[SD][v4.0] RCPs to use num of samples instead iterations

mlperf_logging/rcp_checker/training_4.0.0/rcps_stable_diffusion.json

@@ -0,0 +1,66 @@
+{
+
+  "sd_ref_512":
+  {
+    "Benchmark": "sd",
+    "Creator": "NVIDIA",
+    "When": "Reference RCPs before v3.1",
+    "Platform": "32xDGX-A100",
+    "BS": 512,
+    "Hyperparams": {
+      "opt_adamw_beta_1": 0.9,
+      "opt_adamw_beta_2": 0.999,
+      "opt_adamw_epsilon": 1e-08,
+      "opt_adamw_weight_decay": 0.01,
+      "opt_base_learning_rate": 1.25e-7,
+      "opt_learning_rate_warmup_steps": 1000
+    },
+    "Epochs to converge": [
+      2560000, 2560000, 2560000, 2560000, 2560000,
+      2560000, 2560000, 2560000, 2560000, 2560000,
+      2560000, 2560000, 2560000, 3072000]
+  },
+
+  "sd_ref_1024":
+  {
+    "Benchmark": "sd",
+    "Creator": "NVIDIA",
+    "When": "Reference RCPs before v3.1",
+    "Platform": "32xDGX-A100",
+    "BS": 1024,
+    "Hyperparams": {
+        "opt_adamw_beta_1": 0.9,
+        "opt_adamw_beta_2": 0.999,
+        "opt_adamw_epsilon": 1e-08,
+        "opt_adamw_weight_decay": 0.01,
+        "opt_base_learning_rate": 1.25e-7,
+        "opt_learning_rate_warmup_steps": 1000
+    },
+    "Epochs to converge": [
+      2560000, 2560000, 2560000, 2560000, 2560000,
+      3072000, 3072000, 3072000, 3072000, 3072000,
+      3072000, 3072000, 2560000]
+  },
+
+  "sd_ref_2048":
+  {
+    "Benchmark": "sd",
+    "Creator": "NVIDIA",
+    "When": "Reference RCPs before v3.1",
+    "Platform": "32xDGX-A100",
+    "BS": 2048,
+    "Hyperparams": {
+        "opt_adamw_beta_1": 0.9,
+        "opt_adamw_beta_2": 0.999,
+        "opt_adamw_epsilon": 1e-08,
+        "opt_adamw_weight_decay": 0.01,
+        "opt_base_learning_rate": 1.25e-7,
+        "opt_learning_rate_warmup_steps": 1000
+    },
+    "Epochs to converge": [
+      3584000, 3584000, 3584000, 3584000, 4096000,
+      4096000, 4096000, 4096000, 4096000, 4096000,
+      4096000, 4608000, 4608000]
+  }
+
+}