refactor: remove torch dependency, only install toch for image metrics (

#55)

README.md

@@ -125,15 +125,21 @@ cd camtools
 # Installation mode, if you want to use camtools only.
 pip install .
 
-# Dev mode, if you want to modify camtools on the fly.
+# Editable mode, if you want to modify camtools on the fly.
 pip install -e .
 
-# Dev mode and dev dependencies, if you want to modify camtools and run tests.
+# Editable mode and dev dependencies.
 pip install -e .[dev]
 
 # Help VSCode resolve imports when installed with editable mode.
 # https://stackoverflow.com/a/76897706/1255535
 pip install -e .[dev] --config-settings editable_mode=strict
+
+# Enable torch-related features (e.g. computing image metrics)
+pip install camtools[torch]
+
+# Enable torch-related features in editable mode
+pip install -e .[torch]
 ```
 
 ## Camera coordinate system

camtools/convert.py

@@ -1,6 +1,5 @@
 import cv2
 import numpy as np
-import torch
 
 from . import sanity
 from . import convert
@@ -29,26 +28,15 @@ def pad_0001(array):
  f"Expected array of shape (3, 4) or (N, 3, 4), but got {array.shape}."
  )
 
- if torch.is_tensor(array):
- if array.ndim == 2:
- bottom = torch.tensor([0, 0, 0, 1], dtype=array.dtype, device=array.device)
- return torch.cat([array, bottom[None, :]], dim=0)
- elif array.ndim == 3:
- bottom_single = torch.tensor(
- [0, 0, 0, 1], dtype=array.dtype, device=array.device
- )
- bottom = bottom_single[None, None, :].expand(array.shape[0], 1, 4)
- return torch.cat([array, bottom], dim=-2)
+ if array.ndim == 2:
+ bottom = np.array([0, 0, 0, 1], dtype=array.dtype)
+ return np.concatenate([array, bottom[None, :]], axis=0)
+ elif array.ndim == 3:
+ bottom_single = np.array([0, 0, 0, 1], dtype=array.dtype)
+ bottom = np.broadcast_to(bottom_single, (array.shape[0], 1, 4))
+ return np.concatenate([array, bottom], axis=-2)
  else:
- if array.ndim == 2:
- bottom = np.array([0, 0, 0, 1], dtype=array.dtype)
- return np.concatenate([array, bottom[None, :]], axis=0)
- elif array.ndim == 3:
- bottom_single = np.array([0, 0, 0, 1], dtype=array.dtype)
- bottom = np.broadcast_to(bottom_single, (array.shape[0], 1, 4))
- return np.concatenate([array, bottom], axis=-2)
- else:
- raise ValueError("Should not reach here.")
+ raise ValueError("Should not reach here.")
 
 
 def rm_pad_0001(array, check_vals=False):
@@ -78,42 +66,21 @@ def rm_pad_0001(array, check_vals=False):
 
  # Check vals.
  if check_vals:
- if torch.is_tensor(array):
- if array.ndim == 2:
- bottom = array[3, :]
- if not torch.allclose(
- bottom, torch.tensor([0, 0, 0, 1], dtype=array.dtype)
- ):
- raise ValueError(
- f"Expected bottom row to be [0, 0, 0, 1], but got {bottom}."
- )
- elif array.ndim == 3:
- bottom = array[:, 3:4, :]
- expected_bottom = torch.tensor([0, 0, 0, 1], dtype=array.dtype).expand(
- array.shape[0], 1, 4
+ if array.ndim == 2:
+ bottom = array[3, :]
+ if not np.allclose(bottom, [0, 0, 0, 1]):
+ raise ValueError(
+ f"Expected bottom row to be [0, 0, 0, 1], but got {bottom}."
+ )
+ elif array.ndim == 3:
+ bottom = array[:, 3:4, :]
+ expected_bottom = np.broadcast_to([0, 0, 0, 1], (array.shape[0], 1, 4))
+ if not np.allclose(bottom, expected_bottom):
+ raise ValueError(
+ f"Expected bottom row to be {expected_bottom}, but got {bottom}."
  )
- if not torch.allclose(bottom, expected_bottom):
- raise ValueError(
- f"Expected bottom row to be {expected_bottom}, but got {bottom}."
- )
- else:
- raise ValueError("Should not reach here.")
  else:
- if array.ndim == 2:
- bottom = array[3, :]
- if not np.allclose(bottom, [0, 0, 0, 1]):
- raise ValueError(
- f"Expected bottom row to be [0, 0, 0, 1], but got {bottom}."
- )
- elif array.ndim == 3:
- bottom = array[:, 3:4, :]
- expected_bottom = np.broadcast_to([0, 0, 0, 1], (array.shape[0], 1, 4))
- if not np.allclose(bottom, expected_bottom):
- raise ValueError(
- f"Expected bottom row to be {expected_bottom}, but got {bottom}."
- )
- else:
- raise ValueError("Should not reach here.")
+ raise ValueError("Should not reach here.")
 
  return array[..., :3, :]
 
@@ -363,10 +330,7 @@ def roll_pitch_yaw_to_R(roll, pitch, yaw):
 
 def R_t_to_T(R, t):
  sanity.assert_same_device(R, t)
- if torch.is_tensor(R):
- T = torch.eye(4, device=R.device, dtype=R.dtype)
- else:
- T = np.eye(4)
+ T = np.eye(4)
  T[:3, :3] = R
  T[:3, 3] = t
  return T

camtools/metric.py

@@ -3,8 +3,6 @@
 
 from skimage.metrics import peak_signal_noise_ratio
 from skimage.metrics import structural_similarity
-import torch
-import lpips
 from pathlib import Path
 from typing import Tuple
 
@@ -96,6 +94,9 @@ def image_lpips(
  Returns:
  LPIPS value in float.
  """
+ import torch
+ import lpips
+
  if im_mask is None:
  h, w = im_pd.shape[:2]
  im_mask = np.ones((h, w), dtype=np.float32)

camtools/sanity.py

@@ -1,5 +1,4 @@
 import numpy as np
-import torch
 
 
 def assert_numpy(x, name=None):
@@ -8,12 +7,6 @@ def assert_numpy(x, name=None):
  raise ValueError(f"Expected{maybe_name} to be numpy array, but got {type(x)}.")
 
 
-def assert_torch(x, name=None):
- if not torch.is_tensor(x):
- maybe_name = f" {name}" if name is not None else ""
- raise ValueError(f"Expected{maybe_name} to be torch tensor, but got {type(x)}.")
-
-
 def assert_K(K):
  if K.shape != (3, 3):
  raise ValueError(f"K must has shape (3, 3), but got {K} of shape {K.shape}.")
@@ -22,12 +15,7 @@ def assert_K(K):
 def assert_T(T):
  if T.shape != (4, 4):
  raise ValueError(f"T must has shape (4, 4), but got {T} of shape {T.shape}.")
- if torch.is_tensor(T):
- is_valid = torch.allclose(
- T[3, :], torch.tensor([0, 0, 0, 1], dtype=T.dtype, device=T.device)
- )
- else:
- is_valid = np.allclose(T[3, :], np.array([0, 0, 0, 1]))
+ is_valid = np.allclose(T[3, :], np.array([0, 0, 0, 1]))
  if not is_valid:
  raise ValueError(f"T must has [0, 0, 0, 1] the bottom row, but got {T}.")
 
@@ -37,12 +25,7 @@ def assert_pose(pose):
  raise ValueError(
  f"pose must has shape (4, 4), but got {pose} of shape {pose.shape}."
  )
- if torch.is_tensor(pose):
- is_valid = torch.allclose(
- pose[3, :], torch.tensor([0, 0, 0, 1], dtype=pose.dtype, device=pose.device)
- )
- else:
- is_valid = np.allclose(pose[3, :], np.array([0, 0, 0, 1]))
+ is_valid = np.allclose(pose[3, :], np.array([0, 0, 0, 1]))
  if not is_valid:
  raise ValueError(f"pose must has [0, 0, 0, 1] the bottom row, but got {pose}.")
 
@@ -101,32 +84,3 @@ def assert_shape_3x3(x, name=None):
 
 def assert_shape_3(x, name=None):
  assert_shape(x, (3,), name=name)
-
-
-def assert_same_device(*tensors):
- """
- Args:
- tensors: list of tensors
- """
- if not isinstance(tensors, tuple):
- raise ValueError(f"Unknown input type: {type(tensors)}.")
- if len(tensors) == 0:
- return
- if len(tensors) == 1:
- if torch.is_tensor(tensors[0]) or isinstance(tensors[0], np.ndarray):
- return
- else:
- raise ValueError(f"Unknown input type: {type(tensors)}.")
-
- all_are_torch = all(torch.is_tensor(t) for t in tensors)
- all_are_numpy = all(isinstance(t, np.ndarray) for t in tensors)
-
- if not all_are_torch and not all_are_numpy:
- raise ValueError(f"All tensors must be torch tensors or numpy arrays.")
-
- if all_are_torch:
- devices = [t.device for t in tensors]
- if not all(devices[0] == d for d in devices):
- raise ValueError(
- f"All tensors must be on the same device, bui got {devices}."
- )

camtools/solver.py

@@ -1,5 +1,4 @@
 import numpy as np
-import torch
 
 from camtools import sanity
 
@@ -115,15 +114,9 @@ def closest_points_of_line_pairs(src_os, src_ds, dst_os, dst_ds):
  sanity.assert_shape_nx3(dst_os, "dst_os")
  sanity.assert_shape_nx3(dst_ds, "dst_ds")
 
- is_torch = torch.is_tensor(src_ds) and torch.is_tensor(dst_ds)
- cross = torch.cross if is_torch else np.cross
- norm = torch.linalg.norm if is_torch else np.linalg.norm
- solve = torch.linalg.solve if is_torch else np.linalg.solve
- stack = torch.stack if is_torch else np.stack
-
  # Normalize direction vectors.
- src_ds = src_ds / norm(src_ds, axis=1, keepdims=True)
- dst_ds = dst_ds / norm(dst_ds, axis=1, keepdims=True)
+ src_ds = src_ds / np.linalg.norm(src_ds, axis=1, keepdims=True)
+ dst_ds = dst_ds / np.linalg.norm(dst_ds, axis=1, keepdims=True)
 
  # Find the closest points of the two lines.
  # - src_p = src_o + src_t * src_d is the closest point in src line.
@@ -139,12 +132,12 @@ def closest_points_of_line_pairs(src_os, src_ds, dst_os, dst_ds):
  # │src_d -dst_d mid_d│ │ dst_t │ = │ dst_o │ - │ src_o │
  # │ │ │ │ │ │ mid_t │ │ │ │ │ │ │
  # └ ┘ └ ┘ └ ┘ └ ┘
- mid_ds = cross(src_ds, dst_ds)
- mid_ds = mid_ds / norm(mid_ds, axis=1, keepdims=True)
+ mid_ds = np.cross(src_ds, dst_ds)
+ mid_ds = mid_ds / np.linalg.norm(mid_ds, axis=1, keepdims=True)
 
- lhs = stack((src_ds, -dst_ds, mid_ds), axis=-1)
+ lhs = np.stack((src_ds, -dst_ds, mid_ds), axis=-1)
  rhs = dst_os - src_os
- results = solve(lhs, rhs)
+ results = np.linalg.solve(lhs, rhs)
  src_ts, dst_ts, mid_ts = results[:, 0], results[:, 1], results[:, 2]
  src_ps = src_os + src_ts.reshape((-1, 1)) * src_ds
  dst_ps = dst_os + dst_ts.reshape((-1, 1)) * dst_ds

pyproject.toml

@@ -10,12 +10,11 @@ classifiers = [
  "Programming Language :: Python :: 3",
 ]
 dependencies = [
+ "numpy>=1.15.0",
  "open3d>=0.16.0",
  "opencv-python>=4.5.1.48",
  "matplotlib>=3.3.4",
  "scikit-image>=0.16.2",
- "torch>=1.8.0",
- "lpips>=0.1.4",
  "tqdm>=4.60.0",
 ]
 description = "CamTools: Camera Tools for Computer Vision."
@@ -37,6 +36,10 @@ dev = [
  "pytest>=6.2.2",
  "ipdb",
 ]
+torch = [
+ "torch>=1.8.0",
+ "lpips>=0.1.4",
+]
 
 [tool.setuptools]
 packages = ["camtools", "camtools.tools"]