Support UpSampling1D (#475)

* Support UpSampling1D

* Proper output directory for upsampling tests

Co-authored-by: Javier Duarte <[email protected]>

Author: vloncar

hls4ml/converters/keras/reshape.py

@@ -27,9 +27,34 @@ def parse_reshape_layer(keras_layer, input_names, input_shapes, data_reader, con
 
     return layer, output_shape
 
+@keras_handler('UpSampling1D')
+def parse_upsampling1d_layer(keras_layer, input_names, input_shapes, data_reader, config):
+    assert('UpSampling' in keras_layer['class_name'])
+
+    layer = parse_default_keras_layer(keras_layer, input_names)
+
+    layer['in_height'] = 1
+    (
+        layer['in_width'],
+        layer['n_chan']
+    ) = parse_data_format(input_shapes[0], layer['data_format'])
+
+    layer['algorithm'] = 'nearest'
+
+    layer['width_factor'] = keras_layer['config']['size']
+
+    layer['out_height'] = 1
+    layer['out_width'] = layer['in_width'] * layer['width_factor']
+
+    if layer['data_format'] == 'channels_first':
+        output_shape = [input_shapes[0][0], layer['n_chan'], layer['out_width']]
+    else:
+        output_shape = [input_shapes[0][0], layer['out_width'], layer['n_chan']]
+
+    return layer, output_shape
 
 @keras_handler('UpSampling2D')
-def parse_conv2d_layer(keras_layer, input_names, input_shapes, data_reader, config):
+def parse_upsampling2d_layer(keras_layer, input_names, input_shapes, data_reader, config):
     assert('UpSampling2D' in keras_layer['class_name'])
 
     layer = parse_default_keras_layer(keras_layer, input_names)

hls4ml/model/layers.py

@@ -783,8 +783,12 @@ def initialize(self):
 class Resize(Layer):
     def initialize(self):
         inp = self.get_input_variable()
-        shape = [self.get_attr('out_height'), self.get_attr('out_width'), self.get_attr('n_chan')]
-        dims = ['OUT_HEIGHT_{}'.format(self.index), 'OUT_WIDTH_{}'.format(self.index), 'N_CHAN_{}'.format(self.index)]
+        if len(inp.shape) == 2: # 1D -> width + chan
+            shape = [self.get_attr('out_width'), self.get_attr('n_chan')]
+            dims = ['OUT_WIDTH_{}'.format(self.index), 'N_CHAN_{}'.format(self.index)]
+        elif len(inp.shape) == 3: # 2D -> height + width + chan
+            shape = [self.get_attr('out_height'), self.get_attr('out_width'), self.get_attr('n_chan')]
+            dims = ['OUT_HEIGHT_{}'.format(self.index), 'OUT_WIDTH_{}'.format(self.index), 'N_CHAN_{}'.format(self.index)]
         self.add_output_variable(shape, dims, precision=inp.type.precision)
 
 class Transpose(Layer):
@@ -1012,6 +1016,7 @@ def _initialize_transforms(self):
     'Dot'                    : Dot,
     'Concatenate'            : Concatenate,
     'Resize'                 : Resize,
+    'UpSampling1D'           : Resize,
     'UpSampling2D'           : Resize,
     'Transpose'              : Transpose,
     'GarNet'                 : GarNet,

hls4ml/utils/config.py

@@ -100,20 +100,21 @@ def config_from_keras_model(model, granularity='model', default_precision='ap_fi
         model_arch = json.loads(model.to_json())
 
     #Define supported layers
-    core_layers = ['InputLayer', 'Dropout', 'Flatten', 'Reshape', 'Permute', 'UpSampling2D']
+    core_layers = ['InputLayer', 'Dropout', 'Flatten', 'Reshape', 'Permute']
     dense_layers = ['Dense', 'BinaryDense', 'TernaryDense']
     conv_layers = ['Conv1D', 'Conv2D', 'BinaryConv2D']
     pooling_layers = ['MaxPooling1D', 'MaxPooling2D', 'GlobalMaxPooling1D', 'GlobalMaxPooling2D', 'AveragePooling1D', 'AveragePooling2D', 'GlobalAveragePooling1D', 'GlobalAveragePooling2D']
     norm_layers = ['BatchNormalization']
     activation_layers = ['Activation', 'LeakyReLU', 'ThresholdedReLU', 'ELU', 'PReLU', 'Softmax', 'ReLU']
     merge_layers = ['Add', 'Subtract', 'Multiply', 'Average', 'Maximum', 'Minimum', 'Concatenate', 'Dot']
     qkeras_layers = ['QDense', 'QActivation', 'QConv1D', 'QConv2D', 'QBatchNormalization', 'QConv2DBatchnorm']
+    upsampling_layers = ['UpSampling1D', 'UpSampling2D']
     reshaping_layers = ['ZeroPadding1D', 'ZeroPadding2D']
     graph_layers = ['GarNet', 'GarNetStack']
     #Define layers to skip because they're not configurable or not converted to HLS
     skip_layers = ['Dropout', 'Flatten', 'Reshape', 'Permute']
     #All supported layers
-    supported_layers = core_layers + dense_layers + conv_layers + pooling_layers + norm_layers + activation_layers + merge_layers + qkeras_layers + reshaping_layers + graph_layers + skip_layers
+    supported_layers = core_layers + dense_layers + conv_layers + pooling_layers + norm_layers + activation_layers + merge_layers + qkeras_layers + upsampling_layers + reshaping_layers + graph_layers + skip_layers
 
     keras_layer_config = None
     if model_arch['class_name'] == 'Sequential':

test/pytest/test_upsampling.py

@@ -0,0 +1,67 @@
+import pytest
+from tensorflow.keras.models import Sequential
+from tensorflow.keras.layers import UpSampling1D, UpSampling2D
+import numpy as np
+import hls4ml
+from pathlib import Path
+
+test_root_path = Path(__file__).parent
+
+in_height = 6
+in_width = 8
+in_feat = 4
+
+size = 2
+atol = 5e-3
+
+@pytest.fixture(scope='module')
+def data_1d():
+    X = np.random.rand(100, in_width, in_feat)
+    return X
+
+@pytest.fixture(scope='module')
+def data_2d():
+    X = np.random.rand(100, in_height, in_width, in_feat)
+    return X
+
+
+@pytest.fixture(scope='module')
+def keras_model_1d():
+    model = Sequential()
+    model.add(UpSampling1D(input_shape=(in_width, in_feat), size=size))
+    model.compile()
+    return model
+
+@pytest.fixture(scope='module')
+def keras_model_2d():
+    model = Sequential()
+    model.add(UpSampling2D(input_shape=(in_height, in_width, in_feat), size=(size, size)))
+    model.compile()
+    return model
+
+
+@pytest.mark.parametrize('io_type', ['io_parallel', 'io_stream'])
+@pytest.mark.parametrize('model_type', ['1d', '2d'])
+def test_upsampling(keras_model_1d, keras_model_2d, data_1d, data_2d, model_type, io_type):
+    if model_type == '1d':
+        model = keras_model_1d
+        data = data_1d
+    else:
+        model = keras_model_2d
+        data = data_2d
+
+    config = hls4ml.utils.config_from_keras_model(model,
+                                                  default_precision='ap_fixed<32,1>',
+                                                  granularity='name')
+    odir = str(test_root_path / f'hls4mlprj_upsampling_{model_type}_{io_type}')
+    hls_model = hls4ml.converters.convert_from_keras_model(model,
+                                                           hls_config=config,
+                                                           io_type=io_type,
+                                                           output_dir=odir,
+                                                           part='xcvu9p-flgb2104-2-i')
+    hls_model.compile()
+
+    # Predict
+    y_keras = model.predict(data).flatten()
+    y_hls = hls_model.predict(data).flatten()
+    np.testing.assert_allclose(y_keras, y_hls, rtol=0, atol=atol, verbose=True)