Add global average pooling and multiply to the raster lane, validated live

README.md

@@ -75,7 +75,7 @@ graphs byte for byte; numerical and live evidence have narrower scope, detailed
 | SMP DPT + timm ViT-B/16, 224 × 224 | [DPT conversion](https://github.com/noelEOS/tensor-ee/blob/main/docs/converter.md#smp-050-dpt-at-224--224) | Random weights; full-width graph identity; full-width logits meet the 1e-4 target offline only; live evidence covers reduced twins |
 
 You can also compose a model from the [block catalogue](https://github.com/noelEOS/tensor-ee/blob/main/docs/building-blocks.md).
-It has 88 blocks: 23 in `tensor_ee/blocks/raster` and 65 in `tensor_ee/blocks/window`.
+It has 90 blocks: 25 in `tensor_ee/blocks/raster` and 65 in `tensor_ee/blocks/window`.
 Recent additions include grouped/depthwise convolutions, bilinear upsampling,
 standalone BatchNorm and input standardization. These expand the building blocks;
 whole-model support still depends on the compiler or converter.

docs/api.md

@@ -85,6 +85,20 @@ field reports local bounds but refuses a halo guarantee for tiled execution.
 Merges require identical spatial grids, even when dimensions
 happen to match; residual addition rejects broadcasting.
 
+Global average pooling (`nn.AdaptiveAvgPool2d` and `F.adaptive_avg_pool2d` to output size
+1, `(1, 1)` or `[1, 1]`, and `torch.mean`/`Tensor.mean` over dims `(2, 3)` with
+`keepdim=True` and no `dtype`) uses the `global_avg_pool2d` block: the mean of the whole
+compiled grid, a `[N, C, 1, 1]` tensor on a one-pixel grid that covers exactly the map.
+1 × 1 convolutions and pointwise activations run on that grid unchanged. Elementwise
+multiplication (`*`, `torch.mul`, `Tensor.mul`) uses the `multiply` block: two tensors on
+one grid with equal channel counts, or a map and a one-pixel gate pooled from that same map,
+in either operand order (squeeze-and-excitation's `sigmoid(g) * x`). Adaptive pools to other
+sizes, max pools to one pixel, `mean` without `keepdim` or over other dims, scalar products
+and channel broadcasts (`[N, 1, H, W]`) are refused. A global mean depends on the whole chip:
+the result is per compiled grid, `analyze_receptive_field` lists it in `extent_nodes` and
+refuses a halo, and tiled compilation refuses the model. Both blocks were validated live at
+the map sizes their catalogue rows list; no whole model that uses them has been.
+
 Model and submodules must already be in eval mode. Computational weights,
 biases and BatchNorm statistics are finite float32 with declared shapes;
 nonfloating bookkeeping counters are allowed. BatchNorm folds into an unshared

docs/building-blocks.md

@@ -101,18 +101,18 @@ their own.
 ## Blocks
 
 <!-- BEGIN GENERATED: blocks (tools/block_catalogue.py; edit the code, then rerun) -->
-88 blocks: 23 in `tensor_ee/blocks/raster` and 65 in `tensor_ee/blocks/window`.
+90 blocks: 25 in `tensor_ee/blocks/raster` and 65 in `tensor_ee/blocks/window`.
 
 | Layout in → out | Ready | Fixed | Total |
 |---|---:|---:|---:|
-| R → R | 23 | 0 | 23 |
+| R → R | 25 | 0 | 25 |
 | R → W | 2 | 1 | 3 |
 | W → R | 0 | 1 | 1 |
 | W → W | 23 | 33 | 56 |
 | W → V | 2 | 0 | 2 |
 | V → W | 1 | 0 | 1 |
 | V → V | 1 | 1 | 2 |
-| **All** | **52** | **36** | **88** |
+| **All** | **54** | **36** | **90** |
 
 ### Raster-layout (R) blocks
 
@@ -136,11 +136,13 @@ their own.
 | [`conv2d`](../tensor_ee/blocks/raster/conv.py) | R → R | k×k, radius 3 | `float32_operands` | no | Ready | [cnn/results/parity.json](history/cnn/results/parity.json); [cnn/results/primitives_float32.json](history/cnn/results/primitives_float32.json); [cnn/results/resnet_primitives.json](history/cnn/results/resnet_primitives.json); [cnn/results/resnet18_full64_float32_policy.json](history/cnn/results/resnet18_full64_float32_policy.json); [cnn/results/segwater_js_parity_1x1.json](history/cnn/results/segwater_js_parity_1x1.json); [cnn/results/tiled_resnet18_2x2_parity.json](history/cnn/results/tiled_resnet18_2x2_parity.json) |
 | [`concatenate`](../tensor_ee/blocks/raster/merge.py) | R → R | pointwise | `float32_operands` | no | Ready | [cnn/results/parity.json](history/cnn/results/parity.json); [cnn/results/resnet18_full64_float32_policy.json](history/cnn/results/resnet18_full64_float32_policy.json); [cnn/results/segwater_js_parity_1x1.json](history/cnn/results/segwater_js_parity_1x1.json) |
 | [`residual_add`](../tensor_ee/blocks/raster/merge.py) | R → R | pointwise | `float32_operands` | no | Ready | [cnn/results/resnet_primitives.json](history/cnn/results/resnet_primitives.json); [cnn/results/resnet18_full64_float32_policy.json](history/cnn/results/resnet18_full64_float32_policy.json); [cnn/results/segwater_js_parity_1x1.json](history/cnn/results/segwater_js_parity_1x1.json) |
+| [`multiply`](../tensor_ee/blocks/raster/merge.py) | R → R | pointwise | `float32_operands` | no | Ready — same grid on a 5 × 7 map; one-pixel gate on 5 × 7 and 12 × 20 maps | [evidence/raster-global-pool-multiply-offline.json](evidence/raster-global-pool-multiply-offline.json); [evidence/live-fixtures/20261112T151417Z-global_pool_multiply.json](evidence/live-fixtures/20261112T151417Z-global_pool_multiply.json) |
 | [`batch_norm`](../tensor_ee/blocks/raster/normalization.py) | R → R | pointwise | `float32_operands` | no | Ready | [evidence/raster-batch-norm-offline.json](evidence/raster-batch-norm-offline.json); [evidence/live-fixtures/20261029T070703Z-batch_norm.json](evidence/live-fixtures/20261029T070703Z-batch_norm.json) |
 | [`standardize`](../tensor_ee/blocks/raster/normalization.py) | R → R | pointwise | `round_each_op` | no | Ready — 6 bands with scale 1e-4; 3 bands without scale | [evidence/raster-standardize-offline.json](evidence/raster-standardize-offline.json); [evidence/live-fixtures/20261030T152829Z-standardize.json](evidence/live-fixtures/20261030T152829Z-standardize.json) |
 | [`segmentation_outputs`](../tensor_ee/blocks/raster/outputs.py) | R → R | pointwise | `float32_operands` | no | Ready | [evidence/t03-outputs.json](evidence/t03-outputs.json); [evidence/t05-primary.json](evidence/t05-primary.json) |
 | [`binary_threshold`](../tensor_ee/blocks/raster/outputs.py) | R → R | pointwise | `float32_operands` | no | Ready | [evidence/t03-outputs.json](evidence/t03-outputs.json) |
 | [`pool2d`](../tensor_ee/blocks/raster/pooling.py) | R → R | k×k, radius 3 | `float32_operands` | no | Ready — aligned avg/max k = 2, 3, 4; max 3×3 stride 1 or 2, padding 1; avg 3×3 stride 1 padding 1, both divisor rules | [evidence/live-fixtures/20261013T065257Z-p02.json](evidence/live-fixtures/20261013T065257Z-p02.json); [evidence/p02-raster-offline.json](evidence/p02-raster-offline.json); [cnn/results/parity.json](history/cnn/results/parity.json); [evidence/t05-primary.json](evidence/t05-primary.json); [cnn/results/resnet_primitives.json](history/cnn/results/resnet_primitives.json); [cnn/results/resnet18_full64_float32_policy.json](history/cnn/results/resnet18_full64_float32_policy.json); [cnn/results/segwater_js_parity_1x1.json](history/cnn/results/segwater_js_parity_1x1.json) |
+| [`global_avg_pool2d`](../tensor_ee/blocks/raster/pooling.py) | R → R | extent | `float32_operands` | no | Ready — maps 5 × 7, 8 × 8, 12 × 20, 64 × 64 and 128 × 128 | [evidence/raster-global-pool-multiply-offline.json](evidence/raster-global-pool-multiply-offline.json); [evidence/live-fixtures/20261112T151417Z-global_pool_multiply.json](evidence/live-fixtures/20261112T151417Z-global_pool_multiply.json) |
 | [`bilinear_upsample`](../tensor_ee/blocks/raster/resample.py) | R → R | k×k, radius 1 | `float32_operands` | yes (by setting) | Ready — factors 2, 3, 4 (align_corners=False), 2, 4 (align_corners=True) | [evidence/raster-bilinear-offline.json](evidence/raster-bilinear-offline.json); [evidence/live-fixtures/20261029T094157Z-bilinear.json](evidence/live-fixtures/20261029T094157Z-bilinear.json) |
 | [`nearest_upsample`](../tensor_ee/blocks/raster/resample.py) | R → R | pointwise | `float32_operands` | no | Ready — factors 2, 3, 4 | [evidence/live-fixtures/20261013T065257Z-p02.json](evidence/live-fixtures/20261013T065257Z-p02.json); [evidence/p02-raster-offline.json](evidence/p02-raster-offline.json); [cnn/results/parity.json](history/cnn/results/parity.json); [cnn/results/resnet18_full64_float32_policy.json](history/cnn/results/resnet18_full64_float32_policy.json); [cnn/results/segwater_js_parity_1x1.json](history/cnn/results/segwater_js_parity_1x1.json) |
 
@@ -224,7 +226,7 @@ their own.
 
 ### Validated settings
 
-Each block's `validated` record, the status in machine-readable form: 88 records. Each `{…}` is one configuration, whose values were validated together; a record is their union (∪). A setting no configuration names is not restricted. Names come from the block function, its cost functions and the sources listed.
+Each block's `validated` record, the status in machine-readable form: 90 records. Each `{…}` is one configuration, whose values were validated together; a record is their union (∪). A setting no configuration names is not restricted. Names come from the block function, its cost functions and the sources listed.
 
 | Block | Layout | Validated settings | Names from |
 |---|---|---|---|
@@ -244,11 +246,13 @@ Each block's `validated` record, the status in machine-readable form: 88 records
 | [`conv2d`](../tensor_ee/blocks/raster/conv.py) | R → R | any accepted call | — |
 | [`concatenate`](../tensor_ee/blocks/raster/merge.py) | R → R | {dim 1} | — |
 | [`residual_add`](../tensor_ee/blocks/raster/merge.py) | R → R | any accepted call | — |
+| [`multiply`](../tensor_ee/blocks/raster/merge.py) | R → R | {clip_footprints False} | — |
 | [`batch_norm`](../tensor_ee/blocks/raster/normalization.py) | R → R | {clip_footprints False; eps 1e-05} | `BatchNormConstants` |
 | [`standardize`](../tensor_ee/blocks/raster/normalization.py) | R → R | {channels 6; clip_footprints False; scale 0.0001} ∪ {channels 3; clip_footprints False; scale None} | `StandardizeConstants` |
 | [`segmentation_outputs`](../tensor_ee/blocks/raster/outputs.py) | R → R | any accepted call | — |
 | [`binary_threshold`](../tensor_ee/blocks/raster/outputs.py) | R → R | any accepted call | — |
 | [`pool2d`](../tensor_ee/blocks/raster/pooling.py) | R → R | {kernel 2; padding 0; reducer 'mean' or 'max'; stride 2} ∪ {kernel 3; padding 0; reducer 'mean' or 'max'; stride 3} ∪ {kernel 4; padding 0; reducer 'mean' or 'max'; stride 4} ∪ {kernel 3; padding 1; reducer 'max'; stride 1 or 2} ∪ {count_include_pad True or False; kernel 3; padding 1; reducer 'mean'; stride 1} | `PoolConfig` |
+| [`global_avg_pool2d`](../tensor_ee/blocks/raster/pooling.py) | R → R | {clip_footprints False} | — |
 | [`bilinear_upsample`](../tensor_ee/blocks/raster/resample.py) | R → R | {align_corners False; clip_footprints False; factor 2 or 3 or 4} ∪ {align_corners True; clip_footprints False; factor 2 or 4} | `BilinearConstants` |
 | [`nearest_upsample`](../tensor_ee/blocks/raster/resample.py) | R → R | {scale_factor 2 or 3 or 4} | — |
 | [`gelu`](../tensor_ee/blocks/window/activations.py) | W → W | any accepted call | — |
@@ -440,7 +444,7 @@ generated list cannot drift from the code.
 | Layer | Module | What it holds | Public names |
 |---|---|---|---|
 | Core | [`tensor_ee.core.errors`](../tensor_ee/core/errors.py) | Errors shared by every layer of the library. | `UnsupportedOperation`, `GridMismatch`, `SemanticUnknown`, `reject` |
-| Core | [`tensor_ee.core.grid`](../tensor_ee/core/grid.py) | The three grid kinds behind one protocol, with explicit conversions. | `GridProtocol`, `Grid`, `TokenGrid`, `FeatureGrid`, `to_grid`, `to_feature_grid`, `to_token_grid`, `check_model_grid` |
+| Core | [`tensor_ee.core.grid`](../tensor_ee/core/grid.py) | The three grid kinds behind one protocol, with explicit conversions. | `GridProtocol`, `Grid`, `one_pixel`, `TokenGrid`, `FeatureGrid`, `to_grid`, `to_feature_grid`, `to_token_grid`, `check_model_grid` |
 | Core | [`tensor_ee.core.metadata`](../tensor_ee/core/metadata.py) | Locality, cost and validated-settings metadata for library blocks (docs/library-contract.md, section 9, and its "As built (Phase 3)" section). | `Locality`, `Between`, `Validated`, `SettingsCheck`, `check_settings`, `describe_validated`, `BlockMetadata`, `validate_metadata`, `metadata_at` |
 | Core | [`tensor_ee.core.ops`](../tensor_ee/core/ops.py) | The ops interface: one set of array primitives, two implementations. | `Ops`, `image_reshape`, `ArrayOps`, `ImageOps`, `BlockArrayOps`, `BlockImageOps`, `ValueOps`, `WindowOps` |
 | Core | [`tensor_ee.core.precision`](../tensor_ee/core/precision.py) | Precision policies, and how float32 values enter an Earth Engine graph. | `validate_policy`, `exact_scalar`, `encode_float32`, `constant_value` |