def testTrainSave(self):
model_type = 'mockfasterrcnn'
# Save checkpoints to a temp directory.
tmp_job_dir = tempfile.mkdtemp()
override_params = [
'train.num_epochs={}'.format(self.total_epochs),
'train.job_dir={}'.format(tmp_job_dir),
'train.run_name=test_runname',
]
config = self.get_config(model_type, override_params=override_params)
step = run(config,
get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model)
self.assertEqual(step, 2)
# We have to reset the graph to avoid having duplicate names.
tf.reset_default_graph()
step = run(config,
get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model)
# This is because of a MonitoredTrainingSession "bug".
# When ending training it saves a checkpoint as the next step.
# That causes that we are one step ahead when loading it.
self.assertEqual(step, 5)
def testTrainSave(self):
model_type = 'mockfasterrcnn'
# Save checkpoints to a temp directory.
tmp_job_dir = tempfile.mkdtemp()
override_params = [
'train.num_epochs={}'.format(self.total_epochs),
'train.job_dir={}'.format(tmp_job_dir),
'train.run_name=test_runname',
]
config = self.get_config(model_type, override_params=override_params)
run(config,
get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model)
# Create new graph which will load previously saved checkpoint
tf.reset_default_graph()
new_session = tf.Session()
new_saver = tf.train.import_meta_graph(
tmp_job_dir + '/test_runname/model.ckpt-3.meta')
new_saver.restore(new_session,
tmp_job_dir + '/test_runname/model.ckpt-3')
# Get tensor from graph and run it in session
w_tensor = tf.get_default_graph().get_tensor_by_name(
"mockfasterrcnn/w:0")
w_numpy = new_session.run(w_tensor)
# Assert we correctly loaded the weight
self.assertArrayNear(w_numpy, [2.5, 3.0], err=0.01)
def testTrainSave(self):
model_type = 'mockfasterrcnn'
# Save checkpoints to a temp directory.
tmp_job_dir = tempfile.mkdtemp()
override_params = [
'train.num_epochs={}'.format(self.total_epochs),
'train.job_dir={}'.format(tmp_job_dir),
'train.run_name=test_runname',
]
config = self.get_config(model_type, override_params=override_params)
step = run(
config,
get_dataset_fn=self.get_dataset, get_model_fn=self.get_model
)
self.assertEqual(step, 2)
# We have to reset the graph to avoid having duplicate names.
tf.reset_default_graph()
step = run(
config,
get_dataset_fn=self.get_dataset, get_model_fn=self.get_model
)
# This is because of a MonitoredTrainingSession "bug".
# When ending training it saves a checkpoint as the next step.
# That causes that we are one step ahead when loading it.
self.assertEqual(step, 5)
def testTrain(self):
config = self.config
custom_config = load_config(config.config_files)
# The string we use here is ignored.
model_type = 'mockfasterrcnn'
# This should not fail
run(custom_config, model_type, config.override_params,
get_dataset_fn=self.get_dataset, get_model_fn=self.get_model)
def testTrain(self):
config = self.config
# This should not fail
run(config.model_type,
config.dataset_type,
config.config_file,
config.override_params,
run_name=config.run_name,
save_summaries_secs=config.save_summaries_secs,
get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model)
def testTrain(self):
model_type = 'mockfasterrcnn'
override_params = [
'train.num_epochs={}'.format(self.total_epochs),
'train.job_dir=',
]
config = self.get_config(model_type, override_params=override_params)
# This should not fail
run(config,
get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model)
def testTrain(self):
model_type = 'mockfasterrcnn'
override_params = [
'train.num_epochs={}'.format(self.total_epochs),
'train.job_dir=',
]
config = self.get_config(model_type, override_params=override_params)
# This should not fail
run(
config, get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model
)
def testTrain(self):
custom_config = load_config(self.config.config_files)
# The string we use here is ignored.
model_type = 'mockfasterrcnn'
self.config.override_params = [
'train.num_epochs={}'.format(self.total_epochs),
'train.job_dir=',
]
# This should not fail
run(custom_config,
model_type,
self.config.override_params,
get_dataset_fn=self.get_dataset,
get_model_fn=self.get_model)