update output shape of RNN ops according to ONNX spec (onnx#923)

* update output shape of RNN ops according to ONNX spec

* formatting

* correct test data

onnx/backend/test/case/node/gru.py

@@ -25,17 +25,17 @@ def __init__(self, **params):
         for i in required_inputs:
             assert i in params, "Missing Required Input: {0}".format(i)
 
-        num_directions = params[W].shape[0]
+        self.num_directions = params[W].shape[0]
 
-        if(num_directions == 1):
+        if(self.num_directions == 1):
             for k in params.keys():
                 params[k] = np.squeeze(params[k], axis=0)
 
-            hidden_size = params[R].shape[-1]
-            batch_size = params[X].shape[0]
+            self.hidden_size = params[R].shape[-1]
+            self.batch_size = params[X].shape[0]
 
-            b = params[B] if B in params else np.zeros(2 * number_of_gates * hidden_size)
-            h_0 = params[H_0] if H_0 in params else np.zeros((batch_size, hidden_size))
+            b = params[B] if B in params else np.zeros(2 * number_of_gates * self.hidden_size)
+            h_0 = params[H_0] if H_0 in params else np.zeros((self.batch_size, self.hidden_size))
             lbr = params[LBR] if LBR in params else 0
 
             self.X = params[X]
@@ -65,7 +65,7 @@ def step(self):
         h_linear = self.g(np.dot(self.X, np.transpose(w_h)) + r * (np.dot(self.H_0, r_h) + r_bh) + w_bh)
         h = h_linear if self.LBR else h_default
         H = (1 - z) * h + z * self.H_0
-        return H
+        return np.reshape(H, (self.num_directions, self.batch_size, self.hidden_size))
 
 
 class GRU(Base):

onnx/backend/test/case/node/lstm.py

@@ -27,19 +27,19 @@ def __init__(self, **params):
         for i in required_inputs:
             assert i in params, "Missing Required Input: {0}".format(i)
 
-        num_directions = params[W].shape[0]
+        self.num_directions = params[W].shape[0]
 
-        if(num_directions == 1):
+        if(self.num_directions == 1):
             for k in params.keys():
                 params[k] = np.squeeze(params[k], axis=0)
 
-            hidden_size = params[R].shape[-1]
-            batch_size = params[X].shape[0]
+            self.hidden_size = params[R].shape[-1]
+            self.batch_size = params[X].shape[0]
 
-            b = params[B] if B in params else np.zeros(2 * number_of_gates * hidden_size)
-            p = params[P] if P in params else np.zeros(number_of_peepholes * hidden_size)
-            h_0 = params[H_0] if H_0 in params else np.zeros((batch_size, hidden_size))
-            c_0 = params[C_0] if C_0 in params else np.zeros((batch_size, hidden_size))
+            b = params[B] if B in params else np.zeros(2 * number_of_gates * self.hidden_size)
+            p = params[P] if P in params else np.zeros(number_of_peepholes * self.hidden_size)
+            h_0 = params[H_0] if H_0 in params else np.zeros((self.batch_size, self.hidden_size))
+            c_0 = params[C_0] if C_0 in params else np.zeros((self.batch_size, self.hidden_size))
 
             self.X = params[X]
             self.W = params[W]
@@ -72,7 +72,7 @@ def step(self):
         C = f * self.C_0 + i * c
         o = self.f(np.dot(self.X, np.transpose(w_o)) + np.dot(self.H_0, r_o) + w_bo + r_bo + p_o * C)
         H = o * self.h(C)
-        return H
+        return np.reshape(H, (self.num_directions, self.batch_size, self.hidden_size))
 
 
 class LSTM(Base):

onnx/backend/test/case/node/rnn.py

@@ -23,17 +23,17 @@ def __init__(self, **params):
         for i in required_inputs:
             assert i in params, "Missing Required Input: {0}".format(i)
 
-        num_directions = params[W].shape[0]
+        self.num_directions = params[W].shape[0]
 
-        if(num_directions == 1):
+        if(self.num_directions == 1):
             for k in params.keys():
                 params[k] = np.squeeze(params[k], axis=0)
 
-            hidden_size = params[R].shape[-1]
-            batch_size = params[X].shape[0]
+            self.hidden_size = params[R].shape[-1]
+            self.batch_size = params[X].shape[0]
 
-            b = params[B] if B in params else np.zeros(2 * hidden_size)
-            h_0 = params[H_0] if H_0 in params else np.zeros((batch_size, hidden_size))
+            b = params[B] if B in params else np.zeros(2 * self.hidden_size)
+            h_0 = params[H_0] if H_0 in params else np.zeros((self.batch_size, self.hidden_size))
 
             self.X = params[X]
             self.W = params[W]
@@ -50,7 +50,7 @@ def step(self):
         [w_b, r_b] = np.split(self.B, 2)
 
         H = self.f(np.dot(self.X, np.transpose(self.W)) + np.dot(self.H_0, self.R) + w_b + r_b)
-        return H
+        return np.reshape(H, (self.num_directions, self.batch_size, self.hidden_size))
 
 
 class RNN(Base):

onnx/backend/test/data/node/test_gru_defaults/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�

+backend-test:�

 %
 
X
 
W
@@ -18,8 +18,9 @@
 

 

 
-b
-
Y
-

+b
+
Y
+

+

 
 B

onnx/backend/test/data/node/test_gru_defaults/test_data_set_0/output_0.pb

@@ -1 +1 @@
-
B
YJ<��=��=��=��=��=yYM>yYM>yYM>yYM>yYM>�L>�L>�L>�L>�L>
+

B
YJ<��=��=��=��=��=yYM>yYM>yYM>yYM>yYM>�L>�L>�L>�L>�L>

onnx/backend/test/data/node/test_gru_with_initial_bias/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�

+backend-test:�

 (
 
X
 
W
@@ -23,8 +23,9 @@
 
B
 

 

-b
-
Y
-

+b
+
Y
+

+

 
 B

onnx/backend/test/data/node/test_gru_with_initial_bias/test_data_set_0/output_0.pb

@@ -1 +1 @@
-
B
YJ$yYM>yYM>yYM>�>�>�>)G�=)G�=)G�=
+

B
YJ$yYM>yYM>yYM>�>�>�>)G�=)G�=)G�=

onnx/backend/test/data/node/test_lstm_defaults/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�

+backend-test:�

 &
 
X
 
W
@@ -18,8 +18,9 @@
 

 

 
-b
-
Y
-
+b
+
Y
+
+

 
 B

onnx/backend/test/data/node/test_lstm_defaults/test_data_set_0/output_0.pb

@@ -1 +1 @@
-B
YJH����a�?����a�?����a�?�a{)\c�?�a{)\c�?�a{)\c�?�������?�������?�������?
+
B
YJH����a�?����a�?����a�?�a{)\c�?�a{)\c�?�a{)\c�?�������?�������?�������?

onnx/backend/test/data/node/test_lstm_with_initial_bias/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�

+backend-test:�

 )
 
X
 
W
@@ -23,8 +23,9 @@
 
B
 

 

- b
-
Y
-
+ b
+
Y
+
+

 
 B

onnx/backend/test/data/node/test_lstm_with_initial_bias/test_data_set_0/output_0.pb

@@ -1 +1 @@
-B
YJ`��\c�?��\c�?��\c�?��\c�?Zp��,�?Zp��,�?Zp��,�?Zp��,�?Z?y��Y�?Z?y��Y�?Z?y��Y�?Z?y��Y�?
+
B
YJ`��\c�?��\c�?��\c�?��\c�?Zp��,�?Zp��,�?Zp��,�?Zp��,�?Z?y��Y�?Z?y��Y�?Z?y��Y�?Z?y��Y�?

onnx/backend/test/data/node/test_lstm_with_peepholes/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�
+backend-test:�
 Q
 
X
 
W
@@ -45,8 +45,9 @@ Q
 
P
 

 

-	b
-
Y
-

+	b
+
Y
+

+

 
 B

onnx/backend/test/data/node/test_lstm_with_peepholes/test_data_set_0/output_0.pb

@@ -1 +1 @@
-
B
YJ	�>	�>	�>.?.?.?
+

B
YJ	�>	�>	�>.?.?.?

onnx/backend/test/data/node/test_simple_rnn_defaults/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�

+backend-test:�

 %
 
X
 
W
@@ -18,8 +18,9 @@
 

 

 
-b
-
Y
-

+b
+
Y
+

+

 
 B

onnx/backend/test/data/node/test_simple_rnn_defaults/test_data_set_0/output_0.pb

@@ -1 +1 @@
-
B
YJ0�&�>�&�>�&�>�&�>ٷ?ٷ?ٷ?ٷ?��L?��L?��L?��L?
+

B
YJ0�&�>�&�>�&�>�&�>ٷ?ٷ?ٷ?ٷ?��L?��L?��L?��L?

onnx/backend/test/data/node/test_simple_rnn_with_initial_bias/model.onnx

@@ -1,4 +1,4 @@
-backend-test:�

+backend-test:�

 (
 
X
 
W
@@ -24,8 +24,9 @@
 

 

 
-b
-
Y
-

+b
+
Y
+

+

 
 B

onnx/backend/test/data/node/test_simple_rnn_with_initial_bias/test_data_set_0/output_0.pb

@@ -1 +1 @@
-
B
YJ<ٷ?ٷ?ٷ?ٷ?ٷ?x�k?x�k?x�k?x�k?x�k?��|?��|?��|?��|?��|?
+

B
YJ<ٷ?ٷ?ٷ?ٷ?ٷ?x�k?x�k?x�k?x�k?x�k?��|?��|?��|?��|?��|?