zukakosan / tripletnet Goto Github PK

Image retrieval with triplet loss

Jupyter Notebook 99.98% Python 0.02%

keras triplet-loss deepfashion shop-clothes-retrieval deep-learning

tripletnet's Issues

論文読んで各タスクの取り組み方について理解を深める。

what is L2-norm Layer's role?

import numpy as np
epsilon = 1e-12
x =np.array([0.0000002,0.0000002,0.00000033])
print(sum(x**2))
print(x/np.sqrt(epsilon))
print(x/np.sqrt(sum(x**2)))
x/np.sqrt(max(sum(x**2), epsilon))

output

1.889e-13
[ 0.2   0.2   0.33]
[ 0.46016546  0.46016546  0.75927301]
array([ 0.2 ,  0.2 ,  0.33])

train tripletNet with images of Nogizaka and confirm the behavior

import numpy as np
import os.path

class IdentityMetadata():
    def __init__(self, base, name, file):
        # dataset base directory
        self.base = base
        # identity name
        self.name = name
        # image file name
        self.file = file

    def __repr__(self):
        return self.image_path()

    def image_path(self):
        return os.path.join(self.base, self.name, self.file) 
    
def load_metadata(path):
    metadata = []
    for i in os.listdir(path):
        for f in os.listdir(os.path.join(path, i)):
            metadata.append(IdentityMetadata(path, i, f))
    return np.array(metadata)

metadata = load_metadata('images')

クラスを作成してfilename,とデータを紐付ける。

experiment with one category of clothes dataset

search the best parameter of alpha

for mnist dataset, alpha = 0.05 is currently best.

院ゼミ準備

論文実装いつやんの？

plot mnist_emb with PCA

To reduce dimensions to 2, use PCA and plot mnist_emb

deepfashionからの連絡待ち

DeepFashion読む

データセットを手に入れる

The Last Layer may be L2-normalize?

To compress the value between 0 and 1

Where to Buy it読む

論文まとめ

せっかく読んでも後で見直すのが面倒なので、新しくipadにノート作ってまとめる。

Something wrong with triplet-loss

why loss: 0.0000e+00 ?

Epoch 1/10
400/400 [==============================] - 6s 15ms/step - loss: 3278.5809
Epoch 2/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 3/10
400/400 [==============================] - 4s 9ms/step - loss: 778.3205
Epoch 4/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 5/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 6/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 7/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 8/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 9/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00
Epoch 10/10
400/400 [==============================] - 4s 9ms/step - loss: 0.0000e+00

implementation of FashionNet

implement FashionNet according to its thesis
https://ieeexplore.ieee.org/document/7780493/

Deep Bidirectional Cross-Triplet Embedding For Online Clothing Shopping読む

triplet lossについての理解を深める

somthing wrong with "plt.imshow(img)"

img = np.array(Image.open('path').resize((w,h)))
img = (img/255.).astype(np.float32)
plt.imshow(img)

if I exec this code, normal image comes up...Why?

CNN for positive and negative should have the same one.

If use different one, Network for negative may be completely different.

Clothing Retrieval with Visual Attention Model

データセット入手した。

Where to buy it
http://tamaraberg.com/street2shop/

How can I get the embedded vectors?

embedded = np.zeros((metadata.shape[0], 128))

for i, m in enumerate(metadata):
    img = load_image(m.image_path())
    img = align_image(img)
    # scale RGB values to interval [0,1]
    img = (img / 255.).astype(np.float32)
    # obtain embedding vector for image
    embedded[i] = nn4_small2_pretrained.predict(np.expand_dims(img, axis=0))[0]

predict( )[0]なのでpredictしたうちのanchorに相当する値を取り出しemb_vecとしている？
expand_dims : [a,b,c] -> [[a,b,c]]

zukakosan / tripletnet Goto Github PK

tripletnet's Issues

output

Recommend Projects

Recommend Topics

Recommend Org