First of all I would like to say huge thank you to aicrowd for this unique and super fun challenge! Also congrats to other winners! I consider myself very lucky to land the first place and would love to share my solution and learnings here!
My solution is very simple and straightforward. It is basically “iteratively purchase the next batch of data with the best possible model until run out of purchase budget”. One of the biggest challenges for this competition, imho, is that you cannot get very reliable performance scores locally or on public leaderboard. So if we can filter out more noise from the weak signal of the scores, the chance of overfitting may be much lower. And during my experiments, I was more focused on simple strategies, mainly because more complex strategies require more tuning which means more decisions to make, and higher risk of overfitting (since everytime when making a decision, we may like to refer to the same local and public scores, over and over again).
OK, enough hypothesis and high level talk! Here’s the details (code):
Most importantly, the purchase strategy:
def decide_what_to_purchase(probpred_ary, purchased_labels, num_labels):
"""purchase strategy given the predicted probabilities"""
oneminusprob = 1 - probpred_ary
topk_prob_ind = np.argsort((oneminusprob * np.log(oneminusprob) + probpred_ary * np.log(probpred_ary)).mean(axis=1))
topk_prob_ind = [x for x in topk_prob_ind if x not in set(purchased_labels)][:num_labels]
return set(topk_prob_ind)
basically, select the most uncertain samples based on entropy.
And for each iteration, number of labels to purchase is decided on the fly given the compute and purchase budget:
def get_iteration_numbers(self, purchase_budget, compute_budget):
for ratio in range(30, 100, 5):
ratio /= 100
num_labels_list = self.get_purchase_numbers(purchase_budget, ratio=ratio)
for epochs in ZEWDPCBaseRun.generate_valid_epoch_comb(num_labels_list):
epoch_time = self.calculate_total_time_given_epochs(epochs, num_labels_list)
#print("ratio and time takes!", ratio, epoch_time)
if epoch_time <= compute_budget:
print(f"settle with ratio {ratio}")
return num_labels_list, epochs
return [purchase_budget], [10]
def get_purchase_numbers(self, purchase_budget, ratio):
start_number = int(1000 * ratio)
if start_number >= purchase_budget:
return [purchase_budget]
num_labels_list = [start_number]
remain_budget = purchase_budget - start_number
while remain_budget > 0:
label_to_purchase = min(remain_budget, int(ratio * 1000))
remain_budget -= label_to_purchase
num_labels_list.append(label_to_purchase)
return num_labels_list
basically, we try to see if we can purchase 300 images for each iteration and get the purchase budget exhausted before we run out of time. If not, then we increase it to 350 images (so fewer iterations), and see if that works. And then increase to 400 images… And we do it for each iteration and only take the first element of the purchase list generated by the strategy. Namely, we may have decided to purchase 300 images for each iteration last round, and may increase that to 400 images this iteration. Mainly because we couldn’t accurately estimate the exact time it may take to train the next iteration model so would like to re-estimate each time if we can still finish in time. In fact, I did a moving average (with a extra 0.1 time buffer) to estimate how long it may take to train the next iteration.
self.train_time_per_image = self.train_time_per_image * 0.8 + (train_end_time - train_start_time) / len(tr_dset) / num_epochs * 0.3
Now within each iteration, we need to train the model.
The model I ended up with is
def load_model():
"""load a pretrained model"""
model = models.regnet_y_8gf(pretrained=True)
model.fc = nn.Linear(2016, 6)
return model.to(device)
basically, the most complex model that is still reasonable to train.
Like any other computer vision problems, data augmentation is also very key:
self.train_transform = transforms.Compose([
transforms.ToTensor(),
transforms.ColorJitter(brightness=0.5, contrast=0.8, saturation=0.5, hue=0.5),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomVerticalFlip(p=0.5),
transforms.RandomInvert(p=0.25),
transforms.RandomAdjustSharpness(sharpness_factor=2, p=0.25),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
self.test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
Was trying to also do test time augmentation but I found that prediction takes too much time, and it might not be worth it.
And the optimizer training scheduler:
scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=len(dataloader)*NEPOCH, T_mult=1)
NEPOCH was a tuning parameter, I tried 5, 7, 10, 15. 5 or 7 didn’t seem to be enough, 15 seemed to be a bit too much, and 10 seemed to be pretty good.
So baiscally, the flow works like this:
- train the model, with data aug, 10 epochs, and 1 cycle of cosine annealing lr sheduler
- decide how many images to purchase based on compute and purchase budget
- do one round of prediction to get probabilities and then purchase the most uncertain images based on entropy
- collect the just purchased images, further train the model (load the model and optimizer checkpoints)
- repeat until no purchase budget left
Hopefully this is helpful! And please ask any questions if you have any!
)