[SSD/PyT] New release with 22.10 base image

Author: ahmadki

PyTorch/Detection/SSD/Dockerfile

@@ -1,20 +1,14 @@
-ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:21.07-py3
+ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:22.10-py3
 FROM ${FROM_IMAGE_NAME}
 
 # Set working directory
 WORKDIR /workspace/ssd
 
-# Install nv-cocoapi
-ENV COCOAPI_VERSION=2.0+nv0.6.0
-RUN export COCOAPI_TAG=$(echo ${COCOAPI_VERSION} | sed 's/^.*+n//') \
- && pip install --no-cache-dir pybind11                             \
- && pip install --no-cache-dir git+https://github.com/NVIDIA/cocoapi.git@${COCOAPI_TAG}#subdirectory=PythonAPI
-# Install dllogger
-RUN pip install --no-cache-dir git+https://github.com/NVIDIA/dllogger.git#egg=dllogger
+# Copy the model files
+COPY . .
 
-# Install requirements
-COPY requirements.txt .
-RUN pip install -r requirements.txt
-RUN python3 -m pip install pycocotools==2.0.0
+# Install python requirements
+RUN pip install --no-cache-dir -r requirements.txt
 
-COPY . .
+ENV CUDNN_V8_API_ENABLED=1
+ENV TORCH_CUDNN_V8_API_ENABLED=1

PyTorch/Detection/SSD/README.md

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP16 on 1 GPUs using 256 batch size
 # Usage bash SSD300_FP16_1GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python $1/main.py --backbone resnet50 --warmup 300 --bs 256 --amp --data $2 ${@:3}
+python $1/main.py --backbone resnet50 --warmup 300 --bs 256 --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_A100_FP16_1GPU.sh

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP16 on 4 GPUs using 1024 batch size (256 per GPU)
 # Usage ./SSD300_FP16_4GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python -m torch.distributed.launch --nproc_per_node=4 $1/main.py --backbone resnet50 --learning-rate 2.7e-3 --warmup 1200 --bs 256 --amp --data $2 ${@:3}
+torchrun --nproc_per_node=4 $1/main.py --backbone resnet50 --learning-rate 2.7e-3 --warmup 1200 --bs 256 --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_A100_FP16_4GPU.sh

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP16 on 8 GPUs using 1024 batch size (128 per GPU)
 # Usage ./SSD300_FP16_8GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python -m torch.distributed.launch --nproc_per_node=8 $1/main.py --backbone resnet50 --learning-rate 2.7e-3 --warmup 1200 --bs 128 --amp --data $2 ${@:3}
+torchrun --nproc_per_node=8 $1/main.py --backbone resnet50 --learning-rate 2.7e-3 --warmup 1200 --bs 128 --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_A100_FP16_8GPU.sh

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP32 on 8 GPUs using 1024 batch size (128 per GPU)
 # Usage ./SSD300_FP32_8GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python -m torch.distributed.launch --nproc_per_node=8 $1/main.py --backbone resnet50 --learning-rate 2.7e-3 --warmup 1200 --bs 128 --data $2 ${@:3}
+torchrun --nproc_per_node=8 $1/main.py --backbone resnet50 --learning-rate 2.7e-3 --warmup 1200 --bs 128 --no-amp --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_A100_FP32_8GPU.sh

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP16 on 1 GPUs using 64 batch size
 # Usage bash SSD300_FP16_1GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python $1/main.py --backbone resnet50 --warmup 300 --bs 64 --amp --data $2 ${@:3}
+python $1/main.py --backbone resnet50 --warmup 300 --bs 64 --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_FP16_1GPU.sh

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP16 on 4 GPUs using 256 batch size (64 per GPU)
 # Usage ./SSD300_FP16_4GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python -m torch.distributed.launch --nproc_per_node=4 $1/main.py --backbone resnet50 --warmup 300 --bs 64 --amp --data $2 ${@:3}
+torchrun --nproc_per_node=4 $1/main.py --backbone resnet50 --warmup 300 --bs 64 --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_FP16_4GPU.sh

@@ -1,4 +1,4 @@
 # This script launches SSD300 training in FP16 on 8 GPUs using 512 batch size (64 per GPU)
 # Usage ./SSD300_FP16_8GPU.sh <path to this repository> <path to dataset> <additional flags>
 
-python -m torch.distributed.launch --nproc_per_node=8 $1/main.py --backbone resnet50 --warmup 300 --bs 64 --amp --data $2 ${@:3}
+torchrun --nproc_per_node=8 $1/main.py --backbone resnet50 --warmup 300 --bs 64 --data $2 ${@:3}

PyTorch/Detection/SSD/examples/SSD300_FP16_8GPU.sh

@@ -1,4 +1,4 @@
 # This script evaluates SSD300 model in FP16 using 32 batch size on 1 GPU
 # Usage: ./SSD300_FP16_EVAL.sh <path to this repository> <path to dataset> <path to checkpoint> <additional flags>
 
-python $1/main.py --backbone resnet50 --amp --ebs 32 --data $2 --mode evaluation --checkpoint $3 ${@:4}
+python $1/main.py --backbone resnet50 --ebs 32 --data $2 --mode evaluation --checkpoint $3 ${@:4}