Convert Pretrained Models — detrex documentation (original) (raw)

This document provides a brief intro of how to convert the pretrained model into the format of detrex.

Convert TorchVision Pretrained ResNet Models

To use the detectron2 provided pretrained weights, please refer to ImageNet Pretrained Models. Here we’ve noticed that detectron2 only provided a converted torchvision ResNet-50 model. For more pretrained models like ResNet{101, 152}. You can use the detectron2 provided conversion script to convert the torchvision pretrained weights into the format that can be used in detrex. Here’s the detailed tutorial about the usage the conversion script.

Download Pretrained Weights

Torchvision 0.11.0 was released packed with better pretrained weights on numerous models including ResNet. More details can be found in How to Train State-Of-The-Art Models Using TorchVision’s Latest Primitives, here we collected the download scripts for TorchVision ResNet models.

Note: ImageNet1k-V1 means the old pretrained weights. ImageNet1k-V2 means the improved baseline results.

Run the Conversion

convert-torchvision-to-d2 (borrowed from detectron2)

#!/usr/bin/env python

Copyright (c) Facebook, Inc. and its affiliates.

import pickle as pkl import sys import torch

""" Usage:

download one of the ResNet{18,34,50,101,152} models from torchvision:

wget https://download.pytorch.org/models/resnet50-19c8e357.pth -O r50.pth

run the conversion

./convert-torchvision-to-d2.py r50.pth r50.pkl

Then, use r50.pkl with the following changes in config:

MODEL: WEIGHTS: "/path/to/r50.pkl" PIXEL_MEAN: [123.675, 116.280, 103.530] PIXEL_STD: [58.395, 57.120, 57.375] RESNETS: DEPTH: 50 STRIDE_IN_1X1: False INPUT: FORMAT: "RGB" These models typically produce slightly worse results than the pre-trained ResNets we use in official configs, which are the original ResNet models released by MSRA. """

if name == "main": input = sys.argv[1]

obj = torch.load(input, map_location="cpu")

newmodel = {}
for k in list(obj.keys()):
    old_k = k
    if "layer" not in k:
        k = "stem." + k
    for t in [1, 2, 3, 4]:
        k = k.replace("layer{}".format(t), "res{}".format(t + 1))
    for t in [1, 2, 3]:
        k = k.replace("bn{}".format(t), "conv{}.norm".format(t))
    k = k.replace("downsample.0", "shortcut")
    k = k.replace("downsample.1", "shortcut.norm")
    print(old_k, "->", k)
    newmodel[k] = obj.pop(old_k).detach().numpy()

res = {"model": newmodel, "__author__": "torchvision", "matching_heuristics": True}

with open(sys.argv[2], "wb") as f:
    pkl.dump(res, f)
if obj:
    print("Unconverted keys:", obj.keys())

Firstly, create convert-torchvision-to-d2.py and copy the relative code mentioned above, then run:

python convert-torchvision-to-d2.py
/path/to/r101_v1.pth \ # path to the downloaded pretrained weights ./r101_v1.pkl # where to save the converted weights

Then, change the training configs:

your own config.py

train.init_checkpoint = "path/to/r101_v1.pkl"

make sure that the model config is consistent

with the following settings

model.backbone.stages.depth = 101 model.pixel_mean = [123.675, 116.280, 103.530] model.pixel_std = [58.395, 57.120, 57.375]

Convert DETRs Pretrained Models

We also provides converters for a partial of projects in detrex. These conversions are modified from the detr-d2 conversion script to convert models trained by the original repo into the format of detrex models.

• converter for DETR: convert_detr_to_detrex
• converter for Deformable-DETR: convert_deformable_detr_to_detrex
• converter for ConditionalDETR: convert_conditional_detr_to_detrex
• converter for DN-Deformable-DETR: convert_dn_deformable_detr_to_detrex

All these converters can be runned as:

python converter.py --source_model /path/to/pretrained_weight.pth --output_model converted_model.pth