Update to using pytorch-1.0-rc (#1431)

.travis.yml

@@ -12,11 +12,8 @@ cache:
 
 install:
     - pip install -U pip
-    - if [[ $TRAVIS_PYTHON_VERSION == 2.7 ]]; then
-          pip install http://download.pytorch.org/whl/cpu/torch-0.4.0-cp27-cp27mu-linux_x86_64.whl;
-      else
-          pip install http://download.pytorch.org/whl/cpu/torch-0.4.0-cp36-cp36m-linux_x86_64.whl;
-      fi
+    - pip install torch_nightly==1.0.0.dev20181128 -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
+    - pip install torchvision --no-dependencies
     - pip install .[test]
     - pip freeze
 

Makefile

@@ -57,6 +57,12 @@ test-cuda: lint FORCE
 	CUDA_TEST=1 PYRO_TENSOR_TYPE=torch.cuda.DoubleTensor pytest -vx --stage unit
 	CUDA_TEST=1 pytest -vx tests/test_examples.py::test_cuda
 
+test-jit: FORCE
+	@echo See jit.log
+	pytest -v -n auto --tb=short --runxfail tests/infer/test_jit.py tests/test_examples.py::test_jit | tee jit.log
+	pytest -v -n auto --tb=short --runxfail tests/infer/mcmc/test_hmc.py tests/infer/mcmc/test_nuts.py \
+		-k JIT=True | tee -a jit.log
+
 clean: FORCE
 	git clean -dfx -e pyro-egg.info
 

README.md

@@ -62,27 +62,14 @@ Make sure that the models come from the same release version of the [Pyro source
 
 For recent features you can install Pyro from source.
 
-To install a compatible CPU version of PyTorch on OSX / Linux, you
-could use the PyTorch install helper script.
+To install a compatible version of PyTorch, use the PyTorch nightly
+[build](https://pytorch.org/). We recommend pinning to the specific
+nightly build below that has been well tested.
 
-```
-bash scripts/install_pytorch.sh
-```
-
-Alternatively, build PyTorch following instructions in the PyTorch
-[README](https://github.com/pytorch/pytorch/blob/master/README.md).
-```sh
-git clone --recursive https://github.com/pytorch/pytorch
-cd pytorch
-git checkout 200fb22  # <---- a well-tested commit
-```
-On Linux:
-```sh
-python setup.py install
-```
-On OSX:
 ```sh
-MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
+build_ver=1.0.0.dev20181127  # <---- a well-tested PyTorch build
+pip install torch_nightly==${build_ver} -f \
+    https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
 ```
 
 Finally install Pyro using pip or from source as shown below.

docs/source/conf.py

@@ -205,7 +205,8 @@ def setup(app):
 
 # @jpchen's hack to get rtd builder to install latest pytorch
 if 'READTHEDOCS' in os.environ:
-    os.system('curl -o install.sh https://raw.githubusercontent.com/uber/pyro/dev/scripts/install_pytorch.sh')
-    os.system('curl https://raw.githubusercontent.com/uber/pyro/dev/README.md > README.md')
-    os.system('bash install.sh')
-    os.system('rm -f install.sh')
+    os.system('pip install http://download.pytorch.org/whl/cpu/torch-0.4.1-cp27-cp27mu-linux_x86_64.whl')
+    # for pytorch 1.0 (currently fails with OOM
+    # https://readthedocs.org/projects/pyro-ppl/builds/8159615/
+#     os.system('pip install torch_nightly==1.0.0.dev20181127 -f '
+#               'https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html')

docs/source/distributions.rst

@@ -56,14 +56,6 @@ AVFMultivariateNormal
     :undoc-members:
     :show-inheritance:
 
-Binomial
---------
-
-.. autoclass:: pyro.distributions.Binomial
-    :members:
-    :undoc-members:
-    :show-inheritance:
-
 Delta
 -----
 .. autoclass:: pyro.distributions.Delta
@@ -85,20 +77,6 @@ GaussianScaleMixture
     :undoc-members:
     :show-inheritance:
 
-HalfCauchy
-----------
-.. autoclass:: pyro.distributions.HalfCauchy
-    :members:
-    :undoc-members:
-    :show-inheritance:
-
-LowRankMultivariateNormal
--------------------------
-.. autoclass:: pyro.distributions.LowRankMultivariateNormal
-    :members:
-    :undoc-members:
-    :show-inheritance:
-
 MaskedMixture
 -------------
 .. autoclass:: pyro.distributions.MaskedMixture

docs/source/primitives.rst

@@ -16,4 +16,4 @@ Primitives
 .. autofunction:: pyro.validation_enabled
 .. autofunction:: pyro.enable_validation
 
-.. autofunction:: pyro.ops.jit.compile
+.. autofunction:: pyro.ops.jit.trace

examples/baseball.py

@@ -333,5 +333,7 @@ def main(args):
     parser.add_argument("--num-chains", nargs='?', default=4, type=int)
     parser.add_argument("--warmup-steps", nargs='?', default=100, type=int)
     parser.add_argument("--rng_seed", nargs='?', default=0, type=int)
+    parser.add_argument('--jit', action='store_true', default=False,
+                        help='use PyTorch jit')
     args = parser.parse_args()
     main(args)

examples/eight_schools/mcmc.py

@@ -34,7 +34,7 @@ def conditioned_model(model, sigma, y):
 
 
 def main(args):
-    nuts_kernel = NUTS(conditioned_model, adapt_step_size=True)
+    nuts_kernel = NUTS(conditioned_model, jit_compile=args.jit,)
     posterior = MCMC(nuts_kernel,
                      num_samples=args.num_samples,
                      warmup_steps=args.warmup_steps,
@@ -58,6 +58,7 @@ def main(args):
                         help='number of parallel MCMC chains (default: 1)')
     parser.add_argument('--warmup-steps', type=int, default=1000,
                         help='number of MCMC samples for warmup (default: 1000)')
+    parser.add_argument('--jit', action='store_true', default=False)
     args = parser.parse_args()
 
     main(args)

examples/hmm.py

@@ -324,7 +324,7 @@ def main(args):
     # We'll train on small minibatches.
     logging.info('Step\tLoss')
     for step in range(args.num_steps):
-        loss = svi.step(sequences, lengths, args, batch_size=args.batch_size)
+        loss = svi.step(sequences, lengths, args=args, batch_size=args.batch_size)
         logging.info('{: >5d}\t{}'.format(step, loss / num_observations))
 
     # We evaluate on the entire training dataset,
@@ -340,7 +340,7 @@ def main(args):
     if args.truncate:
         lengths.clamp_(max=args.truncate)
     num_observations = float(lengths.sum())
-    test_loss = elbo.loss(model, guide, sequences, lengths, args, include_prior=False)
+    test_loss = elbo.loss(model, guide, sequences, lengths, args=args, include_prior=False)
     logging.info('test loss = {}'.format(test_loss / num_observations))
 
     # We expect models with higher capacity to perform better,

examples/vae/utils/mnist_cached.py

@@ -9,7 +9,6 @@
 
 from pyro.contrib.examples.util import get_data_directory
 
-
 # This file contains utilities for caching, transforming and splitting MNIST data
 # efficiently. By default, a PyTorch DataLoader will apply the transform every epoch
 # we avoid this by caching the data early on in MNISTCached class

pyro/contrib/autoguide/__init__.py

@@ -642,7 +642,7 @@ def laplace_approximation(self, *args, **kwargs):
         loc = pyro.param("{}_loc".format(self.prefix))
         H = hessian(loss, loc.unconstrained())
         cov = H.inverse()
-        scale_tril = cov.potrf(upper=False)
+        scale_tril = cov.cholesky()
 
         # calculate scale_tril from self.guide()
         scale_tril_name = "{}_scale_tril".format(self.prefix)

pyro/contrib/examples/util.py

@@ -3,9 +3,23 @@
 import os
 import sys
 
+import torchvision
 import torchvision.datasets as datasets
 from torch.utils.data import DataLoader
 from torchvision import transforms
+from torchvision.datasets import MNIST
+
+from pyro.distributions.torch_patch import patch_dependency
+
+
+@patch_dependency('torchvision.datasets.MNIST', torchvision)
+class _MNIST(getattr(MNIST, '_pyro_unpatched', MNIST)):
+    urls = [
+        "https://d2fefpcigoriu7.cloudfront.net/datasets/mnist/train-images-idx3-ubyte.gz",
+        "https://d2fefpcigoriu7.cloudfront.net/datasets/mnist/train-labels-idx1-ubyte.gz",
+        "https://d2fefpcigoriu7.cloudfront.net/datasets/mnist/t10k-images-idx3-ubyte.gz",
+        "https://d2fefpcigoriu7.cloudfront.net/datasets/mnist/t10k-labels-idx1-ubyte.gz",
+    ]
 
 
 def get_data_loader(dataset_name,
@@ -18,11 +32,14 @@ def get_data_loader(dataset_name,
         dataset_transforms = []
     trans = transforms.Compose([transforms.ToTensor()] + dataset_transforms)
     dataset = getattr(datasets, dataset_name)
+    print("downloading data")
+    dset = dataset(root=data_dir,
+                   train=is_training_set,
+                   transform=trans,
+                   download=True)
+    print("download complete.")
     return DataLoader(
-        dataset(root=data_dir,
-                train=is_training_set,
-                transform=trans,
-                download=True),
+        dset,
         batch_size=batch_size,
         shuffle=shuffle
     )

pyro/contrib/gp/models/gplvm.py

@@ -1,14 +1,14 @@
 from __future__ import absolute_import, division, print_function
 
-import torch
 from torch.distributions import constraints
 from torch.nn import Parameter
 
 import pyro
-from pyro.contrib.gp.util import Parameterized
 import pyro.distributions as dist
 import pyro.infer as infer
 import pyro.optim as optim
+from pyro.contrib.gp.util import Parameterized
+from pyro.distributions.util import eye_like
 from pyro.params import param_with_module_name
 
 
@@ -74,7 +74,7 @@ def __init__(self, base_model, name="GPLVM"):
 
         C = self.X_loc.shape[1]
         X_scale_tril_shape = self.X_loc.shape + (C,)
-        Id = torch.eye(C, out=self.X_loc.new_empty(C, C))
+        Id = eye_like(self.X_loc, C)
         X_scale_tril = Id.expand(X_scale_tril_shape)
         self.X_scale_tril = Parameter(X_scale_tril)
         self.set_constraint("X_scale_tril", constraints.lower_cholesky)
@@ -87,7 +87,7 @@ def model(self):
         # sample X from unit multivariate normal distribution
         zero_loc = self.X_loc.new_zeros(self.X_loc.shape)
         C = self.X_loc.shape[1]
-        Id = torch.eye(C, out=self.X_loc.new_empty(C, C))
+        Id = eye_like(self.X_loc, C)
         X_name = param_with_module_name(self.name, "X")
         X = pyro.sample(X_name, dist.MultivariateNormal(zero_loc, scale_tril=Id)
                                     .independent(zero_loc.dim()-1))

pyro/contrib/gp/models/gpr.py

@@ -81,7 +81,7 @@ def model(self):
         N = self.X.shape[0]
         Kff = self.kernel(self.X)
         Kff.view(-1)[::N + 1] += noise  # add noise to diagonal
-        Lff = Kff.potrf(upper=False)
+        Lff = Kff.cholesky()
 
         zero_loc = self.X.new_zeros(self.X.shape[0])
         f_loc = zero_loc + self.mean_function(self.X)
@@ -129,7 +129,7 @@ def forward(self, Xnew, full_cov=False, noiseless=True):
         N = self.X.shape[0]
         Kff = self.kernel(self.X).contiguous()
         Kff.view(-1)[::N + 1] += noise  # add noise to the diagonal
-        Lff = Kff.potrf(upper=False)
+        Lff = Kff.cholesky()
 
         y_residual = self.y - self.mean_function(self.X)
         loc, cov = conditional(Xnew, self.X, self.kernel, y_residual, None, Lff,
@@ -185,7 +185,7 @@ def sample_next(xnew, outside_vars):
             X, y, Kff = outside_vars["X"], outside_vars["y"], outside_vars["Kff"]
 
             # Compute Cholesky decomposition of kernel matrix
-            Lff = Kff.potrf(upper=False)
+            Lff = Kff.cholesky()
             y_residual = y - self.mean_function(X)
 
             # Compute conditional mean and variance

pyro/contrib/gp/models/sgpr.py

@@ -130,7 +130,7 @@ def model(self):
         M = Xu.shape[0]
         Kuu = self.kernel(Xu).contiguous()
         Kuu.view(-1)[::M + 1] += self.jitter  # add jitter to the diagonal
-        Luu = Kuu.potrf(upper=False)
+        Luu = Kuu.cholesky()
         Kuf = self.kernel(Xu, self.X)
         W = Kuf.trtrs(Luu, upper=False)[0].t()
 
@@ -210,7 +210,7 @@ def forward(self, Xnew, full_cov=False, noiseless=True):
 
         Kuu = self.kernel(Xu).contiguous()
         Kuu.view(-1)[::M + 1] += self.jitter  # add jitter to the diagonal
-        Luu = Kuu.potrf(upper=False)
+        Luu = Kuu.cholesky()
         Kus = self.kernel(Xu, Xnew)
         Kuf = self.kernel(Xu, self.X)
 
@@ -225,7 +225,7 @@ def forward(self, Xnew, full_cov=False, noiseless=True):
         W_Dinv = W / D
         K = W_Dinv.matmul(W.t()).contiguous()
         K.view(-1)[::M + 1] += 1  # add identity matrix to K
-        L = K.potrf(upper=False)
+        L = K.cholesky()
 
         # get y_residual and convert it into 2D tensor for packing
         y_residual = self.y - self.mean_function(self.X)

pyro/contrib/gp/models/vgp.py

@@ -8,6 +8,7 @@
 import pyro.distributions as dist
 from pyro.contrib.gp.models.model import GPModel
 from pyro.contrib.gp.util import conditional
+from pyro.distributions.util import eye_like
 from pyro.params import param_with_module_name
 
 
@@ -74,7 +75,7 @@ def __init__(self, X, y, kernel, likelihood, mean_function=None,
         self.f_loc = Parameter(f_loc)
 
         f_scale_tril_shape = self.latent_shape + (N, N)
-        Id = torch.eye(N, out=self.X.new_empty(N, N))
+        Id = eye_like(self.X, N)
         f_scale_tril = Id.expand(f_scale_tril_shape)
         self.f_scale_tril = Parameter(f_scale_tril)
         self.set_constraint("f_scale_tril", constraints.lower_cholesky)
@@ -90,13 +91,13 @@ def model(self):
         N = self.X.shape[0]
         Kff = self.kernel(self.X).contiguous()
         Kff.view(-1)[::N + 1] += self.jitter  # add jitter to the diagonal
-        Lff = Kff.potrf(upper=False)
+        Lff = Kff.cholesky()
 
         zero_loc = self.X.new_zeros(f_loc.shape)
         f_name = param_with_module_name(self.name, "f")
 
         if self.whiten:
-            Id = torch.eye(N, out=self.X.new_empty(N, N))
+            Id = eye_like(self.X, N)
             pyro.sample(f_name,
                         dist.MultivariateNormal(zero_loc, scale_tril=Id)
                             .independent(zero_loc.dim() - 1))

pyro/contrib/gp/models/vsgp.py

@@ -9,6 +9,7 @@
 import pyro.poutine as poutine
 from pyro.contrib.gp.models.model import GPModel
 from pyro.contrib.gp.util import conditional
+from pyro.distributions.util import eye_like
 from pyro.params import param_with_module_name
 
 
@@ -98,7 +99,7 @@ def __init__(self, X, y, kernel, Xu, likelihood, mean_function=None,
         self.u_loc = Parameter(u_loc)
 
         u_scale_tril_shape = self.latent_shape + (M, M)
-        Id = torch.eye(M, out=self.Xu.new_empty(M, M))
+        Id = eye_like(self.Xu, M)
         u_scale_tril = Id.expand(u_scale_tril_shape)
         self.u_scale_tril = Parameter(u_scale_tril)
         self.set_constraint("u_scale_tril", constraints.lower_cholesky)
@@ -115,12 +116,12 @@ def model(self):
         M = Xu.shape[0]
         Kuu = self.kernel(Xu).contiguous()
         Kuu.view(-1)[::M + 1] += self.jitter  # add jitter to the diagonal
-        Luu = Kuu.potrf(upper=False)
+        Luu = Kuu.cholesky()
 
         zero_loc = Xu.new_zeros(u_loc.shape)
         u_name = param_with_module_name(self.name, "u")
         if self.whiten:
-            Id = torch.eye(M, out=Xu.new_empty(M, M))
+            Id = eye_like(Xu, M)
             pyro.sample(u_name,
                         dist.MultivariateNormal(zero_loc, scale_tril=Id)
                             .independent(zero_loc.dim() - 1))

pyro/contrib/gp/util.py

@@ -212,7 +212,7 @@ def conditional(Xnew, X, kernel, f_loc, f_scale_tril=None, Lff=None, full_cov=Fa
     if Lff is None:
         Kff = kernel(X).contiguous()
         Kff.view(-1)[::N + 1] += jitter  # add jitter to diagonal
-        Lff = Kff.potrf(upper=False)
+        Lff = Kff.cholesky()
     Kfs = kernel(X, Xnew)
 
     # convert f_loc_shape from latent_shape x N to N x latent_shape