update 3.28

AlchemistZoro · AlchemistZoro · commit 0f6800c07afd · 2023-03-28T22:31:27.000+08:00
diff --git a/greaselm.py b/greaselm.py
@@ -522,18 +522,18 @@ def evaluate(args, has_test_split, devices, kg):
 
 
 def get_devices(use_cuda):
-    """Get the devices to put the data and the model based on whether to use GPUs and, if so, how many of them are available."""
-    if torch.cuda.device_count() >= 2 and use_cuda:
-        [device0, device1] = nv_usage.get_gpu_index(2)
-        print("device0: {}, device1: {}".format(device0, device1))
-    elif torch.cuda.device_count() == 1 and use_cuda:
-        device0 = torch.device("cuda:0")
-        device1 = torch.device("cuda:0")
-    else:
-        device0 = torch.device("cpu")
-        device1 = torch.device("cpu")
-    # device0 = torch.device("cuda:7")
-    # device1 = torch.device("cuda:7")
+    # """Get the devices to put the data and the model based on whether to use GPUs and, if so, how many of them are available."""
+    # if torch.cuda.device_count() >= 2 and use_cuda:
+    #     [device0, device1] = nv_usage.get_gpu_index(2)
+    #     print("device0: {}, device1: {}".format(device0, device1))
+    # elif torch.cuda.device_count() == 1 and use_cuda:
+    #     device0 = torch.device("cuda:0")
+    #     device1 = torch.device("cuda:0")
+    # else:
+    #     device0 = torch.device("cpu")
+    #     device1 = torch.device("cpu")
+    [device0] = nv_usage.get_gpu_index(1)
+    device1 = device0
     return device0, device1
 
 
diff --git a/run_greaselm.sh b/run_greaselm.sh
@@ -31,6 +31,7 @@ else
   n_epochs=30
   max_epochs_before_stop=10
   ie_dim=400
+  mbs=4
 fi
 
 max_seq_len=100
diff --git a/test.md b/test.md
@@ -0,0 +1,21 @@
+2023&#24180;03&#26376;27&#26085;
+./run_greaselm.sh obqa --data_dir data/ --emp True --use_wandb True 
+./run_greaselm.sh obqa --data_dir data/ --emp False --use_wandb True 
+
+2023&#24180;03&#26376;28&#26085;
+./run_greaselm.sh csqa --data_dir data/ --emp True --use_wandb True 
+./run_greaselm.sh csqa --data_dir data/ --emp False --use_wandb True 
+
+&#30475;obqa&#30340;k&#30340;&#20540;&#30340;&#24433;&#21709;
+./run_greaselm.sh obqa --data_dir data/ --emp True --use_wandb True -k 1
+./run_greaselm.sh obqa --data_dir data/ --emp False --use_wandb True -k 1
+
+./run_greaselm.sh obqa --data_dir data/ --emp True --use_wandb True -k 3
+./run_greaselm.sh obqa --data_dir data/ --emp True --use_wandb True -k 7
+
+
+
+
+
+
+
diff --git a/utils/nv_usage.py b/utils/nv_usage.py
@@ -12,7 +12,9 @@ def get_gpu_index(gpu_number):
         handle = pynvml.nvmlDeviceGetHandleByIndex(i)
         # Get the GPU utilization
         utilization = pynvml.nvmlDeviceGetUtilizationRates(handle).gpu
-        if utilization == 0:
+        meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
+        memrate = meminfo.used/meminfo.total
+        if utilization == 0 and memrate < 0.05:
             gpu_index_list.append(i)
             if len(gpu_index_list) == gpu_number:
                 return gpu_index_list
@@ -24,7 +26,7 @@ def get_gpu_index(gpu_number):
 
 
 if __name__ == "__main__":
-    # print(get_gpu_index(1))
+    print(get_gpu_index(1))
     # print(get_gpu_index(2))
-    print(get_gpu_index(8))
+    # print(get_gpu_index(8))
     
