I trained a BERT based encoder decoder model (EncoderDecoderModel) named ed_model with HuggingFace's transformers module.
I used the BertTokenizer named as input_tokenizer
I tokenized the input with:
txt = "Some wonderful sentence to encode"
inputs = input_tokenizer(txt, return_tensors="pt").to(device)
print(inputs)
The output clearly shows that a input_ids is the return dict
{'input_ids': tensor([[ 101, 5660, 7975, 2127, 2053, 2936, 5061, 102]], device='cuda:0'), 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0]], device='cuda:0'), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1]], device='cuda:0')}
But when I try to predict, I get this error:
ed_model.forward(**inputs)
ValueError: You have to specify either input_ids or inputs_embeds
Any ideas ?
Well, apparently this is a known issue, for example: This issue of T5
The problem is that there's probably a renaming procedure in the code, since we use a encoder-decoder architecture we have 2 types of input ids.
The solution is to explicitly specify the type of input id
ed_model.forward(decoder_input_ids=inputs['input_ids'],**inputs)
I wish it was documented somewhere, but now you know :-)
Related
I am very new to Julia Lang (in fact, just trying it instead of Python for some data analysis). However, I am stuck when loading my data.
My data is from a web-application built using ReactJS/ Python, saved in a csv. I get the data into a Julia DataFrame. The cell in this DataFrame that I need to analyse looks like this:
{'isClicked': [True, True, True, True, True, True, True, True, True, True, True, True, True, True, True], 'continuation': [100, 100, 100, 100, 100, 0, 100, 100, 100, 0, 100, 100, 0, 0, 0, 0, 0, 0, 0, 0, 0]}
This comes from a JS-dictionary. Is there a way to convert it into a dictionary in Julia? I have tried a JSON3 converter (https://discourse.julialang.org/t/why-does-julia-not-support-json-syntax-to-create-a-dict/42873/20), but it seems not to work because of the single quotation mark. I.e., the error I get is:
ArgumentError: invalid JSON at byte position 2 while parsing type
JSON3.Object: ExpectedOpeningQuoteChar {'isClicked': [True, True,
Any suggestion are highly appreciated!
Thanks!
JSON requires double quotes instead of single quotes. Try
replace(text, "'" => "\"")
before sending it to the JSON parser.
That's not a JSON dictionary, it's a python one. In Python you should do
import json
with open('dict_file.json', 'w') as f:
json.dump(my_py_dict, f)
Then in Julia
import JSON
my_julia_dict = JSON.parsefile("dict_file.json")
I have a list of NetCDF files that I would like to open with the xarray.open_mfdataset function.
This would normally be trivial, however I am running into an issue because the files I cam trying to open do not have any "time" dimension included in them:
data
Out[51]:
<xarray.Dataset>
Dimensions: (lat: 850, lon: 1500)
Coordinates:
* lat (lat) float64 54.98 54.94 54.9 54.86 ... 21.14 21.1 21.06 21.02
* lon (lon) float64 -126.0 -125.9 -125.9 -125.9 ... -66.1 -66.06 -66.02
Data variables:
Data (lat, lon) float32 ...
When I try to open my list of files with open_mfdataset, I of course get an error:
xr.open_mfdataset(files)
ValueError: Could not find any dimension coordinates to use to order the datasets for concatenation
I however do have a list of dates corresponding to each file:
dates
Out[54]:
array([datetime.datetime(2009, 1, 1, 0, 0),
datetime.datetime(2009, 1, 2, 0, 0),
datetime.datetime(2009, 1, 3, 0, 0), ...,
datetime.datetime(2019, 12, 29, 0, 0),
datetime.datetime(2019, 12, 30, 0, 0),
datetime.datetime(2019, 12, 31, 0, 0)], dtype=object)
I assume there is some way I add a time dimension to each file and open them all with open_mfdataset, possibly with the "preprocess" argument.
Thanks for any help.
Here is my solution:
Create a function which adds a time dimension to a DataArray, and fill it with a arbitrary date:
def add_time_dim(xda):
xda = xda.expand_dims(time = [datetime.now()])
return xda
Then, pass this function to the preprocess argument when running the open_mfdataset functions:
data = xr.open_mfdataset(files, preprocess = add_time_dim)
Finally, fill the time dimension with my dates:
data['time'] = dates
After training a graph with metrics ops (such as accuracy from tf.python.ops.metrics), I tried to restore the graph and evaluate the accuracy on the test set. However, after restoring the graph with tf.import_meta_graph, when I tried to initialize the local variables (it is necessary) with tf.local_variables_initializer(), I got an error, it said 'Tensor' object has no attribute 'initializer'.
If I print the local variables after restoring, there are two Tensorflow Tensors which may cause the problem.
These two tensorlow Tensors stem from the accuracy metrics:
<tf.Tensor 'accuracy/total:0' shape=() dtype=float32_ref>
<tf.Tensor 'accuracy/count:0' shape=() dtype=float32_ref>
Can someone help me with this? Thank you!
Similar code:
def train():
l_ini = np.array([1, 0, 1, 0, 1, 0], dtype=np.float32)
p_ini = np.array([1, 0, 1, 0, 1, 1], dtype=np.float32)
l = tf.Variable(l_ini, trainable=False)
p = tf.Variable(p_ini, trainable=False)
accuracy = metrics.accuracy(labels=l, predictions=p)
tf.add_to_collection("accuracy", accuracy)
graph = tf.get_default_graph()
sess = tf.Session(graph=graph)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
acc = sess.run(accuracy)
saver = tf.train.Saver()
saver.save(sess, 'test.ckpt')
def restore():
with tf.Session() as sess:
loader = tf.train.import_meta_graph('./test.ckpt.meta')
loader.restore(sess, './test.ckpt')
accuracy = tf.get_collection("accuracy")
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
acc = sess.run(accuracy)
I have a workaround, instead of retrieving the accuracy collection (the get_collection returned an empty list in my case):
Retrieve the logits and label placeholders.
Then compute the accuracy.
Remember to initialize the local running variables after restoring to session as well:
self.running_vars = tf.get_collection(tf.GraphKeys.LOCAL_VARIABLES, scope="your_accuracy_scope_name")
I am new to JavaFX and ControlsFX.
I am trying to create a very basic SpreadsheetView using the ControlsFX library. Following is the function to populate and create the SpreadsheetView:
private fun spreadSheetFunc() : SpreadsheetView {
val rowCount = 15
val columnCount = 10
val grid = GridBase(rowCount, columnCount)
val rows = FXCollections.observableArrayList<ObservableList<SpreadsheetCell>>()
var list = FXCollections.observableArrayList<SpreadsheetCell>()
list.add(SpreadsheetCellType.STRING.createCell(0, 0, 1, 1, "row0-col0"))
list.add(SpreadsheetCellType.STRING.createCell(0, 1, 2, 1, "row0-col1"))
list.add(SpreadsheetCellType.STRING.createCell(0, 2, 1, 1, "row0-col2"))
rows.add(list)
list = FXCollections.observableArrayList()
list.add(SpreadsheetCellType.STRING.createCell(1, 0, 1, 1, "row1-col0"))
//commenting row1-col1 as row0-col1 has a rowspan of 2
//list.add(SpreadsheetCellType.STRING.createCell(1, 1, 1, 1, "row1-col1"))
list.add(SpreadsheetCellType.STRING.createCell(1, 2, 1, 1, "row1-col2"))
rows.add(list)
list = FXCollections.observableArrayList()
list.add(SpreadsheetCellType.STRING.createCell(2, 0, 1, 1, "row2-col0"))
list.add(SpreadsheetCellType.STRING.createCell(2, 1, 1, 1, "row2-col1"))
list.add(SpreadsheetCellType.STRING.createCell(2, 2, 1, 1, "row2-col2"))
rows.add(list)
list = FXCollections.observableArrayList()
list.add(SpreadsheetCellType.STRING.createCell(3, 0, 1, 1, "row3-col0"))
list.add(SpreadsheetCellType.STRING.createCell(3, 1, 1, 1, "row3-col1"))
list.add(SpreadsheetCellType.STRING.createCell(3, 2, 1, 1, "row3-col2"))
rows.add(list)
grid.setRows(rows)
return SpreadsheetView(grid)
}
On running it, I get the following error:
java.lang.IndexOutOfBoundsException: Index: 2, Size: 2 at
java.util.ArrayList.rangeCheck(ArrayList.java:653)
I know its happening because I am not adding any value for rowIndex=1 colIndex=1 (see the commented out line) ... but that is what I want.
The row0-col1 has a rowspan of 2 which should mean that even if my row1-col1 is absent, there shouldn't be any problem.
Why doesn't ControlsFX automatically take care of this?
If I uncomment that line, I get the following output:
Edit 1:
Also, I found another issue, when a colspan/rowspan occupies the whole column/row in the SpreadsheetView and then when one presses arrow key to navigate to cells you get an error:
The above situation arises when you press the right arrow key (Even though their isn't a cell on the right)
Let me apologize because it is not well documented how span must be made in the SpreadsheetView. I will update the documentation.
If you want to span, you have to put the same cell in every cell inside the span.
So either you build your own cell, and then in every place. In your case, you would add the same cell in row 0 column 1 and in row 1 column 1.
Or you could keep your code, and simply call the method spanRow on the Grid. This method will automatically take your cell and place it accordingly.
Regarding the second issue, please submit it to our issue tracker so we can fix it : https://bitbucket.org/controlsfx/controlsfx/issues?status=new&status=open
If you have other issue regarding the SpreadsheetView, consider posting in our Google group where we will get notifications : http://groups.controlsfx.org
When I enter the following commands directly into the R console
library("xts")
mySeries <- xts(c(1.0, 2.0, 3.0, 5.0, 6.0), order.by=c(ISOdatetime(2001, 1, 1, 0, 0, 0), ISOdatetime(2001, 1, 2, 0, 0, 0), ISOdatetime(2001, 1, 3, 0, 0, 0), ISOdatetime(2001, 1, 4, 0, 0, 0), ISOdatetime(2001, 1, 5, 0, 0, 0)))
resultingSeries <- to.monthly(mySeries)
resultingSeries
I will get an output like this
mySeries.Open mySeries.High mySeries.Low mySeries.Close
Jan 2001 1 6 1 6
When I look into the attributes, I see the following output
attributes(resultingSeries)
$dim
[1] 1 4
$dimnames
$dimnames[[1]]
NULL
$dimnames[[2]]
[1] "mySeries.Open" "mySeries.High" "mySeries.Low" "mySeries.Close"
$index
[1] 978307200
attr(,"tclass")
[1] "yearmon"
$tclass
[1] "POSIXct" "POSIXt"
$tzone
[1] ""
$class
[1] "xts" "zoo"
$.indexCLASS
[1] "yearmon"
This is the same I get in Java. I'm wondering where the magic happens so that I see the nice output I get in R. I have no access to the event loop, since I'm using JRI like this (since, it's the recommended way and simplifies error handling):
REngine engine = REngine.engineForClass("org.rosuda.REngine.JRI.JRIEngine");
REXP result = engine.parseAndEval(...)
/edit
In Java I execute each command from above as follows:
REXP result = engine.parseAndEval("resultingSeries") // or any other command
What I get is
org.rosuda.REngine.REXPDouble#4ac66122+[12]
The payload being doubles: 1, 6, 1, 6
The attributes are the same as specified above.
Now R does some magic to display the output above. Is there a way I can get the same output without having to create it manually by myself? Where's the implementation stored, that R gets the above mentioned output?
Here is a piece of code that will work, here i extracted the first element of the field mySeries.Open from the object resultingSeries (which i converted to a data frame) which is equal to 1, notice that you can't pass all of the resultingSeries object strait into Java, you will need to break it down.
package stackoverflow;
import org.rosuda.JRI.REXP;
import org.rosuda.JRI.Rengine;
/**
*
* #author yschellekens
*/
public class StackOverflow {
public static void main(String[] args) throws Exception {
String[] Rargs = {"--vanilla"};
Rengine rengine = new Rengine( Rargs, false, null);
rengine.eval("library('xts')");
rengine.eval("mySeries <- xts(c(1.0, 2.0, 3.0, 5.0, 6.0), order.by=c(ISOdatetime(2001, 1, 1, 0, 0, 0), ISOdatetime(2001, 1, 2, 0, 0, 0), ISOdatetime(2001, 1, 3, 0, 0, 0), ISOdatetime(2001, 1, 4, 0, 0, 0), ISOdatetime(2001, 1, 5, 0, 0, 0)))");
rengine.eval("resultingSeries <- to.monthly(mySeries)");
rengine.eval("resultingSeries<-as.data.frame(resultingSeries)");
REXP result= rengine.eval("resultingSeries$mySeries.Open");
System.out.println("Greeting from R: "+result.asDouble());
}
}
And the Java output:
run:
Greeting from R: 1.0
I figured out the following workaround. The solution is far from perfect.
R offers a command to save its console output as characters vector.
capture.output( {command} )
We can access the output using
REXPString s = rengine.parseAndEval("capture.output( to.monthly(mySeries))")
String[] output = result.asStrings()
The variable output will contain all output lines
[0] mySeries.Open mySeries.High mySeries.Low mySeries.Close
[1]Jan 2001 1 6 1 6
Alternatively you coud use JRIEngine and attack yourself to the event loop, which it did not want in my case (due to the more complicated error handling).