def cifar10_model_fn_helper(self, mode, version, multi_gpu=False):
input_fn = cifar10_main.get_synth_input_fn()
dataset = input_fn(True, '', _BATCH_SIZE)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features, labels, mode, {
'resnet_size': 32,
'data_format': 'channels_last',
'batch_size': _BATCH_SIZE,
'version': version,
'multi_gpu': multi_gpu
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (_BATCH_SIZE, ))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops['accuracy'][0].shape, ())
self.assertAllEqual(eval_metric_ops['accuracy'][1].shape, ())
self.assertEqual(eval_metric_ops['accuracy'][0].dtype, tf.float32)
self.assertEqual(eval_metric_ops['accuracy'][1].dtype, tf.float32)
def cifar10_model_fn_helper(self, mode, resnet_version, dtype):
input_fn = cifar10_main.get_synth_input_fn(dtype)
dataset = input_fn(True, '', _BATCH_SIZE)
iterator = dataset.make_initializable_iterator()
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features, labels, mode, {
'dtype': dtype,
'resnet_size': 32,
'data_format': 'channels_last',
'batch_size': _BATCH_SIZE,
'resnet_version': resnet_version,
'loss_scale': 128 if dtype == tf.float16 else 1,
'fine_tune': False,
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (_BATCH_SIZE, ))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops['accuracy'][0].shape, ())
self.assertAllEqual(eval_metric_ops['accuracy'][1].shape, ())
self.assertEqual(eval_metric_ops['accuracy'][0].dtype, tf.float32)
self.assertEqual(eval_metric_ops['accuracy'][1].dtype, tf.float32)
def cifar10_model_fn_helper(self, mode, resnet_version, dtype):
input_fn = cifar10_main.get_synth_input_fn(dtype)
dataset = input_fn(True, '', _BATCH_SIZE)
iterator = tf.compat.v1.data.make_initializable_iterator(dataset)
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features, labels, mode, {
'dtype': dtype,
'resnet_size': 32,
'data_format': 'channels_last',
'batch_size': _BATCH_SIZE,
'resnet_version': resnet_version,
'loss_scale': 128 if dtype == tf.float16 else 1,
'fine_tune': False,
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (_BATCH_SIZE,))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops['accuracy'][0].shape, ())
self.assertAllEqual(eval_metric_ops['accuracy'][1].shape, ())
self.assertEqual(eval_metric_ops['accuracy'][0].dtype, tf.float32)
self.assertEqual(eval_metric_ops['accuracy'][1].dtype, tf.float32)
def cifar10_model_fn_helper(self, mode, version, multi_gpu=False):
input_fn = cifar10_main.get_synth_input_fn()
dataset = input_fn(True, '', _BATCH_SIZE)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features, labels, mode, {
'resnet_size': 32,
'data_format': 'channels_last',
'batch_size': _BATCH_SIZE,
'version': version,
'multi_gpu': multi_gpu
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (_BATCH_SIZE,))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops['accuracy'][0].shape, ())
self.assertAllEqual(eval_metric_ops['accuracy'][1].shape, ())
self.assertEqual(eval_metric_ops['accuracy'][0].dtype, tf.float32)
self.assertEqual(eval_metric_ops['accuracy'][1].dtype, tf.float32)
def _cifar10_model_fn_helper(self, mode, version, dtype, multi_gpu=False):
with tf.Graph().as_default() as g:
input_fn = cifar10_main.get_synth_input_fn()
dataset = input_fn(True, '', _BATCH_SIZE)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features, labels, mode, {
'dtype': dtype,
'resnet_size': 32,
'data_format': 'channels_last',
'batch_size': _BATCH_SIZE,
'version': version,
'loss_scale': 128 if dtype == tf.float16 else 1,
'multi_gpu': multi_gpu
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (_BATCH_SIZE, ))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops['accuracy'][0].shape, ())
self.assertAllEqual(eval_metric_ops['accuracy'][1].shape, ())
self.assertEqual(eval_metric_ops['accuracy'][0].dtype,
tf.float32)
self.assertEqual(eval_metric_ops['accuracy'][1].dtype,
tf.float32)
for v in tf.trainable_variables():
self.assertEqual(v.dtype.base_dtype, tf.float32)
tensors_to_check = ('initial_conv:0', 'block_layer1:0',
'block_layer2:0', 'block_layer3:0',
'final_reduce_mean:0', 'final_dense:0')
for tensor_name in tensors_to_check:
tensor = g.get_tensor_by_name('resnet_model/' + tensor_name)
self.assertEqual(
tensor.dtype, dtype,
'Tensor {} has dtype {}, while dtype {} was '
'expected'.format(tensor, tensor.dtype, dtype))
def _cifar10_model_fn_helper(self, mode, version, dtype, multi_gpu=False):
with tf.Graph().as_default() as g:
input_fn = cifar10_main.get_synth_input_fn()
dataset = input_fn(True, '', _BATCH_SIZE)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features, labels, mode, {
'dtype': dtype,
'resnet_size': 32,
'data_format': 'channels_last',
'batch_size': _BATCH_SIZE,
'version': version,
'loss_scale': 128 if dtype == tf.float16 else 1,
'multi_gpu': multi_gpu
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (_BATCH_SIZE,))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops['accuracy'][0].shape, ())
self.assertAllEqual(eval_metric_ops['accuracy'][1].shape, ())
self.assertEqual(eval_metric_ops['accuracy'][0].dtype, tf.float32)
self.assertEqual(eval_metric_ops['accuracy'][1].dtype, tf.float32)
for v in tf.trainable_variables():
self.assertEqual(v.dtype.base_dtype, tf.float32)
tensors_to_check = ('initial_conv:0', 'block_layer1:0', 'block_layer2:0',
'block_layer3:0', 'final_reduce_mean:0',
'final_dense:0')
for tensor_name in tensors_to_check:
tensor = g.get_tensor_by_name('resnet_model/' + tensor_name)
self.assertEqual(tensor.dtype, dtype,
'Tensor {} has dtype {}, while dtype {} was '
'expected'.format(tensor, tensor.dtype,
dtype))
def cifar10_model_fn_helper(self, mode, version, multi_gpu=False):
input_fn = cifar10_main.get_synth_input_fn()
dataset = input_fn(True, "", _BATCH_SIZE)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
spec = cifar10_main.cifar10_model_fn(
features,
labels,
mode,
{
"resnet_size": 32,
"data_format": "channels_last",
"batch_size": _BATCH_SIZE,
"version": version,
"multi_gpu": multi_gpu,
},
)
predictions = spec.predictions
self.assertAllEqual(predictions["probabilities"].shape,
(_BATCH_SIZE, 10))
self.assertEqual(predictions["probabilities"].dtype, tf.float32)
self.assertAllEqual(predictions["classes"].shape, (_BATCH_SIZE, ))
self.assertEqual(predictions["classes"].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
self.assertAllEqual(loss.shape, ())
self.assertEqual(loss.dtype, tf.float32)
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = spec.eval_metric_ops
self.assertAllEqual(eval_metric_ops["accuracy"][0].shape, ())
self.assertAllEqual(eval_metric_ops["accuracy"][1].shape, ())
self.assertEqual(eval_metric_ops["accuracy"][0].dtype, tf.float32)
self.assertEqual(eval_metric_ops["accuracy"][1].dtype, tf.float32)
