def build_graph(X, y):
algo = Mlp(input_dim=NUM_FEATURES, num_classes=NUM_CLASSES)
py_x = algo.forward(X)
loss = melt.sparse_softmax_cross_entropy(py_x, y)
tf.scalar_summary('loss_%s' % loss.name, loss)
#tf.scalar_summary('loss', loss)
##loss = tf.reduce_sum(tf.nn.sparse_softmax_cross_entropy_with_logits(py_x, y))
accuracy = melt.precision_at_k(py_x, y, 1)
tf.scalar_summary('precision@1_%s' % accuracy.name, accuracy)
#below will cause tensorflow.python.framework.errors.InvalidArgumentError: Duplicate tag precicsion@1 found in summary inputs
#the problem here is we want to share all other things but without scalar summarys in graph
#so if we want to build more than once ... then scalar_summary must use op.name
#or else do it outof build_graph, setting names by yourself!
#eval_loss, eval_accuracy = build_grapy(X, y)
#tf.scalar_summary('eval_loss', eval_loss)
#tf.scalar_summary('eval_accuracy', eval_accuracy)
#since tensorboard has 'Split on underscores', so for better comparaion side by side
#loss_train, loss_eval, accuracy_train, accuracy_eval is better then train_loss,eval_loss
#tf.scalar_summary('precicsion@1', accuracy)
return loss, accuracy
def build_graph(X, y):
algo = Mlp(NUM_FEATURES, NUM_CLASSES)
py_x = algo.forward(X)
loss = melt.sparse_softmax_cross_entropy(py_x, y)
#loss = tf.reduce_sum(tf.nn.sparse_softmax_cross_entropy_with_logits(py_x, y))
accuracy = melt.precision_at_k(py_x, y, 1)
return loss, accuracy
def build_graph(X, y): #---build forward graph py_x = predict(X) #-----------for classification we can set loss function and evaluation metrics,so only forward graph change #---set loss function loss = melt.sparse_softmax_cross_entropy(py_x, y) #---choose evaluation metrics accuracy = melt.precision_at_k(py_x, y, 1) return loss, accuracy
def build_graph(X, y):
#---build forward graph
algo = Mlp(input_dim=NUM_FEATURES, num_classes=NUM_CLASSES)
py_x = algo.forward(X)
#-----------for classification we can set loss function and evaluation metrics,so only forward graph change
#---set loss function
loss = melt.sparse_softmax_cross_entropy(py_x, y)
#tf.scalar_summary('loss_%s'%loss.name, loss)
#---choose evaluation metrics
accuracy = melt.precision_at_k(py_x, y, 1)
#tf.scalar_summary('precision@1_%s'%accuracy.name, accuracy)
return loss, accuracy