[GraphBolt] disable full neighbor sampling in reference stage for hetero rgcn example (#6396)

examples/sampling/graphbolt/rgcn/README.md

@@ -15,21 +15,21 @@ python3 hetero_rgcn.py --dataset ogbn-mag --num_gups 1
 ```
 
 ### Resource usage and time cost
-Below results are roughly collected from an AWS EC2 **g4dn.metal**, 384GB RAM, 96 vCPUs(Cascade Lake P-8259L), 8 NVIDIA T4 GPUs(16GB RAM). CPU RAM usage is the peak value of `used` and `buff/cache` field of `free` command which are a bit rough. Please refer to `RSS`/`USS`/`PSS` which are more accurate. GPU RAM usage is the peak value recorded by `nvidia-smi` command.
+Below results are roughly collected from an AWS EC2 **g4dn.metal**, 384GB RAM, 96 vCPUs(Cascade Lake P-8259L), 8 NVIDIA T4 GPUs(16GB RAM). CPU RAM usage is the peak value of `used` field of `free` command which is a bit rough. Please refer to `RSS`/`USS`/`PSS` which are more accurate. GPU RAM usage is the peak value recorded by `nvidia-smi` command.
 
 | Dataset Size | CPU RAM Usage | Num of GPUs | GPU RAM Usage | Time Per Epoch(Training) | Time Per Epoch(Inference: train/val/test set)      |
 | ------------ | ------------- | ----------- | ---------- | --------- | ---------------------------    |
-| ~1.1GB       | ~5GB          | 0           |  0GB       | ~4min5s   | ~2min7s + ~0min12s + ~0min8s   |
-| ~1.1GB       | ~4.3GB        | 1           |  4.7GB     | ~1min18s  | ~1min54s + ~0min12s + ~0min8s  |
+| ~1.1GB       | ~4.5GB        | 0           |  0GB       | ~4min14s(615it, 2.41it/s)   | ~0min28s(154it, 5.46it/s) + ~0min2s(16it, 5.48it/s) + ~0min2s(11it, 5.44it/s)   |
+| ~1.1GB       | ~2GB          | 0           |  4.4GB     | ~1min15s(615it, 8.11it/s)   | ~0min27s(154it, 5.63it/s) + ~0min2s(16it, 5.90it/s) + ~0min1s(11it, 5.82it/s)   |
 
 ### Accuracies
 ```
 Final performance: 
 All runs:
-Highest Train: 49.29 ± 0.85
-Highest Valid: 34.69 ± 0.49
-  Final Train: 48.14 ± 1.09
-   Final Test: 33.65 ± 0.63
+Highest Train: 64.66 ± 0.74
+Highest Valid: 41.31 ± 0.12
+  Final Train: 64.66 ± 0.74
+   Final Test: 40.07 ± 0.02
 ```
 
 ## Run on `ogb-lsc-mag240m` dataset
@@ -45,24 +45,19 @@ python3 hetero_rgcn.py --dataset ogb-lsc-mag240m --num_gups 1
 ```
 
 ### Resource usage and time cost
-Below results are roughly collected from an AWS EC2 **g4dn.metal**, 384GB RAM, 96 vCPUs(Cascade Lake P-8259L), 8 NVIDIA T4 GPUs(16GB RAM). CPU RAM usage is the peak value of `used` and `buff/cache` field of `free` command which are a bit rough. Please refer to `RSS`/`USS`/`PSS` which are more accurate. GPU RAM usage is the peak value recorded by `nvidia-smi` command.
+Below results are roughly collected from an AWS EC2 **g4dn.metal**, 384GB RAM, 96 vCPUs(Cascade Lake P-8259L), 8 NVIDIA T4 GPUs(16GB RAM). CPU RAM usage is the peak value of `used` field of `free` command which is a bit rough. Please refer to `RSS`/`USS`/`PSS` which are more accurate. GPU RAM usage is the peak value recorded by `nvidia-smi` command.
 
-Infer with full neighbors on GPU is out of memory on `T4(16GB RAM)``. GPUs with larger memory is required such as `A100(40GB RAM)`.
-```
-Tried to allocate 21.72 GiB (GPU 0; 14.75 GiB total capacity; 12.30 GiB already allocated; 2.02 GiB free; 12.60 GiB reserved in total by PyTorch)
-```
-
-| Dataset Size | CPU RAM Usage(used + buff/cache) | Num of GPUs | GPU RAM Usage | Time Per Epoch(Training) | Time Per Epoch(Inference: train/val/test set)      |
+| Dataset Size | CPU RAM Usage | Num of GPUs | GPU RAM Usage | Time Per Epoch(Training) | Time Per Epoch(Inference: train/val/test set)      |
 | ------------ | ------------- | ----------- | ---------- | --------- | ---------------------------    |
-| ~404GB       | ~110GB + ~250GB       | 0           |  0GB       | ~5min22s(1087it, 3.37it/s)   | ~35min29s(272it, 7.83s/it) + ~6min9s(34it, 10.87s/it) + ~3min32s(22it, 9.66s/it)   |
-| ~404GB       | ~110GB + ~250GB        | 1           |  2.7GB     | ~2min45s(1087it, 6.56it/s)  | ~OOM + ~OOM + ~OOM  |
+| ~404GB       | ~55GB       | 0           |  0GB       | ~3min51s(1087it, 4.70it/s)  | ~2min21s(272it, 1.93it/s) + ~0min22s(34it, 1.48it/s) + ~0min14s(22it, 1.51it/s)   |
+| ~404GB       | ~55GB       | 1           |  7GB       | ~2min41s(1087it, 6.73it/s)  | ~1min52s(272it, 2.41it/s) + ~0min17s(34it, 1.93it/s) + ~0min11s(22it, 1.99it/s)  |
 
 ### Accuracies
 ```
 Final performance: 
 All runs:
-Highest Train: 54.75 ± 0.29
-Highest Valid: 52.08 ± 0.09
-  Final Train: 54.75 ± 0.29
+Highest Train: 54.43 ± 0.39
+Highest Valid: 51.78 ± 0.68
+  Final Train: 54.35 ± 0.51
    Final Test: 0.00 ± 0.00
 ```
\ No newline at end of file

examples/sampling/graphbolt/rgcn/hetero_rgcn.py

@@ -443,8 +443,9 @@ def extract_node_features(name, block, data, node_embed, device):
         }
         # Original feature data are stored in float16 while model weights are
         # float32, so we need to convert the features to float32.
-        # [TODO] Enable mixed precision training on GPU.
-        node_features = {k: v.float() for k, v in node_features.items()}
+        node_features = {
+            k: v.to(device).float() for k, v in node_features.items()
+        }
     return node_features
 
 
@@ -469,18 +470,6 @@ def evaluate(
     else:
         evaluator = MAG240MEvaluator()
 
-    # Initialize a neighbor sampler that samples all neighbors. The model
-    # has 2 GNN layers, so we create a sampler of 2 layers.
-    ######################################################################
-    # [Why we need to sample all neighbors?]
-    # During the testing phase, we use a `MultiLayerFullNeighborSampler` to
-    # sample all neighbors for each node. This is done to achieve the most
-    # accurate evaluation of the model's performance, despite the increased
-    # computational cost. This contrasts with the training phase where we
-    # prefer a balance between computational efficiency and model accuracy,
-    # hence only a subset of neighbors is sampled.
-    ######################################################################
-
     data_loader = create_dataloader(
         name,
         g,
@@ -488,7 +477,7 @@ def evaluate(
         item_set,
         device,
         batch_size=4096,
-        fanouts=[-1, -1],
+        fanouts=[25, 10],
         shuffle=False,
         num_workers=num_workers,
     )