Function get_model

models/conditional.py:46–149

(sess, image_shape=(80, 160, 3), gf_dim=64, df_dim=64, batch_size=64,
              name="transition", gpu=0)

def get_model(sess, image_shape=(80, 160, 3), gf_dim=64, df_dim=64, batch_size=64,
              name="transition", gpu=0):
    K.set_session(sess)
    checkpoint_dir = './results_' + name
    with tf.variable_scope(name):
      # sizes
      ch = image_shape[2]
      rows = [image_shape[0]/i for i in [16, 8, 4, 2, 1]]
      cols = [image_shape[1]/i for i in [16, 8, 4, 2, 1]]

      G = autoencoder.generator(batch_size*out_leng, gf_dim, ch, rows, cols)
      G.compile("sgd", "mse")
      E = autoencoder.encoder(batch_size*(time+out_leng), df_dim, ch, rows, cols)
      E.compile("sgd", "mse")

      G.trainable = False
      E.trainable = False

      # nets
      T = transition(batch_size)
      T.compile("sgd", "mse")
      t_vars = T.trainable_weights
      print "T.shape: ", T.output_shape

      Img = Input(batch_shape=(batch_size, time+out_leng,) + image_shape)
      Z = Input(batch_shape=(batch_size, time+out_leng, 2))  # controls signal
      I = K.reshape(Img, (batch_size*(time+out_leng),)+image_shape)
      code = E(I)[0]
      code = K.reshape(code, (batch_size, time+out_leng, z_dim))
      inp = K.concatenate([Z, code], axis=2)
      target = code[:, time:, :]
      out = T(inp)
      G_dec = G(K.reshape(out, (batch_size*out_leng, z_dim)))

      # costs
      loss = tf.reduce_mean(tf.square(target - out))
      print "Transition variables:"
      for v in t_vars:
        print v.name

      t_optim = tf.train.AdamOptimizer(learning_rate, beta1=beta1).minimize(loss, var_list=t_vars)

      tf.initialize_all_variables().run()

    # summaries
    sum_loss = tf.scalar_summary("loss", loss)
    sum_e_mean = tf.histogram_summary("e_mean", code)
    sum_out = tf.histogram_summary("out", out)
    sum_dec = tf.image_summary("E", G_dec)

    # saver
    saver = tf.train.Saver()
    t_sum = tf.merge_summary([sum_e_mean, sum_out, sum_dec, sum_loss])
    writer = tf.train.SummaryWriter("/tmp/logs/"+name, sess.graph)

    # functions
    def train_d(images, z, counter, sess=sess):
      return 0, 0, 0