EByrdS
commented
4 years ago I am having the same issue.
I am using Google Colab Pro to replicate this project, the provided checkpoint is stored in Google, so I only needed to add a direct access to it in my Drive.
I am using:
from google.colab import drive
drive.mount('/content/drive/', force_remount=True)to access the files from the Colab.
Also, I had to do some changes in the imports, due to deprecations of imread and imresize, expecting the solution to have the exact same behavior (however, I do not think this caused the error):
# from scipy.misc import imread, imresize
from imageio import imread
from skimage.transform import resize as imresizeThe cell that created the same output as the one who oppened this issue was (I used an image that already existed in the COCO14 dataset):
%run 'a-PyTorch-Tutorial-to-Image-Captioning/caption.py' --img='imgs_of_the_wild/coco_val_001.jpg' --model='files for A PyTorch Tutorial For Image Captioning/BEST_checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar' --word_map='files for A PyTorch Tutorial For Image Captioning/WORDMAP_coco_5_cap_per_img_5_min_word_freq.json' --beam_size=3It displayed all the above warnings, and the generated sentence for the image was:
a close up of a black and white background, this output is generated for every image that I tried.
I would prefer not to alter the code that triggered these warnings before ensuring that it works correctly, otherwise I wouldn't know if my changes broke the project or it was already broken. I started to train the model by myself, but it raised another error on the image COCO_val2014_000000320612.jpg, related to its shape:
img = imread(train_image_paths[9020]) # COCO_val2014_000000320612.jpg
img.shape # => (3, 375, 500)
img = imresize(img, (256, 256))
img = img.transpose(2, 0, 1)
img.shape # => (256, 256, 500)
assert img.shape == (3, 256, 256) # Raises AssertionError because (256, 256, 500) != (3, 256, 256)I checked other images and their starting shape has the 3 at the end, like (640, 480, 3). I don't think that I generated this error on my project.
Has anyone replicated this project recently? I need to know if it works before doing any other changes that may or may not break it further.
Note: I will try solving the warnings and check if the checkpoint works correctly. But it may take some time.
shahhaard47
purveshpatel511
shellyztian
ragvri
i get error after running caption.py file:::
here i set
torch.nn.Module.dump_patches=Trueas per first error but both the time i am getting same output for every image.output is :: a close up of a black and white background
can you help me through this error and please you specifi the version of pytorch, torch, torchvision that is compitable with current source code. i think upgraded version of these packages are remove or deprecated some functions.