ChartElementNet-MultiClass / chart_elementnet_swin.py
hanszhu's picture
Upload chart_elementnet_swin.py
164f8a3 verified
# cascade_rcnn_r50_fpn_meta.py - Enhanced config with Swin Transformer backbone
#
# PROGRESSIVE LOSS STRATEGY:
# - All 3 Cascade stages start with SmoothL1Loss for stable initial training
# - At epoch 5, Stage 3 (final stage) switches to GIoULoss via ProgressiveLossHook
# - Stage 1 & 2 remain SmoothL1Loss throughout training
# - This ensures model stability before introducing more complex IoU-based losses
_base_ = [
'../../mmdetection/configs/_base_/datasets/coco_detection.py',
'../../mmdetection/configs/_base_/schedules/schedule_1x.py',
'../../mmdetection/configs/_base_/default_runtime.py'
]
# Custom imports - this registers our modules without polluting config namespace
custom_imports = dict(
imports=[
'legend_match_swin.custom_models.custom_dataset',
'legend_match_swin.custom_models.register',
'legend_match_swin.custom_models.custom_hooks',
'legend_match_swin.custom_models.progressive_loss_hook',
],
allow_failed_imports=False
)
# Add to Python path
import sys
import os
# Use a simpler path approach that doesn't rely on __file__
sys.path.insert(0, os.path.join(os.getcwd(), '..', '..'))
# Custom Cascade model with coordinate handling for chart data
model = dict(
type='CustomCascadeWithMeta', # Use custom model with coordinate handling
coordinate_standardization=dict(
enabled=True,
origin='bottom_left', # Match annotation creation coordinate system
normalize=True,
relative_to_plot=False, # Keep simple for now
scale_to_axis=False # Keep simple for now
),
data_preprocessor=dict(
type='DetDataPreprocessor',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
bgr_to_rgb=True,
pad_size_divisor=32),
# ----- Swin Transformer Base (22K) Backbone + FPN -----
backbone=dict(
type='SwinTransformer',
embed_dims=128, # Swin Base embedding dimensions
depths=[2, 2, 18, 2], # Swin Base depths
num_heads=[4, 8, 16, 32], # Swin Base attention heads
window_size=7,
mlp_ratio=4,
qkv_bias=True,
qk_scale=None,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.3, # Slightly higher for more complex model
patch_norm=True,
out_indices=(0, 1, 2, 3),
with_cp=False,
convert_weights=True,
init_cfg=dict(
type='Pretrained',
checkpoint='https://download.openmmlab.com/mmsegmentation/v0.5/pretrain/swin/swin_base_patch4_window7_224_22k_20220317-4f79f7c0.pth'
)
),
neck=dict(
type='FPN',
in_channels=[128, 256, 512, 1024], # Swin Base: embed_dims * 2^(stage)
out_channels=256,
num_outs=6,
start_level=0,
add_extra_convs='on_input'
),
# Enhanced RPN with smaller anchors for tiny objects + improved losses
rpn_head=dict(
type='RPNHead',
in_channels=256,
feat_channels=256,
anchor_generator=dict(
type='AnchorGenerator',
scales=[1, 2, 4, 8], # Even smaller scales for tiny objects
ratios=[0.5, 1.0, 2.0], # Multiple aspect ratios
strides=[4, 8, 16, 32, 64, 128]), # Extended FPN strides
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[.0, .0, .0, .0],
target_stds=[1.0, 1.0, 1.0, 1.0]),
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=True,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
# Progressive Loss Strategy: Start with SmoothL1 for all 3 stages
# Stage 3 (final stage) will switch to GIoU at epoch 5 via ProgressiveLossHook
roi_head=dict(
type='CascadeRoIHead',
num_stages=3,
stage_loss_weights=[1, 0.5, 0.25],
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),
bbox_head=[
# Stage 1: Always SmoothL1Loss (coarse detection)
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=21, # 21 enhanced categories
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.05, 0.05, 0.1, 0.1]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
# Stage 2: Always SmoothL1Loss (intermediate refinement)
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=21, # 21 enhanced categories
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.033, 0.033, 0.067, 0.067]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
# Stage 3: SmoothL1 → GIoU at epoch 5 (progressive switching)
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=21, # 21 enhanced categories
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.02, 0.02, 0.05, 0.05]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))
]),
train_cfg=dict(
rpn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.7,
neg_iou_thr=0.3,
min_pos_iou=0.3,
match_low_quality=True,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=256,
pos_fraction=0.5,
neg_pos_ub=-1,
add_gt_as_proposals=False),
allowed_border=0,
pos_weight=-1,
debug=False),
rpn_proposal=dict(
nms_pre=2000,
max_per_img=2000,
nms=dict(type='nms', iou_threshold=0.8),
min_bbox_size=0),
rcnn=[
dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.4,
neg_iou_thr=0.4,
min_pos_iou=0.4,
match_low_quality=False,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
pos_weight=-1,
debug=False),
dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.6,
neg_iou_thr=0.6,
min_pos_iou=0.6,
match_low_quality=False,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
pos_weight=-1,
debug=False),
dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.7,
neg_iou_thr=0.7,
min_pos_iou=0.7,
match_low_quality=False,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
pos_weight=-1,
debug=False)
]),
# Enhanced test configuration with soft-NMS and multi-scale support
test_cfg=dict(
rpn=dict(
nms_pre=1000,
max_per_img=1000,
nms=dict(type='nms', iou_threshold=0.7),
min_bbox_size=0),
rcnn=dict(
score_thr=0.005, # Even lower threshold to catch more classes
nms=dict(
type='soft_nms', # Soft-NMS for better small object detection
iou_threshold=0.5,
min_score=0.005,
method='gaussian',
sigma=0.5),
max_per_img=500))) # Allow more detections
# Dataset settings - using cleaned annotations
dataset_type = 'ChartDataset'
data_root = '' # Remove data_root duplication
# Define the 21 chart element classes that match the annotations
CLASSES = (
'title', 'subtitle', 'x-axis', 'y-axis', 'x-axis-label', 'y-axis-label',
'x-tick-label', 'y-tick-label', 'legend', 'legend-title', 'legend-item',
'data-point', 'data-line', 'data-bar', 'data-area', 'grid-line',
'axis-title', 'tick-label', 'data-label', 'legend-text', 'plot-area'
)
# Updated to use cleaned annotation files
train_dataloader = dict(
batch_size=2, # Increased back to 2
num_workers=2,
persistent_workers=True,
sampler=dict(type='DefaultSampler', shuffle=True),
dataset=dict(
type=dataset_type,
data_root=data_root,
ann_file='legend_data/annotations_JSON_cleaned/train_enriched.json', # Full path
data_prefix=dict(img='legend_data/train/images/'), # Full path
metainfo=dict(classes=CLASSES), # Tell dataset what classes to expect
filter_cfg=dict(filter_empty_gt=True, min_size=0, class_specific_min_sizes={
'data-point': 16, # Back to 16x16 from 32x32
'data-bar': 16, # Back to 16x16 from 32x32
'tick-label': 16, # Back to 16x16 from 32x32
'x-tick-label': 16, # Back to 16x16 from 32x32
'y-tick-label': 16 # Back to 16x16 from 32x32
}),
pipeline=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(type='Resize', scale=(1600, 1000), keep_ratio=True), # Higher resolution for tiny objects
dict(type='RandomFlip', prob=0.5),
dict(type='ClampBBoxes'), # Ensure bboxes stay within image bounds
dict(type='PackDetInputs')
]
)
)
val_dataloader = dict(
batch_size=1,
num_workers=2,
persistent_workers=True,
drop_last=False,
sampler=dict(type='DefaultSampler', shuffle=False),
dataset=dict(
type=dataset_type,
data_root=data_root,
ann_file='legend_data/annotations_JSON_cleaned/val_enriched_with_info.json', # Full path
data_prefix=dict(img='legend_data/train/images/'), # All images are in train/images
metainfo=dict(classes=CLASSES), # Tell dataset what classes to expect
test_mode=True,
pipeline=[
dict(type='LoadImageFromFile'),
dict(type='Resize', scale=(1600, 1000), keep_ratio=True), # Base resolution for validation
dict(type='LoadAnnotations', with_bbox=True),
dict(type='ClampBBoxes'), # Ensure bboxes stay within image bounds
dict(type='PackDetInputs', meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'scale_factor'))
]
)
)
test_dataloader = val_dataloader
# Enhanced evaluators with debugging
val_evaluator = dict(
type='CocoMetric',
ann_file='legend_data/annotations_JSON_cleaned/val_enriched_with_info.json', # Using cleaned annotations
metric='bbox',
format_only=False,
classwise=True, # Enable detailed per-class metrics table
proposal_nums=(100, 300, 1000)) # More detailed AR metrics
test_evaluator = val_evaluator
# Add custom hooks for debugging empty results
default_hooks = dict(
timer=dict(type='IterTimerHook'),
logger=dict(type='LoggerHook', interval=50),
param_scheduler=dict(type='ParamSchedulerHook'),
checkpoint=dict(type='CompatibleCheckpointHook', interval=1, save_best='auto', max_keep_ckpts=3),
sampler_seed=dict(type='DistSamplerSeedHook'),
visualization=dict(type='DetVisualizationHook'))
# Add NaN recovery hook for graceful handling like Faster R-CNN
custom_hooks = [
dict(type='SkipBadSamplesHook', interval=1), # Skip samples with bad GT data
dict(type='ChartTypeDistributionHook', interval=500), # Monitor class distribution
dict(type='MissingImageReportHook', interval=1000), # Track missing images
dict(type='NanRecoveryHook', # For logging & monitoring
fallback_loss=1.0,
max_consecutive_nans=100,
log_interval=50),
dict(type='ProgressiveLossHook', # Progressive loss switching
switch_epoch=5, # Switch stage 3 to GIoU at epoch 5
target_loss_type='GIoULoss', # Use GIoU for stage 3 (final stage)
loss_weight=1.0, # Keep same loss weight
warmup_epochs=2, # Monitor for 2 epochs after switch
monitor_stage_weights=True), # Log stage loss details
]
# Training configuration - extended to 40 epochs for Swin Base on small objects
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=40, val_interval=1)
val_cfg = dict(type='ValLoop')
test_cfg = dict(type='TestLoop')
# Optimizer with standard stable settings
optim_wrapper = dict(
type='OptimWrapper',
optimizer=dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001),
clip_grad=dict(max_norm=35.0, norm_type=2)
)
# Extended learning rate schedule with cosine annealing for Swin Base
param_scheduler = [
dict(
type='LinearLR',
start_factor=0.05, # 1e-4 / 2e-2 = 0.05 (warmup from 1e-4 to 2e-2)
by_epoch=False,
begin=0,
end=1000), # 1k iteration warmup
dict(
type='CosineAnnealingLR',
begin=0,
end=40, # Match max_epochs
by_epoch=True,
T_max=40,
eta_min=1e-6, # Minimum learning rate
convert_to_iter_based=True)
]
# Work directory
work_dir = './work_dirs/cascade_rcnn_swin_base_40ep_cosine_fpn_meta'
# Multi-scale test configuration (uncomment to enable)
# img_scales = [(800, 500), (1600, 1000), (2400, 1500)] # 0.5x, 1.0x, 1.5x scales
# tta_model = dict(
# type='DetTTAModel',
# tta_cfg=dict(
# nms=dict(type='nms', iou_threshold=0.5),
# max_per_img=100)
# )
# Fresh start
resume = False
load_from = None