def _get_loss_y_pred_y_true_batch_assign_op(cls, iterator,
inference_mode_var,
batch_size):
"""
get the ResNet50 model as defined in https://arxiv.org/pdf/1512.03385.pdf,
ResNet V2 is used.
overrides :class:`abstract_net_class.get_model()`
:param iterator:
:param inference_mode_var: must hold an element of type :class:`code_.abstract_net_class.inferenceMode`
:param batch_size:
:return: loss , _y_pred, acc_op, acc_update_op, batch_assign_ops
"""
with tf.variable_scope("Res_Net", reuse=tf.AUTO_REUSE):
# Build model
num_classes = 10
x = tf.Variable(tf.zeros([batch_size, 32, 32, 3]),
dtype=tf.float32,
trainable=False)
y_true = tf.Variable(tf.zeros([batch_size, 1], dtype=tf.uint8),
dtype=tf.uint8,
trainable=False)
cx, cy = iterator.get_next()
x_assign = x.assign(cx).op
y_true_assign = y_true.assign(cy).op
batch_assign_ops = (x_assign, y_true_assign)
# params from resnet paper https://arxiv.org/pdf/1512.03385.pdf
res_model = Model(resnet_size=32,
bottleneck=False,
num_classes=num_classes,
num_filters=16,
kernel_size=3,
conv_stride=1,
first_pool_size=None,
first_pool_stride=1,
block_sizes=[5, 5, 5],
block_strides=[1, 2, 2],
resnet_version=2)
logits = res_model(x, (tf.equal(
inference_mode_var, tf.cast(a.InferenceMode.TRAIN, tf.uint8))))
y_pred = logits
y1h = tf.one_hot(y_true, num_classes, on_value=1, name="oneHot")
y1h = tf.cast(tf.reshape(y1h, (batch_size, num_classes)),
tf.float32)
y_pred = tf.cast(tf.reshape(y_pred, (batch_size, num_classes)),
tf.float32)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=y1h,
logits=y_pred))
return loss, y_pred, y_true, batch_assign_ops, None
def _get_loss_y_pred_y_true_batch_assign_op(cls, iterator,
inference_mode_var,
batch_size):
# Build model
x = tf.Variable(tf.zeros([batch_size, 224, 224, 3]),
dtype=tf.float32,
trainable=False)
y_true = tf.Variable(tf.zeros([
batch_size,
], dtype=tf.int32),
dtype=tf.int32,
trainable=False)
cx, cy = iterator.get_next()
x_assign = x.assign(cx).op
y_true_assign = y_true.assign(cy).op
batch_assign_op = tf.group(x_assign, y_true_assign)
# params from resnet paper https://arxiv.org/pdf/1512.03385.pdf
res_model = Model(resnet_size=100,
bottleneck=True,
num_classes=1000,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=[3, 4, 23, 3],
block_strides=[1, 2, 2, 2],
resnet_version=2)
logits = res_model(x, (tf.equal(
inference_mode_var, tf.cast(a.InferenceMode.TRAIN, tf.uint8))))
y_pred = logits
y1h = tf.one_hot(y_true, 1000, on_value=1, name="oneHot")
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=y1h,
logits=y_pred))
return loss, y_pred, y_true, batch_assign_op, None