def create_data(set_name):
if set_name == "training" or set_name == "test":
params = dicto.load_("params.yml")
df = dataset.get_base_df("data/base")
split = int(len(df) * params.preprocessing.split)
training_df = df.iloc[:split]
test_df = df.iloc[split:]
if set_name == "training":
df = training_df
df = pr.grow_dataset(df, params=params)
# df = df.iloc[:100]
generator = utils.dataframe_batch_generator(
df, params.preprocessing.batch_size)
generator = pr.read_images(generator)
generator = pr.data_augmentation(generator, params=params)
pr.create_tfrecords(generator, "training_set.record", params)
else:
df = test_df
generator = utils.dataframe_batch_generator(
df, params.preprocessing.batch_size)
generator = pr.read_images(generator)
pr.create_tfrecords(generator, "training_set.record", params)
else:
raise Exception(
"Please choose 'training' or 'test', got {set_name}".format(
set_name=set_name))
def main():
trainx, trainy = read_images('../input/omniglot/images_background/')
testx, testy = read_images('../input/omniglot/images_evaluation/')
use_gpu = torch.cuda.is_available()
trainx = torch.from_numpy(trainx).float()
testx = torch.from_numpy(testx).float()
if use_gpu:
trainx = trainx.cuda()
testx = testx.cuda()
print(trainx.size(), testx.size())
num_episode = 16000
frame_size = 1000
trainx = trainx.permute(0, 3, 1, 2)
testx = testx.permute(0, 3, 1, 2)
frame_loss = 0
frame_acc = 0
for i in range(num_episode):
loss, acc = train_step(trainx, trainy, 5, 60, 5)
frame_loss += loss.data
frame_acc += acc.data
if((i+1) % frame_size == 0):
print("Frame Number:", ((i+1) // frame_size), 'Frame Loss: ', frame_loss.data.cpu().numpy().tolist() /
frame_size, 'Frame Accuracy:', (frame_acc.data.cpu().numpy().tolist() * 100) / frame_size)
frame_loss = 0
frame_acc = 0
num_test_episode = 2000
avg_loss = 0
avg_acc = 0
for _ in range(num_test_episode):
loss, acc = test_step(testx, testy, 5, 60, 15)
avg_loss += loss.data
avg_acc += acc.data
print('Avg Loss: ', avg_loss.data.cpu().numpy().tolist() / num_test_episode,
'Avg Accuracy:', (avg_acc.data.cpu().numpy().tolist() * 100) / num_test_episode)
def main():
# Reading the data
trainx, trainy = read_images('../input/omniglot/images_background/')
testx, testy = read_images('../input/omniglot/images_evaluation/')
# Checking if GPU is available
use_gpu = torch.cuda.is_available()
# Converting input to pytorch Tensor
trainx = torch.from_numpy(trainx).float()
testx = torch.from_numpy(testx).float()
if use_gpu:
trainx = trainx.cuda()
testx = testx.cuda()
# Priniting the data
print(trainx.size(), testx.size())
# Set training iterations and display period
num_episode = 16000
frame_size = 1000
trainx = trainx.permute(0, 3, 1, 2)
testx = testx.permute(0, 3, 1, 2)
# Initializing prototypical net
protonet = PrototypicalNet(use_gpu)
# Training loop
frame_loss = 0
frame_acc = 0
for i in range(num_episode):
loss, acc = train_step(protonet, trainx, trainy, 5, 60, 5)
frame_loss += loss.data
frame_acc += acc.data
if ((i + 1) % frame_size == 0):
print("Frame Number:", ((i + 1) // frame_size), 'Frame Loss: ',
frame_loss.data.cpu().numpy().tolist() / frame_size,
'Frame Accuracy:',
(frame_acc.data.cpu().numpy().tolist() * 100) / frame_size)
frame_loss = 0
frame_acc = 0
# Test loop
num_test_episode = 2000
avg_loss = 0
avg_acc = 0
for _ in range(num_test_episode):
loss, acc = test_step(protonet, testx, testy, 5, 60, 15)
avg_loss += loss.data
avg_acc += acc.data
print('Avg Loss: ',
avg_loss.data.cpu().numpy().tolist() / num_test_episode,
'Avg Accuracy:',
(avg_acc.data.cpu().numpy().tolist() * 100) / num_test_episode)
# Using Pretrained Model
protonet = load_weights('./protonet.pt', protonet, use_gpu)
def main():
tb_writer = SummaryWriter()
# Reading the data
print("Reading background images")
trainx, trainy = read_images(
r'D:\_hackerreborn\Prototypical-Networks\input\omniglot\images_background'
)
print(trainx.shape)
print(trainy.shape)
print("Reading background images")
testx, testy = read_images(
r'D:\_hackerreborn\Prototypical-Networks\input\omniglot\images_evaluation'
)
print(testx.shape)
print(testy.shape)
# Checking if GPU is available
use_gpu = torch.cuda.is_available()
# Converting input to pytorch Tensor
trainx = torch.from_numpy(trainx).float()
testx = torch.from_numpy(testx).float()
if use_gpu:
trainx = trainx.cuda()
testx = testx.cuda()
# Priniting the data
print(trainx.size(), testx.size())
# Set training iterations and display period
num_episode = 16000
frame_size = 1000
trainx = trainx.permute(0, 3, 1, 2)
testx = testx.permute(0, 3, 1, 2)
# Initializing prototypical net
protonet = PrototypicalNet(use_gpu)
optimizer = optim.SGD(protonet.parameters(), lr=0.01, momentum=0.99)
# Training loop
frame_loss = 0
frame_acc = 0
for i in range(num_episode):
print("Train Episode Number : {0}".format(i))
# if i > 10:
# break
loss, acc = train_step(protonet, trainx, trainy, 5, 60, 5, optimizer)
frame_loss += loss.data
frame_acc += acc.data
if (i + 1) % frame_size == 0:
print("Frame Number:", ((i + 1) // frame_size), 'Frame Loss: ',
frame_loss.data.cpu().numpy().tolist() / frame_size,
'Frame Accuracy:',
(frame_acc.data.cpu().numpy().tolist() * 100) / frame_size)
tb_writer.add_scalar(
'frame_loss',
frame_loss.data.cpu().numpy().tolist() / frame_size,
((i + 1) // frame_size))
tb_writer.add_scalar(
'frame_accuracy',
(frame_acc.data.cpu().numpy().tolist() * 100) / frame_size,
((i + 1) // frame_size))
frame_loss = 0
frame_acc = 0
import torch
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
from matplotlib import pyplot as plt
import cv2
from tqdm import tqdm
import multiprocessing as mp
from preprocessing import read_images
from prototypicalNet import PrototypicalNet, train_step, test_step, load_weights
tqdm.pandas(desc="my bar!")
#def main():
# Reading the data
trainx, trainy = read_images('./omniglot/images_background')
testx, testy = read_images('./omniglot/images_evaluation')
# Checking if GPU is available
print('finish read image')
use_gpu = torch.cuda.is_available()
use_gpu = False
# Converting input to pytorch Tensor
trainx = torch.from_numpy(trainx).float()
testx = torch.from_numpy(testx).float()
if use_gpu:
trainx = trainx.cuda()
testx = testx.cuda()
# Priniting the data
print(trainx.size(), testx.size())
# Set training iterations and display period
num_episode = 16000