Merge pull request #153 from masa-su/develop/v0.3.0

Develop/v0.3.0

README.md

@@ -86,7 +86,7 @@ and
 >>> # inference model (encoder) q(z|x)
 >>> class Inference(Normal):
 ...     def __init__(self):
-...         super(Inference, self).__init__(cond_var=["x"], var=["z"], name="q")  # var: variables of this distribution, cond_var: coditional variables.
+...         super(Inference, self).__init__(var=["z"],cond_var=["x"],name="q")  # var: variables of this distribution, cond_var: coditional variables.
 ...         self.fc1 = nn.Linear(784, 512)
 ...         self.fc21 = nn.Linear(512, 64)
 ...         self.fc22 = nn.Linear(512, 64)
@@ -99,7 +99,7 @@ and
 >>> # generative model (decoder) p(x|z)    
 >>> class Generator(Bernoulli):
 ...     def __init__(self):
-...         super(Generator, self).__init__(cond_var=["z"], var=["x"], name="p")
+...         super(Generator, self).__init__(var=["x"], cond_var=["z"], name="p")
 ...         self.fc1 = nn.Linear(64, 512)
 ...         self.fc2 = nn.Linear(512, 128)
 ... 
@@ -196,7 +196,7 @@ Next, we set the objective (loss) function of the model with defined distributio
 ```python
 >>> from pixyz.losses import KullbackLeibler, LogProb, Expectation as E
 >>> reconst = -E(q, LogProb(p)) # the reconstruction loss (it can also be written as `-p.log_prob().expectation()`)
->>> kl = KullbackLeibler(q, prior) # Kullback-Leibler divergence
+>>> kl = KullbackLeibler(q,prior) # Kullback-Leibler divergence
 >>> loss_cls = (kl + reconst).mean()
 ```
 

docs/source/distributions.rst

@@ -141,10 +141,10 @@ Deterministic
     :members:
     :undoc-members:
 
-DataDistribution
+EmpiricalDistribution
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: DataDistribution
+.. autoclass:: EmpiricalDistribution
     :members:
     :undoc-members:
 

docs/source/losses.rst

@@ -144,6 +144,14 @@ ValueLoss
     :members:
     :undoc-members:
 
+ConstantVar
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. currentmodule:: pixyz.losses.losses
+.. autoclass:: ConstantVar
+    :members:
+    :undoc-members:
+
 
 Operators
 ----------------------------

examples/cvae.ipynb

@@ -79,7 +79,7 @@
     "# inference model q(z|x,y)\n",
     "class Inference(Normal):\n",
     "    def __init__(self):\n",
-    "        super(Inference, self).__init__(cond_var=[\"x\",\"y\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Inference, self).__init__(var=[\"z\"], cond_var=[\"x\",\"y\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim+y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -95,7 +95,7 @@
     "# generative model p(x|z,y)    \n",
     "class Generator(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(Generator, self).__init__(cond_var=[\"z\",\"y\"], var=[\"x\"], name=\"p\")\n",
+    "        super(Generator, self).__init__(var=[\"x\"], cond_var=[\"z\",\"y\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim+y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/gan.ipynb

@@ -76,7 +76,7 @@
     "# generator model p(x|z)    \n",
     "class Generator(Deterministic):\n",
     "    def __init__(self):\n",
-    "        super(Generator, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(Generator, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        def block(in_feat, out_feat, normalize=True):\n",
     "            layers = [nn.Linear(in_feat, out_feat)]\n",
@@ -174,7 +174,7 @@
     "# discriminator model p(t|x)\n",
     "class Discriminator(Deterministic):\n",
     "    def __init__(self):\n",
-    "        super(Discriminator, self).__init__(cond_var=[\"x\"], var=[\"t\"], name=\"d\")\n",
+    "        super(Discriminator, self).__init__(var=[\"t\"], cond_var=[\"x\"], name=\"d\")\n",
     "\n",
     "        self.model = nn.Sequential(\n",
     "            nn.Linear(x_dim, 512),\n",

examples/hierarchical_variational_inference.ipynb

@@ -78,7 +78,7 @@
     "# inference models\n",
     "class Q1(Normal):\n",
     "    def __init__(self):\n",
-    "        super(Q1, self).__init__(cond_var=[\"x\"], var=[\"a\"], name=\"q\")\n",
+    "        super(Q1, self).__init__(var=[\"a\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -93,7 +93,7 @@
     "\n",
     "class Q2(Normal):\n",
     "    def __init__(self):\n",
-    "        super(Q2, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Q2, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -115,7 +115,7 @@
     "# generative models\n",
     "class P2(Normal):\n",
     "    def __init__(self):\n",
-    "        super(P2, self).__init__(cond_var=[\"z\"], var=[\"a\"], name=\"p\")\n",
+    "        super(P2, self).__init__(var=[\"a\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -130,7 +130,7 @@
     "    \n",
     "class P3(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(P3, self).__init__(cond_var=[\"a\"], var=[\"x\"], name=\"p\")\n",
+    "        super(P3, self).__init__(var=[\"x\"], cond_var=[\"a\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(a_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/jmvae.ipynb

@@ -80,7 +80,7 @@
     "# inference model q(z|x,y)\n",
     "class Inference(Normal):\n",
     "    def __init__(self):\n",
-    "        super(Inference, self).__init__(cond_var=[\"x\",\"y\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Inference, self).__init__(var=[\"z\"], cond_var=[\"x\",\"y\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim+y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -96,7 +96,7 @@
     "# inference model q(z|x)\n",
     "class InferenceX(Normal):\n",
     "    def __init__(self):\n",
-    "        super(InferenceX, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(InferenceX, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -112,7 +112,7 @@
     "# inference model q(z|y)\n",
     "class InferenceY(Normal):\n",
     "    def __init__(self):\n",
-    "        super(InferenceY, self).__init__(cond_var=[\"y\"], var=[\"z\"], name=\"q\")\n",
+    "        super(InferenceY, self).__init__(var=[\"z\"], cond_var=[\"y\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -128,7 +128,7 @@
     "# generative model p(x|z)    \n",
     "class GeneratorX(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(GeneratorX, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(GeneratorX, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -143,7 +143,7 @@
     "# generative model p(y|z)    \n",
     "class GeneratorY(Categorical):\n",
     "    def __init__(self):\n",
-    "        super(GeneratorY, self).__init__(cond_var=[\"z\"], var=[\"y\"], name=\"p\")\n",
+    "        super(GeneratorY, self).__init__(var=[\"y\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/jmvae_poe.ipynb

@@ -82,7 +82,7 @@
     "# inference model q(z|x)\n",
     "class InferenceX(Normal):\n",
     "    def __init__(self):\n",
-    "        super(InferenceX, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(InferenceX, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -98,7 +98,7 @@
     "# inference model q(z|y)\n",
     "class InferenceY(Normal):\n",
     "    def __init__(self):\n",
-    "        super(InferenceY, self).__init__(cond_var=[\"y\"], var=[\"z\"], name=\"q\")\n",
+    "        super(InferenceY, self).__init__(var=[\"z\"], cond_var=[\"y\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -114,7 +114,7 @@
     "# generative model p(x|z)    \n",
     "class GeneratorX(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(GeneratorX, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(GeneratorX, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -129,7 +129,7 @@
     "# generative model p(y|z)    \n",
     "class GeneratorY(Categorical):\n",
     "    def __init__(self):\n",
-    "        super(GeneratorY, self).__init__(cond_var=[\"z\"], var=[\"y\"], name=\"p\")\n",
+    "        super(GeneratorY, self).__init__(var=[\"y\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/m2.ipynb

@@ -109,7 +109,7 @@
     "# inference model q(z|x,y)\n",
     "class Inference(Normal):\n",
     "    def __init__(self):\n",
-    "        super().__init__(cond_var=[\"x\",\"y\"], var=[\"z\"], name=\"q\")\n",
+    "        super().__init__(var=[\"z\"], cond_var=[\"x\",\"y\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim+y_dim, 512)\n",
     "        self.fc21 = nn.Linear(512, z_dim)\n",
@@ -123,7 +123,7 @@
     "# generative model p(x|z,y)    \n",
     "class Generator(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super().__init__(cond_var=[\"z\",\"y\"], var=[\"x\"], name=\"p\")\n",
+    "        super().__init__(var=[\"x\"], cond_var=[\"z\",\"y\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim+y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, x_dim)\n",
@@ -136,7 +136,7 @@
     "# classifier p(y|x)\n",
     "class Classifier(RelaxedCategorical):\n",
     "    def __init__(self):\n",
-    "        super(Classifier, self).__init__(cond_var=[\"x\"], var=[\"y\"], name=\"p\")\n",
+    "        super(Classifier, self).__init__(var=[\"y\"], cond_var=[\"x\"], name=\"p\")\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, y_dim)\n",
     "\n",

examples/maximum_likelihood.ipynb

@@ -77,7 +77,7 @@
     "# classifier p(y|x)\n",
     "class Classifier(Categorical):\n",
     "    def __init__(self):\n",
-    "        super(Classifier, self).__init__(cond_var=[\"x\"], var=[\"y\"])\n",
+    "        super(Classifier, self).__init__(var=[\"y\"], cond_var=[\"x\"])\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
     "        self.fc3 = nn.Linear(512, y_dim)\n",

examples/mmd_vae.ipynb

@@ -59,7 +59,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from pixyz.distributions import Normal, Bernoulli, DataDistribution\n",
+    "from pixyz.distributions import Normal, Bernoulli, EmpiricalDistribution\n",
     "from pixyz.losses import CrossEntropy, MMD\n",
     "from pixyz.models import Model\n",
     "from pixyz.utils import print_latex"
@@ -78,7 +78,7 @@
     "# inference model q(z|x)\n",
     "class Inference(Normal):\n",
     "    def __init__(self):\n",
-    "        super(Inference, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Inference, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -94,7 +94,7 @@
     "# generative model p(x|z)    \n",
     "class Generator(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(Generator, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(Generator, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -112,7 +112,7 @@
     "prior = Normal(loc=torch.tensor(0.), scale=torch.tensor(1.),\n",
     "               var=[\"z\"], features_shape=[z_dim], name=\"p_{prior}\").to(device)\n",
     "\n",
-    "p_data = DataDistribution([\"x\"]).to(device)\n",
+    "p_data = EmpiricalDistribution([\"x\"]).to(device)\n",
     "q_mg = (q*p_data).marginalize_var(\"x\")\n",
     "q_mg.name = \"q\""
    ]
@@ -169,7 +169,7 @@
       "Distribution:\n",
       "  q(z) = \\int q(z|x)p_{data}(x)dx\n",
       "Network architecture:\n",
-      "  DataDistribution(\n",
+      "  EmpiricalDistribution(\n",
       "    name=p_{data}, distribution_name=Data distribution,\n",
       "    var=['x'], cond_var=[], input_var=['x'], features_shape=torch.Size([])\n",
       "  )\n",

examples/mvae.ipynb

@@ -128,7 +128,7 @@
     "# inference model q(z|x) for image modality\n",
     "class InferenceX(Normal):\n",
     "    def __init__(self):\n",
-    "        super(InferenceX, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(InferenceX, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -144,7 +144,7 @@
     "# inference model q(z|y) for label modality\n",
     "class InferenceY(Normal):\n",
     "    def __init__(self):\n",
-    "        super(InferenceY, self).__init__(cond_var=[\"y\"], var=[\"z\"], name=\"q\")\n",
+    "        super(InferenceY, self).__init__(var=[\"z\"], cond_var=[\"y\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(y_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -160,7 +160,7 @@
     "# generative model p(x|z) \n",
     "class GeneratorX(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(GeneratorX, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(GeneratorX, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -175,7 +175,7 @@
     "# generative model p(y|z)    \n",
     "class GeneratorY(Categorical):\n",
     "    def __init__(self):\n",
-    "        super(GeneratorY, self).__init__(cond_var=[\"z\"], var=[\"y\"], name=\"p\")\n",
+    "        super(GeneratorY, self).__init__(var=[\"y\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/vae.ipynb

@@ -94,7 +94,7 @@
     "    z ~ N(loc, scale)\n",
     "    \"\"\"\n",
     "    def __init__(self):\n",
-    "        super(Inference, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Inference, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -117,7 +117,7 @@
     "    parameterizes the bernoulli(for MNIST) observation likelihood p(x | z)\n",
     "    \"\"\"\n",
     "    def __init__(self):\n",
-    "        super(Generator, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(Generator, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/vae_with_vae_class.ipynb

@@ -94,7 +94,7 @@
     "    z ~ N(loc, scale)\n",
     "    \"\"\"\n",
     "    def __init__(self):\n",
-    "        super(Inference, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Inference, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -117,7 +117,7 @@
     "    parameterizes the bernoulli(for MNIST) observation likelihood p(x | z)\n",
     "    \"\"\"\n",
     "    def __init__(self):\n",
-    "        super(Generator, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(Generator, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

examples/vi.ipynb

@@ -77,7 +77,7 @@
     "# inference model q(z|x)\n",
     "class Inference(Normal):\n",
     "    def __init__(self):\n",
-    "        super(Inference, self).__init__(cond_var=[\"x\"], var=[\"z\"], name=\"q\")\n",
+    "        super(Inference, self).__init__(var=[\"z\"], cond_var=[\"x\"], name=\"q\")\n",
     "\n",
     "        self.fc1 = nn.Linear(x_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",
@@ -93,7 +93,7 @@
     "# generative model p(x|z)    \n",
     "class Generator(Bernoulli):\n",
     "    def __init__(self):\n",
-    "        super(Generator, self).__init__(cond_var=[\"z\"], var=[\"x\"], name=\"p\")\n",
+    "        super(Generator, self).__init__(var=[\"x\"], cond_var=[\"z\"], name=\"p\")\n",
     "\n",
     "        self.fc1 = nn.Linear(z_dim, 512)\n",
     "        self.fc2 = nn.Linear(512, 512)\n",

pixyz/__init__.py

@@ -1,2 +1,2 @@
 name = "pixyz"
-__version__ = "0.2.1"
+__version__ = "0.3.0"