Clean up code

Add requirements.txt

Author: danielegrattarola

Autoencoder.py

@@ -1,23 +1,24 @@
 import matplotlib.pyplot as plt
 import numpy as np
-from keras.layers import Lambda
 from keras import Input, Model
 from keras import backend as K
 from keras.backend import tf
+from keras.layers import Lambda
+from pygsp import graphs
 from spektral.layers import GraphConvSkip
 from spektral.layers import MinCutPool
 from spektral.utils.convolution import normalized_adjacency
-from utils.misc import sp_matrix_to_sp_tensor_value, get_sw_key
 from tqdm import tqdm
-from pygsp import graphs
+
+from utils.misc import sp_matrix_to_sp_tensor_value, get_sw_key
 
 
 def upsampling_from_mask(inputs):
     X_, A_, I_, M_ = inputs
     S_ = tf.eye(tf.shape(M_)[0])
     S_ = tf.boolean_mask(S_, M_)
     S_t_ = tf.transpose(S_)
-    
+
     X_out_ = K.dot(S_t_, X_)
     A_out_ = K.dot(K.transpose(K.dot(A_, S_)), S_)
     I_out_ = K.dot(
@@ -27,32 +28,32 @@ def upsampling_from_mask(inputs):
     I_out_ = K.cast(I_out_, tf.int32)
     return [X_out_, A_out_, I_out_]
 
+
 def upsampling_from_matrix(inputs):
     X_, A_, I_, S_ = inputs
     X_out_ = K.dot(S_, X_)
-    A_out_ = K.dot(S_, K.transpose(K.dot(S_, A_,)))
+    A_out_ = K.dot(S_, K.transpose(K.dot(S_, A_, )))
     I_out_ = K.dot(
         S_,
         K.cast(I_[:, None], tf.float32)
     )[:, 0]
     I_out_ = K.cast(I_out_, tf.int32)
     return [X_out_, A_out_, I_out_]
 
+
 upsampling_from_mask_op = Lambda(upsampling_from_mask)
 upsampling_from_matrix_op = Lambda(upsampling_from_matrix)
 
-
 # HYPERPARAMS
 ITER = 10000
 ACTIV = 'tanh'
 dataset = 'grid'
 gnn_channels = 32
 es_patience = 1000
 
-
 # LOAD DATASET
 if dataset == 'ring':
-    G = graphs.Ring(N=200) 
+    G = graphs.Ring(N=200)
 elif dataset == 'grid':
     G = graphs.Grid2d(N1=30, N2=30)
 X = G.coords.astype(np.float32)
@@ -62,7 +63,6 @@ def upsampling_from_matrix(inputs):
 n_feat = X.shape[-1]
 n_nodes = A.shape[0]
 
-
 # MODEL DEFINITION
 X_in = Input(tensor=tf.placeholder(tf.float32, shape=(None, n_feat), name='X_in'))
 A_in = Input(tensor=tf.sparse_placeholder(tf.float32, shape=(None, None)), name='A_in')
@@ -76,7 +76,7 @@ def upsampling_from_matrix(inputs):
 X1 = GraphConvSkip(gnn_channels, activation=ACTIV)([X_in, A_in])
 X1 = GraphConvSkip(gnn_channels, activation=ACTIV)([X1, A_in])
 # pooling
-X2, A2, I2, M2 = MinCutPool(k=n_nodes//4, h=gnn_channels)([X1, A_in, I_in])
+X2, A2, I2, M2 = MinCutPool(k=n_nodes // 4, h=gnn_channels)([X1, A_in, I_in])
 # unpooling
 X3, A3, I3 = upsampling_from_matrix_op([X2, A2, I2, M2])
 # decoder
@@ -87,7 +87,6 @@ def upsampling_from_matrix(inputs):
 model = Model([X_in, A_in, I_in], [X3])
 model.compile('adam', 'mse', target_tensors=[X_target])
 
-
 # TRAINING
 sess = K.get_session()
 loss = model.total_loss
@@ -134,11 +133,10 @@ def upsampling_from_matrix(inputs):
 print('MSE', lss_)
 K.clear_session()
 
-
 # PLOTS
 plt.plot(output['loss'])
 plt.title('Loss')
-plt.figure(figsize=(8, 4 ))
+plt.figure(figsize=(8, 4))
 pad = 0.1
 x_min, x_max = X[:, 0].min() - pad, X[:, 0].max() + pad
 y_min, y_max = X[:, 1].min() - pad, X[:, 1].max() + pad
@@ -158,4 +156,4 @@ def upsampling_from_matrix(inputs):
 plt.axvline(0, c='k', alpha=0.2)
 plt.axhline(0, c='k', alpha=0.2)
 plt.tight_layout()
-plt.show()
+plt.show()

Clustering.py

@@ -1,5 +1,5 @@
 from collections import OrderedDict
-from tqdm import tqdm
+
 import matplotlib.pyplot as plt
 import numpy as np
 import pandas as pd
@@ -10,15 +10,18 @@
 from keras.models import Model
 from pygsp import graphs
 from sklearn.cluster import spectral_clustering
-from sklearn.datasets.samples_generator import make_blobs
+from sklearn.datasets import make_blobs
 from sklearn.metrics.cluster import v_measure_score, homogeneity_score, completeness_score
 from sklearn.neighbors import kneighbors_graph
+from spektral.layers import MinCutPool, DiffPool
 from spektral.layers.convolutional import GraphConvSkip
 from spektral.utils import init_logging
 from spektral.utils.convolution import normalized_adjacency
-from spektral.layers import MinCutPool, DiffPool
+from tqdm import tqdm
+
 from utils import citation
 from utils.misc import sp_matrix_to_sp_tensor_value, product_dict
+
 np.random.seed(0)  # for reproducibility
 
 PLOTS_ON = True
@@ -36,15 +39,14 @@
 
 # Tunables
 tunables = OrderedDict([
-    ('dataset', ['cora']), # 'cora', 'citeseer', 'pubmed'
-    ('method', ['mincut_pool']), # 'mincut_pool', 'diff_pool'
+    ('dataset', ['cora']),  # 'cora', 'citeseer', 'pubmed', 'cloud', or 'synth'
+    ('method', ['mincut_pool']),  # 'mincut_pool', 'diff_pool'
     ('H_', [None]),
     ('n_channels', [16]),
     ('learning_rate', [5e-4])
 ])
 
 N_RUNS = 1
-dataset = None
 df_out = None
 for T in product_dict(tunables):
     # Update params with current config
@@ -63,18 +65,16 @@
         A = sp.csr_matrix(A, dtype=np.float32)
         n_clust = y.max() + 1
     elif P['dataset'] == 'cloud':
-        G = graphs.Grid2d(N1=15, N2=10) #Community(N=150, seed=0) #SwissRoll(N=400, seed=0) #Ring(N=100) #TwoMoons()  #Cube(nb_pts=500)  #Bunny()
+        G = graphs.Grid2d(N1=15, N2=10)  # Community(N=150, seed=0) #SwissRoll(N=400, seed=0) #Ring(N=100) #TwoMoons()  #Cube(nb_pts=500)  #Bunny()
         X = G.coords.astype(np.float32)
         A = G.W
         y = np.ones(X.shape[0])  # X[:,0] + X[:,1]
         n_clust = 5
     else:
-        if dataset != P['dataset']:
-            dataset = P['dataset']
-            A, X, _, _, _, _, _, _, y_ohe = citation.load_data(dataset)
-            y = np.argmax(y_ohe, axis=-1)
-            X = X.todense()
-            n_clust = y.max() + 1
+        A, X, _, _, _, _, _, _, y_ohe = citation.load_data(P['dataset'])
+        y = np.argmax(y_ohe, axis=-1)
+        X = X.todense()
+        n_clust = y.max() + 1
 
     # Sort IDs
     if P['dataset'] != 'cloud':
@@ -206,12 +206,12 @@
             plt.scatter(X[:, 0], X[:, 1], c=c)
             plt.title('GNN-pool clustering')
         if P['dataset'] == 'cloud':
-            fig, ax = plt.subplots(1,1,figsize=(3,3))
-            G.plot_signal(c, vertex_size=30,  plot_name='', colorbar=False,ax=ax)
+            fig, ax = plt.subplots(1, 1, figsize=(3, 3))
+            G.plot_signal(c, vertex_size=30, plot_name='', colorbar=False, ax=ax)
             ax.set_xticks([])
             ax.set_yticks([])
             plt.tight_layout()
-            plt.savefig('logs/grid_mincut.pdf', bbox_inches = 'tight', pad_inches = 0)
+            plt.savefig('logs/grid_mincut.pdf', bbox_inches='tight', pad_inches=0)
         plt.show()
 
     # Spectral clustering
@@ -220,7 +220,8 @@
     P['complete_score_sc'] = completeness_score(y, sc)
     P['v_score_sc'] = v_measure_score(y, sc)
 
-    print('Spectral Clust - HOMO: {:.2}, CS: {:2}, NMI: {:2}'.format(P['homo_score_sc'], P['complete_score_sc'], P['v_score_sc']))
+    print('Spectral Clust - HOMO: {:.3f}, CS: {:.3f}, NMI: {:.3f}'
+          .format(P['homo_score_sc'], P['complete_score_sc'], P['v_score_sc']))
 
     if df_out is None:
         df_out = pd.DataFrame([P])
@@ -234,10 +235,10 @@
             plt.title('Spectral clustering')
             plt.show()
         if P['dataset'] == 'cloud':
-            fig, ax = plt.subplots(1,1,figsize=(3,3))
-            G.plot_signal(sc, vertex_size=30,  plot_name='', colorbar=False,ax=ax)
+            fig, ax = plt.subplots(1, 1, figsize=(3, 3))
+            G.plot_signal(sc, vertex_size=30, plot_name='', colorbar=False, ax=ax)
             ax.set_xticks([])
             ax.set_yticks([])
             plt.tight_layout()
-            plt.savefig('logs/grid_spectral.pdf', bbox_inches = 'tight', pad_inches = 0)
-    K.clear_session()
+            plt.savefig('logs/grid_spectral.pdf', bbox_inches='tight', pad_inches=0)
+    K.clear_session()

Graph_Classification.py

@@ -1,6 +1,7 @@
 import itertools
 import time
 from collections import OrderedDict
+
 import keras.backend as K
 import numpy as np
 import pandas as pd
@@ -13,9 +14,10 @@
 from sklearn.model_selection import train_test_split
 from spektral.layers import GraphConv, GlobalAvgPool, ARMAConv, GraphConvSkip
 from spektral.layers import MinCutPool, DiffPool, TopKPool, SAGPool
+from spektral.layers.ops import sp_matrix_to_sp_tensor_value
 from spektral.utils import batch_iterator, log, init_logging
 from spektral.utils.convolution import normalized_adjacency
-from spektral.layers.ops import sp_matrix_to_sp_tensor_value
+
 from utils.dataset_loader import get_graph_kernel_dataset
 
 
@@ -60,7 +62,7 @@ def evaluate(A_list, X_list, y_list, ops):
 # Parameters
 P = OrderedDict(
     runs=10,             # Runs to repeat per config
-    data_mode='bench',   # bench / synth 
+    data_mode='bench',   # bench / synth
     GNN_type='GCS',      # Type of GNN {GCN, GCS, Cheb, ARMA}
     n_channels=32,       # Channels per layer
     activ='relu',        # Activation in GNN and mincut
@@ -77,7 +79,7 @@ def evaluate(A_list, X_list, y_list, ops):
 # Tunables
 tunables = OrderedDict(
     dataset_ID=['PROTEINS'],
-    method=['mincut_pool'] # 'flat', 'dense', 'diff_pool', 'top_k_pool', 'mincut_pool', 'sag_pool'
+    method=['mincut_pool']  # 'flat', 'dense', 'diff_pool', 'top_k_pool', 'mincut_pool', 'sag_pool'
 )
 log(tunables)
 
@@ -99,7 +101,7 @@ def evaluate(A_list, X_list, y_list, ops):
         ########################################################################
         # LOAD DATA
         ########################################################################
-        if  P['data_mode'] == 'bench':
+        if P['data_mode'] == 'bench':
             A, X, y = get_graph_kernel_dataset(P['dataset_ID'], feat_norm='ohe')
             # Train/test split
             A_train, A_test, \
@@ -112,12 +114,12 @@ def evaluate(A_list, X_list, y_list, ops):
             loaded = np.load('data/hard.npz', allow_pickle=True)
             X_train, A_train, y_train = loaded['tr_feat'], list(loaded['tr_adj']), loaded['tr_class']
             X_test, A_test, y_test = loaded['te_feat'], list(loaded['te_adj']), loaded['te_class']
-            X_val, A_val, y_val = loaded['val_feat'], list(loaded['val_adj']), loaded['val_class']            
+            X_val, A_val, y_val = loaded['val_feat'], list(loaded['val_adj']), loaded['val_class']
         else:
             raise ValueError
-            
+
         # Parameters
-        F = X_train[0].shape[-1]      # Dimension of node features
+        F = X_train[0].shape[-1]  # Dimension of node features
         n_out = y_train[0].shape[-1]  # Dimension of the target
         average_N = np.ceil(np.mean([a.shape[-1] for a in A_train]))  # Average number of nodes in dataset
 
@@ -279,7 +281,8 @@ def evaluate(A_list, X_list, y_list, ops):
                 ep = int(current_batch / batches_in_epoch)
                 log('Ep: {:d} - Loss: {:.2f} - Acc: {:.2f} - Val loss: {:.2f} - Val acc: {:.2f}'
                     .format(ep, model_loss, model_acc, val_loss, val_acc))
-                log('{} - Average epoch time: {} +- {}'.format(P['method'], np.mean(epoch_time[1:]), np.std(epoch_time[1:])))
+                log('{} - Average epoch time: {} +- {}'
+                    .format(P['method'], np.mean(epoch_time[1:]), np.std(epoch_time[1:])))
                 epoch_time.append(0)
 
                 if val_loss < best_val_loss: