tensorflow/adanet

AdaNet: Fast and flexible AutoML with learning guarantees

Eugen Hotaj

@EugenHotaj

Hey @jankrynauw, yes unfortunately the candidate evaluations are hard-coded to happen on CPU when we decide which candidate to select to move on to the next iteration

As you mentioned, eval_on_tpu only affects estimator.evaluate()

This hard-coding was necessary based on how previous versions of AdaNet were implemented that we've since moved away from. Looking at the code now, there doesn't seem to be any fundamental reason why we need to evaluate on CPU during the candidate evaluation phase.

Eugen Hotaj

@EugenHotaj

Please feel free to open an issue against us to track this

Jan Krynauw

@jankrynauw

Thank you @EugenHotaj
I've added feature request: #131

Eugen Hotaj

@EugenHotaj

Thanks!

chandramoulirajagopalan

@chandramoulirajagopalan

@cweill can I work on implementing the tf.keras.layers in adanet ?

Charles Weill

@cweill

@chandramoulirajagopalan Thanks for offering, but it's a bit complex at the moment to add support, and we have someone on the team working on it.

chandramoulirajagopalan

@chandramoulirajagopalan

Is there anything I could work on in Adanet ?

Eugen Hotaj

@EugenHotaj

@chandramoulirajagopalan If you're feeling adventurous and what to try out some TPU work, feel free to pick up #131. Unfortunately testing will be pretty hard since we don't have any unit tests on TPU for Travis. You could try manually testing via colab (http://colab.reserach.google.com).

Something easier might be #130 or fixing up the issues with your PR for #119 from August 22

chandramoulirajagopalan

@chandramoulirajagopalan

For PR #119 Issues were not commented. I am still waiting for a review for it to be accepted.

I will work on #131 and try it out.

Feng Li

@Bee-zest

adanet is part of kaggle days google solution?

Feng Li

@Bee-zest

I can't use full cpu resource when I train Adanet on tabular data which has 100000rows and 2000 columns,

palak

@developer22-university

hey who are admin of this group please tell me how can i contribute for gci mentor?

Charles Weill

@cweill

@Bee-zest > I can't use full cpu resource when I train Adanet on tabular data which has 100000rows and 2000 columns

Can you please provide more information?

Could be due to many reasons. Perhaps you dataset input pipeline isn't optimized. Could you please file an issue on our repository on GitHub?

chandramoulirajagopalan

@chandramoulirajagopalan

when I try to install adanet using local package source I get

class TPUEstimator(Estimator, tf.compat.v1.estimator.tpu.TPUEstimator):
AttributeError: module 'tensorflow._api.v2.compat.v2.compat' has no attribute 'v1'

Can someone help me out with this ?

Yu Zhang

@yz-zest

@chandramoulirajagopalan You can change your TF version to 1.1-ish.

🐻

@mt_spt_twitter

AttributeError: module 'tensorflow._api.v1.compat.v1' has no attribute 'GraphKey'

can someone help me with this? i've tried using tf version 2 and 1.

Charles Weill

@cweill

@mt_spt_twitter Please file an issue with the full stacktrace and sample code.

🐻

@mt_spt_twitter


Traceback (most recent call last):
  File "train.py", line 51, in <module>
    from builders import model_builder
  File "C:\tensorflow1\models\research\object_detection\builders\model_builder.py", line 35, in <module>
    from object_detection.models import faster_rcnn_inception_resnet_v2_feature_extractor as frcnn_inc_res
  File "C:\tensorflow1\models\research\object_detection\models\faster_rcnn_inception_resnet_v2_feature_extractor.py", line 29, in <module>
    from nets import inception_resnet_v2
  File "C:\tensorflow1\models\research\slim\nets\inception_resnet_v2.py", line 373, in <module>
    batch_norm_updates_collections=tf.compat.v1.GraphKey.UPDATE_OPS,
AttributeError: module 'tensorflow._api.v1.compat.v1' has no attribute 'GraphKey'

the code are

def inception_resnet_v2_arg_scope(
    weight_decay=0.00004,
    batch_norm_decay=0.9997,
    batch_norm_epsilon=0.001,
    activation_fn=tf.nn.relu,
    batch_norm_updates_collections=tf.compat.v1.GraphKey.UPDATE_OPS,
    batch_norm_scale=False):

🐻

@mt_spt_twitter

i've solved the problem. i just realized that it should be GraphKeys not GraphKey haha mybad

Akiz Uddin Ahmed

@shawpan

hi, for a custom estimator getting this error with tf-serving regress api

Expected output Tensor shape to be either [batch_size] or [batch_size, 1] but got [1,2]\

ijalabt1

@ijalabt1

I am using tensorflow version:1.9 and python version:3.7 while training tensor.api.v1 has no attribute v1 error

Shivam

@andy6975

Hi! I am Shivam from IIT Mandi. Perhaps this is not the best place to ask this but could anyone please tell me how can I contact GSoc 2020 mentors? I am interested in some of the proposed projects and would like to know more about them.

dawg

@le-dawg

Hi all

dawg

@le-dawg

I want to adapt the https://github.com/tensorflow/adanet/blob/master/adanet/examples/tutorials/customizing_adanet.ipynb for use with time series

dawg

@le-dawg

I have an issue with

The input_fn for the estimator. I have created a gf.data.

I have a tf.data.Dataset and an input_fn that returns that dataset. I get a "input_fn" not callable error. But this approach is documented in the tensorflow docs. Can anyone help?

dawg

@le-dawg

I actually created a well-defined SO post: https://stackoverflow.com/q/60437843/2949797

dawg

@le-dawg

I have a tf.data.Dataset and an input_fn that returns that dataset. I get a "input_fn" not callable error. But this approach is documented in the tensorflow docs. Can anyone help?

I may have resolved this (peculiarity of estimator API, input_fn must be argument-free) but there is a problem

FireStorm

@Gyanig

I am unable to come up with any proposals. can someone help?

ROHIT VISHNU GHUMARE

@rohitg00

Hello, I am Rohit Ghumare (final year student of CSE). I have worked on computer vision and graph based networks for Indian Sign language detection project and also worked over image classification models for disease detection for plant leaf images, I would like to contribute to Tensorflow. I would be grateful if someone could guide me on how to start contributing.
LinkedIn: https://www.linkedin.com/in/rohit-ghumare-418375140/

Amisha Shukla

@Amisha-100

Hello World....This is Amisha Shukla. A third year engineering student from India. I had explore few of the datasets, build some data analysis projects..Now, I want to be a part of GSoC and utilise my summer by contributing to TensorFlow. So, I need some guidance!

anudeepSky

@anudeepSky

Hello, Anudeep Kaur from India this side. I happen to be a final year engineering student with an experience in computer vision projects. I am also the recipient of the Intel Edge AI foundation course scholarship. I'd like to contribute to TF, kindly guide me through the same.

Ritvik Sapra

@Ritvik-Sapra

Hi everyone, I am Ritvik Sapra, from Delhi. I am in third year of B.Tech. ECE. I am glad to see I am not the only beginner here! I really want to start my career in AI and contribute to TF. Hi everyone!

@Amisha-100 How did your start doing projects? I mean I try but I am not able to gather the confidence that I will be able to do it..

Ritvik Sapra

@Ritvik-Sapra

@Gyanig I relate! But I am trying to find a subset of what I already know and what I want to learn. By that we can narrow down projects and then check out their websites and projects..

anurag93

@anurag93

Hi, this is Anurag. I am a grad student at the University of California - Irvine and I have research experience and publication in the space of Natural Language Processing. I wanted to get started with open source contribution as I have used tensorflow and its related libraries. Could someone guide me through the process of applying at for the student open source community ?

Ritvik Sapra

@Ritvik-Sapra

Hi @anurag93 I'm glad you asked. Communities and student packs are really very essential. My personal fav is the GitHub student dev pack till now. Got to https://education.github.com and look for students tab. You will definitely get good resources about a no. Of things. I'm also a Microsoft student partner so I can tell MSP program is really good, you must consider applying to it. https://studentpartners.microsoft.com

@anurag93 and the apple student representative program is great too. I don't know much about it, but do consider checking it out. And ofc, the all time fav dsc leads are awesome. Google student clubs are one of the best and you may get to know a lot about tensor flow and other google tech.

nicholasbreckwoldt

@nicholasbreckwoldt

Hi there,

I just converted an existing project from TF 1.14 to TF 2.1 (both having version 0.8.0 of Adanet installed) using adanet.TPUEstimator. After making the switch, testing locally on CPU with use_tpu=False runs successfully. However, I am getting errors when running on TPU (i.e. use_tpu=True). The errors appear to originate in the tpu_estimator.py and error_handling.py scripts seen in the Traceback below:

```File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/tpu_estimator.py", line 3032, in train
rendezvous.record_error('training_loop', sys.exc_info())
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/error_handling.py", line 81, in record_error
if value and value.op and value.op.type == _CHECK_NUMERIC_OP_NAME:
AttributeError: 'RuntimeError' object has no attribute 'op'

During handling of the above exception, another exception occurred:

File "workspace/trainer/train.py", line 331, in <module>
main(args=parsed_args)
File "workspace/trainer/train.py", line 177, in main
run_config=run_config)
File "workspace/trainer/train.py", line 68, in run_experiment
estimator.train(input_fn=train_input_fn, max_steps=total_train_steps)
File "/usr/local/lib/python3.6/site-packages/adanet/core/estimator.py", line 853, in train
saving_listeners=saving_listeners)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/tpu_estimator.py", line 3035, in train
rendezvous.raise_errors()
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/error_handling.py", line 143, in raise_errors
six.reraise(typ, value, traceback)
File "/usr/local/lib/python3.6/site-packages/six.py", line 703, in reraise
raise value
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/tpu_estimator.py", line 3030, in train
saving_listeners=saving_listeners)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 374, in train
loss = self._train_model(input_fn, hooks, saving_listeners)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 1164, in _train_model
return self._train_model_default(input_fn, hooks, saving_listeners)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 1194, in _train_model_default
features, labels, ModeKeys.TRAIN, self.config)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/tpu_estimator.py", line 2857, in _call_model_fn
config)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 1152, in _call_model_fn
model_fn_results = self._model_fn(features=features, **kwargs)
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/tpu_estimator.py", line 3186, in _model_fn
host_ops = host_call.create_tpu_hostcall()
File "/usr/local/lib/python3.6/site-packages/tensorflow_estimator/python/estimator/tpu/tpu_estimator.py", line 2226, in create_tpu_hostcall
'dimension, but got scalar {}'.format(dequeue_ops[i][0]))
RuntimeError: All tensors outfed from TPU should preserve batch size dimension, but got scalar Tensor("OutfeedDequeueTuple:1", shape=(), dtype=int64, device=/job:tpu_worker/task:0/device:CPU:0)
'''

The previous version of the project using TF 1.14 and Adanet 0.8.0 runs both locally and on TPU using adanet.TPUEstimator without issues. Is there something obvious I am potentially missing for the switch over to TF 2.1 when using TPUEstimator?

chandramoulirajagopalan

@chandramoulirajagopalan

How can adanet be used for CNN based image level classification ? Can we implement a block of CNN like one Inception Module and then start ensembling using Adanet ?

@cweill

Where communities thrive

People

Repo info

Activity