resolve merge conflict

Author: rasdani

data/toy/train-set/poses0.npy

@@ -1,6 +1,6 @@
 # might move this to a yaml or json
 CONFIG = {
-    "pose_dim": 3,
+    "pose_dim": 6,
     "num_attention_heads": 4,
     "num_transformer_layers": 4,
     # "latent_dim": 8,
@@ -39,8 +39,10 @@
     # "lr": 5e-4,
     "batch_size": 8,
     # "batch_size": 16,
-    "N_train": 200000,
-    "N_val": 40000,
+    # "N_train": 200000,
+    "N_train": 2,
+    # "N_val": 40000,
+    "N_val": 2,
     "sequence_length": 128,
     #"epochs": 295,
     "epochs": 230,

data/toy/train-set/poses1.npy

@@ -14,32 +14,23 @@
 
 TOY_DATA_PATH = "data/toy/train-set"
 
-global counter
-counter = 0
-
 
 def normalize_pose(pose):
-    """Radians to [0, 1]"""
-    global counter
-    # print('before')
-    if not np.all(-np.pi <= pose) and np.all(pose <= np.pi):
-        print(pose, end="\n")
-        print(np.where(~(-np.pi <= pose)))
-        print(np.where(~(pose <= np.pi)))
-        counter += 1
-        print(counter)
-    pose = 0.5 + (pose / (2 * np.pi))
-    # print('after')
-    # print(pose)
-    # assert np.all(0 <= pose) and np.all(pose <= 1)
+    """Transform [-1, 1] to [0, 1]"""
+    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)
     return pose
 
 
 def unnormalize_pose(pose):
-    """[0, 1] to radians"""
-    assert np.all(0 <= pose) and np.all(pose <= 1)
-    pose = (pose - 0.5) * (2 * np.pi)
-    assert np.all(-np.pi <= pose) and np.all(pose <= np.pi)
+    """Transform [0, 1] to [-1, 1]"""
+    assert np.all(pose >= 0) and np.all(pose <= 1)
+    pose = pose * 2 - 1
+    assert np.all(pose >= -1) and np.all(pose <= 1)
     return pose
 
 
@@ -103,6 +94,7 @@ def __getitem__(self, idx):
 
     def _get_single_item(self, idx):
         pose = self.poses[idx]
+        # print(f"{idx=}")
         pose = normalize_pose(pose)
         pose = torch.tensor(pose).to(torch.float32)
         ret = {"pose": pose}
@@ -120,6 +112,7 @@ def _get_single_item(self, idx):
         return ret
 
     def _get_segment(self, idx):
+        print(f"{idx=}")
         images = []
         poses = []
         timestamps = []

data/toy/train-set/poses10.npy

@@ -248,8 +248,9 @@ def forward(self, timestamps, latent_primitives):
         recons_sequence = self.masked_average(
             recons_subseqs, repeated_transfo, gaussian_masks
         )
-        print(f"{recons_sequence.max()=}")
-        print(f"{recons_sequence.min()=}")
+        if recons_sequence.min() < 0 or recons_sequence.max() > 1:
+            print(f"{recons_sequence.max()=}")
+            print(f"{recons_sequence.min()=}")
 
         return {
             "recons_sequence": recons_sequence,

data/toy/train-set/poses11.npy

@@ -141,6 +141,7 @@ def compute_primitive_features(self, poses, timestamps):
             torch.tensor: tensor of size BSxMxD of features with M the number of primitives
         """
 
+        # breakpoint()
         # Get the batch size from the input data
         batch_size = poses.shape[0]
 

data/toy/train-set/poses12.npy

@@ -115,6 +115,7 @@ def main(config, no_log=False, debug=False, checkpoint_path=None):
             log_every_n_steps=1,
             gpus=gpus,
             resume_from_checkpoint=checkpoint_path,
+            num_sanity_val_steps=0,
         )
 
     trainer.fit(

data/toy/train-set/poses13.npy

@@ -23,78 +23,9 @@
 #     return bone_lengths, key_marker_width
 
 
-# def generate_gaussian_process_poses(N, bone_lengths, train_or_val):
-#     """Uses gaussian processes of two speeds to animate different parts of the limb."""
-
-#     print(f"Drawing {N} samples, might take a while...")
-
-#     if train_or_val == "train":
-#         rbf = gaussian_process.kernels.RBF(length_scale=0.25)
-#         rbf_slow = gaussian_process.kernels.RBF(length_scale=0.6)
-#     elif train_or_val == "val":
-#         rbf = gaussian_process.kernels.RBF(length_scale=0.15)  # val-set
-#         rbf_slow = gaussian_process.kernels.RBF(length_scale=0.55)  # val-set
-#     else:
-#         raise ValueError(f"{train_or_val} must be 'train' or 'val' in Gaussian Process.")
-
-#     GP = gaussian_process.GaussianProcessRegressor(kernel=rbf)
-#     GP_slow = gaussian_process.GaussianProcessRegressor(kernel=rbf_slow)
-
-#     t = np.linspace(0, 120, N)
-#     # TODO: hardcode if I end up using only 3 joints
-#     y = np.empty((N, len(bone_lengths)))
-
-#     y[:, 0] = 3 * GP_slow.sample_y(t[:, None], random_state=None)[:, 0]
-
-#     y_samples = 0.7 * GP.sample_y(t[:, None], n_samples=len(bone_lengths) - 1, random_state=None).T
-#     y[:, 1:] = np.column_stack(y_samples)
-
-#     print("RAW")
-#     print(f"y.min() = {y.min()}, y.max() = {y.max()}")
-
-#     # Wrap angles between -pi and pi
-#     y = (y + np.pi) % (2 * np.pi) - np.pi
-#     print("FIRT WRAP")
-#     print(f"y.min() = {y.min()}, y.max() = {y.max()}")
-
-#     # Unwrap handles the case where the angle jumps from -pi to pi
-#     y = np.unwrap(y, axis=0)
-#     print("UNWRAP")
-#     print(f"y.min() = {y.min()}, y.max() = {y.max()}")
-
-
-#     # Wrap again since unwrapping can lead to values outside of [-pi, pi]
-#     y = (y + np.pi) % (2 * np.pi) - np.pi
-#     print("SECOND WRAP")
-#     print(f"y.min() = {y.min()}, y.max() = {y.max()}")
-
-#     # Check for discontinuities
-#     y_diff = np.diff(y, axis=0)
-#     print("DIFF")
-#     print(y_diff)
-#     print(f"y_diff.min() = {y_diff.min()}, y_diff.max() = {y_diff.max()}")
-
-#     max_jump = np.pi
-#     indices_of_jumps = np.where(np.abs(y_diff) > max_jump)
-#     print(f"indices_of_jumps = {indices_of_jumps}")
-
-#     # if np.any(abs(y_diff) > max_jump):
-#     #     raise ValueError("Detected a discontinuity in the pose sequence.")
-
-#     window_size = 2  # Adjust as needed
-
-#     for idx in indices_of_jumps[0]:
-#         print(f"Jump at index {idx} in column {indices_of_jumps[1][0]}:")
-#         print(f"Values before: {y[idx-window_size:idx+1]}")
-#         print(f"Values after: {y[idx+1:idx+window_size+1]}")
-#         print(f"Difference causing the jump: {y_diff[idx]}")
-#         print("---")
-
-#     return y
-
-
 def generate_gaussian_process_poses(N, bone_lengths, train_or_val):
-    """Uses gaussian processes of two speeds to animate different parts of the limb."""
+    """Uses gaussian processes of two speeds to animate different parts of the limb.
+    To avoid discontinuities angles are represented as sin and cos."""
 
     print(f"Drawing {N} samples, might take a while...")
 
@@ -158,46 +89,6 @@ def generate_gaussian_process_poses(N, bone_lengths, train_or_val):
     return y
 
 
-# def forward(angles, bone_lenghts=BONE_LENGTHS):
-#     """forward
-#     Compute forward kinematics
-#     angles --> cartesian coordinates
-
-#     :param angles:
-#       List of angles for each bone_length in the hierarchy
-#       relative to its parent bone_length
-#     :param bone_lengths: List of bone_length lengths
-#     """
-#     bone_lengths = np.array(bone_lenghts)
-#     if bone_lengths is None:
-#         bone_lengths = np.ones_like(angles)
-#     elif len(bone_lengths) == 2:
-#         bone_lengths = bone_lengths * np.ones_like(angles)
-#     else:
-#         try:
-#             assert len(angles) == len(bone_lengths)
-#             # assert angles.shape == bone_lengths.shape
-#         except AssertionError as excp:
-#             raise Exception(
-#                 f"Number of angles and bone_lengths should be the same"
-#                 f" but: {len(angles)} is not {len(bone_lengths)}"
-#                 # f" but: {angles.shape} is not {bone_lengths.shape}"
-#             ) from excp
-
-#     coordinates = [(0, 0)]
-#     cumulative_angle = 0
-#     for angle, bone_length in zip(angles, bone_lengths):
-#         offs = coordinates[-1]
-#         cumulative_angle += angle
-#         coordinates += [
-#             (
-#                 bone_length * np.cos(cumulative_angle) + offs[0],
-#                 bone_length * np.sin(cumulative_angle) + offs[1],
-#             )
-#         ]
-#     return coordinates
-
-
 def forward(angles, bone_lengths=BONE_LENGTHS):
     """
     Compute forward kinematics