ADD gnerale bidirectional wrapper

Author: Enrico Lupi

hls4ml/model/layers.py

@@ -1538,6 +1538,82 @@ def initialize(self):
         self.add_weights_variable(name='recurrent_bias_b', var_name='br_b{index}')
 
 
+class Bidirectional(Layer):
+    _expected_attributes = [
+        Attribute('n_out'),
+        Attribute('return_sequences', value_type=bool, default=False),
+        Attribute('return_state', value_type=bool, default=False),
+        Attribute('pass_initial_states', value_type=bool, default=False),
+        Attribute('time_major', value_type=bool, default=False),
+        Attribute('forward_activation', value_type=str),
+        Attribute('forward_recurrent_activation', value_type=str),
+        WeightAttribute('forward_weight'),
+        WeightAttribute('forward_bias'),
+        WeightAttribute('forward_recurrent_weight'),
+        WeightAttribute('forward_recurrent_bias'),
+        TypeAttribute('forward_weight'),
+        TypeAttribute('forward_bias'),
+        TypeAttribute('forward_recurrent_weight'),
+        TypeAttribute('forward_recurrent_bias'),
+        Attribute('backward_activation', value_type=str),
+        Attribute('backward_recurrent_activation', value_type=str),
+        WeightAttribute('backward_weight'),
+        WeightAttribute('backward_bias'),
+        WeightAttribute('backward_recurrent_weight'),
+        WeightAttribute('backward_recurrent_bias'),
+        TypeAttribute('backward_weight'),
+        TypeAttribute('backward_bias'),
+        TypeAttribute('backward_recurrent_weight'),
+        TypeAttribute('backward_recurrent_bias'),
+    ]
+
+    def initialize(self):
+        if self.attributes['return_sequences']:
+            shape = [self.attributes['n_timesteps'], self.attributes['n_out']]
+            dims = [f'N_TIME_STEPS_{self.index}', f'N_OUT_{self.index}']
+        else:
+            shape = [self.attributes['n_out']]
+            dims = [f'N_OUT_{self.index}']
+
+        self.add_output_variable(shape, dims)
+
+        if self.attributes['return_state']:
+            state_shape = [self.attributes['n_out']]
+            state_dims = [f'N_OUT_{self.index}']
+            self.add_output_variable(
+                state_shape, state_dims, out_name=self.outputs[1], var_name='layer{index}_h', type_name='layer{index}_h_t'
+            )
+            self.add_output_variable(
+                state_shape, state_dims, out_name=self.outputs[2], var_name='layer{index}_c', type_name='layer{index}_c_t'
+            )
+
+        for dir in ['forward', 'backward']:
+            # weights
+            self.add_weights_variable(name=f'{dir}_weight', var_name=(f'w_{dir[0]}_' + '{index}'))
+
+            # recurrent weights
+            recurrent_weight = self.get_attr(f'{dir}_recurrent_weight_data')
+            self.add_weights_variable(
+                name=f'{dir}_recurrent_weight', var_name=(f'wr_{dir[0]}_' + '{index}'), data=recurrent_weight
+            )
+
+            # biases
+            self.add_weights_variable(name=f'{dir}_bias', var_name=(f'b_{dir[0]}_' + '{index}'))
+
+            if self.attributes[f'{dir}_class_name'] == 'LSTM':
+                if "pytorch" in self.attributes.keys():
+                    self.add_weights_variable(name=f'{dir}_recurrent_bias', var_name=(f'br_{dir[0]}_' + '{index}'))
+                else:
+                    recurrent_bias = np.zeros(recurrent_weight.shape[1])
+                    self.add_weights_variable(
+                        name=f'{dir}_recurrent_bias', var_name=(f'br_{dir[0]}_' + '{index}'), data=recurrent_bias
+                    )
+            else:
+                self.add_weights_variable(
+                    name=f'{dir}_recurrent_bias', var_name=(f'br_{dir[0]}_' + '{index}'), data=recurrent_bias
+                )
+
+
 class GarNet(Layer):
     ref_impl = False
 
@@ -1828,6 +1904,7 @@ def initialize(self):
     'GRU': GRU,
     'BidirectionalLSTM': BidirectionalLSTM,
     'BidirectionalGRU': BidirectionalGRU,
+    'Bidirectional': Bidirectional,
     'QSimpleRNN': SimpleRNN,
     'QLSTM': LSTM,
     'QGRU': GRU,