def model_fn(inputs, mode, **kwargs):
# In train or eval, id_or_labels represents labels. In predict, id_or_labels represents id.
images, id_or_labels, angles = inputs
# Reshape angles from [batch_size] to [batch_size, 1]
angles = tf.expand_dims(angles, 1)
# Apply your version of model
logits = model_v1(images, angles, mode)
if mode == mox.ModeKeys.PREDICT:
logits = tf.nn.softmax(logits)
# clip logits to get lower loss value.
logits = tf.clip_by_value(logits,
clip_value_min=0.05,
clip_value_max=0.95)
model_spec = mox.ModelSpec(output_info={
'id': id_or_labels,
'logits': logits
})
elif mode == mox.ModeKeys.EXPORT:
predictions = tf.nn.softmax(logits)
export_spec = mox.ExportSpec(inputs_dict={
'images': images,
'angles': angles
},
outputs_dict={'predictions': predictions},
version='model')
model_spec = mox.ModelSpec(export_spec=export_spec)
else:
labels_one_hot = slim.one_hot_encoding(id_or_labels, 2)
loss = tf.losses.softmax_cross_entropy(logits=logits,
onehot_labels=labels_one_hot,
label_smoothing=0.0,
weights=1.0)
model_spec = mox.ModelSpec(loss=loss, log_info={'loss': loss})
return model_spec
def model_fn(inputs, mode, **kwargs):
images, labels = inputs
mox_model_fn = mox.get_model_fn(name=flags.model_name,
run_mode=mode,
num_classes=data_meta.num_classes,
weight_decay=0.00004,
data_format='NCHW',
batch_norm_fused=True)
images = tf.cast(images, tf.float16)
with mox.var_scope(force_dtype=tf.float32):
logits, _ = mox_model_fn(images)
labels_one_hot = slim.one_hot_encoding(labels, data_meta.num_classes)
loss = tf.losses.softmax_cross_entropy(labels_one_hot, logits=logits)
regularization_losses = mox.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
regularization_loss = tf.add_n(regularization_losses)
loss = loss + regularization_loss
logits_fp32 = tf.cast(logits, tf.float32)
accuracy = tf.reduce_mean(
tf.cast(tf.nn.in_top_k(logits_fp32, labels, 1), tf.float32))
export_spec = mox.ExportSpec(inputs_dict={'images': images},
outputs_dict={'logits': logits_fp32})
return mox.ModelSpec(loss=loss,
log_info={
'loss': loss,
'accuracy': accuracy
},
export_spec=export_spec)
def model_fn(inputs, run_mode, **kwargs):
x, y_ = inputs
W = tf.get_variable(name='W', initializer=tf.zeros([784, 10]))
b = tf.get_variable(name='b', initializer=tf.zeros([10]))
y = tf.matmul(x, W) + b
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))
predictions = tf.argmax(y, 1)
correct_predictions = tf.equal(predictions, tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))
export_spec = mox.ExportSpec(inputs_dict={'images': x}, outputs_dict={'predictions': predictions}, version='model')
return mox.ModelSpec(loss=cross_entropy, log_info={'loss': cross_entropy, 'accuracy': accuracy},
export_spec=export_spec)
def model_fn(inputs, run_mode, **kwargs):
x, y_ = inputs
y = tf.keras.layers.Dense(128, activation='relu')(x)
y = tf.keras.layers.Dense(10)(y)
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
export_spec = mox.ExportSpec(inputs_dict={'images': x},
outputs_dict={'logits': y},
version='model')
return mox.ModelSpec(loss=cross_entropy,
log_info={
'loss': cross_entropy,
'accuracy': accuracy
},
export_spec=export_spec)
