make tf and qkeras optionl, stop assuming keras is tf.keras

Author: calad0i

hls4ml/converters/__init__.py

@@ -93,10 +93,10 @@ def parse_yaml_config(config_file):
     """
 
     def construct_keras_model(loader, node):
-        from tensorflow.keras.models import load_model
-
         model_str = loader.construct_scalar(node)
-        return load_model(model_str)
+        import keras
+
+        return keras.models.load_model(model_str)
 
     yaml.add_constructor('!keras_model', construct_keras_model, Loader=yaml.SafeLoader)
 

hls4ml/model/optimizer/passes/qkeras.py

@@ -1,5 +1,4 @@
 import numpy as np
-import tensorflow as tf
 
 from hls4ml.model.layers import ApplyAlpha
 from hls4ml.model.optimizer import ConfigurableOptimizerPass, OptimizerPass, register_pass
@@ -113,6 +112,8 @@ def match(self, node):
     def transform(self, model, node):
         # The quantizer has to be applied to set the scale attribute
         # This must be applied to the _unquantized_ weights to obtain the correct scale
+        import tensorflow as tf
+
         quantizer = node.weights['weight'].quantizer.quantizer_fn  # get QKeras quantizer
         weights = node.weights['weight'].data_unquantized  # get weights
         qweights = quantizer(tf.convert_to_tensor(weights))

hls4ml/model/profiling.py

@@ -13,12 +13,11 @@
 from hls4ml.model.layers import GRU, LSTM, SeparableConv1D, SeparableConv2D
 
 try:
-    import qkeras
-    from tensorflow import keras
+    import keras
 
-    __tf_profiling_enabled__ = True
+    __keras_profiling_enabled__ = True
 except ImportError:
-    __tf_profiling_enabled__ = False
+    __keras_profiling_enabled__ = False
 
 try:
     import torch
@@ -27,6 +26,19 @@
 except ImportError:
     __torch_profiling_enabled__ = False
 
+try:
+    import qkeras
+
+    __qkeras_profiling_enabled__ = True
+except ImportError:
+    __qkeras_profiling_enabled__ = False
+
+_activations = list()
+if __keras_profiling_enabled__:
+    _activations.append(keras.layers.Activation)
+if __qkeras_profiling_enabled__:
+    _activations.append(qkeras.qactivations)
+
 
 def get_unoptimized_hlsmodel(model):
     from hls4ml.converters import convert_from_config
@@ -482,7 +494,7 @@ def numerical(model=None, hls_model=None, X=None, plot='boxplot'):
     if hls_model_present:
         data = weights_hlsmodel(hls_model_unoptimized, fmt='summary', plot=plot)
     elif model_present:
-        if __tf_profiling_enabled__ and isinstance(model, keras.Model):
+        if __keras_profiling_enabled__ and isinstance(model, keras.Model):
             data = weights_keras(model, fmt='summary', plot=plot)
         elif __torch_profiling_enabled__ and isinstance(model, torch.nn.Sequential):
             data = weights_torch(model, fmt='summary', plot=plot)
@@ -520,7 +532,7 @@ def numerical(model=None, hls_model=None, X=None, plot='boxplot'):
     if X is not None:
         print("Profiling activations" + before)
         data = None
-        if __tf_profiling_enabled__ and isinstance(model, keras.Model):
+        if __keras_profiling_enabled__ and isinstance(model, keras.Model):
             data = activations_keras(model, X, fmt='summary', plot=plot)
         elif __torch_profiling_enabled__ and isinstance(model, torch.nn.Sequential):
             data = activations_torch(model, X, fmt='summary', plot=plot)
@@ -590,7 +602,7 @@ def get_ymodel_keras(keras_model, X):
         if (
             hasattr(layer, 'activation')
             and layer.activation is not None
-            and not isinstance(layer, (keras.layers.Activation, qkeras.qlayers.QActivation))
+            and not isinstance(layer, _activations)
             and layer.activation.__name__ != 'linear'
         ):
             tmp_activation = layer.activation

hls4ml/optimization/dsp_aware_pruning/keras/__init__.py

@@ -4,9 +4,6 @@
 import numpy as np
 import tensorflow as tf
 
-# Enables printing of loss tensors during custom training loop
-from tensorflow.python.ops.numpy_ops import np_config
-
 import hls4ml.optimization.dsp_aware_pruning.keras.utils as utils
 from hls4ml.optimization.dsp_aware_pruning.config import SUPPORTED_STRUCTURES
 from hls4ml.optimization.dsp_aware_pruning.keras.builder import build_optimizable_model, remove_custom_regularizers
@@ -15,7 +12,6 @@
 from hls4ml.optimization.dsp_aware_pruning.keras.reduction import reduce_model
 from hls4ml.optimization.dsp_aware_pruning.scheduler import OptimizationScheduler
 
-np_config.enable_numpy_behavior()
 default_regularization_range = np.logspace(-6, -2, num=16).tolist()
 
 

hls4ml/utils/config.py

@@ -1,7 +1,5 @@
 import json
 
-import qkeras
-
 import hls4ml
 
 
@@ -48,6 +46,8 @@ def create_config(output_dir='my-hls-test', project_name='myproject', backend='V
 
 def _get_precision_from_quantizer(quantizer):
     if isinstance(quantizer, str):
+        import qkeras
+
         quantizer_obj = qkeras.get_quantizer(quantizer)
         quantizer = {}
         # Some activations are classes with get_config method

hls4ml/writer/catapult_writer.py

@@ -889,7 +889,9 @@ def keras_model_representer(dumper, keras_model):
             return dumper.represent_scalar('!keras_model', model_path)
 
         try:
-            from tensorflow.keras import Model as KerasModel
+            import keras
+
+            KerasModel = keras.models.Model
 
             yaml.add_multi_representer(KerasModel, keras_model_representer)
         except Exception:

hls4ml/writer/quartus_writer.py

@@ -1327,7 +1327,9 @@ def keras_model_representer(dumper, keras_model):
             return dumper.represent_scalar('!keras_model', model_path)
 
         try:
-            from tensorflow.keras import Model as KerasModel
+            import keras
+
+            KerasModel = keras.models.Model
 
             yaml.add_multi_representer(KerasModel, keras_model_representer)
         except Exception:

hls4ml/writer/vivado_writer.py

@@ -815,7 +815,9 @@ def keras_model_representer(dumper, keras_model):
             return dumper.represent_scalar('!keras_model', model_path)
 
         try:
-            from tensorflow.keras import Model as KerasModel
+            import keras
+
+            KerasModel = keras.models.Model
 
             yaml.add_multi_representer(KerasModel, keras_model_representer)
         except Exception: