Downloading of sentence transformers

[MNLubov]

Basic downloading doesn't work for sentence-transformers. Is it possible to download model using direct url or url for config.json?

[chengchingwen]

Which code are you testing with?

[MNLubov]

I tested hgf"model_name:model_type" approach described in README.
Also I tried another approach

model_name = "all-MiniLM-L6-v2"
file_name = "config.json"
path = "d:/models/sentence transformers/all-MiniLM-L6-v2"
file_hash = Transformers.HuggingFace.find_or_register_hgf_file_hash(path, model_name, file_name)

Which allows to register config file. Then

cfg = load_config(model_name)
model_cons = Transformers.HuggingFace.get_model_type((Val ∘ Symbol ∘ lowercase)(model_type), (Val ∘ Symbol ∘ lowercase)(item))
model = load_model(model_cons, model_name; config=cfg)

Which gives me following error: Model Val{:model_task}() doesn't support this kind of task: Val{:model_task}(). I tried different model tasks. I also used "BertModel" as a task, which is given in architectures field of config file.
As far as I understand there is no model types/model tasks for sentence transformers

[chengchingwen]

Generally, It might be better to wait for #103.

Which gives me following error: Model Val{:model_task}() doesn't support this kind of task: Val{:model_task}(). I tried different model tasks. I also used "BertModel" as a task, which is given in architectures field of config file.

For "BertModel", you would need to pass Val(:model) to the function, or one of the following:

julia> map(first, HuggingFace.get_model_type(:bert))
(:model, :forpretraining, :lmheadmodel, :formaskedlm, :fornextsentenceprediction, :forsequenceclassification, :formultiplechoice, :fortokenclassification, :forquestionanswering)

[chengchingwen]

It is now merged and released in v0.1.19, please give it a try.

Let me know if there are any problems, feature requests, or any inconvenience when using that.

[MNLubov]

Hi, Peter. I've tested downloading sentence transformers. Approach described above works fine with BERT-like sentence-transformers. In this approach I download and register config file and model:

path = "d:/models/sentence transformers/all-MiniLM-L6-v2"
model_name = "all-MiniLM-L6-v2" 
config_file = "config.json"  
model_file = "pytorch_model.bin"

config_hash = Transformers.HuggingFace.find_or_register_hgf_file_hash(path, model_name, config_file)
model_hash = Transformers.HuggingFace.find_or_register_hgf_file_hash(path, model_name, model_file)

cfg = Transformers.HuggingFace.load_config(model_name)
model_cons = Transformers.HuggingFace.HGFBertModel
model = Transformers.HuggingFace.load_model(model_cons, model_name; config=cfg)

Using macro doesn't work for sentence transformers, only for "token" transformers. Macro works only if I downloaded and register config file and model file. If I use macro without pre-download sentence-transformer:

model_hfapi = hgf"all-distilroberta-v1:model"

I get:

 Info: No local config.json found. downloading...
ERROR: HTTP.Exceptions.StatusError(404, "GET", "/all-distilroberta-v1-config.json", HTTP.Messages.Response:
"""
HTTP/1.1 404 Not Found
Content-Type: application/xml
Transfer-Encoding: chunked
Connection: keep-alive
Date: Tue, 23 Aug 2022 14:18:41 GMT
Server: AmazonS3
X-Cache: Error from cloudfront
Via: 1.1 bccded73b8b9a1d038e5d874cf586402.cloudfront.net (CloudFront)
X-Amz-Cf-Pop: HEL50-C1
X-Amz-Cf-Id: dldXOXWkBtTB714zBVOMwiL4xIDRiGrjzc5BB46_yPYXker_lPo4dA== """)

In the case of using macro with adding 'sentence-transformers' prefix to model name:

[ Info: No local config.json found. downloading...
ERROR: HTTP.Exceptions.ConnectError("https://cdn.huggingface.co\\sentence-transformers/all-distilroberta-v1\\config.json", DNSError: cdn.huggingface.co\sentence-transformers, unknown node or service (EAI_NONAME))

[chengchingwen]

You seems to be using the old version, please update to the latest one. And it should be used like this:

using Transformers
using Transformers.Basic
using Transformers.HuggingFace

textenc = hgf"sentence-transformers/all-MiniLM-L6-v2:tokenizer"
model = hgf"sentence-transformers/all-MiniLM-L6-v2:model"

sentences = ["This is an example sentence", "Each sentence is converted"]
a = encode(textenc, sentences)
model_outputs = model(a.input.tok; token_type_ids = a.input.segment, attention_mask = a.mask) 

[MNLubov]

@chengchingwen, How could I load HuggingFace model from file stored locally? Is loading from HuggingFace portal and then saving it as .bson file the best way to do it?

[chengchingwen]

It really depends on the scenario. Would need to expand on that (e.g. what do you have and what do you want to achieve).