Benign huggingface Download

logeshsenthil13/ConvFormer main

Shell changes working directoryPython interpreter runs local script
SHA-256ac96c97d3eadb98fd71f6d7c8dccf81c52157b3337113eee6a21be26aeaa26aa

Evidence

Contains common English function words lines 1–9
⋯4 lines
5#
6# Anything can be re-run in isolation; every step writes its own results/*.json
7# and 8_report.py just collects whatever is there.
8#
9# Budget: ConvFormer-B36 is 99.9 M params / ~23 …
Contains common English function words lines 9–11
9:66… - four times
10# Inception-V3's parameter count. The five engine builds dominate the wall clock,
11# and the FP32 build is the heaviest. OPTLEVEL=1 cu …
Contains common English function words lines 16–18
16:33… calibration npz on top of a 400 MB model. If
17# that is too much on an Orin Nano, run `python 4_make_calib.py --num 512
18# --batch 8` first and then `bash run …
Shell changes working directory lines 23–37
23:9… 6_build_engines.sh fp16 int8 int8_bf16 int4
24
25set -euo pipefail
26cd "$(dirname "$0")"
27source ./env.sh
28
29MODEL=convformer_b36
30
31echo; echo "##### 0) dataset splits + preprocessing cache (224 bicubic -> 224)"
32python 0_prepare_data.py
33
34echo; echo "##### 0b) verify preprocessing matches timm's own transform"
35python verify.py
36
37echo; echo "##### …
Contains common English function words lines 38–43
38:29…
39
40echo; echo "##### 2) ONNX export FP32 + FP16 (with the measured fp16 block list)"
41python 2_export_onnx.py
42
43echo; echo "##### 3) ONNX Runtime eval …

Keyboard shortcuts on this page: j for the next sample, k for the previous one, x to go back to the feed, d to download the original bytes, r to re-queue the sample for analysis.