Version mismatch at installation & encoding issue

[Krak91]

I get the following error when trying to ask the model a question. My pipeline looks like this:

def pipeline(docs, questions):

document_store = InMemoryDocumentStore()
document_store.write_documents(docs)
retriever = TfidfRetriever(document_store=document_store)
reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2", use_gpu=False)
finder = Finder(reader, retriever)
predictions = {}
for question in questions:
	predictions[question] = finder.get_answers(question=question, top_k_retriever=5, top_k_reader=3)

return predictions

Error message
File ".../model.py", line 23, in pipeline
predictions[question] = finder.get_answers(question=question, top_k_retriever=5, top_k_reader=3)
File "...venv\lib\site-packages\haystack\finder.py", line 57, in get_answers
top_k=top_k_reader) # type: Dict[str, Any]
File "...\venv\lib\site-packages\haystack\reader\farm.py", line 259, in predict
if self._check_no_answer(ans):
File "...\venv\lib\site-packages\haystack\reader\farm.py", line 416, in _check_no_answer
assert d["answer"] == "is_impossible", f"Check for no answer is not working"
AssertionError: Check for no answer is not working

Expected behavior
Return model predictions

┆Issue is synchronized with this Jira Task by Unito

[anirbansaha96]

Did you try doing this:

for question in questions:
	prediction = finder.get_answers(question=question, top_k_retriever=5, top_k_reader=3)
        predictions.append((question,prediction))

[Krak91]

yes, the error is still the same at the "prediction = finder.get_answers(question=question, top_k_retriever=5, top_k_reader=3)"
argument

[tholor]

Hey @Krak91 ,

What versions of Haystack and FARM are you using?

The error makes me believe that you are using Haystack from latest master + FARM 0.4.6.
These two versions are not compatible yet (we'll update to FARM 0.4.6 shortly in #172 ).

To resolve this you can do one of these options:

• Last stable release: pip install haystack==0.2.1
• Latest code:

git clone https://github.com/deepset-ai/haystack.git
cd haystack
pip install -e .

(which will install FARM 0.4.5)

Let me know if this resolves your issue!

[Krak91]

Hi, I've installed the right versions of farm and haystack and it worked. Thanks.

Some more information:
I've noticed that all the scripts are in the 'haystack' folder of the project. As I'm just trying to add its functionality to my existing project, I simply move that folder to my virtual environment so that the imports in the scripts say valid. A few other things I've noticed:

• Always during installation (either by cloning or pip) it shows this error: "ERROR: Could not find a version that satisfies the requirement torch==1.4.0 (from farm==0.4.3->farm-haystack) (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)
  ERROR: No matching distribution found for torch==1.4.0 (from farm==0.4.3->farm-haystack)'. Versions vary depending on release. Need to go install torch manually.
• I always need to modify haystack/indexing/utils line 36 whenever I install the package to avoid errors when trying to open documents with funny characters (encoding='x', errors='ignore'). I like the new functions for passing the document store raw text (dicts of title, text) rather than a physical file's path.

[tholor]

Always during installation (either by cloning or pip) it shows this error: "ERROR: Could not find a version that satisfies the requirement torch==1.4.0 (from farm==0.4.3->farm-haystack) (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)
ERROR: No matching distribution found for torch==1.4.0 (from farm==0.4.3->farm-haystack)'. Versions vary depending on release. Need to go install torch manually.

This should be resolved in the latest version as we upgraded to torch 1.5.1

I always need to modify haystack/indexing/utils line 36 whenever I install the package to avoid errors when trying to open documents with funny characters (encoding='x', errors='ignore'). I like the new functions for passing the document store raw text (dicts of title, text) rather than a physical file's path.

Can you please elaborate on what you are exactly changing there? Are you using utf-8 encoding there? What errors come up that you want to ignore?
As encodings are always a big pain, we could think of adding an arg to convert_files_to_dicts() to specify the encoding or at least put a better default there ...

[Krak91]

I'm changing with open(path) as doc: to with open(path, encoding='utf-8', errors='ignore') as doc:
It's a unicode error:
UnicodeDecodeError: 'charmap' codec can't decode byte 0x90 in position 24072: character maps to <undefined>
The actual character in this case is: ‐

[tanaysoni]

Hi @Krak91, this issue should be resolved by #478. I'm closing the thread but please feel free to update here if you still face any encoding errors.