Input configuration for ImportData Action.
The format of input depends on dataset_metadata the Dataset into which
the import is happening has. As input source the
gcs_source
is expected, unless specified otherwise. Additionally any input .CSV file
by itself must be 100MB or smaller, unless specified otherwise.
If an "example" file (that is, image, video etc.) with identical content
(even if it had different GCS_FILE_PATH) is mentioned multiple times, then
its label, bounding boxes etc. are appended. The same file should be always
provided with the same ML_USE and GCS_FILE_PATH, if it is not, then
these values are nondeterministically selected from the given ones.
The formats are represented in EBNF with commas being literal and with
non-terminal symbols defined near the end of this comment. The formats are:
For Image Classification:
CSV file(s) with each line in format:
ML_USE,GCS_FILE_PATH,LABEL,LABEL,...
GCS_FILE_PATH leads to image of up to 30MB in size. Supported
extensions: .JPEG, .GIF, .PNG, .WEBP, .BMP, .TIFF, .ICO
For MULTICLASS classification type, at most one LABEL is allowed
per image. If an image has not yet been labeled, then it should be
mentioned just once with no LABEL.
Some sample rows:
TRAIN,gs://folder/image1.jpg,daisy
TEST,gs://folder/image2.jpg,dandelion,tulip,rose
UNASSIGNED,gs://folder/image3.jpg,daisy
UNASSIGNED,gs://folder/image4.jpg
For Image Object Detection:
CSV file(s) with each line in format:
ML_USE,GCS_FILE_PATH,(LABEL,BOUNDING_BOX | ,,,,,,,)
GCS_FILE_PATH leads to image of up to 30MB in size. Supported
extensions: .JPEG, .GIF, .PNG.
Each image is assumed to be exhaustively labeled. The minimum
allowed BOUNDING_BOX edge length is 0.01, and no more than 500
BOUNDING_BOX-es per image are allowed (one BOUNDING_BOX is defined
per line). If an image has not yet been labeled, then it should be
mentioned just once with no LABEL and the ",,,,,,," in place of the
BOUNDING_BOX. For images which are known to not contain any
bounding boxes, they should be labelled explictly as
"NEGATIVE_IMAGE", followed by ",,,,,,," in place of the
BOUNDING_BOX.
Sample rows:
TRAIN,gs://folder/image1.png,car,0.1,0.1,,,0.3,0.3,,
TRAIN,gs://folder/image1.png,bike,.7,.6,,,.8,.9,,
UNASSIGNED,gs://folder/im2.png,car,0.1,0.1,0.2,0.1,0.2,0.3,0.1,0.3
TEST,gs://folder/im3.png,,,,,,,,,
TRAIN,gs://folder/im4.png,NEGATIVE_IMAGE,,,,,,,,,
For Video Classification:
CSV file(s) with each line in format:
ML_USE,GCS_FILE_PATH
where ML_USE VALIDATE value should not be used. The GCS_FILE_PATH
should lead to another .csv file which describes examples that have
given ML_USE, using the following row format:
GCS_FILE_PATH,(LABEL,TIME_SEGMENT_START,TIME_SEGMENT_END | ,,)
Here GCS_FILE_PATH leads to a video of up to 50GB in size and up
to 3h duration. Supported extensions: .MOV, .MPEG4, .MP4, .AVI.
TIME_SEGMENT_START and TIME_SEGMENT_END must be within the
length of the video, and end has to be after the start. Any segment
of a video which has one or more labels on it, is considered a
hard negative for all other labels. Any segment with no labels on
it is considered to be unknown. If a whole video is unknown, then
it shuold be mentioned just once with ",," in place of LABEL,
TIME_SEGMENT_START,TIME_SEGMENT_END.
Sample top level CSV file:
TRAIN,gs://folder/train_videos.csv
TEST,gs://folder/test_videos.csv
UNASSIGNED,gs://folder/other_videos.csv
Sample rows of a CSV file for a particular ML_USE:
gs://folder/video1.avi,car,120,180.000021
gs://folder/video1.avi,bike,150,180.000021
gs://folder/vid2.avi,car,0,60.5
gs://folder/vid3.avi,,,
For Video Object Tracking:
CSV file(s) with each line in format:
ML_USE,GCS_FILE_PATH
where ML_USE VALIDATE value should not be used. The GCS_FILE_PATH
should lead to another .csv file which describes examples that have
given ML_USE, using one of the following row format:
GCS_FILE_PATH,LABEL,[INSTANCE_ID],TIMESTAMP,BOUNDING_BOX
or
GCS_FILE_PATH,,,,,,,,,,
Here GCS_FILE_PATH leads to a video of up to 50GB in size and up
to 3h duration. Supported extensions: .MOV, .MPEG4, .MP4, .AVI.
Providing INSTANCE_IDs can help to obtain a better model. When
a specific labeled entity leaves the video frame, and shows up
afterwards it is not required, albeit preferable, that the same
INSTANCE_ID is given to it.
TIMESTAMP must be within the length of the video, the
BOUNDING_BOX is assumed to be drawn on the closest video's frame
to the TIMESTAMP. Any mentioned by the TIMESTAMP frame is expected
to be exhaustively labeled and no more than 500 BOUNDING_BOX-es per
frame are allowed. If a whole video is unknown, then it should be
mentioned just once with ",,,,,,,,,," in place of LABEL,
[INSTANCE_ID],TIMESTAMP,BOUNDING_BOX.
Sample top level CSV file:
TRAIN,gs://folder/train_videos.csv
TEST,gs://folder/test_videos.csv
UNASSIGNED,gs://folder/other_videos.csv
Seven sample rows of a CSV file for a particular ML_USE:
gs://folder/video1.avi,car,1,12.10,0.8,0.8,0.9,0.8,0.9,0.9,0.8,0.9
gs://folder/video1.avi,car,1,12.90,0.4,0.8,0.5,0.8,0.5,0.9,0.4,0.9
gs://folder/video1.avi,car,2,12.10,.4,.2,.5,.2,.5,.3,.4,.3
gs://folder/video1.avi,car,2,12.90,.8,.2,,,.9,.3,,
gs://folder/video1.avi,bike,,12.50,.45,.45,,,.55,.55,,
gs://folder/video2.avi,car,1,0,.1,.9,,,.9,.1,,
gs://folder/video2.avi,,,,,,,,,,,
For Text Extraction:
CSV file(s) with each line in format:
ML_USE,GCS_FILE_PATH
GCS_FILE_PATH leads to a .JSONL (that is, JSON Lines) file which
either imports text in-line or as documents. Any given
.JSONL file must be 100MB or smaller.
The in-line .JSONL file contains, per line, a proto that wraps a
TextSnippet proto (in json representation) followed by one or more
AnnotationPayload protos (called annotations), which have
display_name and text_extraction detail populated. The given text
is expected to be annotated exhaustively, for example, if you look
for animals and text contains "dolphin" that is not labeled, then
"dolphin" is assumed to not be an animal. Any given text snippet
content must be 10KB or smaller, and also be UTF-8 NFC encoded
(ASCII already is).
The document .JSONL file contains, per line, a proto that wraps a
Document proto. The Document proto must have either document_text
or input_config set. In document_text case, the Document proto may
also contain the spatial information of the document, including
layout, document dimension and page number. In input_config case,
only PDF documents are supported now, and each document may be up
to 2MB large. Currently, annotations on documents cannot be
specified at import.
Three sample CSV rows:
TRAIN,gs://folder/file1.jsonl
VALIDATE,gs://folder/file2.jsonl
TEST,gs://folder/file3.jsonl
Sample in-line JSON Lines file for entity extraction (presented here
with artificial line breaks, but the only actual line break is
denoted by \n).:
{
"document": {
"document_text": {"content": "dog cat"}
"layout": [
{
"text_segment": {
"start_offset": 0,
"end_offset": 3,
},
"page_number": 1,
"bounding_poly": {
"normalized_vertices": [
{"x": 0.1, "y": 0.1},
{"x": 0.1, "y": 0.3},
{"x": 0.3, "y": 0.3},
{"x": 0.3, "y": 0.1},
],
},
"text_segment_type": TOKEN,
},
{
"text_segment": {
"start_offset": 4,
"end_offset": 7,
},
"page_number": 1,
"bounding_poly": {
"normalized_vertices": [
{"x": 0.4, "y": 0.1},
{"x": 0.4, "y": 0.3},
{"x": 0.8, "y": 0.3},
{"x": 0.8, "y": 0.1},
],
},
"text_segment_type": TOKEN,
}
],
"document_dimensions": {
"width": 8.27,
"height": 11.69,
"unit": INCH,
}
"page_count": 1,
},
"annotations": [
{
"display_name": "animal",
"text_extraction": {"text_segment": {"start_offset": 0,
"end_offset": 3}}
},
{
"display_name": "animal",
"text_extraction": {"text_segment": {"start_offset": 4,
"end_offset": 7}}
}
],
}\n
{
"text_snippet": {
"content": "This dog is good."
},
"annotations": [
{
"display_name": "animal",
"text_extraction": {
"text_segment": {"start_offset": 5, "end_offset": 8}
}
}
]
}
Sample document JSON Lines file (presented here with artificial line
breaks, but the only actual line break is denoted by \n).:
{
"document": {
"input_config": {
"gcs_source": { "input_uris": [ "gs://folder/document1.pdf" ]
}
}
}
}\n
{
"document": {
"input_config": {
"gcs_source": { "input_uris": [ "gs://folder/document2.pdf" ]
}
}
}
}
For Text Classification:
CSV file(s) with each line in format:
ML_USE,(TEXT_SNIPPET | GCS_FILE_PATH),LABEL,LABEL,...
TEXT_SNIPPET and GCS_FILE_PATH are distinguished by a pattern. If
the column content is a valid gcs file path, i.e. prefixed by
"gs://", it will be treated as a GCS_FILE_PATH, else if the content
is enclosed within double quotes (""), it is
treated as a TEXT_SNIPPET. In the GCS_FILE_PATH case, the path
must lead to a .txt file with UTF-8 encoding, for example,
"gs://folder/content.txt", and the content in it is extracted
as a text snippet. In TEXT_SNIPPET case, the column content
excluding quotes is treated as to be imported text snippet. In
both cases, the text snippet/file size must be within 128kB.
Maximum 100 unique labels are allowed per CSV row.
Sample rows:
TRAIN,"They have bad food and very rude",RudeService,BadFood
TRAIN,gs://folder/content.txt,SlowService
TEST,"Typically always bad service there.",RudeService
VALIDATE,"Stomach ache to go.",BadFood
For Text Sentiment:
CSV file(s) with each line in format:
ML_USE,(TEXT_SNIPPET | GCS_FILE_PATH),SENTIMENT
TEXT_SNIPPET and GCS_FILE_PATH are distinguished by a pattern. If
the column content is a valid gcs file path, that is, prefixed by
"gs://", it is treated as a GCS_FILE_PATH, otherwise it is treated
as a TEXT_SNIPPET. In the GCS_FILE_PATH case, the path
must lead to a .txt file with UTF-8 encoding, for example,
"gs://folder/content.txt", and the content in it is extracted
as a text snippet. In TEXT_SNIPPET case, the column content itself
is treated as to be imported text snippet. In both cases, the
text snippet must be up to 500 characters long.
Sample rows:
TRAIN,"@freewrytin this is way too good for your product",2
TRAIN,"I need this product so bad",3
TEST,"Thank you for this product.",4
VALIDATE,gs://folder/content.txt,2
- For Tables:
Either
gcs_source or
bigquery_source
can be used. All inputs is concatenated into a single
primary_table
For gcs_source:
CSV file(s), where the first row of the first file is the header,
containing unique column names. If the first row of a subsequent
file is the same as the header, then it is also treated as a
header. All other rows contain values for the corresponding
columns.
Each .CSV file by itself must be 10GB or smaller, and their total
size must be 100GB or smaller.
First three sample rows of a CSV file:
"Id","First Name","Last Name","Dob","Addresses"
"1","John","Doe","1968-01-22","[{"status":"current","address":"123_First_Avenue","city":"Seattle","state":"WA","zip":"11111","numberOfYears":"1"},{"status":"previous","address":"456_Main_Street","city":"Portland","state":"OR","zip":"22222","numberOfYears":"5"}]"
"2","Jane","Doe","1980-10-16","[{"status":"current","address":"789_Any_Avenue","city":"Albany","state":"NY","zip":"33333","numberOfYears":"2"},{"status":"previous","address":"321_Main_Street","city":"Hoboken","state":"NJ","zip":"44444","numberOfYears":"3"}]}
For bigquery_source:
An URI of a BigQuery table. The user data size of the BigQuery
table must be 100GB or smaller.
An imported table must have between 2 and 1,000 columns, inclusive,
and between 1000 and 100,000,000 rows, inclusive. There are at most 5
import data running in parallel.
Definitions:
ML_USE = "TRAIN" | "VALIDATE" | "TEST" | "UNASSIGNED"
Describes how the given example (file) should be used for model
training. "UNASSIGNED" can be used when user has no preference.
GCS_FILE_PATH = A path to file on GCS, e.g. "gs://folder/image1.png".
LABEL = A display name of an object on an image, video etc., e.g. "dog".
Must be up to 32 characters long and can consist only of ASCII
Latin letters A-Z and a-z, underscores(_), and ASCII digits 0-9.
For each label an AnnotationSpec is created which display_name
becomes the label; AnnotationSpecs are given back in predictions.
INSTANCE_ID = A positive integer that identifies a specific instance of a
labeled entity on an example. Used e.g. to track two cars on
a video while being able to tell apart which one is which.
BOUNDING_BOX = VERTEX,VERTEX,VERTEX,VERTEX | VERTEX,,,VERTEX,,
A rectangle parallel to the frame of the example (image,
video). If 4 vertices are given they are connected by edges
in the order provided, if 2 are given they are recognized
as diagonally opposite vertices of the rectangle.
VERTEX = COORDINATE,COORDINATE
First coordinate is horizontal (x), the second is vertical (y).
COORDINATE = A float in 0 to 1 range, relative to total length of
image or video in given dimension. For fractions the
leading non-decimal 0 can be omitted (i.e. 0.3 = .3).
Point 0,0 is in top left.
TIME_SEGMENT_START = TIME_OFFSET
Expresses a beginning, inclusive, of a time segment
within an example that has a time dimension
(e.g. video).
TIME_SEGMENT_END = TIME_OFFSET
Expresses an end, exclusive, of a time segment within
an example that has a time dimension (e.g. video).
TIME_OFFSET = A number of seconds as measured from the start of an
example (e.g. video). Fractions are allowed, up to a
microsecond precision. "inf" is allowed, and it means the end
of the example.
TEXT_SNIPPET = A content of a text snippet, UTF-8 encoded, enclosed within
double quotes ("").
SENTIMENT = An integer between 0 and
Dataset.text_sentiment_dataset_metadata.sentiment_max
(inclusive). Describes the ordinal of the sentiment - higher
value means a more positive sentiment. All the values are
completely relative, i.e. neither 0 needs to mean a negative or
neutral sentiment nor sentiment_max needs to mean a positive one
- it is just required that 0 is the least positive sentiment
in the data, and sentiment_max is the most positive one.
The SENTIMENT shouldn't be confused with "score" or "magnitude"
from the previous Natural Language Sentiment Analysis API.
All SENTIMENT values between 0 and sentiment_max must be
represented in the imported data. On prediction the same 0 to
sentiment_max range will be used. The difference between
neighboring sentiment values needs not to be uniform, e.g. 1 and
2 may be similar whereas the difference between 2 and 3 may be
huge.
Errors:
If any of the provided CSV files can't be parsed or if more than certain
percent of CSV rows cannot be processed then the operation fails and
nothing is imported. Regardless of overall success or failure the per-row
failures, up to a certain count cap, is listed in
Operation.metadata.partial_failures.
