added necessary weight sources and non default precision convert for GarNet

Author: joshlerner

.gitignore

@@ -13,3 +13,4 @@ docs/_build
 docs/autodoc/*
 hls4mlprj_*
 *~
+*.ipynb_checkpoints/

hls4ml/backends/vivado/passes/garnet_templates.py

@@ -114,7 +114,8 @@ def format(self, node):
             params[f'{vname}_t'], type_name = node.model.config.get_precision(node, var=vname)
             if type_name.endswith('default_t'):
                 params[f'{vname}_t'] = precision_converter.convert(default_precision).definition_cpp()
-
+            else:
+                params[f'{vname}_t'] = precision_converter.convert(params[f'{vname}_t']).definition_cpp()
         params['output_t'] = node.get_output_variable().type.name
 
         if node.attributes['collapse'] in ['mean', 'max']:

hls4ml/converters/keras/graph.py

@@ -46,13 +46,17 @@ def parse_garnet_layer(keras_layer, input_names, input_shapes, data_reader):
         layer['n_sublayers'] = keras_layer['config']['n_sublayers']
         layer['n_in_features'] = [input_shapes[0][2]]
 
-        for il in range(1, layer['n_sublayers']):
-            layer['n_in_features'].append(layer['n_out_features'][il - 1])
+        for il in range(layer['n_sublayers']):
+            if il > 0:
+                layer['n_in_features'].append(layer['n_out_features'][il - 1])
 
             weights_source = [
-                f'S{il}_kernel',
-                f'S{il}_bias',
-                f'Fout{il}_bias',
+            f'FLR{il}_kernel',
+            f'FLR{il}_bias',
+            f'S{il}_kernel',
+            f'S{il}_bias',
+            f'Fout{il}_kernel',
+            f'Fout{il}_bias',
             ]
             for weight in weights_source:
                 layer[weight + '_data'] = get_weights_data(data_reader, layer['name'], weight)

hls4ml/model/layers.py

@@ -1182,10 +1182,7 @@ def _initialize_transforms(self):
 
     def _make_input_transform_weights(self, n_propagate, n_aggregators, n_out_features, quantize=False, sublayer=''):
         # Due to linearity of the input transform, input weights and biases can be contracted away at conversion time
-
-        output_transform_kernel = self.get_attr(
-            f'Fout{sublayer}_kernel_data'
-        )  # [(n_aggregators, n_propagate), n_out_features]
+        output_transform_kernel = self.get_attr(f'Fout{sublayer}_kernel_data')  # [(n_aggregators, n_propagate), n_out_features]
         output_transform_kernel = output_transform_kernel.reshape((n_aggregators, n_propagate, n_out_features))
         if quantize:
             output_transform_kernel = self.get_attr('quantizer')(output_transform_kernel)