feat(7cam): add matching FOV30 model config

Configuration paired with the seven-camera iter-14000 checkpoint.

configs/stream_petr_r50_nucarla_7cam_front_narrow_fov30_finetune.py

@@ -0,0 +1,80 @@
+_base_ = ["./stream_petr_r50_nucarla_full_90e.py"]
+
+# Fine-tune the mature six-view nuCarla model with one extra synchronized,
+# co-located 30-degree front view. The model has no camera-count-specific
+# parameters: view count is inferred from the image and calibration tensors.
+source_data_root = "./data/nucarla_full/"
+seven_camera_info_root = "./data/nucarla_full_7cam_fov30/"
+six_camera_checkpoint = (
+    "./work_dirs/stream_petr_r50_nucarla_full_90e/"
+    "best_NuCarla_NDS_iter_116650.pth"
+)
+
+num_train_samples = 28000
+num_val_samples = 6000
+num_epochs = 15
+num_gpus = 1
+
+# Seven views at batch 8 use fewer total image activations than the verified
+# six-view batch 12 run (56 versus 72 views), leaving room for dense frames.
+batch_size = 8
+global_batch_size = num_gpus * batch_size
+num_iters_per_epoch = num_train_samples // global_batch_size
+
+data = dict(
+    samples_per_gpu=batch_size,
+    workers_per_gpu=4,
+    train=dict(
+        data_root=source_data_root,
+        ann_file=(
+            seven_camera_info_root
+            + "nucarla_full_7cam_fov30_temporal_infos_train.pkl"
+        ),
+    ),
+    val=dict(
+        samples_per_gpu=1,
+        data_root=source_data_root,
+        ann_file=(
+            seven_camera_info_root
+            + "nucarla_full_7cam_fov30_temporal_infos_val.pkl"
+        ),
+    ),
+    test=dict(
+        samples_per_gpu=1,
+        data_root=source_data_root,
+        ann_file=(
+            seven_camera_info_root
+            + "nucarla_full_7cam_fov30_temporal_infos_val.pkl"
+        ),
+    ),
+)
+
+# Conservative fine-tuning rate for the already-converged six-camera model.
+optimizer = dict(lr=1.0e-4)
+# ``load_from`` intentionally does not restore optimizer state. Start at the
+# loss scale stored in the mature six-camera checkpoint (512) instead of
+# MMCV's dynamic-scaler default (65,536), which would skip several warm-up
+# steps while backing off from predictable fp16 overflows.
+optimizer_config = dict(
+    _delete_=True,
+    type="Fp16OptimizerHook",
+    loss_scale=dict(
+        init_scale=512.0,
+        growth_factor=2.0,
+        backoff_factor=0.5,
+        growth_interval=2000,
+    ),
+    grad_clip=dict(max_norm=35, norm_type=2),
+)
+lr_config = dict(warmup_iters=500)
+runner = dict(max_iters=num_epochs * num_iters_per_epoch)
+checkpoint_config = dict(interval=num_iters_per_epoch, max_keep_ckpts=10)
+evaluation = dict(
+    interval=num_iters_per_epoch,
+    save_best="NuCarla_NDS",
+    rule="greater",
+)
+
+load_from = six_camera_checkpoint
+resume_from = None
+work_dir = "./work_dirs/stream_petr_r50_nucarla_full_7cam_fov30_finetune"