TensorFlow high-level APIs: Part 1 - loading data

Welcome to Part 1 of our mini-series on TensorFlow high-level APIs! In this 3 part mini-series, TensorFlow Engineering Manager Karmel Allison runs us through different scenarios using TensorFlow’s high-level APIs. Building a ML model takes a lot of time, effort, and often involves multiple stages. Luckily, TensorFlow high-level APIs aim to help you along with each stage, from the start of your idea, to training and serving large scale applications. Watch to discover the key steps in developing machine learning models, where TensorFlow comes in for each step, and lastly how to prepare and load your data!

And...stay tuned for Part 2 & 3!

[wrightw]

Super helpful series. Doesn't just tell you how to build SOME PARTICULAR model, but how to build any basic/canned model. Best mini-tutorial I've seen so far.

[MikeDownes]

Thanks this was incredibly well put together easy to understand and so so useful.

[mdrayedbinwahed]

Typo in 5:24. 'row_vals' should be 'vals'

[bhesht]

Immaculately described and concise at once. Thank you!

[andrassteger]

A lot of changes happened in the last 3 years. Maybe some update would be appreciated :-) Thank You!

[nunodias]

Many things to do only for read a datafile. It will be more easy do a function to aggreate all possibilities and the end user set the options in according with your own needs. In fact this demo is useful only if you knows the structure of your data. If you can not open the file with a simple editor you don't know the structure of your data, so you can not setup TF to read the data if you do not know the structure of the data.

[writesekhar]

Good and well presented
looking for coode and examples.

[soumyodey]

There is a big mistake in the code shown in the video. The function _parse_csv_rows() is incorrect, it should be


def _parse_csv_rows(*vals):
soil_type_t =
feat_vals = vals[:10] + (soil_type_t, vals[54])
features = dict(zip(col_names_features, feat_vals))
class_label = tf.argmax(vals[10:14], axis=0)
labels = {col_names_label: class_label}
return features, labels

where col_names_features, col_names_label are

col_names_features = [
"Elevation",
"Aspect",
"Slope",
"Horizontal_Distance_To_Hydrology",
"Vertical_Distance_To_Hydrology",
"Horizontal_Distance_To_Roadways",
"Hillshade_9am",
"Hillshade_Noon",
"Hillshade_3pm",
"Horizontal_Distance_To_Fire_Points",
"Soil_Type (40 binary columns)",
"Cover_Type (7 types)"
]

col_names_label = "Wilderness_Area (4 binary columns)"

[gowthamkumar]

In Tensorflow updated versions, numpy is removed from tensor, then how we will extract value from that tensor

[tingnews]

great vedio. Finally tensorflow make some vedio I can unstand. Good job

[qiyang]

What can we do to solve the problem of memory leaking when cooperating tf.keras with tf?

[sachavanweeren]

There is a bug in the code snippet on Parsing data - 'row_vals' is not defined, this should be replace by ' vals' . Better to share working code and provide links with the video.

[neerajgobari]

How can I use model final output to compute a prediction..so that I can compare prediction to ground truth value to compute loss function and fine tune my model

[BadriNathJK]

Can you tell me how to get each token vector when using both feature column and TF HUB? Because feature column gives average of token vector in a sentence. I am struck with the problem now

[shionwill]

Please share your code and the data so that we can follow along this great tutorial!

[huqiao]

This is not a very good example as you were using tf.contrib which will soon be deprecated in TensorFlow 2.0.

[anubhav]

hello,
thanks for the wonderful explanation of tensorflow usage. Could you please point me to the guide for i/o optimization @5:51 ?

[aonoymousandy]

my csv files have a string at the end of each row, and this is causing error in my code, how would I setup defaults to account that I have a row of 10 floats and the very last feature is a string?

[fusedglass]

Is this supposed to be TF 2.0? when I issue a print version I get Unbelievable how much pain these guys cause, by changing stuff every release. Google... can you please Google the term "backwards compatibility?" I'm just trying to read in a two column csv. I agree it looked great while I was watching it, but its next to useless when I get errors about contrib and enable_eager arnt part of the tf module. Guys please release more of this style, but add more detail around what versions you need and at the end show us 100% of the code.

[shaktirajsinhjadeja]

Hi, Great video to learn from :) Will it be possible to get the access of dataset ?