sync with Threadripper

Author: rasdani

mp_transformer/config.py

@@ -8,17 +8,17 @@
     #  "latent_dim": 32,
     # "latent_dim": 48,
     # "latent_dim": 64,
-    #  "latent_dim": 128,
-     "latent_dim": 256,
-    #"num_primitives": 4,
+    "latent_dim": 128,
+    #  "latent_dim": 256,
+    # "num_primitives": 4,
     "num_primitives": 6,
     # "num_primitives": 8,
     # "hidden_dim": 16,
     #  "hidden_dim": 32,
     # "hidden_dim": 48,
     # "hidden_dim": 64,
-    #  "hidden_dim": 128,
-     "hidden_dim": 256,
+    "hidden_dim": 128,
+    #  "hidden_dim": 256,
     "learn_segmentation": True,
     "masking_slope": 1,
     # "masking_slope": 0.75,
@@ -27,7 +27,9 @@
     # "kl_weight": 5e-3,
     # "kl_weight": 1e-2,
     # "kl_weight": 2e-2,
-    "kl_weight": 1e-3,
+    # "kl_weight": 5e-2,
+    "kl_weight": 1e-1,
+    # "kl_weight": 1e-3,
     # "kl_weight": 1e-4,
     "anneal_start": 10,
     # "anneal_start": 5,
@@ -41,7 +43,7 @@
     # "cycle_length": 100,
     # "cycle_length": 200,
     # "durations_weight": 1e-6,
-    #"durations_weight": 1e-4,
+    # "durations_weight": 1e-4,
     "durations_weight": 0,
     # "durations_weight": 1e-5,
     "lr": 1e-4,
@@ -53,10 +55,10 @@
     "N_val": 40000,
     # "N_val": 2,
     "sequence_length": 128,
-    #"epochs": 200,
+    # "epochs": 200,
     # "epochs": 250,
     # "epochs": 230,
-    # "epochs": 3000,
+    # "epochs": 2000,
     # "epochs": 1000,
     # "epochs": 300,
     "epochs": 500,
@@ -67,14 +69,18 @@
     # "run_name": "fresh-Transformer",
     # "run_name": "smol-Transformer",
     # "run_name": "resume-Transformer",
-    #"run_name": "smol-Transformer",
-    #"run_name": "notrafo-Transformer",
+    # "run_name": "smol-Transformer",
+    # "run_name": "notrafo-Transformer",
     # "run_name": "notrafo-sigmoid-Transformer",
     # "run_name": "noanneal-sigmoid-Transformer",
     # "run_name": "sigmoid-Transformer",
     # "run_name": "midKL-Transformer",
-    "run_name": "lowKL-Transformer",
+    # "run_name": "lowKL-Transformer",
+    # "run_name": "highKL-Transformer",
+    "run_name": "veryhighKL-Transformer",
+    # "run_name": "noanneal-highKL-Transformer",
     # "run_name": "cyclical-Transformer",
+    # "run_name": "cyclical-lowKL-Transformer",
     # "run_name": "nosigmoid-Transformer",
     # "run_name": "nosigmoid-Transformer",
     # "run_name": "relu-sigmoid-Transformer",

mp_transformer/datasets/toy_dataset.py

@@ -9,8 +9,6 @@
 from torch.utils.data import Dataset
 from torchvision import transforms
 
-from mp_transformer.utils.generate_toydata import forward
-
 PIL.PILLOW_VERSION = PIL.__version__  # torchvision bug
 
 
@@ -22,18 +20,17 @@ def normalize_pose(pose):
     # print(pose)
     # print(np.where(pose < -1))
     # print(np.where(pose > 1))
-    # assert np.all(pose >= -1) and np.all(pose <= 1)
-    # pose = (pose + 1) / 2
-    # assert np.all(pose >= 0) and np.all(pose <= 1)
-    
-    pose = forward(pose)
+    assert np.all(pose >= -1) and np.all(pose <= 1)
+    pose = (pose + 1) / 2
+    assert np.all(pose >= 0) and np.all(pose <= 1)
+
     return pose
 
 
 def unnormalize_pose(pose):
     """Transform [0, 1] to [-1, 1]"""
     # assert np.all(pose >= 0) and np.all(pose <= 1)
-    # pose = pose * 2 - 1
+    pose = pose * 2 - 1
     # assert np.all(pose >= -1) and np.all(pose <= 1)
     return pose
 

mp_transformer/utils/generate_toy_data.py

@@ -241,6 +241,6 @@ def main(iterations=1, train_or_val="train", gen_images=True, N=5000):
     # ITERATIONS = 16
     ITERATIONS = 20
     TRAIN_OR_VAL = "both"
-    #TRAIN_OR_VAL = "train"
+    # TRAIN_OR_VAL = "train"
     # TRAIN_OR_VAL = "val"
     main(iterations=ITERATIONS, train_or_val=TRAIN_OR_VAL, gen_images=GEN_IMAGES, N=N)