Adds notebooks/colab_train.ipynb covering the full pipeline (clone,
TF 2.10 pin, Drive mount, data unzip, seg + class training with
checkpoint backup to Drive) so an 8 GB local GPU isn't a blocker.

Pins TF 2.10.1 + keras 2.10 + protobuf 3.19.6 in the install cell —
Colab's bundled TF (2.15 with Keras 3) breaks `tf.keras.activations.softmax`
calls and a few other patterns this codebase relies on.

notebooks/README.md captures the data zip layout, why TF 2.10, and a
VRAM cheat sheet for the common Colab GPUs.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Restructures the notebook so the entire data prep pipeline runs in Colab
from the raw FPIC archive (~7 GB) instead of requiring a pre-zipped
processed dataset (~18 GB):

- §4 unzips data_raw.zip (pcb_image + smd_annotation)
- §5 runs create_mask.py (GPU-accelerated EDSR upscaling)
- §6 runs create_patches.py (768 px patches + 80/20 train/val split)
- §§7-10 unchanged training/eval flow with section numbers shifted

Caps full training at 40 epochs for both segmentation and classification.
Colab Pro caps a single session at 24 h with a 90-min idle limit and no
background execution; Seg 100 + Class 100 (~30-37 h) cannot fit.
Seg 40 + Class 40 fits comfortably in roughly 12 h on a T4.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The latest condacolab defaults to Python 3.11, which TF 2.10 also has no
wheels for (only 3.7–3.10). Pass python_version="3.10" so the kernel
restart lands on a Python 3.10 base that the TF 2.10 install can match.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Drop the condacolab Python 3.10 dance. Colab's default Python keeps
moving past TF 2.10's wheel matrix (now 3.11/3.12), and the latest
condacolab doesn't accept python_version on install_miniforge. TF 2.15
is the last TF release on Keras 2 (Keras 3 starts at TF 2.16) and
ships wheels for the Python versions Colab actually serves, so the
codebase's tf.keras.backend.{dot,transpose} usage keeps working with
no source changes.

Also rewrites the notebook from scratch to clean up duplicate cells
that crept in during incremental NotebookEdit changes (two ## 6 / ## 7
sections, both 100- and 40-epoch training cells, missing sanity cells).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Plain `pip install tensorflow==2.15.0` on Colab falls back to CPU because
Colab's bundled CUDA libs are pinned to whatever TF version Colab ships,
not 2.15. The `[and-cuda]` extra pulls in matching nvidia-cudnn-cu12 /
cublas-cu12 / etc. wheels alongside TF, which is what TF's GPU loader
actually expects to dlopen.

Without this, training falls back to CPU and create_mask.py / train_*.py
take ~10× longer with periodic "Cannot dlopen some GPU libraries"
warnings in stderr.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…oken)

tensorflow[and-cuda]==2.15.0 fails to resolve because the extra pins
tensorrt-libs==8.6.1, which has been removed from PyPI (only 9.x is
still available). Drop the bracket extra and install nvidia-cudnn-cu12,
nvidia-cublas-cu12, etc. by name in a separate pip call. TF needs them
at dlopen time but doesn't actually use TensorRT for training.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Colab's notebook kernel runs on Python 3.12, but TF 2.15 only ships
wheels for Python 3.9–3.11. Colab images already include
/usr/local/bin/python3.11; install the TF 2.15 stack into that
interpreter and run create_mask.py / create_patches.py /
train_*.py via !python3.11 instead of !python.

The notebook kernel itself stays on Python 3.12 — we never
import tensorflow from kernel cells, just shell-out to
python3.11 for everything that touches TF.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…cell

DISLoss's SSIM gradient backward path spikes a 416 MB tensor
([batch=16, 26 classes, 512, 512]) that fragments allocator on T4 16 GB
GPUs and OOMs even though plenty of free memory exists. TF itself
recommends `cuda_malloc_async` in this case. Add it as a prefix to
every train/eval invocation so the recommendation actually fires;
on L4 24 GB it's redundant but harmless.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

L4 Colab images don't ship python3.11 (T4 ones do). Add a guard that
installs python3.11 from deadsnakes PPA when it's missing.

Pin every nvidia-*-cu12 wheel to the version TF 2.15 expects to dlopen:
- nvidia-cudnn-cu12==8.9.4.25 (latest is 9.x; TF 2.15 needs libcudnn.so.8)
- nvidia-cublas-cu12==12.2.5.6 etc.

Without these pins TF 2.15 falls back to CPU on a fresh runtime because
it can't find the right .so versions, and the warnings are easy to miss.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

40 was too aggressive a cut from the paper's 100. 80 is the sweet
spot: enough room for ReduceLROnPlateau (patience=15) to fire and
fine-tune, while still fitting inside Colab Pro's 24 h session limit
(~9h per model on L4 = 18h total + preprocessing buffer).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The first 80-epoch segmentation run lands val_dice around 0.71 with train
dice 0.92 (clear overfit) and lr already at min_lr=1e-5. Add a second
optional stage that resumes from best_seg.h5 with a lower lr range so
ReduceLROnPlateau can keep stepping down past 1e-5.

Changes:
- train_segmentation.py: -resume CLI flag; when set, model.load_weights
  is called on the configured checkpoint path before fit().
- src/cfs/pscn_seg_finetune.yml: same architecture as pscn_seg.yml but
  lr=1e-5 (where the first run left off) and min_lr=1e-6.
- notebooks/colab_train.ipynb: new §8b that restores best_seg.h5 from
  Drive if missing, runs 20 epochs with -resume + the finetune config,
  then re-mirrors the best checkpoint.
- .gitignore: ignore /best_*.h5 and root-level *.zip (Colab artifacts
  that landed in the working tree).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Brings the fine-tune second stage into the main Colab branch so users
running the notebook always have §8b available without switching
branches.

TF 2.15 lazy-loads tf.keras, and accessing __version__ on it raises
AttributeError mid-cell, swallowing the GPU print that follows. Print
TF version + GPU list only; users who specifically need the keras
version can run it in a separate cell.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>