small changes

evaluate.py

@@ -4,7 +4,7 @@
 
 import torch
 import numpy as np
-from torch.autograd import Variable
+from collections import OrderedDict
 from src.data_loader import fetch_dataloaders
 
 
@@ -57,20 +57,6 @@ def evaluate(model, loss_fn, meta_classes, task_lr, task_type, metrics, params,
             X_sup, Y_sup = X_sup.cuda(async=True), Y_sup.cuda(async=True)
             X_que, Y_que = X_que.cuda(async=True), Y_que.cuda(async=True)
 
-        # # Adapt parameters by single gradient step
-        # Y_sup_hat = model(X_sup)
-        # loss = loss_fn(Y_sup_hat, Y_sup)
-        # optimizer.zero_grad()
-        # loss.backward()
-
-        # # follows train_single_task
-        # adapted_state_dict = model.cloned_state_dict()
-        # for key, val in model.named_parameters():
-        #     adapted_state_dict[key] = val - task_lr * val.grad
-
-        # # compute preditions for query set
-        # Y_que_hat = model(X_que, adapted_state_dict)
-
         # Direct optimization
         net_clone = copy.deepcopy(model)
         optim = torch.optim.SGD(net_clone.parameters(), lr=task_lr)
@@ -83,6 +69,27 @@ def evaluate(model, loss_fn, meta_classes, task_lr, task_type, metrics, params,
         Y_que_hat = net_clone(X_que)
         loss = loss_fn(Y_que_hat, Y_que)
 
+        # # clear previous gradients, compute gradients of all variables wrt loss
+        # def zero_grad(params):
+        #     for p in params:
+        #         if p.grad is not None:
+        #             p.grad.zero_()
+
+        # # NOTE In Meta-SGD paper, num_eval_updates=1 is enough
+        # for _ in range(num_eval_updates):
+        #     Y_sup_hat = model(X_sup)
+        #     loss = loss_fn(Y_sup_hat, Y_sup)
+        #     zero_grad(model.parameters())
+        #     grads = torch.autograd.grad(loss, model.parameters())
+        #     # step() manually
+        #     adapted_state_dict = model.cloned_state_dict()
+        #     adapted_params = OrderedDict()
+        #     for (key, val), grad in zip(model.named_parameters(), grads):
+        #         adapted_params[key] = val - task_lr * grad
+        #         adapted_state_dict[key] = adapted_params[key]
+        # Y_que_hat = model(X_que, adapted_state_dict)
+        # loss = loss_fn(Y_que_hat, Y_que)  # NOTE !!!!!!!!
+
         # extract data from torch Variable, move to cpu, convert to numpy arrays
         Y_que_hat = Y_que_hat.data.cpu().numpy()
         Y_que = Y_que.data.cpu().numpy()

experiments/base_model/params.json

@@ -2,13 +2,13 @@
     "SEED": 1,
     "dataset": "Omniglot",
     "meta_lr": 1e-3,
-    "task_lr": 0.4,
+    "task_lr": 1e-1,
     "num_episodes": 10000,
     "num_classes": 5,
-    "num_samples": 5,
+    "num_samples": 1,
     "num_query": 10,
     "num_steps": 100,
-    "num_inner_tasks": 32,
+    "num_inner_tasks": 8,
     "num_train_updates": 1,
     "num_eval_updates": 3,
     "save_summary_steps": 100,