如何将我的训练数据上传到谷歌以进行 Tensorflow 云训练

Question

I want to train my keras model in gcp.

My code:

this is how I load the dataset

dataset = pandas.read_csv('USDJPY.fx5.csv', usecols=[2, 3, 4, 5], engine='python')

this is how i trigger cloud training

job_labels = {"job": "forex-usdjpy", "team": "xxx", "user": "xxx"}
tfc.run(requirements_txt="./requirements.txt",
        job_labels=job_labels,
        stream_logs=True
        )

Right before my model, which shouldn't make much of a difference

model = Sequential()
model.add(LSTM(4, input_shape=(1, 4)))
model.add(Dropout(0.2))
model.add(Dense(4))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(trainX, trainY, epochs=1, batch_size=1, verbose=2)

Everything is working, docker image for my model is being created, but the USDJPY.fx5.csv file is not being uploaded. So I get file not found error

What is the proper way of loading custom files into the training job? I uploaded the train data to s3 bucket but I wasn't able to tell google to look there.

Answer 1

Turns out it was a problem with my GCP configuration Here are the steps I made to make it work:

• Create an s3 bucket and make all files inside it public so the train job can access them

• Include these two in the requirements fsspec and gcsfs

• remove the 'engine' parameter from panda.readCsv like so

  dataset = pandas.read_csv('gs:///USDJPY.fx5.csv', usecols=[2, 3, 4, 5])

Since you are uploading the python file to GCP a good way to organize your code it to put all of the training logic into a method and then called it conditionally on the cloud train flag:

if tfc.remote():
    train()

Here is the whole working code if someone is interested

import pandas
import numpy
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Dropout
from sklearn.preprocessing import MinMaxScaler
import tensorflow_cloud as tfc
import os

os.environ["PATH"] = os.environ["PATH"] + ":<path to google-cloud-sdk/bin"
os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "<path to google credentials json (you can generate this through their UI"


def create_dataset(data):
    dataX = data[0:len(data) - 1]
    dataY = data[1:]
    return numpy.array(dataX), numpy.array(dataY)

def train():
    dataset = pandas.read_csv('gs://<bucket>/USDJPY.fx5.csv', usecols=[2, 3, 4, 5])

    scaler = MinMaxScaler(feature_range=(-1, 1))
    scaler = scaler.fit(dataset)

    dataset = scaler.transform(dataset)

    # split into train and test sets
    train_size = int(len(dataset) * 0.67)
    train, test = dataset[0:train_size], dataset[train_size:len(dataset)]

    trainX, trainY = create_dataset(train)

    trainX = numpy.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))

    model = Sequential()
    model.add(LSTM(4, input_shape=(1, 4)))
    model.add(Dropout(0.2))
    model.add(Dense(4))
    model.compile(loss='mean_squared_error', optimizer='adam')
    model.fit(trainX, trainY, epochs=1000, verbose=1)


job_labels = {"job": "forex-usdjpy", "team": "zver", "user": "zver1"}
tfc.run(requirements_txt="./requirements.txt",
        job_labels=job_labels,
        stream_logs=True
        )

if tfc.remote():
    train()

NOTE: This is probably not an optimal LSTM config, take it with a grain of salt