rnn_bidir_layer_gradients('rnn_relu', input_size, hidden_size, loss)
+        check_rnn_unidir_layer_gradients('rnn_relu', input_size, hidden_size, num_layers, loss)
+        check_rnn_bidir_layer_gradients('rnn_relu', input_size, hidden_size, num_layers, loss)
 
 
 @with_seed()
 @assert_raises_cudnn_not_satisfied(min_version='5.1.10')
 def test_fused_rnntanh_layer():
     input_sizes = [8]
     hidden_sizes = [8, 16]
-    for input_size, hidden_size in product(input_sizes, hidden_sizes):
+    num_layers = [1, 2, 3, 4]
+    for input_size, hidden_size, num_layers in product(input_sizes, hidden_sizes, num_layers):
         loss = mx.gluon.loss.L2Loss()
-        check_rnn_unidir_layer_gradients('rnn_tanh', input_size, hidden_size, loss)
-        check_rnn_bidir_layer_gradients('rnn_tanh', input_size, hidden_size, loss)
+        check_rnn_unidir_layer_gradients('rnn_tanh', input_size, hidden_size, num_layers, loss)
+        check_rnn_bidir_layer_gradients('rnn_tanh', input_size, hidden_size, num_layers, loss)
 
 
 def test_rnn_unroll_variant_length():