Move to LFNA

exps/LFNA/vis-synthetic.py

@@ -1,7 +1,7 @@
 #####################################################
 # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.02 #
 ############################################################################
-# CUDA_VISIBLE_DEVICES=0 python exps/synthetic/baseline.py                 #
+# CUDA_VISIBLE_DEVICES=0 python exps/LFNA/vis-synthetic.py                 #
 ############################################################################
 import os, sys, copy, random
 import torch
@@ -31,17 +31,19 @@ from datasets.synthetic_example import create_example_v1
 from utils.temp_sync import optimize_fn, evaluate_fn
 
 
-def draw_fig(save_dir, timestamp, scatter_list):
+def draw_multi_fig(save_dir, timestamp, scatter_list, fig_title=None):
     save_path = save_dir / "{:04d}".format(timestamp)
     # print('Plot the figure at timestamp-{:} into {:}'.format(timestamp, save_path))
-    dpi, width, height = 40, 1500, 1500
+    dpi, width, height = 40, 2000, 1300
     figsize = width / float(dpi), height / float(dpi)
     LabelSize, LegendFontsize, font_gap = 80, 80, 5
 
     fig = plt.figure(figsize=figsize)
+    if fig_title is not None:
+        fig.suptitle(fig_title, fontsize=LegendFontsize)
 
-    cur_ax = fig.add_subplot(1, 1, 1)
-    for scatter_dict in scatter_list:
+    for idx, scatter_dict in enumerate(scatter_list):
+        cur_ax = fig.add_subplot(len(scatter_list), 1, idx + 1)
         cur_ax.scatter(
             scatter_dict["xaxis"],
             scatter_dict["yaxis"],
@@ -50,15 +52,15 @@ def draw_fig(save_dir, timestamp, scatter_list):
             alpha=scatter_dict["alpha"],
             label=scatter_dict["label"],
         )
-    cur_ax.set_xlabel("X", fontsize=LabelSize)
-    cur_ax.set_ylabel("f(X)", rotation=0, fontsize=LabelSize)
-    cur_ax.set_xlim(-6, 6)
-    cur_ax.set_ylim(-40, 40)
-    for tick in cur_ax.xaxis.get_major_ticks():
-        tick.label.set_fontsize(LabelSize - font_gap)
-        tick.label.set_rotation(10)
-    for tick in cur_ax.yaxis.get_major_ticks():
-        tick.label.set_fontsize(LabelSize - font_gap)
+        cur_ax.set_xlabel("X", fontsize=LabelSize)
+        cur_ax.set_ylabel("f(X)", rotation=0, fontsize=LabelSize)
+        cur_ax.set_xlim(scatter_dict["xlim"][0], scatter_dict["xlim"][1])
+        cur_ax.set_ylim(scatter_dict["ylim"][0], scatter_dict["ylim"][1])
+        for tick in cur_ax.xaxis.get_major_ticks():
+            tick.label.set_fontsize(LabelSize - font_gap)
+            tick.label.set_rotation(10)
+        for tick in cur_ax.yaxis.get_major_ticks():
+            tick.label.set_fontsize(LabelSize - font_gap)
 
     plt.legend(loc=1, fontsize=LegendFontsize)
     fig.savefig(str(save_path) + ".pdf", dpi=dpi, bbox_inches="tight", format="pdf")
@@ -66,7 +68,7 @@ def draw_fig(save_dir, timestamp, scatter_list):
     plt.close("all")
 
 
-def main(save_dir):
+def compare_cl(save_dir):
     save_dir = Path(str(save_dir))
     save_dir.mkdir(parents=True, exist_ok=True)
     dynamic_env, function = create_example_v1(100, num_per_task=1000)
@@ -74,6 +76,10 @@ def main(save_dir):
     additional_xaxis = np.arange(-6, 6, 0.2)
     models = dict()
 
+    cl_function = copy.deepcopy(function)
+    cl_function.set_timestamp(0)
+    cl_xaxis_all = None
+
     for idx, (timestamp, dataset) in enumerate(tqdm(dynamic_env, ncols=50)):
         xaxis_all = dataset[:, 0].numpy()
         # xaxis_all = np.concatenate((additional_xaxis, xaxis_all))
@@ -81,51 +87,46 @@ def main(save_dir):
         function.set_timestamp(timestamp)
         yaxis_all = function.noise_call(xaxis_all)
 
-        # split the dataset
-        indexes = list(range(xaxis_all.shape[0]))
-        random.shuffle(indexes)
-        train_indexes = indexes[: len(indexes) // 2]
-        valid_indexes = indexes[len(indexes) // 2 :]
-        train_xs, train_ys = xaxis_all[train_indexes], yaxis_all[train_indexes]
-        valid_xs, valid_ys = xaxis_all[valid_indexes], yaxis_all[valid_indexes]
+        # create CL data
+        if cl_xaxis_all is None:
+            cl_xaxis_all = xaxis_all
+        else:
+            cl_xaxis_all = np.concatenate((cl_xaxis_all, xaxis_all + timestamp * 0.2))
+        cl_yaxis_all = cl_function(cl_xaxis_all)
 
-        model, loss_fn, train_loss = optimize_fn(train_xs, train_ys)
-        # model, loss_fn, train_loss = optimize_fn(xaxis_all, yaxis_all)
-        pred_valid_ys, valid_loss = evaluate_fn(model, valid_xs, valid_ys, loss_fn)
-        print(
-            "[{:03d}] T-{:03d}, train-loss={:.5f}, valid-loss={:.5f}".format(
-                idx, timestamp, train_loss, valid_loss
-            )
-        )
-
-        # the first plot
         scatter_list = []
         scatter_list.append(
             {
-                "xaxis": valid_xs,
-                "yaxis": valid_ys,
+                "xaxis": xaxis_all,
+                "yaxis": yaxis_all,
                 "color": "k",
                 "s": 10,
                 "alpha": 0.99,
-                "label": "Timestamp={:02d}".format(timestamp),
+                "xlim": (-6, 6),
+                "ylim": (-40, 40),
+                "label": "LFNA",
             }
         )
 
         scatter_list.append(
             {
-                "xaxis": valid_xs,
-                "yaxis": pred_valid_ys,
+                "xaxis": cl_xaxis_all,
+                "yaxis": cl_yaxis_all,
                 "color": "r",
                 "s": 10,
-                "alpha": 0.5,
-                "label": "MLP at now",
+                "xlim": (-6, 6 + timestamp * 0.2),
+                "ylim": (-200, 40),
+                "alpha": 0.99,
+                "label": "Continual Learning",
             }
         )
 
-        draw_fig(save_dir, timestamp, scatter_list)
+        draw_multi_fig(
+            save_dir, timestamp, scatter_list, "Timestamp={:03d}".format(timestamp)
+        )
     print("Save all figures into {:}".format(save_dir))
     save_dir = save_dir.resolve()
-    cmd = "ffmpeg -y -i {xdir}/%04d.png -pix_fmt yuv420p -vf fps=2 -vf scale=1000:1000 -vb 5000k {xdir}/vis.mp4".format(
+    cmd = "ffmpeg -y -i {xdir}/%04d.png -pix_fmt yuv420p -vf fps=2 -vf scale=1500:1000 -vb 5000k {xdir}/vis.mp4".format(
         xdir=save_dir
     )
     os.system(cmd)
@@ -133,7 +134,7 @@ def main(save_dir):
 
 if __name__ == "__main__":
 
-    parser = argparse.ArgumentParser("Baseline")
+    parser = argparse.ArgumentParser("Visualize synthetic data.")
     parser.add_argument(
         "--save_dir",
         type=str,
@@ -142,4 +143,4 @@ if __name__ == "__main__":
     )
     args = parser.parse_args()
 
-    main(args.save_dir)
+    compare_cl(os.path.join(args.save_dir, "compare-cl"))

lib/xlayers/super_activations.py

@@ -17,8 +17,7 @@ from .super_module import BoolSpaceType
 class SuperReLU(SuperModule):
     """Applies a the rectified linear unit function element-wise."""
 
-    def __init__(
-        self, inplace=False) -> None:
+    def __init__(self, inplace=False) -> None:
         super(SuperReLU, self).__init__()
         self._inplace = inplace
 
@@ -33,4 +32,4 @@ class SuperReLU(SuperModule):
         return F.relu(input, inplace=self._inplace)
 
     def extra_repr(self) -> str:
-        return 'inplace=True' if self._inplace else ''
+        return "inplace=True" if self._inplace else ""

lib/xlayers/super_core.py

@@ -18,4 +18,3 @@ from .super_activations import SuperReLU
 
 from .super_trade_stem import SuperAlphaEBDv1
 from .super_positional_embedding import SuperPositionalEncoder
-

tests/test_super_container.py

@@ -79,6 +79,7 @@ def test_super_sequential_v1():
         super_core.SuperSimpleNorm(1, 1),
         torch.nn.ReLU(),
         super_core.SuperLinear(10, 10),
+        super_core.SuperReLU()
     )
     inputs = torch.rand(10, 10)
     print(model)