Corrected the code #19
Conversation
|
@take2rohit , hi! Have you managed to reproduce the results of the paper (or at least close to it)? I tried several times in this repository and tried to integrate this approach (ML GCN model) into my project, but the model hasn't been trained (VOC dataset, the metric is about 50-60 mAP) |
|
Hi @kprokofi . I haven't tried reproducing results. I was just using this code as a boilerplate for other code. |
Hi @kprokofi , did you manage to reproduce the results afterwards? |
I tried to reproduce the result, but the mAP is noly about 10. Could you provide me with the training command? |
|
https://github.com/kprokofi/ML-GCN - I couldn't reproduce the author's result, but it is got better. 93+ mAP |
Hi @kprokofi, Any advise apart from the standard configs that you adopt? Appreciate advices, thanks. |
Gradient clipping was the bottleneck. Also, you could play with it and the learning rate. |
Could you provide the command for training? Thanks. |
|
pretrained=True、lr=0.01，I got 93.4 mAP |
|
Hi, I tried to reproduce the VOC2007 project but found that if I train the model from scratch, the mAP will rise to 17 and keep still after 50 epochs. However, if I use the pre-trained model, and using the command provided from Github. The mAP will reach 90 in about 5 epochs quickly. |
excuse me ，i have tried to run this code recently but there are some problem when i run this on my computer can you help me？ |
hi，i have tried to run this code recently but there are some problem when i run this on my computer can you help me？ |
I was running this on the server, what kind of problem did you meet? |
pretrained=True |
Appreciate! Now the training process looks good! |
In VOC2007, ResNet101+GMP also achieves desirable results(93.*) |
|
Hi, How could I load my own dataset to test with the pre-trained model? Thanks. |
|
Refer to this repository for more information. |
thanks for your help. |
|
Hello, I am very interested in this project and would like to know how to train and test on my own data set and finally output visual results |
How to set pretrained = True |
def gcn_resnet101(num_classes, t, pretrained=False, adj_file=None, in_channel=300):
model = models.resnet101(pretrained=True) # set pretrained = True
return GCNResnet(model, num_classes, t=t, adj_file=adj_file, in_channel=in_channel) |
Thank you so much！ |
|
|
@take2rohit 我想问一下，作者的源文件没有train文件，你有吗 可以分享给我一下吗 我想尝试着复现这篇论文！谢谢 |
你可以看看 https://github.com/yu-gi-oh-leilei/ML-GCN_cvpr2019/tree/main 里的trainer和main |
如果有复现代码的问题可以问我，就在那个项目下面。 |
请问环境条件是什么，我打算在服务器上尝试复现 |
|
How to create this file by our self? [coco_glove_word2vec.pkl) |
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Created by: jasonseu
# Created on: 2022-3-22
# Email: [email protected]
#
# Copyright © 2022 - CPSS Group
# https://github.com/jasonseu/SALGL/blob/main/scripts/embedding.py
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
import os
import argparse
import numpy as np
import torch
from transformers import BertTokenizer, BertModel
model = BertModel.from_pretrained('bert-base-uncased', output_hidden_states=True)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
token_mapping = {
'diningtable': 'dining table',
'pottedplant': 'potted plant',
'tvmonitor': 'tv monitor'
}
def bert_text_preparation(text):
marked_text = "[CLS] " + text
tokenized_text = tokenizer.tokenize(marked_text)
indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
segments_ids = [1] * len(indexed_tokens)
tokens_tensor = torch.tensor([indexed_tokens])
segments_tensors = torch.tensor([segments_ids])
return tokenized_text, tokens_tensor, segments_tensors
def get_bert_embeddings(tokens_tensor, segments_tensors):
"""Get embeddings from an embedding model
Args:
tokens_tensor (obj): Torch tensor size [n_tokens]
with token ids for each token in text
segments_tensors (obj): Torch tensor size [n_tokens]
with segment ids for each token in text
model (obj): Embedding model to generate embeddings
from token and segment ids
Returns:
list: List of list of floats of size
[n_tokens, n_embedding_dimensions]
containing embeddings for each token
"""
# Gradient calculation id disabled
# Model is in inference mode
with torch.no_grad():
outputs = model(tokens_tensor, segments_tensors)
# Removing the first hidden state
# The first state is the input state
hidden_states = outputs[2][1:]
# Getting embeddings from the final BERT layer
token_embeddings = hidden_states[-1]
# Collapsing the tensor into 1-dimension
token_embeddings = torch.squeeze(token_embeddings, dim=0)
# Converting torchtensors to lists
list_token_embeddings = [token_embed.tolist() for token_embed in token_embeddings]
return list_token_embeddings
def main(data):
label_path = os.path.join('data', data, 'label.txt')
labels = [t.strip() for t in open(label_path)]
labels = [token_mapping[t] if t in token_mapping else t for t in labels]
# labels = ['an image of {}'.format(t) for t in labels]
bert_embeddings = []
for t in labels:
_, tokens_tensor, segments_tensors = bert_text_preparation(t)
list_token_embeddings = get_bert_embeddings(tokens_tensor, segments_tensors)
token_embedding = list_token_embeddings[1] if len(list_token_embeddings) == 2 else list_token_embeddings[0]
bert_embeddings.append(token_embedding)
# bert_embeddings.append(list_token_embeddings[0])
bert_embeddings = np.array(bert_embeddings)
np.save(os.path.join('data', data, 'bert.npy'), bert_embeddings)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, default='voc2007')
args = parser.parse_args()
main(args.data) |
你可以参考issue。 |
Hi Dr, thank you for your reply, but what is label.txt? Can you explain more about this 2 files How to generate or make this file by ourself? Thanks |
A1: Use "glove" to convert label_name into a vector, and you can get coco_glove_word2vec. A2: “coco_adj.pkl” is the co-occurrence matrix, which is obtained by counting the training set labels of COCO. I'm in the China time zone, at 2:30 PM. If you have any questions, please let me know. Get the train label.import json
import os
import argparse
import numpy as np
pp = argparse.ArgumentParser(description='Format COCO2014 metadata.')
pp.add_argument('--load-path', type=str, default='.', help='Path to a directory containing a copy of the COCO2014 dataset.')
pp.add_argument('--save-path', type=str, default='.', help='Path to output directory.')
args = pp.parse_args()
def parse_categories(categories):
category_list = []
id_to_index = {}
for i in range(len(categories)):
category_list.append(categories[i]['name'])
id_to_index[categories[i]['id']] = i
return (category_list, id_to_index)
# initialize metadata dictionary:
meta = {}
meta['category_id_to_index'] = {}
meta['category_list'] = []
for split in ['train', 'val']:
with open(os.path.join(args.load_path, 'annotations', 'instances_' + split + '2014.json'), 'r') as f:
D = json.load(f)
if len(meta['category_list']) == 0:
# parse the category data:
(meta['category_list'], meta['category_id_to_index']) = parse_categories(D['categories'])
else:
# check that category lists are consistent for train2014 and val2014:
(category_list, id_to_index) = parse_categories(D['categories'])
assert category_list == meta['category_list']
assert id_to_index == meta['category_id_to_index']
image_id_list = sorted(np.unique([str(D['annotations'][i]['image_id']) for i in range(len(D['annotations']))]))
image_id_list = np.array(image_id_list, dtype=int)
# sorting as strings for backwards compatibility
image_id_to_index = {image_id_list[i]: i for i in range(len(image_id_list))}
num_categories = len(D['categories'])
num_images = len(image_id_list)
label_matrix = np.zeros((num_images,num_categories))
image_ids = np.zeros(num_images)
for i in range(len(D['annotations'])):
image_id = int(D['annotations'][i]['image_id'])
row_index = image_id_to_index[image_id]
category_id = int(D['annotations'][i]['category_id'])
category_index = int(meta['category_id_to_index'][category_id])
label_matrix[row_index][category_index] = 1
image_ids[row_index] = int(image_id)
image_ids = np.array(['{}2014/COCO_{}2014_{}.jpg'.format(split, split, str(int(x)).zfill(12)) for x in image_ids])
# save labels and corresponding image ids:
np.save(os.path.join(args.save_path, 'formatted_' + split + '_labels.npy'), label_matrix)
np.save(os.path.join(args.save_path, 'formatted_' + split + '_images.npy'), image_ids)Get the coco_adj.pklimport os
import pickle
import numpy as np
# _adj = np.array(result['adj'])
# _nums = np.array(result['nums'])
def getCoOccurrenceLabel(path, mode, data_name=None):
object_dit = {}
assert mode in ('train', 'val')
if mode == 'train':
label_path = os.path.join(path, 'formatted_train_labels.npy')
else:
label_path = os.path.join(path, 'formatted_val_labels.npy')
labels = np.load(label_path).astype(np.float64)
num_sample = labels.shape[0]
num_label = labels.shape[1]
# print(num_sample, num_label)
nums_matrix = np.zeros(shape=(num_label), dtype=np.int64)
coOccurrencegraph = np.zeros((labels.shape[1], labels.shape[1]), dtype=np.int64)
for index in range(num_sample):
data = labels[index]
for i in range(num_label):
if data[i] == 1:
nums_matrix[i] += 1
for j in range(num_label):
if j != i:
if data[j] == 1:
coOccurrencegraph[i][j] += 1
object_dit.update({'nums': nums_matrix})
object_dit.update({'adj': coOccurrencegraph})
# print(object_dit['nums'])
return object_dit
# np.save('./data/coco/{}_co-occurrence_label_vectors.npy'.format(mode), coOccurrenceLabel)
def main():
root_path = '/media/data2/maleilei/MLIC/DDP-VTPMOD/data'
path = os.path.join(root_path, 'vg256')
object_dit = getCoOccurrenceLabel(path=path, mode='train')
with open('./pkl_and_json/vg256_adj.pkl', 'wb') as file:
pickle.dump(object_dit, file)
path = os.path.join(root_path, 'voc2007')
object_dit = getCoOccurrenceLabel(path=path, mode='train')
with open('./pkl_and_json/voc2007_adj.pkl', 'wb') as file:
pickle.dump(object_dit, file)
path = os.path.join(root_path, 'nus')
object_dit = getCoOccurrenceLabel(path=path, mode='train')
with open('./pkl_and_json/nus_adj.pkl', 'wb') as file:
pickle.dump(object_dit, file)
path = os.path.join(root_path, 'coco')
object_dit = getCoOccurrenceLabel(path=path, mode='train')
with open('./pkl_and_json/coco_adj.pkl', 'wb') as file:
pickle.dump(object_dit, file)
path = os.path.join(root_path, 'cub')
object_dit = getCoOccurrenceLabel(path=path, mode='train')
with open('./pkl_and_json/cub_adj.pkl', 'wb') as file:
pickle.dump(object_dit, file)
path = os.path.join(root_path, 'objects365')
object_dit = getCoOccurrenceLabel(path=path, mode='train')
with open('./pkl_and_json/objects356_adj.pkl', 'wb') as file:
pickle.dump(object_dit, file)
if __name__ == '__main__':
main() |
Code wasn't running so i debugged the code and now its working (tested on torch version 1.7.1 )
async=Trueis deprecated therefor usednon_blocking=True