^jH h U d Z ddlmZ ddlZddlZddlmZ ddlmZm Z m
Z
erddl m Z ej d ZdZ ddlZd Zd
ed < ed
ddZddZd dZd dZd dZd dZd dZ d!dZ d" d#dZ d$ d%dZ d&dZ d" d#dZ y# e$ r d ZY fw xY w)'u\ Post-training metrics plotting utilities.
Reads the ``metrics.csv`` written by PTL's ``CSVLogger`` (always present after a ``build_trainer``-based run) and builds
seaborn figures grouped by metric type (Loss, AP@0.50, AP@0.50:0.95, AR).
Loss panel shows aggregate and component loss scalars. AP/AR panels show all ``train/``, ``val/``, and ``test/``
columns for each group — both the base and EMA series when EMA is enabled, so both are visible in the legend.
Usage::
from rfdetr.visualize.training import plot_metrics
fig = plot_metrics("output/rfdetr_base/metrics.csv")
plt.show(fig)
) annotationsN)Path)
TYPE_CHECKINGAnyOptional)Figurez_\d+$ Tbool_IS_SEABORN_AVAILABLEF) n_columnsc | j \ }}|sy| j ||ddt |t dt | dd y)z valtesttrain ) partition)columnsplit separator metric_names r _split_metric_columnr+ : s9 $*$4$4S$9!E9kU 66k!!
v:r c | dk( ry| dk( ryy)z2Return the plotting line style for a metric split.r$ :r# z-.- )r( s r _line_style_for_splitr0 B s r c
` ddl m} |j d j j dg }i }|D ] }t
| \ } }
|
|vr!|r|t | t | z nd||
< ||
} t | } t rEddl
}
|d|g j |g }|j ru |
j |d|| d | | d
| | j |d || d
| | | y# t $ r}t d |d}~ww xY w# t $ r |
j |d|| d | | d
|
Y w xY w)u
Plot columns with color by metric name and line style by split.
When seaborn is installed, draws mean ± 1 std-dev bands computed from within-epoch step-level rows in ``raw_df``.
Falls back to epoch-averaged lines from ``epoch_df`` when seaborn is absent.
r NYmatplotlib is required for training metric plots. Install it with: pip install matplotlibzaxes.prop_cyclecolorC0epoch)subset)sdr g333333?) dataxyr errorbarr3 linestyle linewidthlabelr7 ) r8 r9 r: r cir3 r< r= r> )r= r< r3 r> )matplotlib.pyplotpyplot ImportErrorrcParamsby_keygetr+ r r0 r seaborndropnaemptylineplot TypeErrorplot)r raw_dfepoch_dfmetric_columnspltexc color_cycle
metric_colorsr' r( r* r3 r< snscol_datas r _plot_columns_on_axesrU K s ' ,, 0188:>>wKK$&M - 1& 9{ m +_jS 5G#kJZ5Z)[ptM+ & k*)%0 ! w /077x7HH>>LL %! !* #"+"% $ !
2
GG! #
M- g
@ ! LL %! #"+"% $ !
s) C* /D* D3 C??D#D-,D-c d fd }d d} j D cg c]8 } |t | s | j j s7|: }} |dd D cg c] }d|vsd|vs
| }} |dd } |ddd D cg c]
}d|vsd|v s | }} |d
d D cg c] }d |vs| } } |d
d D cg c] }d|vsd|vs
| }
} |dd } |dd } |dd }
|ddd }||||| |
| | |
|d
}|j D ci c] \ }}|s ||
c}}S c c}w c c}w c c}w c c}w c c}w c c}}w ) a Build plot groups from numeric PTL CSVLogger metrics.
Args:
df: DataFrame-like object with metric columns.
Returns:
Non-empty metric groups keyed by subplot title.
Raises:
AttributeError: If ``df`` does not provide DataFrame-like columns.
c d}j D cg c]M j | st fd| D s* j j sLO c}S c c}w )zLReturn split-prefixed columns whose name contains any of the given patterns.)train/val/test/c 3 & K | ] }|v
y wNr/ ).0pcs r z<_build_metric_groups.._split_cols.. s C]qAFC]s )columns
startswithanynotna)patternsprefixesr_ dfs `r _split_colsz)_build_metric_groups.._split_cols s] .zz
Q\\(%;C]T\C]@]bdefbgbmbmbobsbsbuA
s A)A) !A)"A)c | dv ry| j dd d j }t j | ryd|v xs |j d S ) z0Return whether a metric column is a loss scalar.> stepr5 Fr! r )maxsplitlosskp_)rsplitlower_AUXILIARY_LOSS_SUFFIX_REsearchrb )nameleafs r _is_loss_colz*_build_metric_groups.._is_loss_col sW $ ${{3{+B/557 $ + +D 1~7 !77r mAP_50
ema_mAP_50 mAP_50_95mAP_75
ema_mAP_75
ema_mAP_50_95z/AP/mARema_mAR keypoint_keypoint_map_50ema_keypoint_map_50 map_50_95map_75keypoint_map_75ema_keypoint_map_75keypoint_map_50_95ema_keypoint_map_50_95 keypoint_mARema_keypoint_mARF1 precisionrecall)
LossDetection AP@0.50Detection AP@0.50:0.95Detection AP@0.75z Detection ARKeypoint AP@0.50Keypoint AP@0.50:0.95Keypoint AP@0.75z Keypoint ARzF1 / Precision / Recall)re strreturn list[str])rs r r r
)ra r rd rc items)rg rh ru rs loss_colsr_ detection_map_50detection_map_75detection_map_50_95
detection_marr r r keypoint_marf1_precision_recall
metric_groupsra s ` r _build_metric_groupsr s
8 #%**c$ SY0GBtHNNL\L`L`LbcIc#.x #Fua+]^J^ckstctuu"8\: {OVDXYHY]hlm]m !,E9 =V1TUAUQVMV 02GHK_`L`emuvevO ""35JKO$%9;ST ~/ABL%dKB -"5 -% +!3 +##6 +M 0=/B/B/D PmdGD'M PP7 du W( QsX D6!D6D6$ D;.D;3D;
E E - E7E E
E
E
"
E-Ec h ddl }t | }|j st d| |j | }d|j vr t d t | }|j d j d j }||fS # t $ r}t d |d}~ww xY w) aR Read a PTL CSVLogger metrics file and return both step-level and epoch-averaged DataFrames.
Returns:
A ``(raw_df, epoch_df)`` pair where ``raw_df`` contains every logged row
(one per training step or validation epoch) and ``epoch_df`` is the per-epoch
mean used for column detection and log-scale checks.
r NzQpandas is required for training metric plots. Install it with: pip install pandaszmetrics.csv not found: r5 z/metrics.csv does not contain an 'epoch' column.T) numeric_only) pandasrB r existsFileNotFoundErrorread_csvra
ValueError%_drop_trailing_validation_only_epochsgroupbymean reset_index) metrics_csvpdrP csv_pathrg rL rM s r _read_metrics_csvr s x K H ?? "9( DEE X Bbjj JKK
22
6F~~g&+++>JJLH 8 xmntwwxs B B1 B,,B1c | j D cg c] }t | j d s |! }}| j D cg c] }t | j d s |! }}|r|s| S | | j j d s| S | }t | dkD r|d j d }|d |k( }|j | }|j | j r|S || j j d }|| j j d } |s| s|S |j | }t | dkD r|S c c}w c c}w )a Remove post-fit validation rows that Lightning logs as a synthetic final epoch.
The Roboflow finetune demos run ``trainer.validate(...)`` after ``trainer.fit(...)`` to write a final metrics JSON.
PTL appends that validation pass to the same ``CSVLogger`` file using ``epoch == max_epochs``. That row is useful as
a standalone final validation result, but it is not part of the training curve and can create a misleading last-
epoch jump in plots. Only trailing epochs with validation/test metrics and no training metrics are removed, and only
when the CSV also contains real training rows. Pure validation CSV files are preserved.
rX )rY rZ Naxisr r5 rl ) ra r rb rd rc r ilocr rH )
rg r'
train_columns eval_columnscleaned
last_epoch
epoch_mask
epoch_rowshas_train_metricshas_eval_metricss
r r r s\ +-**YF 8N8Nx8XVYMY)+avs6{7M7MN_7`FaLa
m " " $ ( (d ( 3 G
g,
W%**2.
W%3
[[,
;;
{ # ) )N &} 5 ; ; = A At A L%l399;??T?J $4N++zk* g,
N' Zas EE E#Ec ddl }|j d ddlm} |s t d t
| } | dk( rdnd}
| |
z dz
|
z } |j | |
dd | z fd
\ } }
|
j }t |j D ] \ }\ }}|| }t || || |j |d d
|j dd |j |d |j dd |dk( rJ|rH|| }|dk j! d r t# j$ dt& d n|j) d |dk( st+ | t- | t
| D ] }|| j/ d
| j1 |d | j3 d|v r| j5 | dk( rdnd || j7 |d d | S # t $ r}t d |d}~ww xY w)!z#Build a figure for grouped metrics.r NAggr2 zDmetrics.csv does not contain any supported non-empty metric columns.r F)figsizesqueeze
boldr
fontweightEpoch r T333333?alphar r zvloss_log_scale=True was requested, but at least one loss value is non-positive; using linear scale for the Loss panel.)
stacklevellog Q?g Q?bottom tightdpi bbox_inches)
matplotlibuser@ rA rB r r subplotsflatten enumerater rU set_title
set_xlabel
set_ylabelgridrc warningswarn UserWarning
set_yscaler range set_visiblesuptitle tight_layoutsubplots_adjustsavefig)rL rM r title output_pathloss_log_scaler rO rP n_groupsn_colsn_rowsfigaxes axes_flatidx
subplot_title metric_listr
group_datas r _plot_metric_groupsr s u ' _``=!Ha-QQF !#
.F VVb!f*5Eu UIC I-6}7J7J7L-M ))
)m[
s^ b&(K@
]RF C
g
+
mb
1
C F "~!+.Ja$$$$/
=
e$ F " $R (')* Xs9~. *#""5)* LLL$
8q=4dC KSg > J[ g
s G G:) G55G:c l t | \ }}t | }t ||d|v rd|d ini d|| S )uE Plot aggregate and component training losses from a PTL ``metrics.csv`` file.
Reads the CSV written by PyTorch Lightning's ``CSVLogger``, groups loss columns
(aggregate loss, per-component scalars, and keypoint NLL terms) into a single panel,
and renders an optional seaborn error-band overlay when seaborn is installed.
Args:
metrics_csv: Path to the ``metrics.csv`` written by PTL's ``CSVLogger``.
output_path: Optional filesystem path to save the rendered figure (PNG, PDF, …).
When ``None`` the figure is returned but not written to disk.
loss_log_scale: When ``True``, the loss y-axis uses a logarithmic scale.
Useful when loss components span several orders of magnitude.
Returns:
A ``matplotlib.figure.Figure`` containing the loss panel.
Raises:
FileNotFoundError: When ``metrics_csv`` does not exist.
ImportError: When ``pandas`` or ``matplotlib`` is not installed.
ValueError: When ``metrics_csv`` contains no loss columns.
Examples:
.. code-block:: python
from rfdetr.visualize.training import plot_loss_metrics
fig = plot_loss_metrics("output/rfdetr_small/metrics.csv")
fig = plot_loss_metrics("output/rfdetr_small/metrics.csv", output_path="loss.png", loss_log_scale=True)
r zRF-DETR Loss Metricsr r r r r r )r r r rL rM groupss r plot_loss_metricsr = sO B )5FH
!(
+F $*f$4 "$ %
r c t | \ }}t | }|j D cg c] \ }}|dv r |D ] }| }}}}|s t d t |||| S c c}}}w )u Plot train/val/test detection and keypoint mAP metrics from a PTL ``metrics.csv`` file.
Reads the CSV written by PyTorch Lightning's ``CSVLogger``, selects all AP-family
columns (AP@0.50, AP@0.75, AP@0.50:0.95 for both detection and keypoints), and renders
them in a single combined panel. EMA series are included when present so both live and
EMA metric trajectories are visible in the legend.
Args:
metrics_csv: Path to the ``metrics.csv`` written by PTL's ``CSVLogger``.
output_path: Optional filesystem path to save the rendered figure (PNG, PDF, …).
When ``None`` the figure is returned but not written to disk.
Returns:
A ``matplotlib.figure.Figure`` containing the combined mAP panel.
Raises:
FileNotFoundError: When ``metrics_csv`` does not exist.
ImportError: When ``pandas`` or ``matplotlib`` is not installed.
ValueError: When ``metrics_csv`` contains no supported mAP metric columns.
Examples:
.. code-block:: python
from rfdetr.visualize.training import plot_map_metrics
fig = plot_map_metrics("output/rfdetr_small/metrics.csv")
fig = plot_map_metrics("output/rfdetr_small/metrics.csv", output_path="map.png")
> r r r r r r zHmetrics.csv does not contain any supported non-empty mAP metric columns.)r )r r r r _plot_map_columns) r r rL rM r
group_namera r' map_columnss r plot_map_metricsr j s > )5FH(2M $1#6#6#8
J
K
cdd VX{ TT!
s A%c ddl }|j d ddlm} |j d \ }}t
|| || |j ddd
|j d d
|j dd
|j dd t | |j |j d ||j |dd |S # t $ r}t d |d}~ww xY w)z;Plot mAP metrics on a single axes with line style by split.r Nr zTmatplotlib is required for mAP metric plots. Install it with: pip install matplotlib)r )r zRF-DETR mAP Metricsr r r r r r mAPTr r r r r r r )r r r@ rA rB r rU r r r r r r r r ) rL rM rN r r rO rP r r s r r r s u ' ll7l+GC"fh?LL&LGMM'BM'MM%"M%GGDG R t$ KSg > J# b
s C C, C''C,c V t | \ }}t | }t |||d|| S )u Read a PTL ``CSVLogger`` metrics file and build a training plot.
The figure contains one subplot per metric group (loss, detection metrics,
keypoint metrics, and F1/precision/recall), arranged in a 2-column grid.
Only groups with at least one non-NaN column are shown.
When seaborn is installed, each series is drawn as mean ± 1 std-dev band
computed from the within-epoch step-level rows logged by PTL's
``CSVLogger``. Metrics recorded only once per epoch (e.g. ``val/mAP_50``)
show a plain line because their per-epoch std is zero. When seaborn is
absent the plot falls back to epoch-averaged lines.
Args:
metrics_csv: Path to the ``metrics.csv`` file produced by
``CSVLogger``.
output_path: Optional destination for the PNG file. If omitted, the
figure is returned without saving.
loss_log_scale: If ``True``, use a logarithmic y-axis for the Loss
panel when all loss values are positive.
Returns:
The matplotlib figure. The figure is left open so notebook cells can
display it inline.
Raises:
ImportError: If ``matplotlib`` or ``pandas`` are not installed.
FileNotFoundError: If ``metrics_csv`` does not exist.
zRF-DETR Training Metricsr r )r r r rL rM r s r plot_metricsr s= B )5FH(2M ( %
r )r r r intr None)r' r r ztuple[str, str])r( r r r )
r r rL r rM r rN r r r )rg r r dict[str, list[str]])r r r ztuple[Any, Any])rg r r r )rL r rM r r r r r r
Optional[str]r r
r r )NF)r r r r
r r
r r r\ )r r r r
r r )
rL r rM r rN r r r
r r ) __doc__
__future__r rer pathlibr typingr r r matplotlib.figurer compilerq _LEGEND_COLUMNSrF r __annotations__rB r r+ r0 rU r r r r r r r r r/ r r r sj
# / /( &BJJx 0 ""&4&
;J
=@8Qv 2 >< << (<
<
< < <B "& ** * * *^ "&1U1U 1U 1Uh: "& ** * * *g " !"s B' 'B10B1