[Frontend][MxNet] Support bidirectional RNN layer (#3397)

* Support bidirectional RNN layer

* tweak

* tweak

python/tvm/relay/frontend/mxnet.py

@@ -748,13 +748,12 @@ def _mx_rnn_layer(inputs, attrs):
 
     num_layers = attrs.get_int("num_layers", 1)
     mode = attrs.get_str("mode")
+    output_states = attrs.get_bool("state_outputs", False)
     if mode.startswith("rnn"):
         mode, activation = mode.split('_')
     assert mode in ["rnn", "gru", "lstm"]
     bidirectional = attrs.get_bool("bidirectional", False)
-    if bidirectional:
-        raise tvm.error.OpAttributeUnimplemented(
-            "Bidirectional RNN op is not supported yet")
+    direct = 2 if bidirectional else 1
     layout = attrs.get_str("layout", "TNC")
     if layout != "TNC":
         raise tvm.error.OpAttributeUnimplemented(
@@ -765,11 +764,10 @@ def _mx_rnn_layer(inputs, attrs):
     seq_data = inputs[0]
     concat_weight = inputs[1]
     init_states = inputs[2:]
-
     data_shape = ir_pass.infer_type(seq_data).checked_type.shape
     seq_len = int(data_shape[0])
-    assert len(concat_weight) == num_layers * 4
-    output_states = True
+    assert len(concat_weight) == num_layers * 4 * direct
+
     for idx, state in enumerate(init_states[:]):
         if isinstance(state, dict):
             node = state
@@ -787,43 +785,76 @@ def _mx_rnn_layer(inputs, attrs):
                     assert axis >= 0
                     new_shape[i] = int(data_shape[axis])
             init_states[idx] = _op.zeros(new_shape, dtype)
-            output_states = False
 
     weights = []
     bias = []
     states = []
+    back_weights = []
+    back_bias = []
+    back_states = []
     for i in range(num_layers):
-        w = []
-        b = []
+        weights.append([concat_weight[i*2*direct].args[0],
+                        concat_weight[i*2*direct + 1].args[0]])
+        bias.append([concat_weight[(num_layers+i)*2*direct].args[0],
+                     concat_weight[(num_layers+i)*2*direct + 1].args[0]])
         s = []
-        for j in range(2):
-            w.append(concat_weight[i*2 + j].args[0])
-            b.append(concat_weight[num_layers*2 + i*2 + j].args[0])
         for state in init_states:
-            s.append(_op.take(state, _expr.const(i, "int32"), axis=0))
-        weights.append(w)
-        bias.append(b)
+            s.append(_op.take(state, _expr.const(i*direct, "int32"), axis=0))
         states.append(s)
-
-    seq_output = []
-    for t in range(seq_len):
-        data = _op.take(seq_data, _expr.const(t, "int32"), axis=0)
-        for l in range(num_layers):
+        if bidirectional:
+            back_weights.append([concat_weight[i*2*direct + 2].args[0],
+                                 concat_weight[i*2*direct + 3].args[0]])
+            back_bias.append([concat_weight[(num_layers+i)*2*direct + 2].args[0],
+                              concat_weight[(num_layers+i)*2*direct + 3].args[0]])
+            s = []
+            for state in init_states:
+                s.append(_op.take(state, _expr.const(i*direct+1, "int32"), axis=0))
+            back_states.append(s)
+
+    xs = [_op.take(seq_data, _expr.const(t, "int32"), axis=0) for t in range(seq_len)]
+    for l in range(num_layers):
+        outputs = []
+        back_outputs = []
+        for x in xs:
             if mode == "rnn":
-                out, new_states = _rnn_cell(data, states[l], *weights[l], *bias[l], activation)
+                out, new_states = _rnn_cell(x, states[l], *weights[l], *bias[l], activation)
             elif mode == "gru":
-                out, new_states = _gru_cell(data, states[l], *weights[l], *bias[l])
+                out, new_states = _gru_cell(x, states[l], *weights[l], *bias[l])
             else: # mode == "lstm"
-                out, new_states = _lstm_cell(data, states[l], *weights[l], *bias[l])
+                out, new_states = _lstm_cell(x, states[l], *weights[l], *bias[l])
             states[l] = new_states
-            data = out
-        seq_output.append(out)
-
-    outputs = [_op.stack(seq_output, axis=0)]
+            outputs.append(out)
+        if bidirectional:
+            for x in reversed(xs):
+                if mode == "rnn":
+                    out, new_states = _rnn_cell(
+                        x, back_states[l], *back_weights[l], *back_bias[l], activation)
+                elif mode == "gru":
+                    out, new_states = _gru_cell(
+                        x, back_states[l], *back_weights[l], *back_bias[l])
+                else: # mode == "lstm"
+                    out, new_states = _lstm_cell(
+                        x, back_states[l], *back_weights[l], *back_bias[l])
+                back_states[l] = new_states
+                back_outputs.append(out)
+            back_outputs.reverse()
+            concat_outputs = []
+            for t, out in enumerate(outputs):
+                new_out = _op.concatenate([out, back_outputs[t]], axis=-1)
+                concat_outputs.append(new_out)
+            outputs = concat_outputs
+        xs = outputs
+
+    ret = [_op.stack(outputs, axis=0)]
     if output_states:
         for i in range(num_states):
-            outputs.append(_op.stack([s[i] for s in states], axis=0))
-    return outputs
+            inputs = []
+            for l, s in enumerate(states):
+                inputs.append(s[i])
+                if bidirectional:
+                    inputs.append(back_states[l][i])
+            ret.append(_op.stack(inputs, axis=0))
+    return ret
 
 
 # Note: due to attribute conversion constraint

tests/python/frontend/mxnet/test_forward.py

@@ -536,29 +536,31 @@ def test_forward_bilinear_resize():
     verify_mxnet_frontend_impl(mx_sym, (1, 2, 3, 4), (1, 2, 5, 10))
 
 def test_forward_rnn_layer():
-    def verify(mode, input_size, seq_len, hidden_size, num_layers, init_states=True):
+    def verify(mode, seq_len, input_size, hidden_size, num_layers,
+               batch=1, init_states=True, bidirectional=False):
         if mode == "rnn":
-            layer = gluon.rnn.RNN(hidden_size, num_layers)
+            layer = gluon.rnn.RNN(hidden_size, num_layers, bidirectional=bidirectional)
         elif mode == "gru":
-            layer = gluon.rnn.GRU(hidden_size, num_layers)
+            layer = gluon.rnn.GRU(hidden_size, num_layers, bidirectional=bidirectional)
         else: # mode == "lstm"
-            layer = gluon.rnn.LSTM(hidden_size, num_layers)
+            layer = gluon.rnn.LSTM(hidden_size, num_layers, bidirectional=bidirectional)
         num_states = 2 if mode == "lstm" else 1
         layer.initialize()
         layer.hybridize()
 
         dtype = "float32"
-        batch = 1
+        directions = 2 if bidirectional else 1
         data_np = np.random.uniform(size=(seq_len, batch, input_size)).astype(dtype)
         data_mx = mx.nd.array(data_np)
 
         if init_states:
             shape_dict = {'data0': data_np.shape}
             inputs = {'data0': data_np}
+            state_shape = (num_layers*directions, batch, hidden_size)
             states_np = []
             states_mx = []
             for i in range(num_states):
-                s = np.random.uniform(size=(num_layers, batch, hidden_size)).astype(dtype)
+                s = np.random.uniform(size=state_shape).astype(dtype)
                 states_np.append(s)
                 states_mx.append(mx.nd.array(s))
                 shape_dict['data%s' % (i+1)] = s.shape
@@ -592,10 +594,13 @@ def test_forward_rnn_layer():
                         op_res.asnumpy(), mx_res.asnumpy(), rtol=1e-3)
 
     for mode in ["rnn", "gru", "lstm"]:
-        verify(mode, 64, 10, 64, 1)
-        verify(mode, 64, 10, 64, 2)
-        verify(mode, 64, 10, 32, 2)
-        verify(mode, 64, 10, 64, 2, init_states=False)
+        verify(mode, 1, 64, 64, 1)
+        verify(mode, 10, 64, 64, 2)
+        verify(mode, 10, 64, 32, 2)
+        verify(mode, 10, 64, 32, 2, batch=2)
+        verify(mode, 10, 64, 64, 3, init_states=False)
+        verify(mode, 10, 32, 64, 1, bidirectional=True)
+        verify(mode, 10, 64, 64, 3, batch=2, bidirectional=True, init_states=False)
 
 def test_forward_Crop():
     def verify(xshape, yshape, offset=None):