def _ReluGrad(op, grad): return gen_nn_ops.relu_grad(grad, op.outputs[0])
def _ReluGradGrad(op, grad):
x = op.inputs[1]
return (gen_nn_ops.relu_grad(grad, x),
array_ops.zeros(shape=array_ops.shape(x), dtype=x.dtype))
def _ReluGradGrad(op, grad):
x = op.inputs[1]
return (gen_nn_ops.relu_grad(grad, x),
array_ops.zeros(shape=array_ops.shape(x), dtype=x.dtype))
def _ReluGrad(op, grad):
return gen_nn_ops.relu_grad(grad, op.outputs[0])
def _GuidedReluGrad(op, grad):
return tf.where(0. < grad, gen_nn_ops.relu_grad(grad, op.outputs[0]),
tf.zeros(tf.shape(grad)))
pconv1 = activ(conv1)
conv2 = conv2d(pconv1, weights['wc2'], biases['bc2'], 'SAME')
pool1 = maxpool(conv2, k=2)
pconv2 = activ(pool1)
conv3 = conv2d(pconv2, weights['wc3'], biases['bc3'], 'SAME')
pconv3 = activ(conv3)
conv4 = conv2d(pconv3, weights['wc4'], biases['bc4'], 'SAME')
pool2 = maxpool(conv4, k=2)
pconv4 = activ(pool2)
conv5 = conv2d(pconv4, weights['wc5'], biases['bc5'], 'SAME')
pconv5 = activ(conv5)
conv6 = conv2d(pconv5, weights['wc6'], biases['bc6'], 'VALID')
pool3 = maxpool(conv6, k=2)
pconv6 = activ(pool3)
grad_pre_relu6 = tf.where(0. < grad, gen_nn_ops.relu_grad(grad, pool3),
tf.zeros(grad.get_shape()))
grad_conv5 = tf.gradients(pool3, pconv5, grad_ys=grad_pre_relu6)[0]
grad_pre_relu5 = tf.where(0. < grad_conv5,
gen_nn_ops.relu_grad(grad_conv5, conv5),
tf.zeros((1, 16, 16, 64)))
grad_conv4 = tf.gradients(conv5, pconv4, grad_ys=grad_pre_relu5)[0]
grad_pre_relu4 = tf.where(0. < grad_conv4,
gen_nn_ops.relu_grad(grad_conv4, pool2),
tf.zeros((1, 16, 16, 64)))
grad_conv3 = tf.gradients(pool2, pconv3, grad_ys=grad_pre_relu4)[0]
grad_pre_relu3 = tf.where(0. < grad_conv3,
gen_nn_ops.relu_grad(grad_conv3, conv3),
def _ReluGradGrad(op, grad):
x = op.inputs[1]
return (gen_nn_ops.relu_grad(grad, x), array_ops.zeros_like(x))
