Add more models

Also add .so and tuning data for easier reproducibility.

.gitignore

@@ -4,3 +4,4 @@ mod.json
 mod.params
 model.tar
 tuning*.json
+.nfs*

a100_superresolution_1_tuned/README.md

@@ -0,0 +1,12 @@
+This is an extend version of the superresolution model.
+Instead of using a fixed 3 upscale ratio with an input of 224x224,
+we feed the model 640x480 inputs, similar to what would be a 480p video,
+and by changing the value in UPSCALE_FACTOR, we can simulate various resolution enanchement.
+
+For instance, a 6x upscaling would be close to a 4K output.
+A 4x would be similar to a FullHD output.
+A 12x would be a 8K output
+
+
+Original code available at https://raw.githubusercontent.com/pytorch/tutorials/main/advanced_source/super_resolution_with_onnxruntime.py
+Surrounding scripts released under GNU GPL v3. See [../LICENSE](../LICENSE).

a100_superresolution_1_tuned/UPSCALE_FACTOR

@@ -0,0 +1 @@
+1

a100_superresolution_1_tuned/compile.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+TUNING=$(ls tuning* --sort=time -1 | head -n1)
+
+set -x
+python3 -m tvm.driver.tvmc compile \
+        --target cuda
+        --target-cuda-arch sm_80 \
+        --output model.tar \
+        --input-shape 'input:[1,1,640,480]' \
+        --tuning-records $TUNING \
+        super_resolution.onnx $@ \
+
+        #--output-shape 'input:[1,1,4096,3112]' \
+
+tar -xvf model.tar

a100_superresolution_1_tuned/convert_docker.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+UPSCALE_FACTOR=$(cat UPSCALE_FACTOR)
+#DIR=$(basename "$PWD")
+DIR=a100_superresolution_${UPSCALE_FACTOR}_tuned
+CMD="../a100_superresolution_1_tuned/super_resolution_with_onnxruntime.py $UPSCALE_FACTOR"
+
+echo $CMD
+set -x
+sudo docker run --rm --gpus 1 \
+        -v $(pwd):/semla/modelzoo/$DIR \
+        -v $(pwd)/../a100_superresolution_1_tuned:/semla/modelzoo/a100_superresolution_1_tuned \
+        --workdir /semla/modelzoo/$DIR \
+        --entrypoint "/usr/bin/python3" \
+        semla/tvm_pytorch \
+        $CMD

a100_superresolution_1_tuned/do.sh

@@ -0,0 +1,4 @@
+../a100_superresolution_1_tuned/convert.sh
+../a100_superresolution_1_tuned/tune.sh
+../a100_superresolution_1_tuned/compile.sh
+../a100_superresolution_1_tuned/fix_sizes.sh

a100_superresolution_1_tuned/download.sh

@@ -0,0 +1,2 @@
+#!/bin/bash
+wget https://github.com/onnx/models/raw/main/vision/super_resolution/sub_pixel_cnn_2016/model/super-resolution-10.onnx

a100_superresolution_1_tuned/fix_sizes.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+UPSCALE_FACTOR=$(cat UPSCALE_FACTOR)
+W_OUTPUT=$(($UPSCALE_FACTOR * 640))
+H_OUTPUT=$(($UPSCALE_FACTOR * 480))
+MODEL_NAME=a100_superresolution_${UPSCALE_FACTOR}_tuned
+
+
+cat ../a100_superresolution_1_tuned/metadata.template.json \
+        | sed -e "s/W_OUTPUT/${W_OUTPUT}/g" \
+        | sed -e "s/H_OUTPUT/${H_OUTPUT}/g" \
+        | sed -e "s/MODEL_NAME/${MODEL_NAME}/g" \
+        | tee metadata.json

a100_superresolution_1_tuned/metadata.json

@@ -0,0 +1,16 @@
+{
+        "load_time": 1000000000,
+        "exec_time": [1000000,2000000,4000000,8000000],
+        "weights_size": 100000000,
+        "workspace_size": 500000000,
+        "input_name" : "input",
+        "output_name" : "output",
+        "input_shape": [1,1,640,480],
+        "output_shape": [1,1,640,480],
+        "input_type": "FP32",
+        "output_type": "FP32",
+        "device": "CUDA",
+        "model" : "a100_superresolution_1_tuned",
+        "architecture" : "CUDA_80"
+
+}

a100_superresolution_1_tuned/metadata.template.json

@@ -0,0 +1,16 @@
+{
+        "load_time": 1000000000,
+        "exec_time": [1000000,2000000,4000000,8000000],
+        "weights_size": 100000000,
+        "workspace_size": 500000000,
+        "input_name" : "input",
+        "output_name" : "output",
+        "input_shape": [1,1,640,480],
+        "output_shape": [1,1,W_OUTPUT,H_OUTPUT],
+        "input_type": "FP32",
+        "output_type": "FP32",
+        "device": "CUDA",
+        "model" : "MODEL_NAME",
+        "architecture" : "CUDA_80"
+
+}