Update documentation.

csaroff · Jan 21, 2019 · 874e217 · 874e217
1 parent 0e74eb5
commit 874e217
Show file tree

Hide file tree

Showing 5 changed files with 81 additions and 79 deletions.
diff --git a/docs/api.md b/docs/api.md
@@ -1,4 +1,5 @@
 Here is the only method `pytorch_to_keras` from `pytorch2keras` module.
+
 ```
 def pytorch_to_keras(
     model, args, input_shapes,
@@ -8,10 +9,10 @@ def pytorch_to_keras(
 
 Options:
 
-* model -- a PyTorch module to convert;
-* args -- list of dummy variables with proper shapes;
-* input_shapes -- list with shape tuples;
-* change_ordering -- boolean, if enabled, the converter will try to change `BCHW` to `BHWC`
-* training -- boolean, switch model to training mode (never use it)
-* verbose -- boolean, verbose output
-* names -- choice from [`keep`, `short`, `random`]. The selector set the target layer naming policy.
+	* `model` -- a PyTorch module to convert;
+	* `args` -- list of dummy variables with proper shapes;
+	* `input_shapes` -- list with shape tuples;
+	* `change_ordering` -- boolean, if enabled, the converter will try to change `BCHW` to `BHWC`
+	* `training` -- boolean, switch model to training mode (never use it)
+	* `verbose` -- boolean, verbose output
+	* `names` -- choice from [`keep`, `short`, `random`]. The selector set the target layer naming policy.
diff --git a/docs/getting_started.md b/docs/getting_started.md
@@ -4,48 +4,48 @@ It's the converter of PyTorch graph to a Keras (Tensorflow backend) model.
 
 Firstly, we need to load (or create) a valid PyTorch model:
 
-    ```
-    class TestConv2d(nn.Module):
-        """
-        Module for Conv2d testing
-        """
+```
+class TestConv2d(nn.Module):
+    """
+    Module for Conv2d testing
+    """
 
-        def __init__(self, inp=10, out=16, kernel_size=3):
-            super(TestConv2d, self).__init__()
-            self.conv2d = nn.Conv2d(inp, out, stride=1, kernel_size=kernel_size, bias=True)
+    def __init__(self, inp=10, out=16, kernel_size=3):
+        super(TestConv2d, self).__init__()
+        self.conv2d = nn.Conv2d(inp, out, stride=1, kernel_size=kernel_size, bias=True)
 
-        def forward(self, x):
-            x = self.conv2d(x)
-            return x
+    def forward(self, x):
+        x = self.conv2d(x)
+        return x
 
-    model = TestConv2d()
+model = TestConv2d()
 
-    # load weights here
-    # model.load_state_dict(torch.load(path_to_weights.pth))
-    ```
+# load weights here
+# model.load_state_dict(torch.load(path_to_weights.pth))
+```
 
 The next step - create a dummy variable with correct shape:
 
-    ```
-    input_np = np.random.uniform(0, 1, (1, 10, 32, 32))
-    input_var = Variable(torch.FloatTensor(input_np))
-    ```
+```
+input_np = np.random.uniform(0, 1, (1, 10, 32, 32))
+input_var = Variable(torch.FloatTensor(input_np))
+```
 
 We use the dummy-variable to trace the model (with jit.trace):
 
-    ```
-    from converter import pytorch_to_keras
-    # we should specify shape of the input tensor
-    k_model = pytorch_to_keras(model, input_var, [(10, 32, 32,)], verbose=True)  
-    ```
+```
+from converter import pytorch_to_keras
+# we should specify shape of the input tensor
+k_model = pytorch_to_keras(model, input_var, [(10, 32, 32,)], verbose=True)  
+```
 
 You can also set H and W dimensions to None to make your model shape-agnostic (e.g. fully convolutional netowrk):
 
-    ```
-    from converter import pytorch_to_keras
-    # we should specify shape of the input tensor
-    k_model = pytorch_to_keras(model, input_var, [(10, None, None,)], verbose=True)  
-    ```
+```
+from converter import pytorch_to_keras
+# we should specify shape of the input tensor
+k_model = pytorch_to_keras(model, input_var, [(10, None, None,)], verbose=True)  
+```
 
 That's all! If all the modules have converted properly, the Keras model will be stored in the `k_model` variable.
 
@@ -63,7 +63,6 @@ Here is a short instruction how to get a tensorflow.js model:
 
 2. Now you have Keras model. You can save it as h5 file and then convert it with `tensorflowjs_converter` but it doesn't work sometimes. As alternative, you may get Tensorflow Graph and save it as a frozen model:
 
-        ```
         # Function below copied from here:
         # https://stackoverflow.com/questions/45466020/how-to-export-keras-h5-to-tensorflow-pb 
         def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
@@ -103,22 +102,21 @@ Here is a short instruction how to get a tensorflow.js model:
 
         tf.train.write_graph(frozen_graph, ".", "my_model.pb", as_text=False)
         print([i for i in k_model.outputs])
-        ```
 
 3. You will see the output layer name, so, now it's time to convert `my_model.pb` to tfjs model:
 
-        ```
-        tensorflowjs_converter  \
-            --input_format=tf_frozen_model \
-            --output_node_names='TANHTObs/Tanh' \
-            my_model.pb \
-            model_tfjs
-        ```
+    ```
+    tensorflowjs_converter  \
+        --input_format=tf_frozen_model \
+        --output_node_names='TANHTObs/Tanh' \
+        my_model.pb \
+        model_tfjs
+    ```
 
 4. Thats all!
 
-        ```
-        const MODEL_URL = `model_tfjs/tensorflowjs_model.pb`;
-        const WEIGHTS_URL = `model_tfjs/weights_manifest.json`;
-        cont model = await tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL);
-        ```
+    ```
+    const MODEL_URL = `model_tfjs/tensorflowjs_model.pb`;
+    const WEIGHTS_URL = `model_tfjs/weights_manifest.json`;
+    cont model = await tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL);
+    ```
diff --git a/docs/index.md b/docs/index.md
@@ -7,32 +7,3 @@
 ![PyPI](https://img.shields.io/pypi/v/pytorch2keras.svg)
 
 PyTorch to Keras model converter. This project is created to make a model conversation easier, so, the converter API is developed with maximal simplicity.
-
-
-## PyTorch 0.4.1 and greater
-
-There is [the problem related to a new version](https://github.com/pytorch/pytorch/issues/13963):
-
-To make it work, please, cast all your `.view()` parameters to `int`. For example:
-
-```
-class ResNet(torchvision.models.resnet.ResNet):
-    def __init__(self, *args, **kwargs):
-        super(ResNet, self).__init__(*args, **kwargs)
-
-    def forward(self, x):
-        x = self.conv1(x)
-        x = self.bn1(x)
-        x = self.relu(x)
-        x = self.maxpool(x)
-
-        x = self.layer1(x)
-        x = self.layer2(x)
-        x = self.layer3(x)
-        x = self.layer4(x)
-
-        x = self.avgpool(x)
-        x = x.view(int(x.size(0)), -1)  #  << Here
-        x = self.fc(x)
-        return x
-```
diff --git a/docs/known_problems.md b/docs/known_problems.md
@@ -0,0 +1,31 @@
+## Linear layer problem with PyTorch 0.4.1 and greater
+
+There is [the problem related to a new version](https://github.com/pytorch/pytorch/issues/13963):
+
+To make it work, please, cast all your `.view()` parameters to `int`. For example:
+
+```
+class ResNet(torchvision.models.resnet.ResNet):
+    def __init__(self, *args, **kwargs):
+        super(ResNet, self).__init__(*args, **kwargs)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        x = x.view(int(x.size(0)), -1)  #  << Here
+        x = self.fc(x)
+        return x
+```
+
+## Recurrent layers
+
+The recurrent layers are not supported due to complicated onnx-translation. The support is planned, but haven't implemented yet.
diff --git a/mkdocs.yml b/mkdocs.yml
@@ -7,4 +7,5 @@ pages:
 - Getting started: getting_started.md
 - Supported layers and models: supported_layers_and_models.md
 - API: api.md
+- Known problems: known_problems.md
 - License: license.md