7. Packaging and CI

.github/workflows/tests.yml

@@ -0,0 +1,27 @@
+name: tests
+
+on:
+  push:
+  pull_request:
+
+jobs:
+  pytest:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install
+        run: |
+          python -m pip install --upgrade pip
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
+          # auton-survival 0.1.0 (release commit); --no-deps because its metadata pins torch<2.0
+          pip install --no-deps "auton-survival @ git+https://github.com/autonlab/auton-survival@5dde465f7223601717abddc1d075e837707c403b"
+          pip install -e ".[test]"
+      - name: Test
+        run: python -m pytest -q

.gitignore

@@ -127,3 +127,15 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# Patient-level data and anything derived from it never belong in this
+# repository. Point the evaluation scripts at a directory outside it.
+*.csv
+*.parquet
+*.feather
+*.pkl
+*.pickle
+*.pt
+*.npz
+data/
+outputs/

ddh/losses.py

@@ -73,8 +73,8 @@ def build_landmark_prefixes(x, s_all, visit_days=None, decay_days=None,
         A patient with V_i observed visits becomes V_i rows: row v keeps visits
         1..v+1 and NaNs out everything after, exactly the input the model would
         receive if that visit were the patient's last observation.  This is the
-        training-time counterpart of build_intermediate_landmark_dataset in the
-        DDH_implement pipeline, which does the same truncation at evaluation time.
+        training-time counterpart of landmark evaluation, which applies the same
+        truncation to every patient at test time.
 
         Args:
             x      : [N, T, D] padded features, NaN after each patient's last visit.
@@ -217,7 +217,7 @@ def landmark_nll_loss(model, x_exp, t_exp, e_exp, s_exp, weight,
         a single (final-visit) landmark.  Rows whose landmark bin is at or beyond
         the patient's event bin carry no information about the remaining horizon
         and are dropped; the 1 / V_i weight is deliberately left as-is, so those
-        few patients (0.5% of rows on the VTE cohort) count slightly under 1.
+        few patients count slightly under 1.
 
         With `horizon_spec` and gamma > 0 the same forward pass also supplies
         horizon_likelihood_loss, and the returned scalar is
@@ -272,7 +272,7 @@ def ranking_loss(cif, t, e, sigma, s=None, cause_weights=None):
         cause_weights: list of K non-negative floats (alpha_k in the DDH paper)
            scaling each cause's ranking contribution. Defaults to equal weights
            1/K. Increase a cause's weight to prioritise its ranking accuracy
-           (e.g. [0.8, 0.2] to penalise wrong VTE ordering more than death).
+           (e.g. [0.8, 0.2] to penalise wrong ordering of cause 1 more than cause 2).
 
         Uses torch.sum over valid pairs per the DDH paper (Σ_{i≠j}).
         Because the sum scales with batch size, set alpha in total_loss small
@@ -413,12 +413,12 @@ def total_loss(model, x, t, e, alpha, beta, sigma, s=None, cause_weights=None, i
 # Binary log likelihood is a strictly proper scoring rule, so its minimiser is
 # the true conditional probability at h -- the level, not just the ordering.
 #
-# LABELS AND WEIGHTS mirror V3's evaluation code exactly (Graf/Gerds-Schumacher):
+# LABELS AND WEIGHTS mirror the landmark evaluation exactly (Graf/Gerds-Schumacher):
 #     T <= h, cause 1   -> y = 1, weight G(s) / G(T-)
-#     T <= h, cause 2   -> y = 0, weight G(s) / G(T-)  (death without VTE is a true 0)
+#     T <= h, cause 2   -> y = 0, weight G(s) / G(T-)  (a competing event is a true 0)
 #     T >  h            -> y = 0, weight G(s) / G(h)
 #     censored before h -> weight 0                    (status unknown)
-# with the G(s) factor conditioning on the landmark, which V3 does not need
+# with the G(s) factor conditioning on the landmark, which the evaluation does not need
 # because it scores one landmark at a time and the factor divides out there.
 # Matching the metric matters more than picking a different defensible estimator:
 # a training term and a reported metric that disagree cannot be read together.

pyproject.toml

@@ -0,0 +1,32 @@
+[build-system]
+requires = ["setuptools>=64"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "ddh"
+version = "0.2.0"
+description = "Dynamic-DeepHit in PyTorch, extended for landmark-based dynamic prediction"
+readme = "README.md"
+license = { file = "LICENSE" }
+requires-python = ">=3.10"
+authors = [{ name = "Vincent Jeanselme" }, { name = "Dulin Wang" }]
+dependencies = [
+    "torch>=2.0",
+    "numpy",
+    "tqdm",
+    # Imported when auton-survival loads. auton-survival itself is installed
+    # with --no-deps (see README): its PyPI release pins torch<2.0.
+    "pandas",
+    "scikit-learn",
+    "scikit-survival",
+    "matplotlib",
+]
+
+[project.optional-dependencies]
+test = ["pytest"]
+
+[tool.setuptools]
+packages = ["ddh"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

tests/test_hazard_head.py

@@ -210,8 +210,8 @@ def test_proportional_keeps_risk_ordering_constant_across_bins():
     step = pmf.sum(-1)
     survival = 1.0 - step.cumsum(1)
     s_before = torch.cat([torch.ones(N, 1, dtype=torch.float64), survival[:, :-1]], 1)
-    lam_vte = (pmf[:, :, 0] / s_before)                       # [N, B]
-    order = lam_vte.argsort(dim=0)
+    lam_1 = (pmf[:, :, 0] / s_before)                         # [N, B]
+    order = lam_1.argsort(dim=0)
     for b in range(1, B):
         assert torch.equal(order[:, b], order[:, 0]), f"ranking changed at bin {b}"
 

tests/test_horizon_likelihood.py

@@ -1,5 +1,5 @@
 """The horizon-specific likelihood term must be (a) inert at gamma=0, (b) scored on
-exactly the labels and weights V3's evaluation uses, and (c) minimised at the truth."""
+exactly the labels and weights the landmark evaluation uses, and (c) minimised at the truth."""
 import sys
 from pathlib import Path
 
@@ -104,8 +104,8 @@ def test_horizon_inside_the_landmark_bin_contributes_nothing():
     assert parts["horizon"] == 0.0
 
 
-def test_labels_and_weights_match_v3_evaluation():
-    """Graf/Gerds-Schumacher, as V3's parts(): cause-2 deaths before h are true
+def test_labels_and_weights_match_landmark_evaluation():
+    """Graf/Gerds-Schumacher, as the landmark evaluation: cause-2 deaths before h are true
     zeros carrying 1/G(T-), survivors past h carry 1/G(h), censored-before-h
     carry nothing."""
     n_bins = 10
@@ -114,7 +114,7 @@ def test_labels_and_weights_match_v3_evaluation():
     s_all = torch.zeros(4, 1).long()
     visit_days = torch.zeros(4, 1, dtype=torch.double)
     t = torch.tensor([3, 3, 9, 3]).long()          # bins
-    e = torch.tensor([1, 2, 0, 0]).int()           # VTE, death, censored late, censored early
+    e = torch.tensor([1, 2, 0, 0]).int()           # cause 1, cause 2, censored late, censored early
     split_time = torch.linspace(0, 1000, n_bins + 1).double()   # bin width 100
     # horizon 500 days -> bucketize(500) - 1 = bin 4
     mass = torch.full((2, n_bins), 0.02, dtype=torch.double)
@@ -130,8 +130,8 @@ def test_labels_and_weights_match_v3_evaluation():
     b_h = int(torch.bucketize(torch.tensor([500.0]), split_time, right=False).item()) - 1
     p = float((cif[0][0, b_h] - cif[0][0, 0]) / (1 - cif[0][0, 0] - cif[1][0, 0]))
     want = 0.0
-    want += -np.log(p) * float(1 / g[2])            # VTE at bin 3 <= 4, y=1, 1/G(T-)
-    want += -np.log(1 - p) * float(1 / g[2])        # death at bin 3, y=0, 1/G(T-)
+    want += -np.log(p) * float(1 / g[2])            # cause 1 at bin 3 <= 4, y=1, 1/G(T-)
+    want += -np.log(1 - p) * float(1 / g[2])        # cause 2 at bin 3, y=0, 1/G(T-)
     want += -np.log(1 - p) * float(1 / g[b_h])      # alive past h, y=0, 1/G(h)
     want += 0.0                                     # censored at bin 3 < h: dropped
     assert abs(val.item() - want / 2) < 1e-9, (val.item(), want / 2)