WIP - Gated pixelcnn

1 - Autoregressive Models - PixelCNN/pixelCNN.py

@@ -197,7 +197,7 @@ def train_step(batch_x, batch_y):
         with tf.GradientTape() as ae_tape:
             logits = pixelcnn(batch_x, training=True)
 
-            loss = compute_loss(tf.one_hot(batch_y, q_levels), logits)
+            loss = compute_loss(tf.squeeze(tf.one_hot(batch_y, q_levels)), logits)
 
         gradients = ae_tape.gradient(loss, pixelcnn.trainable_variables)
         gradients, _ = tf.clip_by_global_norm(gradients, 1.0)
@@ -218,14 +218,15 @@ def train_step(batch_x, batch_y):
             loss = train_step(batch_x, batch_y)
 
             progbar.add(1, values=[('loss', loss)])
+
     # ------------------------------------------------------------------------------------
     # Test set performance
     test_loss = []
     for batch_x, batch_y in test_dataset:
         logits = pixelcnn(batch_x, training=False)
 
         # Calculate cross-entropy (= negative log-likelihood)
-        loss = compute_loss(tf.one_hot(batch_y, q_levels), logits)
+        loss = compute_loss(tf.squeeze(tf.one_hot(batch_y, q_levels)), logits)
 
         test_loss.append(loss)
     print('nll : {:} nats'.format(np.array(test_loss).mean()))

README.md

@@ -15,12 +15,15 @@ Clone the git repository :
 Python 3 with [TensorFlow 2.0+](https://www.tensorflow.org/) are the primary requirements.
 Install virtualenv and creat a new virtual environment:
 
-    sudo apt-get install -y python3-venv
-    python3 -m venv ./venv
+    sudo apt update
+    sudo apt install python3-dev python3-pip
+    sudo pip3 install -U virtualenv  # system-wide install
+    virtualenv --system-site-packages -p python3 ./venv
 
 Then, install requirements
 
-    source venv/bin/activate
+    source ./venv/bin/activate
+    pip3 install --upgrade pip
     pip3 install -r requirements.txt
 
 <h1 align="center"> 1. Autoregressive Models — PixelCNN </h1>