Created YML files to run ofa and added a fix for resize_crop script

README.md

@@ -184,6 +184,49 @@ horovodrun -np 32 -H <server1_ip>:8,<server2_ip>:8,<server3_ip>:8,<server4_ip>:8
 * ImageNet Dataset 
 * Horovod
 
+## Installation
+
+The first thing to do is install [OpenMPI](http://www.open-mpi.org), see the [OpenMPI installation guide](https://edu.itp.phys.ethz.ch/hs12/programming_techniques/openmpi.pdf).
+
+We need to install NCC as well:
+
+We also may need to set the following environment variables to fit your CUDA environment:
+
+```bash
+  export CUDA=/usr/local/cuda
+  export HOROVOD_CUDA_HOME=/usr/local/cuda
+  export CUDACXX=/usr/local/cuda/bin/nvcc
+  export CUDA_INCLUDE_DIRS=/usr/local/cuda
+  export LD_LIBRARY_PATH=$CUDA/lib64:$LD_LIBRARY_PATH
+  export PATH=$CUDA/bin:$PATH
+
+  export HOROVOD_WITH_PYTORCH=1 
+  export HOROVOD_NCCL_INCLUDE=/usr/include 
+  export HOROVOD_GPU=CUDA 
+  export HOROVOD_WITHOUT_TENSORFLOW=1 
+  export HOROVOD_WITHOUT_MXNET=1 
+  export HOROVOD_WITHOUT_GLOO=1
+  export HOROVOD_GPU_OPERATIONS=NCCL
+```
+
+Next, go ahead and run:
+
+```bash
+  conda env create -f conda-gpu.yml --force
+```
+
+This enables Horovod with GPU support. 
+
+Next activate the environment:
+
+```bash
+  conda activate ofa-gpu
+  pip install .
+```
+
+You can now run ofa with GPU support. 
+
+
 ## Related work on automated and efficient deep learning:
 [ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware](https://arxiv.org/pdf/1812.00332.pdf) (ICLR’19)
 

conda-cpu.yml

@@ -0,0 +1,20 @@
+name: ofa-cpu
+
+channels:
+  - pytorch
+dependencies:
+  - python==3.7
+  - pip
+  - matplotlib
+  - opencv
+  - pip:
+    - numpy
+    - torch==1.8.1
+    - torchaudio==0.8.1
+    - torchvision==0.9.1
+    - torchsummary
+    - filelock
+    - PyYaml
+    - tqdm
+    - thop
+    - -e .

conda-gpu.yml

@@ -0,0 +1,36 @@
+name: ofa-gpu
+
+channels:
+  - pytorch
+  - conda-forge
+  - defaults
+dependencies:
+  - python==3.7
+  - ccache
+  - cmake
+  - cudatoolkit=10.1
+  - cudnn=7.6.5
+  - cxx-compiler
+  - gcc_linux-64
+  - gxx_linux-64
+  - jupyterlab
+  - mpi4py # installs cuda-aware openmpi
+  - nccl
+  - nvcc_linux-64=10.1
+  - openmpi
+  - pip
+  - matplotlib
+  - opencv
+  - pip:
+    - numpy
+    - torch==1.8.1
+    - torchaudio==0.8.1
+    - torchvision==0.9.1
+    - torchsummary
+    - filelock
+    - PyYaml
+    - tqdm
+    - thop
+    - horovod==0.19.*
+    - -e .
+    - --no-binary=horovod

ofa/utils/my_dataloader/my_random_resize_crop.py

@@ -6,6 +6,7 @@
 
 import torchvision.transforms.functional as F
 import torchvision.transforms as transforms
+from torchvision.transforms import InterpolationMode
 
 __all__ = ['MyRandomResizedCrop', 'MyResizeRandomCrop', 'MyResize']
 
@@ -69,7 +70,11 @@ def sample_image_size(batch_id=None):
 		MyRandomResizedCrop.ACTIVE_SIZE = random.choices(candidate_sizes, weights=relative_probs)[0]
 
 	def __repr__(self):
-		interpolate_str = _pil_interpolation_to_str[self.interpolation]
+		if isinstance(self.interpolation, InterpolationMode):
+			interpolate_str = self.interpolation.value
+		else:
+			interpolate_str = _pil_interpolation_to_str[self.interpolation]
+
 		format_string = self.__class__.__name__ + '(size={0}'.format(MyRandomResizedCrop.IMAGE_SIZE_LIST)
 		if MyRandomResizedCrop.CONTINUOUS:
 			format_string += '@continuous'
@@ -113,9 +118,14 @@ def __call__(self, img):
 		return F.crop(img, i, j, h, w)
 
 	def __repr__(self):
+		if isinstance(self.interpolation, InterpolationMode):
+			interpolate_str = self.interpolation.value
+		else:
+			interpolate_str = _pil_interpolation_to_str[self.interpolation]
+
 		return 'MyResizeRandomCrop(size=%s%s, interpolation=%s, use_padding=%s, fill=%s)' % (
 			MyRandomResizedCrop.IMAGE_SIZE_LIST, '@continuous' if MyRandomResizedCrop.CONTINUOUS else '',
-			_pil_interpolation_to_str[self.interpolation], self.use_padding, self.fill,
+			interpolate_str, self.use_padding, self.fill,
 		)
 
 
@@ -130,7 +140,12 @@ def __call__(self, img):
 		return img
 
 	def __repr__(self):
+		if isinstance(self.interpolation, InterpolationMode):
+			interpolate_str = self.interpolation.value
+		else:
+			interpolate_str = _pil_interpolation_to_str[self.interpolation]
+
 		return 'MyResize(size=%s%s, interpolation=%s)' % (
 			MyRandomResizedCrop.IMAGE_SIZE_LIST, '@continuous' if MyRandomResizedCrop.CONTINUOUS else '',
-			_pil_interpolation_to_str[self.interpolation]
-		)
+			interpolate_str
+		)