0

I'm currently trying to run a deep learning tool software that was previously created by someone else a few years ago. While trying to load a class called Evaluator which wraps all of the important mmdetection functions, I keep getting the following error: enter image description here

The model was downloaded automatically while running the code due to the following part of the config file:

model = dict(
        type='FCOS',
        pretrained='open-mmlab://detectron/resnet101_caffe',
        backbone=dict(
            type='ResNet',
            depth=101,
            num_stages=4,
            out_indices=(0, 1, 2, 3),
            frozen_stages=1,
            norm_cfg=dict(type='BN', requires_grad=False),
            norm_eval=True,
            style='caffe'),
        neck=dict(
            type='FPN',
            in_channels=[256, 512, 1024, 2048],
            out_channels=256,
            start_level=1,
            add_extra_convs=True,
            extra_convs_on_inputs=False,
            num_outs=5,
            relu_before_extra_convs=True),
        bbox_head=dict(
            type='FCOSHead',
            num_classes=15,
            in_channels=256,
            stacked_convs=4,
            feat_channels=256,
            strides=[8, 16, 32, 64, 128],
            loss_cls=dict(
                type='FocalLoss',
                use_sigmoid=True,
                gamma=2.0,
                alpha=0.25,
                loss_weight=1.0),
            loss_bbox=dict(type='IoULoss', loss_weight=1.0),
            loss_centerness=dict(
                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0)))

I'm not sure how to determine if the model I'm trying to load and the state dictionary are compatible or how to fix this problem. I'm new to deep learning and using MMdetection.

Here is part of the code from the utils.py file that contains the Evaluator class:

from skimage.draw import rectangle_perimeter
    import skimage.io as io
    from skimage.transform import resize
    import numpy as np
    import skimage
    import pickle

    import torch

    from mmcv import Config, DictAction
    from mmdet.models import build_detector
    from mmcv.runner import load_checkpoint
    import mmcv

    from mmdet.datasets.pipelines import Compose  # TO LOOK AT
    from mmcv.parallel import collate, scatter

    from mmdet.core import bbox2result
    from skimage import data, io, filters
from matplotlib.pyplot import figure

    import os


    class_to_number = {"Yeast White": 0, "Budding White": 1, "Yeast Opaque": 2,
                           "Budding Opaque":3,"Yeast Gray": 4, "Budding Gray": 5,
                            "Shmoo":6,"Artifact": 7, "Unknown ": 8,
                           "Pseudohyphae": 9, "Hyphae": 10, "H-junction": 11,
                           "P-junction":12,"P-Start":13,"H-Start":14}
    number_to_class = {y:x for x,y in class_to_number.items()}
class Evaluator():
        def __init__(self,config,checkpoint_file):
            self.cfg = Config.fromfile(config)
            self.cfg["gpu-ids"] = 6
            self.model = build_detector(
            self.cfg.model, train_cfg=self.cfg.train_cfg, test_cfg=self.cfg.test_cfg)
            checkpoint_dict = load_checkpoint(self.model,checkpoint_file)
            state_dict = checkpoint_dict["state_dict"]
            self.model.CLASSES = checkpoint_dict['meta']['CLASSES']
            self.model.load_state_dict(state_dict)
            self.model.eval()

I looked at the version of mmdet, mmcv, and pytorch to ensure they were the same versions that were used by the original creator of the software. I redownloaded the model file to ensure that it wasn't corrupted. `

jjacob
  • 1
  • 2

2 Answers2

0

It is normal that the model and loaded state dict do not match exactly, because the fully connected layers in the pretrained models are unused. It will not affect the training. If it causes any further issues while testing, then this is a problem otherwise you should be good.

Refer to the issue here.

Pranav Sharma
  • 34
  • 2
0

While it is normal for the warning to compare head weights, in your case, it looks like you are passing RESNET 101 backbone weights to the entire model. remove it from the model and add it to backbone with key as pretrained or init_cfg as per your mmdet version

Mohit Burkule
  • 136
  • 1
  • 6