Merge pull request #4 from kengz/initfn

allow arc.init to use dict for init kwargs

README.md

@@ -83,7 +83,10 @@ arc = {
     'batch_norm': True,
     'activation': 'ReLU',
     'dropout': 0.2,
-    'init': 'kaiming_uniform_',
+    'init': {
+      'type': 'normal_',
+      'std': 0.01,
+    },
 }
 model = torcharc.build(arc)
 

setup.py

@@ -35,7 +35,7 @@ def run_tests(self):
 
 setup(
     name='torcharc',
-    version='0.0.4',
+    version='0.0.5',
     description='Build PyTorch networks by specifying architectures.',
     long_description='https://github.com/kengz/torcharc',
     keywords='torcharc',
@@ -57,6 +57,8 @@ def run_tests(self):
     extras_require={},
     classifiers=[],
     tests_require=[
+        'autopep8==1.5.3',
+        'flake8==3.8.3',
         'flaky==3.6.1',
         'pytest==5.4.1',
         'pytest-cov==2.8.1',

test/test_module_builder.py

@@ -88,6 +88,7 @@ def test_build_module(arc, nn_class):
     'kaiming_uniform_',
     'kaiming_normal_',
     'orthogonal_',
+    {'type': 'normal_', 'std': 0.01},  # when using init kwargs
 ])
 @pytest.mark.parametrize('activation', [
     'ReLU',

torcharc/arc_ref.py

@@ -43,7 +43,10 @@
         'batch_norm': True,
         'activation': 'ReLU',
         'dropout': 0.2,
-        'init': 'kaiming_uniform_',
+        'init': {
+            'type': 'normal_',
+            'std': 0.01,
+        },
     },
     'tstransformer': {
         'type': 'TSTransformer',

torcharc/module_builder.py

@@ -24,23 +24,34 @@
 setattr(torch.optim, 'RAdam', optim.RAdam)
 
 
-def get_init_fn(init: str, activation: Optional[str] = None) -> Callable:
+def get_init_fn(init: Union[str, dict], activation: Optional[str] = None) -> Callable:
     '''Get init function that can be called as `module.apply(init_fn)`. Initializes weights only. Internally this also takes care of gain and nonlinearity args. Ref: https://pytorch.org/docs/stable/nn.init.html'''
     def init_fn(module: nn.Module) -> None:
-        fn = getattr(nn.init, init)
+        if init is None:
+            return
+        elif ps.is_string(init):
+            init_type = init
+            init_kwargs = {}
+        else:
+            assert ps.is_dict(init)
+            init_type = init['type']
+            init_kwargs = ps.omit(init, 'type')
+        fn = getattr(nn.init, init_type)
         args = inspect.getfullargspec(fn).args
         try:
             try:
                 # first try with gain/activation args
                 if 'gain' in args:
                     gain = nn.init.calculate_gain(activation)
-                    fn(module.weight, gain=gain)
+                    ext_init_kwargs = {'gain': gain, **init_kwargs}
+                    fn(module.weight, **ext_init_kwargs)
                 elif 'nonlinearity' in args:
-                    fn(module.weight, nonlinearity=activation)
+                    ext_init_kwargs = {'nonlinearity': activation, **init_kwargs}
+                    fn(module.weight, **ext_init_kwargs)
                 else:
-                    fn(module.weight)
+                    fn(module.weight, **init_kwargs)
             except Exception:  # first fallback to plain init
-                fn(module.weight)
+                fn(module.weight, **init_kwargs)
         except Exception:  # second fallback: module weight cannot be initialized, ok
             pass
     return init_fn
@@ -51,7 +62,7 @@ def build_module(arc: dict) -> nn.Module:
     if arc.get('layers'):  # if given layers, build as sequential
         module = sequential.build(arc)
     else:
-        kwargs = ps.omit(arc, 'type', 'in_names')
+        kwargs = ps.omit(arc, 'type', 'in_names', 'init')
         module = getattr(nn, arc['type'])(**kwargs)
     # initialize weights if 'init' is given
     if arc.get('init'):