Merge branch 'master' into feature/1947_load_disparity

Author: tchaton

CHANGELOG.md

@@ -75,6 +75,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed that metrics do not store computational graph for all seen data ([#4313](https://github.com/PyTorchLightning/pytorch-lightning/pull/4313))
 - Fixed AMP unscale for `on_after_backward` ([#4439](https://github.com/PyTorchLightning/pytorch-lightning/pull/4439))
 - Fixed TorchScript export when module includes Metrics ([#4428](https://github.com/PyTorchLightning/pytorch-lightning/pull/4428))
+- Fixed TorchScript trace method's data to device and docstring ([#4360](https://github.com/PyTorchLightning/pytorch-lightning/pull/4360))
 - Fixed CSV logger warning ([#4419](https://github.com/PyTorchLightning/pytorch-lightning/pull/4419))
 - Fixed skip DDP parameter sync ([#4301](https://github.com/PyTorchLightning/pytorch-lightning/pull/4301))
 

dockers/README.md

@@ -14,10 +14,10 @@ or with specific arguments
 ```bash
 git clone <git-repository>
 docker image build \
-    -t pytorch-lightning:py3.8 \
-    -f dockers/conda/Dockerfile \
+    -t pytorch-lightning:py3.8-pt1.6 \
+    -f dockers/base-cuda/Dockerfile \
     --build-arg PYTHON_VERSION=3.8 \
-    --build-arg PYTORCH_VERSION=1.4 \
+    --build-arg PYTORCH_VERSION=1.6 \
     .
 ```
 

dockers/base-conda/Dockerfile

@@ -35,7 +35,8 @@ SHELL ["/bin/bash", "-c"]
 
 ENV PATH="$PATH:/root/.local/bin"
 
-RUN apt-get update && apt-get install -y --no-install-recommends \
+RUN apt-get update -qq && \
+    apt-get install -y --no-install-recommends \
         build-essential \
         cmake \
         git \
@@ -104,6 +105,7 @@ RUN \
     # Install remaining requirements
     pip install -r requirements-extra.txt --upgrade-strategy only-if-needed && \
     pip install -r requirements-test.txt --upgrade-strategy only-if-needed && \
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-cuda${CUDA_VERSION%%.*}0 && \
     rm requirements*
 
 RUN \
@@ -118,4 +120,4 @@ RUN \
     conda info && \
     pip list && \
     python -c "import sys; assert sys.version[:3] == '$PYTHON_VERSION', sys.version" && \
-    python -c "import torch; assert torch.__version__[:3] == '$PYTORCH_VERSION', torch.__version__"
+    python -c "import torch; assert torch.__version__[:3] == '$PYTORCH_VERSION', torch.__version__"

dockers/base-cuda/Dockerfile

@@ -28,6 +28,7 @@ FROM nvidia/cuda:${CUDA_VERSION}-cudnn${CUDNN_VERSION}-devel-ubuntu18.04
 
 ARG PYTHON_VERSION=3.7
 ARG PYTORCH_VERSION=1.6
+ARG CMAKE_VERSION=3.18.4
 
 SHELL ["/bin/bash", "-c"]
 # https://techoverflow.net/2019/05/18/how-to-fix-configuring-tzdata-interactive-input-when-building-docker-images/
@@ -37,7 +38,7 @@ ENV TZ=Europe/Prague
 ENV PATH="$PATH:/root/.local/bin"
 ENV CUDA_TOOLKIT_ROOT_DIR="/usr/local/cuda"
 
-RUN apt-get update && \
+RUN apt-get update -qq && \
     apt-get install -y --no-install-recommends \
         build-essential \
         pkg-config \
@@ -93,6 +94,7 @@ RUN \
 
     # Install all requirements
     pip install -r requirements/devel.txt --upgrade-strategy only-if-needed --use-feature=2020-resolver && \
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-cuda${CUDA_VERSION%%.*}0 && \
     rm -rf requirements*
 
 RUN \
@@ -105,5 +107,6 @@ RUN \
     # Show what we have
     pip --version && \
     pip list && \
+    python -c 'from nvidia.dali.pipeline import Pipeline' && \
     python -c "import sys; assert sys.version[:3] == '$PYTHON_VERSION', sys.version" && \
-    python -c "import torch; assert torch.__version__[:3] == '$PYTORCH_VERSION', torch.__version__"
+    python -c "import torch; assert torch.__version__[:3] == '$PYTORCH_VERSION', torch.__version__"

dockers/base-xla/Dockerfile

@@ -31,7 +31,7 @@ ENV CONDA_ENV=lightning
 # show system inforation
 RUN lsb_release -a && cat /etc/*-release
 
-RUN apt-get update && \
+RUN apt-get update -qq && \
     apt-get install -y --no-install-recommends \
         build-essential \
         cmake \

docs/source/community_examples.rst

@@ -16,4 +16,5 @@ Community Examples
 - `VAE Library of over 18+ VAE flavors <https://github.com/AntixK/PyTorch-VAE>`_.
 - `Transformers Question Answering (SQuAD) <https://github.com/tshrjn/Finetune-QA/>`_.
 - `Atlas: End-to-End 3D Scene Reconstruction from Posed Images <https://github.com/magicleap/atlas>`_.
-- `Self-Supervised Representation Learning (MoCo and BYOL) <https://github.com/untitled-ai/self_supervised>`_.
+- `Self-Supervised Representation Learning (MoCo and BYOL) <https://github.com/untitled-ai/self_supervised>`_.
+- `pytorch-forecasting: Time series forecasting package <https://github.com/jdb78/pytorch-forecasting>`_.

docs/source/metrics.rst

@@ -78,6 +78,26 @@ If ``on_epoch`` is True, the logger automatically logs the end of epoch metric v
         self.valid_acc(logits, y)
         self.log('valid_acc', self.valid_acc, on_step=True, on_epoch=True)
 
+.. note::
+    If using metrics in data parallel mode (dp), the metric update/logging should be done
+    in the ``<mode>_step_end`` method (where ``<mode>`` is either ``training``, ``validation``
+    or ``test``). This is due to metric states else being destroyed after each forward pass,
+    leading to wrong accumulation. In practice do the following:
+
+    .. code-block:: python
+
+        def training_step(self, batch, batch_idx):
+            data, target = batch
+            pred = self(data)
+            ...
+            return {'loss' : loss, 'preds' : preds, 'target' : target}
+
+        def training_step_end(self, outputs):
+            #update and log
+            self.metric(outputs['preds'], outputs['target'])
+            self.log('metric', self.metric)
+
+
 This metrics API is independent of PyTorch Lightning. Metrics can directly be used in PyTorch as shown in the example:
 
 .. code-block:: python

pytorch_lightning/core/lightning.py

@@ -1399,7 +1399,6 @@ def _verify_is_manual_optimization(self, fn_name):
 
     @classmethod
     def _auto_collect_arguments(cls, frame=None) -> Tuple[Dict, Dict]:
-        """"""
         """
         Collect all module arguments in the current constructor and all child constructors.
         The child constructors are all the ``__init__`` methods that reach the current class through