Merge branch 'master' into feat/change-installation

.github/CONTRIBUTING.md

@@ -103,7 +103,7 @@ Want to add a new test case and not sure how? [Talk to us!](https://join.slack.c
 
 ## Guidelines
 
-For this section, we refer to read the [parent PL guidelines](https://pytorch-lightning.readthedocs.io/en/latest/CONTRIBUTING.html)
+For this section, we refer to read the [parent PL guidelines](https://pytorch-lightning.readthedocs.io/en/latest/generated/CONTRIBUTING.html)
 
 **Reminder**
 

CHANGELOG.md

@@ -15,6 +15,12 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Changed the installation command for extra features ([#346](https://github.com/PyTorchLightning/lightning-flash/pull/346))
 
 
+- Fixed a bug where the translation task wasn't decoding tokens properly ([#332](https://github.com/PyTorchLightning/lightning-flash/pull/332))
+
+
+- Fixed a bug where huggingface tokenizers were sometimes being pickled ([#332](https://github.com/PyTorchLightning/lightning-flash/pull/332))
+
+
 ## [0.3.0] - 2021-05-20
 
 ### Added

docs/source/conf.py

@@ -80,15 +80,15 @@ def _load_py_module(fname, pkg="flash"):
 #
 source_suffix = [".rst", ".md"]
 
-needs_sphinx = '3.4'
+needs_sphinx = "4.0"
 
 # -- Options for intersphinx extension ---------------------------------------
 
 # Example configuration for intersphinx: refer to the Python standard library.
 intersphinx_mapping = {
     "python": ("https://docs.python.org/3", None),
     "torch": ("https://pytorch.org/docs/stable/", None),
-    "numpy": ("https://docs.scipy.org/doc/numpy/", None),
+    "numpy": ("https://numpy.org/doc/stable/", None),
     "PIL": ("https://pillow.readthedocs.io/en/stable/", None),
     "pytorchvideo": ("https://pytorchvideo.readthedocs.io/en/latest/", None),
     "pytorch_lightning": ("https://pytorch-lightning.readthedocs.io/en/stable/", None),

flash/image/detection/model.py

@@ -94,7 +94,7 @@ def __init__(
     ):
 
         if not _IMAGE_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install . '[image]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[image]'")
 
         self.save_hyperparameters()
 

flash/image/embedding/model.py

@@ -60,7 +60,7 @@ def __init__(
         pooling_fn: Callable = torch.max
     ):
         if not _IMAGE_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[image]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[image]'")
 
         super().__init__(
             model=None,

flash/image/segmentation/data.py

@@ -80,7 +80,7 @@ class SemanticSegmentationPathsDataSource(PathsDataSource):
 
     def __init__(self):
         if not _IMAGE_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[image]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[image]'")
         super().__init__(IMG_EXTENSIONS)
 
     def load_data(self, data: Union[Tuple[str, str], Tuple[List[str], List[str]]],
@@ -172,7 +172,7 @@ def __init__(
             image_size: A tuple with the expected output image size.
         """
         if not _IMAGE_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[image]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[image]'")
         self.image_size = image_size
         self.num_classes = num_classes
         if num_classes:

flash/image/segmentation/model.py

@@ -86,7 +86,7 @@ def __init__(
     ) -> None:
 
         if isinstance(backbone, str) and (not _TORCHVISION_AVAILABLE or not _TIMM_AVAILABLE):
-            raise ModuleNotFoundError("Please, pip install -e '.[image]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[image]'")
 
         if metrics is None:
             metrics = IoU(num_classes=num_classes)

flash/image/style_transfer/model.py

@@ -94,7 +94,7 @@ def __init__(
     ):
 
         if not _IMAGE_STLYE_TRANSFER:
-            raise ModuleNotFoundError("Please, pip install -e '.[image_style_transfer]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[image_style_transfer]'")
 
         self.save_hyperparameters(ignore="style_image")
 

flash/tabular/classification/model.py

@@ -57,7 +57,7 @@ def __init__(
         **tabnet_kwargs,
     ):
         if not _TABULAR_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[tabular]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[tabular]'")
 
         self.save_hyperparameters()
 

flash/text/classification/data.py

@@ -36,8 +36,9 @@ def __init__(self, backbone: str, max_length: int = 128):
         super().__init__()
 
         if not _TEXT_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[text]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
 
+        self.backbone = backbone
         self.tokenizer = AutoTokenizer.from_pretrained(backbone, use_fast=True)
         self.max_length = max_length
 
@@ -55,6 +56,15 @@ def _transform_label(self, label_to_class_mapping: Dict[str, int], target: str,
         ex[target] = label_to_class_mapping[ex[target]]
         return ex
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class TextFileDataSource(TextDataSource):
 
@@ -115,18 +125,45 @@ def load_data(
     def predict_load_data(self, data: Any, dataset: AutoDataset):
         return self.load_data(data, dataset, columns=["input_ids", "attention_mask"])
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class TextCSVDataSource(TextFileDataSource):
 
     def __init__(self, backbone: str, max_length: int = 128):
         super().__init__("csv", backbone, max_length=max_length)
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class TextJSONDataSource(TextFileDataSource):
 
     def __init__(self, backbone: str, max_length: int = 128):
         super().__init__("json", backbone, max_length=max_length)
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class TextSentencesDataSource(TextDataSource):
 
@@ -143,6 +180,15 @@ def load_data(
             data = [data]
         return [self._tokenize_fn(s, ) for s in data]
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class TextClassificationPreprocess(Preprocess):
 
@@ -157,7 +203,7 @@ def __init__(
     ):
 
         if not _TEXT_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[text]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
 
         self.backbone = backbone
         self.max_length = max_length

flash/text/classification/model.py

@@ -50,7 +50,7 @@ def __init__(
         serializer: Optional[Union[Serializer, Mapping[str, Serializer]]] = None,
     ):
         if not _TEXT_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[text]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
 
         self.save_hyperparameters()
 

flash/text/seq2seq/core/data.py

@@ -21,7 +21,7 @@
 import flash
 from flash.core.data.data_module import DataModule
 from flash.core.data.data_source import DataSource, DefaultDataSources
-from flash.core.data.process import Preprocess
+from flash.core.data.process import Postprocess, Preprocess
 from flash.core.data.properties import ProcessState
 from flash.core.utilities.imports import _TEXT_AVAILABLE
 
@@ -43,9 +43,10 @@ def __init__(
         super().__init__()
 
         if not _TEXT_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[text]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
 
-        self.tokenizer = AutoTokenizer.from_pretrained(backbone, use_fast=True)
+        self.backbone = backbone
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
         self.max_source_length = max_source_length
         self.max_target_length = max_target_length
         self.padding = padding
@@ -71,6 +72,15 @@ def _tokenize_fn(
             padding=self.padding,
         )
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class Seq2SeqFileDataSource(Seq2SeqDataSource):
 
@@ -112,6 +122,15 @@ def load_data(
     def predict_load_data(self, data: Any) -> Union['datasets.Dataset', List[Dict[str, torch.Tensor]]]:
         return self.load_data(data, columns=["input_ids", "attention_mask"])
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class Seq2SeqCSVDataSource(Seq2SeqFileDataSource):
 
@@ -130,6 +149,15 @@ def __init__(
             padding=padding,
         )
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class Seq2SeqJSONDataSource(Seq2SeqFileDataSource):
 
@@ -148,6 +176,15 @@ def __init__(
             padding=padding,
         )
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 class Seq2SeqSentencesDataSource(Seq2SeqDataSource):
 
@@ -161,6 +198,15 @@ def load_data(
             data = [data]
         return [self._tokenize_fn(s) for s in data]
 
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
 
 @dataclass(unsafe_hash=True, frozen=True)
 class Seq2SeqBackboneState(ProcessState):
@@ -190,7 +236,7 @@ def __init__(
         self.padding = padding
 
         if not _TEXT_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[text]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
 
         super().__init__(
             train_transform=train_transform,
@@ -240,7 +286,47 @@ def collate(self, samples: Any) -> Tensor:
         return default_data_collator(samples)
 
 
+class Seq2SeqPostprocess(Postprocess):
+
+    def __init__(self):
+        super().__init__()
+
+        if not _TEXT_AVAILABLE:
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
+
+        self._backbone = None
+        self._tokenizer = None
+
+    @property
+    def backbone(self):
+        backbone_state = self.get_state(Seq2SeqBackboneState)
+        if backbone_state is not None:
+            return backbone_state.backbone
+
+    @property
+    def tokenizer(self):
+        if self.backbone is not None and self.backbone != self._backbone:
+            self._tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+            self._backbone = self.backbone
+        return self._tokenizer
+
+    def uncollate(self, generated_tokens: Any) -> Any:
+        pred_str = self.tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+        pred_str = [str.strip(s) for s in pred_str]
+        return pred_str
+
+    def __getstate__(self):  # TODO: Find out why this is being pickled
+        state = self.__dict__.copy()
+        state.pop("_tokenizer")
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self._tokenizer = AutoTokenizer.from_pretrained(self.backbone, use_fast=True)
+
+
 class Seq2SeqData(DataModule):
     """Data module for Seq2Seq tasks."""
 
     preprocess_cls = Seq2SeqPreprocess
+    postprocess_cls = Seq2SeqPostprocess

flash/text/seq2seq/core/model.py

@@ -66,7 +66,7 @@ def __init__(
         num_beams: Optional[int] = None,
     ):
         if not _TEXT_AVAILABLE:
-            raise ModuleNotFoundError("Please, pip install -e '.[text]'")
+            raise ModuleNotFoundError("Please, pip install 'lightning-flash[text]'")
 
         os.environ["TOKENIZERS_PARALLELISM"] = "TRUE"
         # disable HF thousand warnings