femalegeekinsv / street2shoptriplet

Challenge the customer to shop task with tripletNet

Street2Shop implementation with keras

Task

Clothing image retrieval from consumer(street) image to shop(online) image.

Goal

Compare Models for clothing retrieval. Main: VGG16 + Triplet Loss/Improved Triplet Loss

1. SingleNetwork for one input
2. DualNetwork for one input

Dataset

I used DeepFashion Dataset to train and test my model.

• anchor = street image
• positive = shop image consistent to anchor
• negative = shop image not consistent to anchor

data example

Retrieval Sample

T_Shirt

Dress

Coat

Pants

Implementation

These notebooks are implemented by python3.5 or 3.6, keras 2.2.1. Sorry, I use Japanese for comment in my notebook, haha. Please execute this command for building my environment. I trained models for one category respectively.

$ conda env create -n {new environment name} -f env_zkk.yml

Reference

• DeepFashion(https://liuziwei7.github.io/projects/DeepFashion.html)

notion

If you encount the error like below, please Quite jupyter process.

Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.

If you close tab without Quit, the jupyter-process would be zombie process. GPU would be occupied by that process and you can't initialize a new process.

reference(https://qiita.com/Ka-k/items/cb942855ab669ff60630)

　TensorFlow-GPUは、デフォルトだとGPUの空いている領域をすべて確保しようとする。すべてです。ALL。 　つまり、TensorFlow-GPUを使った機械学習プログラムを複数同時に走らせると1つめは普通に通るけど2つめはGPUを確保できないので初期化に失敗する。これがcuDNN failed to initializeの正体。 　そしてJupyterをタブ閉じで終わらせるとGPUを確保したままゾンビ化する。その後Jupyterを立ち上げ直してもGPUはゾンビ化したコードに専有されたままなのでこれも初期化失敗する。これが1回目でも学習失敗する症状の正体。 　試しに1つめのコードを走らせたあと、Jupyter上でコードを停止させてから2つめを走らせると、普通に動いた。CUDAもcuDNNもAnacondaも悪くない。悪いのはGPUをひとり占めしようとするTensorFlowとタブを消したらゾンビ化するJupyterやったんや。