def newsgl(X,R,layers=[512],eg=None,l1=None,dropout=None):
layers = [l for l in layers]
eg_strength = eg if eg else 0.0
l1_strength = tf.constant(l1,dtype=tf.float64) if l1 else tf.constant(0.0,dtype=tf.float64)
dropout_strength = dropout if dropout else 0.0
explainer = TFOpsExplainer()
cond_input_op, train_eg = explainer.input_to_samples_delta(X,lambda: R)
if len(layers)>0:
kernels = []
biases = []
firstlayer = layers.pop(0)
hid_layer = tf.layers.Dense(firstlayer,activation=tf.nn.relu)
hid = hid_layer(cond_input_op)
if dropout: hid = tf.nn.dropout(hid,keep_prob=1-dropout_strength)
kernels.append(hid_layer.kernel)
biases.append(hid_layer.bias)
for l in layers:
hid_layer = tf.layers.Dense(l,activation=tf.nn.relu)
hid = hid_layer(hid)
if dropout: hid = tf.nn.dropout(hid,keep_prob=1-dropout_strength)
kernels.append(hid_layer.kernel)
biases.append(hid_layer.bias)
output = tf.layers.Dense(2)(hid)
else:
output_layer = tf.layers.Dense(2)
output = output_layer(cond_input_op)
kernels = [output_layer.kernel]
biases = [output_layer.bias]
kernel = kernels[0]
y_pred = tf.nn.sigmoid(output)
eg_op = explainer.shap_value_op(output, cond_input_op, y[:,1])
cost = tf.nn.sigmoid_cross_entropy_with_logits(logits=y_pred,labels=y)
total_cost = tf.reduce_mean(cost)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0).minimize(total_cost)
l1_reg = (tf.reduce_sum([sgl_penalty_layer(k) for k in kernels])+
tf.reduce_sum([2*tf.reduce_sum(tf.abs(b)) for b in biases]))*l1_strength
weighted_abs_loss = tf.reshape(tf.reduce_mean(tf.abs(tf.cast(eg_op,tf.float32)),axis=0),(1,-1))
mad_loss = tf.abs(tf.transpose(weighted_abs_loss)-weighted_abs_loss)
rmad_loss = tf.reduce_mean(mad_loss)/tf.reduce_mean(weighted_abs_loss)
reg_lambda = tf.constant(eg_strength,dtype=tf.float32)
reg_loss_op = -rmad_loss * reg_lambda
reg_loss_op = reg_loss_op + tf.cast(l1_reg,dtype=tf.float32)
train_eg_op = tf.train.GradientDescentOptimizer(learning_rate=1.0).minimize(reg_loss_op)
return (y_pred, eg_op, total_cost, optimizer, train_eg_op, train_eg, explainer, reg_lambda, l1_strength)
def ginigrad(X,R,layers=[512],eg=None,l1=None,dropout=None,inputs=False):
layers = [l for l in layers]
eg_strength = eg if eg else 0.0
l1_strength = tf.constant(l1,dtype=tf.float64) if l1 else tf.constant(0.0,dtype=tf.float64)
dropout_strength = dropout if dropout else 0.0
explainer = TFOpsExplainer()
cond_input_op, train_eg = explainer.input_to_samples_delta(X,lambda: R)
if len(layers)>0:
firstlayer = layers.pop(0)
hid_layer = tf.layers.Dense(firstlayer,activation=tf.nn.relu)
hid = hid_layer(cond_input_op)
if dropout: hid = tf.nn.dropout(hid,keep_prob=1-dropout_strength)
l1_reg = tf.reduce_mean(tf.abs(hid_layer.kernel)) * l1_strength
for l in layers:
hid = tf.layers.Dense(l,activation=tf.nn.relu)(hid)
if dropout: hid = tf.nn.dropout(hid,keep_prob=1-dropout_strength)
output = tf.layers.Dense(2)(hid)
else:
output_layer = tf.layers.Dense(2)
output = output_layer(cond_input_op)
l1_reg = tf.reduce_mean(tf.abs(output_layer.kernel)) * l1_strength
y_pred = tf.nn.sigmoid(output)#model(cond_input_op)
eg_op = explainer.shap_value_op(output, cond_input_op, y[:,1]) # Train with gradients but use EG for explanations (you should probably use this)
grads = tf.gradients(output[:,1],X) # Train with gradients but use EG for explanations (you should probably use this)
if inputs: grads *= X
cost = tf.nn.sigmoid_cross_entropy_with_logits(logits=y_pred,labels=y)
total_cost = tf.reduce_mean(cost)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0).minimize(total_cost)
weighted_abs_loss = tf.reshape(tf.reduce_mean(tf.abs(tf.cast(grads,tf.float32)),axis=0),(1,-1))
mad_loss = tf.abs(tf.transpose(weighted_abs_loss)-weighted_abs_loss)
rmad_loss = tf.reduce_mean(mad_loss)/tf.reduce_mean(weighted_abs_loss)
reg_lambda = tf.constant(eg_strength,dtype=tf.float32)
reg_loss_op = -rmad_loss * reg_lambda
reg_loss_op = reg_loss_op + tf.cast(l1_reg,dtype=tf.float32)
train_eg_op = tf.train.GradientDescentOptimizer(learning_rate=1.0).minimize(reg_loss_op)
return (y_pred, eg_op, total_cost, optimizer, train_eg_op, train_eg, explainer, reg_lambda, l1_strength)
def pipeline(use_old_model=False):
# Data iterators
training_iterator, validation_iterator, test_iterator, handle, x, y_ = inputs(
FLAGS.batch_size, FLAGS.epochs, VALIDATION_SIZE, TEST_BATCH_SIZE)
# Get explainer conditional input
explainer = AttributionPriorExplainer()
cond_input_op, train_eg = explainer.input_to_samples_delta(x)
# Predict
y, train_pl = get_model(cond_input_op)
# Get explanations
expected_gradients_op = explainer.shap_value_op(y,
cond_input_op,
sparse_labels_op=y_)
expected_gradients_op = batch_standardize(expected_gradients_op)
learning_rate = get_learning_rate(FLAGS.eta, FLAGS.batch_size)
global_step = tf.train.get_or_create_global_step()
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_, logits=y))
train_step = tf.train.AdamOptimizer(learning_rate).minimize(
(1.0 - FLAGS.lamb) * loss, global_step=global_step)
eg_loss = tf.reduce_mean(
tf.image.total_variation(expected_gradients_op, name='variation'))
eg_train = tf.train.AdamOptimizer(learning_rate).minimize(
FLAGS.lamb * eg_loss, global_step=global_step)
y_pred = tf.argmax(y, 1)
accuracy_op, accuracy_update_op = tf.metrics.accuracy(y_pred, y_)
reset_metrics_op = tf.variables_initializer(
tf.get_collection(tf.GraphKeys.METRIC_VARIABLES))
return cond_input_op, x, y, y_, train_pl, loss, train_step, eg_loss, eg_train, train_eg, accuracy_op, accuracy_update_op, reset_metrics_op, training_iterator, validation_iterator, test_iterator, handle