Bump to Torch 1.7 and Transformers 4.1.1 (#665)

* bump torch version

* update to transformers 4.4.1

* move CI from azure to github actions

* modify and fix tests

Co-authored-by: Timo Moeller <timo.moeller@deepset.ai>

.github/workflows/ci.yml

@@ -0,0 +1,34 @@
+name: Build
+
+on:
+  push:
+    branches: [ master ]
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  build:
+
+    runs-on: ubuntu-20.04
+
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: Set up Python 3.8
+      uses: actions/setup-python@v2
+      with:
+        python-version: 3.8
+
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install pytest
+        pip install -r requirements.txt
+        pip install onnxruntime
+        pip install -e .
+
+    - name: Run pytest - only "conversion" marker
+      run: cd test && pytest -m "conversion"
+
+    - name: Run Pytest - all except conversion marker
+      run: cd test &&  pytest -m "not conversion"

farm/data_handler/processor.py

@@ -13,7 +13,7 @@
 import torch
 from numpy.random import random as random_float
 from sklearn.preprocessing import StandardScaler
-from transformers.configuration_auto import AutoConfig
+from transformers import AutoConfig
 from tokenizers import Encoding
 
 from farm.data_handler.dataset import convert_features_to_dataset
@@ -111,8 +111,8 @@ def __init__(
                     Format as in  "requests" library: https://2.python-requests.org//en/latest/user/advanced/#proxies
         :type proxies: dict
         :param multithreading_rust: Whether to allow multithreading in Rust, e.g. for FastTokenizers.
-                                    Note: Enabling multithreading in Rust AND multiprocessing in python can cause
-                                    trouble incl. deadlocks.
+                                    Note: Enabling multithreading in Rust AND multiprocessing in python might cause
+                                    deadlocks.
         :type multithreading_rust: bool
         """
         if not multithreading_rust:

farm/data_handler/samples.py

@@ -1,4 +1,4 @@
-from transformers.tokenization_bert import whitespace_tokenize
+from transformers.models.bert.tokenization_bert import whitespace_tokenize
 from farm.visual.ascii.images import SAMPLE
 import numpy as np
 

farm/infer.py

@@ -170,6 +170,7 @@ def load(
         tokenizer_class=None,
         use_fast=True,
         tokenizer_args=None,
+        multithreading_rust=True,
         dummy_ph=False,
         benchmarking=False,
     ):
@@ -218,13 +219,17 @@ def load(
         :type disable_tqdm: bool
         :param tokenizer_class: (Optional) Name of the tokenizer class to load (e.g. `BertTokenizer`)
         :type tokenizer_class: str
-        :param use_fast: (Optional, False by default) Indicate if FARM should try to load the fast version of the tokenizer (True) or
+        :param use_fast: (Optional, True by default) Indicate if FARM should try to load the fast version of the tokenizer (True) or
             use the Python one (False).
+        :type use_fast: bool
         :param tokenizer_args: (Optional) Will be passed to the Tokenizer ``__init__`` method.
             See https://huggingface.co/transformers/main_classes/tokenizer.html and detailed tokenizer documentation
             on `Hugging Face Transformers <https://huggingface.co/transformers/>`_.
         :type tokenizer_args: dict
-        :type use_fast: bool
+        :param multithreading_rust: Whether to allow multithreading in Rust, e.g. for FastTokenizers.
+                                    Note: Enabling multithreading in Rust AND multiprocessing in python might cause
+                                    deadlocks.
+        :type multithreading_rust: bool
         :param dummy_ph: If True, methods of the prediction head will be replaced
                              with a dummy method. This is used to isolate lm run time from ph run time.
         :type dummy_ph: bool
@@ -250,14 +255,6 @@ def load(
             else:
                 processor = Processor.load_from_dir(model_name_or_path)
 
-            # override processor attributes loaded from config file with inferencer params
-            processor.max_seq_len = max_seq_len
-            if hasattr(processor, "doc_stride"):
-                assert doc_stride < max_seq_len, "doc_stride is longer than max_seq_len. This means that there will be gaps " \
-                                                 "as the passage windows slide, causing the model to skip over parts of the document. "\
-                                                 "Please set a lower value for doc_stride (Suggestions: doc_stride=128, max_seq_len=384) "
-                processor.doc_stride = doc_stride
-
         # b) or from remote transformers model hub
         else:
             if not task_type:
@@ -278,8 +275,15 @@ def load(
                                                             tokenizer_args=tokenizer_args,
                                                             use_fast=use_fast)
 
-        if not isinstance(model,ONNXAdaptiveModel):
-            model, _ = optimize_model(model=model, device=device, local_rank=-1, optimizer=None)
+        # override processor attributes loaded from config or HF with inferencer params
+        processor.max_seq_len = max_seq_len
+        processor.multithreading_rust = multithreading_rust
+        if hasattr(processor, "doc_stride"):
+            assert doc_stride < max_seq_len, "doc_stride is longer than max_seq_len. This means that there will be gaps " \
+                                             "as the passage windows slide, causing the model to skip over parts of the document. " \
+                                             "Please set a lower value for doc_stride (Suggestions: doc_stride=128, max_seq_len=384) "
+            processor.doc_stride = doc_stride
+
         return cls(
             model,
             processor,

farm/modeling/adaptive_model.py

@@ -8,7 +8,7 @@
 import numpy
 import torch
 from torch import nn
-from transformers.configuration_auto import AutoConfig
+from transformers import AutoConfig
 from transformers.convert_graph_to_onnx import convert, quantize as quantize_model
 
 
@@ -455,11 +455,11 @@ def forward_lm(self, **kwargs):
 
         # Run forward pass of language model
         if extraction_layer == -1:
-            sequence_output, pooled_output = self.language_model(**kwargs, output_all_encoded_layers=False)
+            sequence_output, pooled_output = self.language_model(**kwargs, return_dict=False, output_all_encoded_layers=False)
         else:
             # get output from an earlier layer
             self.language_model.enable_hidden_states_output()
-            sequence_output, pooled_output, all_hidden_states = self.language_model(**kwargs)
+            sequence_output, pooled_output, all_hidden_states = self.language_model(**kwargs, return_dict=False)
             sequence_output = all_hidden_states[extraction_layer]
             pooled_output = None #not available in earlier layers
             self.language_model.disable_hidden_states_output()

farm/modeling/language_model.py

@@ -34,17 +34,20 @@
 
 logger = logging.getLogger(__name__)
 
-from transformers.modeling_bert import BertModel, BertConfig
-from transformers.modeling_roberta import RobertaModel, RobertaConfig
-from transformers.modeling_xlnet import XLNetModel, XLNetConfig
-from transformers.modeling_albert import AlbertModel, AlbertConfig
-from transformers.modeling_xlm_roberta import XLMRobertaModel, XLMRobertaConfig
-from transformers.modeling_distilbert import DistilBertModel, DistilBertConfig
-from transformers.modeling_electra import ElectraModel, ElectraConfig
-from transformers.modeling_camembert import CamembertModel, CamembertConfig
-from transformers.modeling_auto import AutoModel, AutoConfig
+from transformers import (
+    BertModel, BertConfig,
+    RobertaModel, RobertaConfig,
+    XLNetModel, XLNetConfig,
+    AlbertModel, AlbertConfig,
+    XLMRobertaModel, XLMRobertaConfig,
+    DistilBertModel, DistilBertConfig,
+    ElectraModel, ElectraConfig,
+    CamembertModel, CamembertConfig
+)
+
+from transformers import AutoModel, AutoConfig
 from transformers.modeling_utils import SequenceSummary
-from transformers.tokenization_bert import load_vocab
+from transformers.models.bert.tokenization_bert import load_vocab
 import transformers
 
 from farm.modeling import wordembedding_utils

farm/modeling/prediction_head.py

@@ -4,8 +4,8 @@
 import numpy as np
 
 from pathlib import Path
-from transformers.modeling_bert import BertForPreTraining, ACT2FN
-from transformers.modeling_auto import AutoModelForQuestionAnswering, AutoModelForTokenClassification, AutoModelForSequenceClassification
+from transformers.models.bert.modeling_bert import BertForPreTraining, ACT2FN
+from transformers import AutoModelForQuestionAnswering, AutoModelForTokenClassification, AutoModelForSequenceClassification
 from typing import List, Tuple
 
 import torch

farm/modeling/tokenization.py

@@ -23,26 +23,22 @@
 from pathlib import Path
 
 import numpy as np
-from transformers.tokenization_albert import AlbertTokenizer
-from transformers.tokenization_albert_fast import AlbertTokenizerFast
-from transformers.tokenization_bert import BertTokenizer, load_vocab
-from transformers.tokenization_bert_fast import BertTokenizerFast
-from transformers.tokenization_distilbert import DistilBertTokenizer
-from transformers.tokenization_distilbert_fast import DistilBertTokenizerFast
-from transformers.tokenization_electra import ElectraTokenizer
-from transformers.tokenization_electra_fast import ElectraTokenizerFast
-from transformers.tokenization_roberta import RobertaTokenizer
-from transformers.tokenization_roberta_fast import RobertaTokenizerFast
+from transformers import (
+    AlbertTokenizer, AlbertTokenizerFast,
+    BertTokenizer, BertTokenizerFast,
+    DistilBertTokenizer, DistilBertTokenizerFast,
+    ElectraTokenizer, ElectraTokenizerFast,
+    RobertaTokenizer, RobertaTokenizerFast,
+    XLMRobertaTokenizer, XLMRobertaTokenizerFast,
+    XLNetTokenizer, XLNetTokenizerFast,
+    CamembertTokenizer, CamembertTokenizerFast,
+    DPRContextEncoderTokenizer, DPRContextEncoderTokenizerFast,
+    DPRQuestionEncoderTokenizer, DPRQuestionEncoderTokenizerFast
+)
+from transformers.models.bert.tokenization_bert import load_vocab
 from transformers.tokenization_utils import PreTrainedTokenizer
-from transformers.tokenization_xlm_roberta import XLMRobertaTokenizer
-from transformers.tokenization_xlm_roberta_fast import XLMRobertaTokenizerFast
-from transformers.tokenization_xlnet import XLNetTokenizer
-from transformers.tokenization_xlnet_fast import XLNetTokenizerFast
-from transformers.tokenization_camembert import CamembertTokenizer
-from transformers.tokenization_camembert_fast import CamembertTokenizerFast
-from transformers.modeling_auto import AutoConfig
-from transformers import DPRContextEncoderTokenizer, DPRQuestionEncoderTokenizer
-from transformers import DPRContextEncoderTokenizerFast, DPRQuestionEncoderTokenizerFast
+from transformers import AutoConfig
+
 
 from farm.data_handler.samples import SampleBasket
 from farm.modeling.wordembedding_utils import load_from_cache, EMBEDDING_VOCAB_FILES_MAP, run_split_on_punc

farm/modeling/wordembedding_utils.py

@@ -10,7 +10,7 @@
 import numpy as np
 import pandas as pd
 from tqdm import tqdm
-from transformers.tokenization_bert import BertTokenizer
+from transformers import BertTokenizer
 from sklearn.decomposition import TruncatedSVD
 from sklearn.cluster import KMeans
 from collections import Counter

requirements.txt

@@ -2,8 +2,9 @@
 setuptools
 wheel
 # PyTorch
---find-links=https://download.pytorch.org/whl/torch_stable.html
-torch>1.5,<1.7
+# Temp. disabled the next line as it gets currently resolved to https://download.pytorch.org/whl/rocm3.8/torch-1.7.1%2Brocm3.8-cp38-cp38-linux_x86_64.whl
+#--find-links=https://download.pytorch.org/whl/torch_stable.html
+torch>1.5,<1.8
 # progress bars in model download and training scripts
 tqdm
 # Accessing files from S3 directly.
@@ -17,7 +18,7 @@ sklearn
 seqeval==0.0.12
 mlflow==1.0.0
 # huggingface repository
-transformers==3.5.1
+transformers==4.1.1
 # accessing dictionary elements with dot notation
 dotmap==1.3.0
 # for inference-rest-apis
@@ -32,3 +33,4 @@ dill # pickle extension for (de-)serialization
 #onnxruntime
 #onnxruntime_tools
 psutil
+sentencepiece

test/benchmarks/question_answering_components.py

@@ -160,7 +160,7 @@ def prepare_dict(sample_file, q, document_size):
         if sample_file[-3:] == "txt":
             text = f.read()[:document_size]
             assert len(text) == document_size
-            dicts = [{"qas": [q], "context": text}]
+            dicts = [{"questions": [q], "text": text}]
         elif sample_file[-4:] == "json":
             data = json.load(f)
             dicts = []

test/conftest.py

@@ -33,7 +33,13 @@ def pytest_generate_tests(metafunc):
             metafunc.parametrize("use_gpu", [False], scope="session")
 
 
-@pytest.fixture()
+def pytest_collection_modifyitems(items):
+    for item in items:
+        if "conversion" in item.nodeid:
+            item.add_marker(pytest.mark.conversion)
+
+
+@pytest.fixture(scope="module")
 def adaptive_model_qa(use_gpu, num_processes):
     """
     PyTest Fixture for a Question Answering Inferencer based on PyTorch.
@@ -61,20 +67,21 @@ def adaptive_model_qa(use_gpu, num_processes):
     assert len(children) == 0
 
 
-@pytest.fixture()
+@pytest.fixture(scope="module")
 def bert_base_squad2(request):
     model = QAInferencer.load(
-            "deepset/bert-base-cased-squad2",
+            "deepset/minilm-uncased-squad2",
             task_type="question_answering",
-            batch_size=16,
+            batch_size=4,
             num_processes=0,
+            multithreading_rust=False,
             use_fast=True # TODO parametrize this to test slow as well
     )
     return model
 
 # TODO add other model types (roberta, xlm-r, albert) here as well
 
-@pytest.fixture()
+@pytest.fixture(scope="module")
 def distilbert_squad(request):
     set_all_seeds(seed=42)
     device, n_gpu = initialize_device_settings(use_cuda=False)