I read the files with the following codes:

import pickle as pkl

with open("val_images_png.pkl", "rb") as f:
    data = pkl.load(f)

Then I found that data is a list of (n, 1) arrays. For example,

print(data[0].shape) # (18068, 1)

I think each array corresponds to an image. However, 18068 is not even divisible by 3, which means that it is not an RGB image?

How could I convert each array to an image with shape (H, W, 3)? Thanks.

The general labels in each split seem to have an extra, corrupted(?) value that many data points take on. For instance in the training set there are 20 labels {0, ..., 19} but there is a label value 20 used by 341546 data points.

Should all of these data points be excluded, or is there a way to generate correct labels for these?

Apologies if I am confused in my understanding of your dataset. Thank you for working on a hierarchical few-shot dataset.

Here is some illustrative code from my exploration of the data:

>>> train['label_general_str']
['garment',
 'musical instrument, instrument',
 'restraint, constraint',
 'feline, felid',
 'instrument',
 'hound, hound dog',
 'electronic equipment',
 'passerine, passeriform bird',
 'ungulate, hoofed mammal',
 'aquatic bird',
 'snake, serpent, ophidian',
 'primate',
 'protective covering, protective cover, protect',
 'terrier',
 'saurian',
 'building, edifice',
 'establishment',
 'tool',
 'craft',
 'game equipment']
>>> len(train['label_general_str'])
20
>>> train['label_general'].max()  # should be 19
20
>>> uniq, count = np.unique(train['label_general'], return_counts=True)
>>> count
array([  1300,   1300,   1216,   1300,   1300,   1300,   1300,   1300,
         1300,   2600,   1300,   2449,   2600,  11700,   2590,  10258,
        13587,  13000,  24158,  11291, 341546])  # many invalid points with last value

Hi, Thankyou for sharing the code and data. I have a question about the tieredImagenet data. When I loaded the .pkl file, I found the data was a list of numpy array. But the array is not of the shape (img_size * img_size * 3) and can't not be reshaped into this form. A numpy array should be a image right? Besides, I find that the shape of images change, from near 8000 to near 19000?

Hey @renmengye, I have a doubt regarding the paper. Suppose we trained the model by taking 5 classes in each episode. I have total of 40 classes. After training, I have test data where I have to classify each image to one of the 40 categories. Now how can I do that? The most logical way is to calculate a noormalized probability for each class and then assigned an image to a class with the highest probability. But this somehow looks not the right way as we trained the model for only 5 classes. Kindly help me here Regards

Hi,

Thank you for the paper and sharing the code with community.

When running 'run_exp.py' on mini-imagenet, I get the following error and have not been able to resolve it: No such file or directory: 'data/mini-imagenet/images/n0153282900000005.jpg'

I have downloaded and saved .pkl files as instructed but this error shows that probably unpickling is not done successfully, I guess. I would appreciate if you help me through.

I am trying to reproduce your ssl-experiments and am a bit confused about the exact number of unlabelled images.

In your paper you write:

We used M = 5 for training and M = 20 for testing in most cases, thus measuring the ability of the models to generalize to a larger unlabeled set size.

In your README.md under 'Core Experiments' it says:

Add additional flags --num_unlabel 20 --num_test 20 for testing mini-imagenet and tiered-imagenet models, so that each episode contains 20 unlabeled images per class and 20 query images per class.

Could you please specify which experiments were carried out with M=5 and which with M=20. Moreover, for which experiments is there a difference in M between meta-train and meta-test time?

How to specify the number of distractors? I tried following modification, in run_exp.py line 329 meta_train_dataset = get_dataset( FLAGS.dataset, train_split_name, nclasses_train, nshot, num_distractor=NUMBER_OF_DISTRACTORS,

but the batch size of unlabelled data became nclasses_train*num_unlabel +num_unlabel*num_distractor. According your paper, the unlabelled batch size should be nclasses_train*(num_unlabel + num_distractor).

Any suggestion?

 File "/home/dd/few-shot-ssl-public/run_baselines_exp.py", line 57, in <module>
    from fewshot.data.omniglot import OmniglotEpisode
ImportError: cannot import name OmniglotEpisode

How can i use "DATASET_REGISTRY" and "CONFIG_REGISTRY" in the code for this paper "META-LEARNING FOR SEMI-SUPERVISED FEW-SHOT CLASSIFICATION"?

When I run the baseline file, it shows an error : cannot reshape array of size 0 into shape (0,newaxis) in the File "/home/xxx/few-shot/run_baselines_exp.py", line 413, in get_nn_fit x_test_ = x_test.reshape([x_test[ii].shape[0], -1]). Is the flags I set not wrong? My settings are as following, flags.DEFINE_integer("nclasses_eval", default=5, help="Number of classes for testing") flags.DEFINE_integer("nclasses_train", default=5, help="Number of classes for training") flags.DEFINE_integer("nshot", default=1, help="1 nshot") flags.DEFINE_integer("num_eval_episode", default=600, help="Number of evaluation episodes") flags.DEFINE_integer("num_test", default=-1, help="-1 Number of test images per episode") flags.DEFINE_integer("num_unlabel", default=5, help="5 Number of unlabeled for training") flags.DEFINE_integer("seed", default=0, help="Random seed")

Hi, I have a question for Omniglot dataset setting. In the paper, for test episode,

We used M = 5 for training and M = 20 for testing in most cases, thus measuring the ability of the models to generalize to a larger unlabeled set size.

But in Omniglot dataset, there are only 20 samples per class, so I think M(the number of unlabeled data in support set for one class) should be less than 20. How did you set M for experiments with Omniglot dataset?

In clustering, I don't understand this code:

# Run clustering.
for tt in range(num_cluster_steps):
      protos_1 = tf.expand_dims(protos, 2)
      protos_2 = tf.expand_dims(h_unlabel, 1)
      pair_dist = tf.reduce_sum((protos_1 - protos_2)**2, [3])  # [B, K, N]
      m_dist = tf.reduce_mean(pair_dist, [2])  # [B, K]
      m_dist_1 = tf.expand_dims(m_dist, 1)  # [B, 1, K]
      m_dist_1 += tf.to_float(tf.equal(m_dist_1, 0.0))

Does m_dist_1 += tf.to_float(tf.equal(m_dist_1, 0.0)) mean that if the distance from the center of the cluster is 1 then add 1. But why add 1?

In this repository, quite a few cvs files are provided. Which three files are the oringal split proposed in the paper and should be used for reporting performance in new papers?