I still want to know how I can predict the value of an image after training the network, but it seems like it is not supported yet. Any idea for a workaround (taken from the mnist_tpu.py)?
if mode == tf.estimator.ModeKeys.PREDICT:
raise RuntimeError("mode {} is not supported yet".format(mode))
Besides Stackoverflow - anywhere else I can get support for the implementing my models using TPU?
Here is a Python program that sends an image to a TPU-trained model (ResNet in this case) and gets back a classification:
with tf.gfile.FastGFile('/some/path.jpg', 'r') as ifp:
credentials = GoogleCredentials.get_application_default()
api = discovery.build('ml', 'v1', credentials=credentials,
discoveryServiceUrl='https://storage.googleapis.com/cloud-ml/discovery/ml_v1_discovery.json')
request_data = {'instances':
[
{"image_bytes": {"b64": base64.b64encode(ifp.read())}}
]
}
parent = 'projects/%s/models/%s/versions/%s' % (PROJECT, MODEL, VERSION)
response = api.projects().predict(body=request_data, name=parent).execute()
print("response={0}".format(response))
Full code is here: https://github.com/GoogleCloudPlatform/training-data-analyst/blob/master/quests/tpu/flowers_resnet.ipynb
This article documents the process of writing a model for the Cloud TPU: https://medium.com/tensorflow/how-to-write-a-custom-estimator-model-for-the-cloud-tpu-7d8bd9068c26
It is supported now. Changes have been done to https://github.com/tensorflow/models/blob/master/official/mnist/mnist_tpu.py to make it working.
Besides stackoverflow, you can add your issues on github https://github.com/tensorflow/tpu/issues.
According to the documentation, you can choose online or batch modes for prediction, but you can't select the target device. As stated, "the prediction service allocates resources to run your job."
The documentation says that prediction is performed by nodes. I thought I'd read somewhere that prediction nodes are always CPUs in the Google Compute Engine, but I can't find a clear reference.
Related
I'm trying to reload or access the Keras-Tuner Trials after the Tuner's search has completed for inspecting the results. I'm not able to find any documentation or answers related to this issue.
For example, I set up BayesianOptimization to search for the best hyper-parameters as follows:
## Build Hyper Parameter Search
tuner = kt.BayesianOptimization(build_model,
objective='val_categorical_accuracy',
max_trials=10,
directory='kt_dir',
project_name='lstm_dense_bo')
tuner.search((X_train_seq, X_train_num), y_train_cat,
epochs=30,
batch_size=64,
validation_data=((X_val_seq, X_val_num), y_val_cat),
callbacks=[callbacks.EarlyStopping(monitor='val_loss', patience=3,
restore_best_weights=True)])
I see this creates trial files in the directory kt_dir with project name lstm_dense_bo such as below:
Now, if I restart my Jupyter kernel, how can I reload these trials into a Tuner object and subsequently inspect the best model or the best hyperparameters or the best trial?
I'd very much appreciate your help. Thank you
I was trying to do the same thing. I was looking into the keras docs for an easier way than this but could not find one - so if any other SO-ers have a better idea, please let us know!
Load the previous tuner. Make sure overwrite=False or else you'll delete your trials.
workdir = "mlp_202202151345"
obj = "val_recall"
tuner = kt.Hyperband(
hypermodel=build_model,
metrics=metrics,
objective=kt.Objective(obj, direction="max"),
executions_per_trial=1,
overwrite=False,
directory=workdir,
project_name="keras_tuner",
)
Look for a trial you want to load. Note that TensorBoard works really well for this. In this example, I'm loading 1a38ebaba07b77501999cb1c4ab9413e.
Here's the part that I could not find in Keras docs. This might be dependent on the tuner you use (I am using Hyperband):
tuner.oracle.get_trial('1a38ebaba07b77501999cb1c4ab9413e')
Returns a Trial object (also could not find in the docs). The Trial object has a hyperparameters attribute that will return that trial's hyperparameters. Now:
tuner.hypermodel.build(trial.hyperparameters)
Gives you the trial's model for training, evaluation, predictions, etc.
NOTE This seems convuluted and hacky, would love to see a better way.
j7skov has correctly mentioned that you need to reload previous tuner and set the parameter overwrite=False(so that tuner will not overwrite already generated trials).
Further if you want to load first K best models then we need to use tuner's get_best_models method as below
# This will load 10 best hyper tuned models with the weights
# corresponding to their best checkpoint (at the end of the best epoch of best trial).
best_model_count = 10
bo_tuner_best_models = tuner.get_best_models(num_models=best_model_count)
Then you can access a specific best model as below
best_model_id = 7
model = bo_tuner_best_models[best_model_id]
This method is for querying the models trained during the search. For best performance, it is recommended to retrain your Model on the full dataset using the best hyperparameters found during search, which can be obtained using tuner.get_best_hyperparameters().
tuner_best_hyperparameters = tuner.get_best_hyperparameters(num_trials=best_model_count)
best_hp = tuner_best_hyperparameters[best_model_id]
model = tuner.hypermodel.build(best_hp)
If you want to just display hyperparameters for the K best models then use tuner's results_summary method as below
tuner.results_summary(num_trials=best_model_count)
For further reference visit this page.
Inspired by j7skov, I found that the models can be reloaded
by manipulating tuner.oracle.trials and tuner.load_model.
By assigning tuner.oracle.trials to a variable, we can find that it is a dict object containing all relavant trials in the tuning process.
The keys of the dictionary are the trial_id, and the values of the
dictionary are the instance of the Trial object.
Alternatively, we can return the best few trials by using tuner.oracle.get_best_trials.
To inspect the hyperparameters of the trial, we can use the summary method of the instance.
To load the model, we can pass the trial instance to tuner.load_model.
Beware that different versions can lead to incompatibilities.
For example the directory structure is a little different between keras-tuner==1.0 and keras-tuner==1.1 as far as I know.
Using your example, the working flow may be summarized as follows.
# Recreate the tuner object
tuner = kt.BayesianOptimization(build_model,
objective='val_categorical_accuracy',
max_trials=10,
directory='kt_dir',
project_name='lstm_dense_bo',
overwrite=False)
# Return all trials from the oracle
trials = tuner.oracle.trials
# Print out the ID and the score of all trials
for trial_id, trial in trials.items():
print(trial_id, trial.score)
# Return best 5 trials
best_trials = tuner.oracle.get_best_trials(num_trials=5)
for trial in best_trials:
trial.summary()
model = tuner.load_model(trial)
# Do some stuff to the model
using
tuner = kt.BayesianOptimization(build_model,
objective='val_categorical_accuracy',
max_trials=10,
directory='kt_dir',
project_name='lstm_dense_bo')
will load the tuner again.
Problem
I'm working on training and deploying an instance segmentation model using TF's object detection API. I'm able to successfully train the model, package it into a TF Serving Docker image (latest tag as of Oct 2020), and process inference requests via the REST interface. However, the amount of data returned from an inference request is very large (hundreds of Mb). This is a big problem when the inference request and processing don't happen on the same machine because all that returned data has to go over the network.
Is there a way to trim down the number of outputs (either during model export or within the TF Serving image) so allow faster round trip times during inference?
Details
I'm using TF OD API (with TF2) to train a Mask RCNN model, which is a modified version of this config. I believe the full list of outputs is described in code here. The list of items I get during inference is also pasted below. For a model with 100 object proposals, that information is ~270 Mb if I just write the returned inference as json to disk.
inference_payload['outputs'].keys()
dict_keys(['detection_masks', 'rpn_features_to_crop', 'detection_anchor_indices', 'refined_box_encodings', 'final_anchors', 'mask_predictions', 'detection_classes', 'num_detections', 'rpn_box_predictor_features', 'class_predictions_with_background', 'proposal_boxes', 'raw_detection_boxes', 'rpn_box_encodings', 'box_classifier_features', 'raw_detection_scores', 'proposal_boxes_normalized', 'detection_multiclass_scores', 'anchors', 'num_proposals', 'detection_boxes', 'image_shape', 'rpn_objectness_predictions_with_background', 'detection_scores'])
I already encode the images within my inference requests as base64, so the request payload is not too large when going over the network. It's just that the inference response is gigantic in comparison. I only need 4 or 5 of the items out of this response, so it'd be great to exclude the rest and avoid passing such a large package of bits over the network.
Things I've tried
I've tried setting the score_threshold to a higher value during the export (code example here) to reduce the number of outputs. However, this seems to just threshold the detection_scores. All the extraneous inference information is still returned.
I also tried just manually excluding some of these inference outputs by adding the names of keys to remove here. That also didn't seem to have any effect, and I'm worried this is a bad idea because some of those keys might be needed during scoring/evaluation.
I also searched here and on tensorflow/models repo, but I wasn't able to find anything.
I was able to find a hacky workaround. In the export process (here), some of the components of the prediction dict are deleted. I added additional items to the non_tensor_predictions list, which contains all keys that will get removed during the postprocess step. Augmenting this list cut down my inference outputs from ~200MB to ~12MB.
Full code for the if self._number_of_stages == 3 block:
if self._number_of_stages == 3:
non_tensor_predictions = [
k for k, v in prediction_dict.items() if not isinstance(v, tf.Tensor)]
# Add additional keys to delete during postprocessing
non_tensor_predictions = non_tensor_predictions + ['raw_detection_scores', 'detection_multiclass_scores', 'anchors', 'rpn_objectness_predictions_with_background', 'detection_anchor_indices', 'refined_box_encodings', 'class_predictions_with_background', 'raw_detection_boxes', 'final_anchors', 'rpn_box_encodings', 'box_classifier_features']
for k in non_tensor_predictions:
tf.logging.info('Removing {0} from prediction_dict'.format(k))
prediction_dict.pop(k)
return prediction_dict
I think there's a more "proper" way to deal with this using signature definitions during the creation of the TF Serving image, but this worked for a quick and dirty fix.
I've ran into the same problem. In the exporter_main_v2 code there is stated that the outputs should be:
and the following output nodes returned by the model.postprocess(..):
* `num_detections`: Outputs float32 tensors of the form [batch]
that specifies the number of valid boxes per image in the batch.
* `detection_boxes`: Outputs float32 tensors of the form
[batch, num_boxes, 4] containing detected boxes.
* `detection_scores`: Outputs float32 tensors of the form
[batch, num_boxes] containing class scores for the detections.
* `detection_classes`: Outputs float32 tensors of the form
[batch, num_boxes] containing classes for the detections.
I've submitted an issue on the tensorflow object detection github repo, I hope we will get feedback from the tensorflow dev team.
The github issue can be found here
If you are using exporter_main_v2.py file to export your model, you can try this hack way to solve this problem.
Just add following codes in the function _run_inference_on_images of exporter_lib_v2.py file:
detections[classes_field] = (
tf.cast(detections[classes_field], tf.float32) + label_id_offset)
############# START ##########
ignored_model_output_names = ["raw_detection_boxes", "raw_detection_scores"]
for key in ignored_model_output_names:
if key in detections.keys(): del detections[key]
############# END ##########
for key, val in detections.items():
detections[key] = tf.cast(val, tf.float32)
Therefore, the generated model will not output the values of ignored_model_output_names.
Please let me know if this can solve your problem.
Another approach would be to alter the signatures of the saved model:
model = tf.saved_model.load(path.join("models", "efficientdet_d7_coco17_tpu-32", "saved_model"))
infer = model.signatures["serving_default"]
outputs = infer.structured_outputs
for o in ["raw_detection_boxes", "raw_detection_scores"]:
outputs.pop(o)
tf.saved_model.save(
model,
export_dir="export",
signatures={"serving_default" : infer},
options=None
)
I successfully managed to implement learning to rank by following the tutorial TF-Ranking for sparse features using the ANTIQUE question answering dataset.
Now my goal is to successfully save the learned model to disk so that I can easily load it without training again. Due to the Tensorflow docs, the estimator.export_saved_model() method seems to be the way to go. But I can't wrap my head around how to tell Tensorflow how my feature structure looks like. Due to the docs here the easiest way seems to be calling tf.estimator.export.build_parsing_serving_input_receiver_fn(), which returns me the required inpur receiver function which I have to pass to the export_saved_model function. But how do I tell Tensorflow how my features from my learning to rank model look like?
From my current understanding the model has context feature specs and example feature specs. So I guess I somehow have to combine those two specs into one feature description, which I then can pass to the build_parsing_serving_input_receiver_fn function?
So I think you are on the right track;
You can get a build_ranking_serving_input_receiver_fn like this: (substitue context_feature_columns(...) and example_feature_columns(...) with defs you probably have for creating your own context and example structures for your training data):
def example_serving_input_fn():
context_feature_spec = tf.feature_column.make_parse_example_spec(
context_feature_columns(_VOCAB_PATHS).values())
example_feature_spec = tf.feature_column.make_parse_example_spec(
list(example_feature_columns(_VOCAB_PATHS).values()))
servingInputReceiver = tfr.data.build_ranking_serving_input_receiver_fn(
data_format=tfr.data.ELWC,
context_feature_spec=context_feature_spec,
example_feature_spec=example_feature_spec,
list_size=_LIST_SIZE,
receiver_name="input_ranking_data",
default_batch_size=None)
return servingInputReceiver
And then pass this to export_saved_model like this:
ranker.export_saved_model('path_to_save_model', example_serving_input_fn())
(ranker here is a tf.estimator.Estimator, maybe you called this 'estimator' in your code)
ranker = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=_MODEL_DIR,
config=run_config)
I'm training MobileNet on WIDER FACE dataset and I encountered problem I couldn't solve. TF Object Detection API stores only last 5 checkpoints in train dir, but what I would like to do, is to save best models relative to mAP metric (or at least leave many more models in train dir before deletion).
For example, today I've looked at Tensorboard after next night of training and I see that overnight model has over-fitted and I can't restore best checkpoint, because it was deleted already.
EDIT: I just use Tensorflow Object Detection API, it by default saves last 5 checkpoints in train dir which I point. I look for some configuration parameter or anything that will change this behavior.
Has anyone have some fix in code/config param to set/workaround for that? It seems like I'm missing something, it should be obvious that what's in fact important is the best model, not the newest one (which can overfit).
Thanks!
You can modify (hardcoding in your fork or opening a pull request and adding the options to protos) the arguments passed to tf.train.Saver in:
https://github.com/tensorflow/models/blob/master/research/object_detection/legacy/trainer.py#L376-L377
You will probably want to set:
max_to_keep: Maximum number of recent checkpoints to keep. Defaults to 5.
keep_checkpoint_every_n_hours: How often to keep checkpoints. Defaults to 10,000 hours.
You can change config.
in run_config.py
class RunConfig(object):
"""This class specifies the configurations for an `Estimator` run."""
def __init__(self,
model_dir=None,
tf_random_seed=None,
save_summary_steps=100,
save_checkpoints_steps=_USE_DEFAULT,
save_checkpoints_secs=_USE_DEFAULT,
session_config=None,
keep_checkpoint_max=10,
keep_checkpoint_every_n_hours=10000,
log_step_count_steps=100,
train_distribute=None,
device_fn=None,
protocol=None,
eval_distribute=None,
experimental_distribute=None):
You may be interested by this Tf github thread that tackles the newest/best checkpoint issue. A user developed his own wrapper, chekmate, around tf.Saver to keep track of the best checkpoints.
You can follow up this PR. Here your best checkpoint is saved within your checkpoint directory, sub-directory named as best.
You just need to integrate the best_saver() and (method call in _run_checkpoint_once()) inside ../object_detection/eval_util.py
Additionally it will also create a json for all_evaluation_metrices.
For saving more checkpoints, you can write a simple python script that will store the checkpoints in a timely manner to a specific.
import os
import shutil
import time
while True:
training_file = '/home/vignesh/training' # path of your train directory
archive_file = 'home/vignesh/training/archive' #path of the directory where you want to save your checkpoints
files_to_save = []
for files in os.listdir(training_file):
if files.rsplit('.')[0]=='model':
files_to_save.append(files)
for files in files_to_save:
if files in os.listdir(archive_file):
pass
else:
shutil.copy2(training_file+'/'+files,archive_file)
time.sleep(600) # This will make the script run for every 600 seconds, modify it for your need
The tensorflow documentation does not provide any example of how to perform a periodic evaluation of the model on an evaluation set.
Some people suggested the use of an Experiment, which sounds great but unfortunately does not work (depreciation and triggers an error).
Others suggested the use of SummarySaverHook, but I don't see how you can use that with an evaluation set (as opposed to the training set).
A solution would be to do the following
for i in range(number_of_epoch):
estimator.train(...) // on training set
estimator.evaluate(...) // on evaluation set
This architecture is explicitly discouraged in this paper (page 4 top right).
Any other idea/implementation?
EDIT:
The error message when running the experiment is the following:
File ".../anaconda2/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 253, in train if (config.environment != run_config.Environment.LOCAL and
AttributeError: 'RunConfig' object has no attribute 'environment'
Tensorflow version 1.3
Only a few parameters/options of Experiment are deprecated (what specific errors are you seeing?). If you create an Estimator that will do periodic checkpoints (using options in RunConfig) and an Experiment using it, you will get evaluation for each checkpoint by default when using train_and_evaluate method.
EDIT: As Maxime pointed out in the comments. He needed to add the following lines to get rid of his error:
os.environ['TF_CONFIG'] = json.dumps({'environment': 'local'})
config = tf.contrib.learn.RunConfig()