Add save_hyperparam to pytorch sequential to enable save/load

amber/__init__.py

@@ -8,7 +8,7 @@
 from .wrapper import Amber
 from . import architect, modeler, utils, plots
 
-__version__ = "0.1.4"
+__version__ = "0.1.5"
 
 __all__ = [
     'Amber',

amber/backend/pytorch/model.py

@@ -25,6 +25,7 @@ def __init__(self, *args, **kwargs):
         self.valid_metrics = {}
         self.trainer = None
         self.task = None
+        self.save_hyperparameters()
 
     def compile(self, loss, optimizer, metrics=None, *args, **kwargs):
         if callable(loss):
@@ -258,9 +259,10 @@ def test_epoch_end(self, outputs):
 
 class Sequential(Model):
     def __init__(self, layers=None):
-        layers = layers or []
         super().__init__()
+        layers = layers or []
         self.layers = torch.nn.ModuleList(layers)
+        self.save_hyperparameters()
 
     def add(self, layer):
         self.layers.append(layer)

amber/modeler/sequential/pytorch_sequential.py

@@ -29,7 +29,7 @@ def __init__(self, inputs_op, output_op, model_compile_dict, model_space, custom
         assert self.session is None or isinstance(self.session, F.SessionType)
 
     def build(self, model_states):
-        model = F.Sequential()
+        layers = []
         curr_shape = self.input_node.Layer_attributes['shape']
         x = torch.empty(*curr_shape, device='cpu')
         # add a batch dim
@@ -38,7 +38,7 @@ def build(self, model_states):
         if len(curr_shape) > 1:
             dims = [0, len(curr_shape)] + np.arange(1, len(curr_shape)).tolist()
             layer = F.get_layer(op=F.Operation('permute', dims=dims))
-            model.add(layer)
+            layers.append(layer)
             x = layer(x)
         for i, state in enumerate(model_states):
             if issubclass(type(state), int) or np.issubclass_(type(state), np.integer):
@@ -51,9 +51,10 @@ def build(self, model_states):
                 )
             layer = F.get_layer(torch.squeeze(x, dim=0), op, custom_objects=self.custom_objects)
             x = layer(x)
-            model.add(layer)
+            layers.append(layer)
         out = F.get_layer(torch.squeeze(x, dim=0), op=self.output_node, custom_objects=self.custom_objects)
-        model.add(out)
+        layers.append(out)
+        model = F.Sequential(layers=layers)
         return model
 
     def __call__(self, model_states):

amber/utils/io.py

@@ -20,9 +20,12 @@ def read_history_set(fn_list):
 
 
 def read_history(fn_list,
-                 metric_name_dict={'acc': 0, 'knowledge': 1, 'loss': 2}):
+                 metric_name_dict=None):
+    if not isinstance(fn_list, (list, tuple)):
+        fn_list = [fn_list]
     d = read_history_set(fn_list)
     d.columns = ['ID', 'metrics', 'reward'] + ['L%i' % i for i in range(1, d.shape[1] - 3)] + ['dir']
+    metric_name_dict = {} if metric_name_dict is None else metric_name_dict
     metrics = {x: [] for x in metric_name_dict}
     for i in range(d.shape[0]):
         tmp = d.iloc[i, 1].split(',')

setup.py

@@ -3,7 +3,7 @@
 config = {
     'description': 'Automated Modelling in Biological Evidence-based Research',
     'download_url': 'https://github.com/zj-zhang/AMBER',
-    'version': '0.1.4',
+    'version': '0.1.5',
     'packages': find_packages(exclude=["*.tests", "*.tests.*", "tests.*", "tests"]),
     'include_package_data': True,
     'setup_requires': [],