Fix bugs

xautodl/datasets/math_adv_funcs.py

@@ -64,65 +64,29 @@ class ComposedSinFunc(FitFunc):
         )
 
 
-class ComposedSinFuncV2(FitFunc):
+class ComposedCosFunc(FitFunc):
     """The composed sin function that outputs:
-      f(x) = amplitude-scale-of(x) * sin( period-phase-shift-of(x) )
-    - the amplitude scale is a quadratic function of x
-    - the period-phase-shift is another quadratic function of x
+    f(x) = a * cos( b*x ) + c
     """
 
-    def __init__(self, **kwargs):
-        super(ComposedSinFuncV2, self).__init__(0, None)
-        self.fit(**kwargs)
+    def __init__(self, params, xstr="x"):
+        super(ComposedCosFunc, self).__init__(3, None, params, xstr)
 
     def __call__(self, x):
         self.check_valid()
-        scale = self._params["amplitude_scale"](x)
-        period_phase = self._params["period_phase_shift"](x)
-        return scale * math.sin(period_phase)
-
-    def fit(self, **kwargs):
-        num_sin_phase = kwargs.get("num_sin_phase", 7)
-        sin_speed_use_power = kwargs.get("sin_speed_use_power", True)
-        min_amplitude = kwargs.get("min_amplitude", 1)
-        max_amplitude = kwargs.get("max_amplitude", 4)
-        phase_shift = kwargs.get("phase_shift", 0.0)
-        # create parameters
-        if kwargs.get("amplitude_scale", None) is None:
-            amplitude_scale = QuadraticFunc(
-                [(0, min_amplitude), (0.5, max_amplitude), (1, min_amplitude)]
-            )
-        else:
-            amplitude_scale = kwargs.get("amplitude_scale")
-        if kwargs.get("period_phase_shift", None) is None:
-            fitting_data = []
-            if sin_speed_use_power:
-                temp_max_scalar = 2 ** (num_sin_phase - 1)
-            else:
-                temp_max_scalar = num_sin_phase - 1
-            for i in range(num_sin_phase):
-                if sin_speed_use_power:
-                    value = (2 ** i) / temp_max_scalar
-                    next_value = (2 ** (i + 1)) / temp_max_scalar
-                else:
-                    value = i / temp_max_scalar
-                    next_value = (i + 1) / temp_max_scalar
-                for _phase in (0, 0.25, 0.5, 0.75):
-                    inter_value = value + (next_value - value) * _phase
-                    fitting_data.append((inter_value, math.pi * (2 * i + _phase)))
-            period_phase_shift = QuarticFunc(fitting_data)
-        else:
-            period_phase_shift = kwargs.get("period_phase_shift")
-        self.set(
-            dict(amplitude_scale=amplitude_scale, period_phase_shift=period_phase_shift)
-        )
+        a = self._params[0]
+        b = self._params[1]
+        c = self._params[2]
+        return a * math.cos(b * x) + c
 
     def _getitem(self, x, weights):
         raise NotImplementedError
 
     def __repr__(self):
-        return "{name}({amplitude_scale} * sin({period_phase_shift}))".format(
+        return "{name}({a} * sin({b} * {x}) + {c})".format(
             name=self.__class__.__name__,
-            amplitude_scale=self._params["amplitude_scale"],
-            period_phase_shift=self._params["period_phase_shift"],
+            a=self._params[0],
+            b=self._params[1],
+            c=self._params[2],
+            x=self.xstr,
         )

xautodl/datasets/math_core.py

@@ -5,5 +5,5 @@ from .math_base_funcs import LinearFunc, QuadraticFunc, CubicFunc, QuarticFunc
 from .math_dynamic_funcs import DynamicLinearFunc
 from .math_dynamic_funcs import DynamicQuadraticFunc
 from .math_adv_funcs import ConstantFunc
-from .math_adv_funcs import ComposedSinFunc
+from .math_adv_funcs import ComposedSinFunc, ComposedCosFunc
 from .math_dynamic_generator import GaussianDGenerator

xautodl/datasets/synthetic_core.py

@@ -4,7 +4,11 @@ from .synthetic_env import SyntheticDEnv
 from .math_core import LinearFunc
 from .math_core import DynamicLinearFunc
 from .math_core import DynamicQuadraticFunc
-from .math_core import ConstantFunc, ComposedSinFunc as SinFunc
+from .math_core import (
+    ConstantFunc,
+    ComposedSinFunc as SinFunc,
+    ComposedCosFunc as CosFunc,
+)
 from .math_core import GaussianDGenerator
 
 
@@ -50,6 +54,25 @@ def get_synthetic_env(total_timestamp=1600, num_per_task=1000, mode=None, versio
         dynamic_env = SyntheticDEnv(
             data_generator, oracle_map, time_generator, num_per_task
         )
+    elif version.lower() == "v3":
+        mean_generator = SinFunc(params={0: 1, 1: 1, 2: 0})  # sin(t)
+        std_generator = CosFunc(params={0: 0.5, 1: 1, 2: 1})  # 0.5 cos(t) + 1
+        data_generator = GaussianDGenerator(
+            [mean_generator], [[std_generator]], (-2, 2)
+        )
+        time_generator = TimeStamp(
+            min_timestamp=0, max_timestamp=max_time, num=total_timestamp, mode=mode
+        )
+        oracle_map = DynamicQuadraticFunc(
+            params={
+                0: LinearFunc(params={0: 0.1, 1: 0}),  # 0.1 * t
+                1: SinFunc(params={0: 1, 1: 1, 2: 0}),  # sin(t)
+                2: ConstantFunc(0),
+            }
+        )
+        dynamic_env = SyntheticDEnv(
+            data_generator, oracle_map, time_generator, num_per_task
+        )
     else:
         raise ValueError("Unknown version: {:}".format(version))
     return dynamic_env

xautodl/models/xcore.py

@@ -39,9 +39,9 @@ def get_model(config: Dict[Text, Any], **kwargs):
         norm_cls = super_name2norm[kwargs["norm_cls"]]
         sub_layers, last_dim = [], kwargs["input_dim"]
         for i, hidden_dim in enumerate(kwargs["hidden_dims"]):
+            sub_layers.append(SuperLinear(last_dim, hidden_dim))
             if hidden_dim > 1:
                 sub_layers.append(norm_cls(hidden_dim, elementwise_affine=False))
-            sub_layers.append(SuperLinear(last_dim, hidden_dim))
             sub_layers.append(act_cls())
             last_dim = hidden_dim
         sub_layers.append(SuperLinear(last_dim, kwargs["output_dim"]))

xautodl/nas_infer_model/DXYs/__init__.py

@@ -1,5 +1,5 @@
 # Performance-Aware Template Network for One-Shot Neural Architecture Search
-from .CifarNet  import NetworkCIFAR as CifarNet
-from .ImageNet  import NetworkImageNet as ImageNet
+from .CifarNet import NetworkCIFAR as CifarNet
+from .ImageNet import NetworkImageNet as ImageNet
 from .genotypes import Networks
 from .genotypes import build_genotype_from_dict

xautodl/nas_infer_model/__init__.py

@@ -8,24 +8,44 @@
 
 import os, torch
 
+
 def obtain_nas_infer_model(config, extra_model_path=None):
-  
-  if config.arch == 'dxys':
-    from .DXYs import CifarNet, ImageNet, Networks
-    from .DXYs import build_genotype_from_dict
-    if config.genotype is None:
-      if extra_model_path is not None and not os.path.isfile(extra_model_path):
-        raise ValueError('When genotype in confiig is None, extra_model_path must be set as a path instead of {:}'.format(extra_model_path))
-      xdata = torch.load(extra_model_path)
-      current_epoch = xdata['epoch']
-      genotype_dict = xdata['genotypes'][current_epoch-1]
-      genotype = build_genotype_from_dict(genotype_dict)
+
+    if config.arch == "dxys":
+        from .DXYs import CifarNet, ImageNet, Networks
+        from .DXYs import build_genotype_from_dict
+
+        if config.genotype is None:
+            if extra_model_path is not None and not os.path.isfile(extra_model_path):
+                raise ValueError(
+                    "When genotype in confiig is None, extra_model_path must be set as a path instead of {:}".format(
+                        extra_model_path
+                    )
+                )
+            xdata = torch.load(extra_model_path)
+            current_epoch = xdata["epoch"]
+            genotype_dict = xdata["genotypes"][current_epoch - 1]
+            genotype = build_genotype_from_dict(genotype_dict)
+        else:
+            genotype = Networks[config.genotype]
+        if config.dataset == "cifar":
+            return CifarNet(
+                config.ichannel,
+                config.layers,
+                config.stem_multi,
+                config.auxiliary,
+                genotype,
+                config.class_num,
+            )
+        elif config.dataset == "imagenet":
+            return ImageNet(
+                config.ichannel,
+                config.layers,
+                config.auxiliary,
+                genotype,
+                config.class_num,
+            )
+        else:
+            raise ValueError("invalid dataset : {:}".format(config.dataset))
     else:
-      genotype = Networks[config.genotype]
-    if config.dataset == 'cifar':
-      return CifarNet(config.ichannel, config.layers, config.stem_multi, config.auxiliary, genotype, config.class_num)
-    elif config.dataset == 'imagenet':
-      return ImageNet(config.ichannel, config.layers, config.auxiliary, genotype, config.class_num)
-    else: raise ValueError('invalid dataset : {:}'.format(config.dataset))
-  else:
-    raise ValueError('invalid nas arch type : {:}'.format(config.arch))
+        raise ValueError("invalid nas arch type : {:}".format(config.arch))