def loss(self, truth, prediction):
y = prediction[0]
mu = truth
sigma = op.get(prediction[1], self.inputs.actions)
# Gaussian log-likelihood
result = op.tofloat(y - mu) # Primarily to prevent under/overflow since they are already float16
result = tf.cast(result, 'float32') * tf.inv(sigma)
result = -tf.square(result) / 2
result = result + tf.log(tf.inv(sigma))
return tf.reduce_mean(-result)
def image():
try:
size = int(request.args.get('size'))
size_all = int(request.args.get('size_all'))
num_child = int(request.args.get('num_child'))
except Exception as e:
return (jsonify({
'error': "parameter error",
'code': 400,
'detail': str(e)
}), 400)
response = make_response()
response.data = ops.get(size, size_all, num_child)
response.headers['Content-Disposition'] = 'attachment; filename=image.jpg'
response.mimetype = 'image/jpeg'
return response
def prediction(self, train_output_states):
return op.get(op.tofloat(train_output_states), self.inputs.actions)
def truth(self, train_output_states, train_output_next_states, target_output_next_states):
# Double DQN - http://arxiv.org/pdf/1509.06461v3.pdf
double_q_next = op.get(target_output_next_states, op.argmax(train_output_next_states))
return (op.tofloat(self.inputs.rewards) + self.args.discount *
(1.0 - op.tofloat(self.inputs.terminals)) * op.tofloat(double_q_next))