def evaluate_fine_tune(model, df_val, top_K, random_samples, num_items):
model.eval()
avg_HR = np.zeros((len(df_val), top_K))
avg_NDCG = np.zeros((len(df_val), top_K))
for i in range(len(df_val)):
test_user_input, test_item_input = get_test_instances_with_random_samples(
df_val[i], random_samples, num_items, device)
y_hat = model(test_user_input, test_item_input)
y_hat = y_hat.cpu().detach().numpy().reshape((-1, ))
test_item_input = test_item_input.cpu().detach().numpy().reshape(
(-1, ))
map_item_score = {}
for j in range(len(y_hat)):
map_item_score[test_item_input[j]] = y_hat[j]
for k in range(top_K):
# Evaluate top rank list
ranklist = heapq.nlargest(k,
map_item_score,
key=map_item_score.get)
gtItem = test_item_input[0]
avg_HR[i, k] = getHitRatio(ranklist, gtItem)
avg_NDCG[i, k] = getNDCG(ranklist, gtItem)
avg_HR = np.mean(avg_HR, axis=0)
avg_NDCG = np.mean(avg_NDCG, axis=0)
return avg_HR, avg_NDCG
#%% evaluate
testData = Variable(testData.float())
with torch.no_grad():
avg_HR = np.zeros((len(test_features), top_K))
avg_NDCG = np.zeros((len(test_features), top_K))
for i in range(len(test_features)):
y_hat = clf_model(testData[i])
# _, predicted = torch.max(y_hat.data, 0)
for ki in range(top_K):
# Evaluate top rank list
idx = torch.topk(y_hat.data, k=ki, dim=0)[1]
ranklist = idx.tolist()
gtItem = test_careers['like_id'][i]
avg_HR[i, ki] = getHitRatio(ranklist, gtItem)
avg_NDCG[i, ki] = getNDCG(ranklist, gtItem)
avg_HR = np.mean(avg_HR, axis=0)
avg_NDCG = np.mean(avg_NDCG, axis=0)
np.savetxt('results/avg_HR_CLF.txt', avg_HR)
np.savetxt('results/avg_NDCG_CLF.txt', avg_NDCG)
#%% evaluate fairness
import sys
sys.stdout = open("dnnClf_output.txt", "w")
with torch.no_grad():
y_hat = clf_model(testData)
device = torch.device("cpu")