do_forward -> train_step and test_step;fix some typos

EdisonLeeeee · Aug 20, 2020 · 5bf82a5 · 5bf82a5
1 parent 7b74d97
commit 5bf82a5
Show file tree

Hide file tree

Showing 8 changed files with 89 additions and 59 deletions.
diff --git a/README.md b/README.md
@@ -109,25 +109,32 @@ Otherwise you can also use `model.show('model')` or `model.show('train')` to sho
 NOTE: you should install texttable first.
 
 ## Visualization
+NOTE: you must install [SciencePlots](https://github.com/garrettj403/SciencePlots) package for a better preview.
 + Accuracy
 ```python
 import matplotlib.pyplot as plt
-plt.plot(his.history['acc'])
-plt.plot(his.history['val_acc'])
-plt.legend(['Accuracy', 'Val Accuracy'])
-plt.xlabel('Epochs')
-plt.show()
+with plt.style.context(['science', 'no-latex']):
+    plt.plot(his.history['acc'])
+    plt.plot(his.history['val_acc'])
+    plt.legend(['Train Accuracy', 'Val Accuracy'])
+    plt.ylabel('Accuracy')
+    plt.xlabel('Epochs')
+    plt.autoscale(tight=True)
+    plt.show()    
 ```
 ![visualization](https://github.com/EdisonLeeeee/GraphGallery/blob/master/imgs/visualization_acc.png)
 
 + Loss
 ```python
 import matplotlib.pyplot as plt
-plt.plot(his.history['loss'])
-plt.plot(his.history['val_loss'])
-plt.legend(['Loss', 'Val Loss'])
-plt.xlabel('Epochs')
-plt.show()
+with plt.style.context(['science', 'no-latex']):
+    plt.plot(his.history['loss'])
+    plt.plot(his.history['val_loss'])
+    plt.legend(['Train Loss', 'Val Loss'])
+    plt.ylabel('Loss')
+    plt.xlabel('Epochs')
+    plt.autoscale(tight=True)
+    plt.show()    
 ```
 ![visualization](https://github.com/EdisonLeeeee/GraphGallery/blob/master/imgs/visualization_loss.png)
 

diff --git a/graphgallery/__init__.py b/graphgallery/__init__.py
@@ -23,6 +23,6 @@
 from graphgallery import data
 
 
-__version__ = '0.1.9'
+__version__ = '0.1.10'
 
 __all__ = ['graphgallery', 'nn', 'utils', 'sequence', 'data', '__version__']
diff --git a/graphgallery/nn/models/basemodel.py b/graphgallery/nn/models/basemodel.py
@@ -2,13 +2,16 @@
 import random
 import logging
 
+import os.path as osp
 import numpy as np
 import tensorflow as tf
 import scipy.sparse as sp
 
 from graphgallery import config, check_and_convert, asintarr, Bunch
 from graphgallery.utils.type_check import is_list_like
 from graphgallery.utils.misc import print_table
+from graphgallery.data.utils import makedirs
+
 
 
 class BaseModel:
@@ -97,8 +100,8 @@ def __init__(self, adj, x, labels=None, device="CPU:0", seed=None, name=None, **
         self.__sparse = is_adj_sparse
 
         # log path
-        self.weight_dir = "/tmp/weight"
-        self.weight_path = f"{self.weight_dir}/{name}_weights"
+        self.weight_dir = osp.expanduser(osp.normpath("/tmp/weight"))
+        self.weight_path = osp.join(self.weight_dir, f"{name}_weights")
 
         # data types, default: `float32` and `int64`
         self.floatx = config.floatx()
@@ -170,20 +173,19 @@ def _check_inputs(self, adj, x):
             raise RuntimeError(f"The adjacency matrix should be N by N square matrix.")
         return adj, x
 
-    @property
-    def model(self):
-        return self.__model
-
-    ################### TODO: This may cause ERROR #############
     def __getattr__(self, attr):
+    ################### TODO: This may cause ERROR #############
         try:
             return self.__dict__[attr]
         except KeyError:
             if hasattr(self.model, attr):
                 return getattr(self.model, attr)
             raise AttributeError(f"'{self.name}' and '{self.name}.model' objects have no attribute '{attr}'")
 
-
+    @property
+    def model(self):
+        return self.__model
+
     @model.setter
     def model(self, m):
         # Back up
@@ -194,8 +196,8 @@ def model(self, m):
 
 
     def save(self, path=None, as_model=False):
-        if not os.path.exists(self.weight_dir):
-            os.makedirs(self.weight_dir)
+        if not osp.exists(self.weight_dir):
+            makedirs(self.weight_dir)
             logging.log(logging.WARNING, f"Make Directory in {self.weight_dir}")
 
         if path is None:
@@ -216,6 +218,7 @@ def save(self, path=None, as_model=False):
     def load(self, path=None, as_model=False):
         if not path:
             path = self.weight_path
+
         if not path.endswith('.h5'):
             path += '.h5'
         if as_model:

diff --git a/graphgallery/nn/models/semisupervised/sbvat.py b/graphgallery/nn/models/semisupervised/sbvat.py
@@ -128,13 +128,6 @@ def build(self, hiddens=[16], activations=['relu'], dropouts=[0.5],
         self.epsilon = epsilon  # Norm length for (virtual) adversarial training
         self.n_power_iterations = n_power_iterations  # Number of power iterations
 
-    # def propagation(self, x, adj, training=True):
-    #     h = x
-    #     for layer in self.GCN_layers:
-    #         h = self.dropout_layer(h, training=training)
-    #         h = layer([h, adj])
-    #     return h
-
     def propagation(self, x, adj, training=True):
         h = x
         for dropout_layer, GCN_layer in zip(self.dropout_layers, self.GCN_layers[:-1]):
@@ -145,7 +138,7 @@ def propagation(self, x, adj, training=True):
         return h
 
     @tf.function
-    def do_train_forward(self, sequence):
+    def train_step(self, sequence):
 
         with tf.device(self.device):
             self.train_metric.reset_states()
@@ -171,7 +164,7 @@ def do_train_forward(self, sequence):
         return loss, self.train_metric.result()
 
     @tf.function
-    def do_test_forward(self, sequence):
+    def test_step(self, sequence):
 
         with tf.device(self.device):
             self.test_metric.reset_states()
@@ -186,14 +179,6 @@ def do_test_forward(self, sequence):
 
         return loss, self.test_metric.result()
 
-    def do_forward(self, sequence, training=True):
-        if training:
-            loss, accuracy = self.do_train_forward(sequence)
-        else:
-            loss, accuracy = self.do_test_forward(sequence)
-
-        return loss.numpy(), accuracy.numpy()
-
     def virtual_adversarial_loss(self, x, adj, logit, adv_mask):
         d = tf.random.normal(shape=tf.shape(x), dtype=self.floatx)
 

diff --git a/graphgallery/nn/models/semisupervised/semi_supervised_model.py b/graphgallery/nn/models/semisupervised/semi_supervised_model.py
@@ -220,7 +220,7 @@ def train_v1(self, idx_train, idx_val=None,
             if self.do_before_train:
                 self.do_before_train()
 
-            loss, accuracy = self.do_forward(train_data)
+            loss, accuracy = self.train_step(train_data)
             train_data.on_epoch_end()
 
             history.add_results(loss, 'loss')
@@ -230,7 +230,7 @@ def train_v1(self, idx_train, idx_val=None,
                 if self.do_before_validation:
                     self.do_before_validation()
 
-                val_loss, val_accuracy = self.do_forward(val_data, training=False)
+                val_loss, val_accuracy = self.test_step(val_data)
 
                 history.add_results(val_loss, 'val_loss')
                 history.add_results(val_accuracy, 'val_acc')
@@ -409,7 +409,7 @@ def train(self, idx_train, idx_val=None,
                 self.do_before_train()
 
             callbacks.on_train_batch_begin(0)
-            loss, accuracy = self.do_forward(train_data)
+            loss, accuracy = self.train_step(train_data)
             train_data.on_epoch_end()
 
             training_logs = {'loss': loss, 'acc': accuracy}
@@ -419,7 +419,7 @@ def train(self, idx_train, idx_val=None,
                 if self.do_before_validation:
                     self.do_before_validation()
 
-                val_loss, val_accuracy = self.do_forward(val_data, training=False)
+                val_loss, val_accuracy = self.test_step(val_data)
                 training_logs.update({'val_loss': val_loss, 'val_acc': val_accuracy})
 
             callbacks.on_epoch_end(epoch, training_logs)
@@ -584,7 +584,7 @@ def train_v2(self, idx_train, idx_val=None,
                 self.do_before_train()
 
             callbacks.on_train_batch_begin(0)
-            loss, accuracy = self.do_forward(train_data)
+            loss, accuracy = self.train_step(train_data)
             train_data.on_epoch_end()
 
             training_logs = {'loss': loss, 'acc': accuracy}
@@ -594,7 +594,7 @@ def train_v2(self, idx_train, idx_val=None,
                 if self.do_before_validation:
                     self.do_before_validation()
 
-                val_loss, val_accuracy = self.do_forward(val_data, training=False)
+                val_loss, val_accuracy = self.test_step(val_data)
                 training_logs.update({'val_loss': val_loss, 'val_acc': val_accuracy})
 
             callbacks.on_epoch_end(epoch, training_logs)
@@ -650,16 +650,15 @@ def test(self, index, **kwargs):
         if self.do_before_test:
             self.do_before_test(**kwargs)
 
-        loss, accuracy = self.do_forward(test_data, training=False)
+        loss, accuracy = self.test_step(test_data)
 
         return loss, accuracy
 
-    def do_forward(self, sequence, training=True):
+    def train_step(self, sequence):
         """
             Forward propagation for the input `sequence`. This method will be called 
-            in `train` and `test`, you can rewrite it for you customized training/testing 
-            process. If you want to specify your customized data during traing/testing/predicting, 
-            you can implement a sub-class of `graphgallery.NodeSequence`, wich is iterable 
+            in `train`. If you want to specify your customized data during traing/testing/predicting, 
+            you can implement a subclass of `graphgallery.NodeSequence`, wich is iterable 
             and yields `inputs` and `labels` in each iteration. 
 
 
@@ -672,8 +671,6 @@ def do_forward(self, sequence, training=True):
         ----------
             sequence: `graphgallery.NodeSequence`
                 The input `sequence`.    
-            trainng (Boolean, optional): 
-                Indicating training or test procedure. (default: :obj:`True`)
 
         Return:
         ----------
@@ -684,20 +681,49 @@ def do_forward(self, sequence, training=True):
 
         """
         model = self.model
+        model.reset_metrics()
 
-        if training:
-            forward_fn = model.train_on_batch
-        else:
-            forward_fn = model.test_on_batch
+        with tf.device(self.device):
+            for inputs, labels in sequence:
+                loss, accuracy = model.train_on_batch(x=inputs, y=labels, reset_metrics=False)
+
+        return loss, accuracy
+
+    def test_step(self, sequence):
+        """
+            Forward propagation for the input `sequence`. This method will be called 
+            in `test`. If you want to specify your customized data during traing/testing/predicting, 
+            you can implement a subclass of `graphgallery.NodeSequence`, wich is iterable 
+            and yields `inputs` and `labels` in each iteration. 
+
+
+        Note:
+        ----------
+            You must compile your model before training/testing/predicting. 
+            Use `model.build()`.
 
+        Arguments:
+        ----------
+            sequence: `graphgallery.NodeSequence`
+                The input `sequence`.    
+
+        Return:
+        ----------
+            loss: Float scalar
+                Output loss of forward propagation.
+            accuracy: Float scalar
+                Output accuracy of prediction.
+
+        """
+        model = self.model
         model.reset_metrics()
 
         with tf.device(self.device):
             for inputs, labels in sequence:
-                loss, accuracy = forward_fn(x=inputs, y=labels, reset_metrics=False)
+                loss, accuracy = model.test_on_batch(x=inputs, y=labels, reset_metrics=False)
 
         return loss, accuracy
-
+    
     def predict(self, index, **kwargs):
         """
             Predict the output probability for the `index` of nodes.

diff --git a/graphgallery/nn/models/semisupervised/sgc.py b/graphgallery/nn/models/semisupervised/sgc.py
@@ -70,10 +70,19 @@ def preprocess(self, adj, x):
         if self.norm_x:
             x = normalize_x(x, norm=self.norm_x)
 
-# InvalidArgumentError: Cannot use GPU when output.shape[1] * nnz(a) > 2^31 [Op:SparseTensorDenseMatMul]
-        with tf.device(self.device):
+
+        # To avoid this tensorflow error in large dataset:
+        # InvalidArgumentError: Cannot use GPU when output.shape[1] * nnz(a) > 2^31 [Op:SparseTensorDenseMatMul]
+        if self.n_features*adj.nnz>2**31:
+            device = "CPU"
+        else:
+            device = self.device
+
+        with tf.device(device):
             x, adj = astensors([x, adj])
             x = SGConvolution(order=self.order)([x, adj])
+
+        with tf.device(self.device):
             self.x_norm, self.adj_norm = x, adj
 
     def build(self, lr=0.2, l2_norms=[5e-5], use_bias=True):

diff --git a/imgs/visualization_acc.png b/imgs/visualization_acc.png
diff --git a/imgs/visualization_loss.png b/imgs/visualization_loss.png