criterion = nn.CrossEntropyLoss(size_average=True)
optimizer_ft = optim.SGD(list(model_ft.parameters())[-32:],
lr=0.01,
momentum=0.9)
# optimizer_ft = optim.Adam(model_ft.parameters(), lr=0.001)
train_model(model_ft,
criterion,
optimizer_ft,
exp_lr_scheduler,
num_epochs=n_epoch,
kid=kid)
for i in range(kfold):
train_val_split(kfold, i)
train_val_folder = {
x: datasets.ImageFolder(os.path.join(DIR_BASE, x), data_transforms[x])
for x in ['train', 'val']
}
train_val_loader = {
x: torch.utils.data.DataLoader(train_val_folder[x],
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=12)
for x in ['train', 'val']
}
train_val_size = {x: len(train_val_folder[x]) for x in ['train', 'val']}
classes = train_val_folder['train'].classes
# print('There are {} classes, {} images in train set and {} images in validation set.'
import os
from scipy import ndimage, misc
import scipy.misc
import re, glob
import cv2
from train_val_split import train_val_split
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
root_train = ''
root_val = ''
root_total = ''
#classes = ['',''.....]
train_val_split(prop=0.8)
train_imgs = glob.glob('')
for filename in train_imgs:
image = ndimage.imread(filename, mode="RGB")
image_resized = misc.imresize(image, (299, 299))
out_file = filename.split('.')[0].split('/')[1]
cv2.imwrite(filename, image_resized)
learning_rate = 0.0001
img_width = 299
img_height = 299
nbr_train_samples = ''
nbr_validation_samples = ''
nbr_epochs = 50
def eval_CNN_regression(experiment_name, trials, inFile, image_path, categoryName, gpus, model_type, epoch_number, augment_number, learning_rate, decay, regularization, batch_size, img_height, img_width):
#Experiment folder
try:
os.mkdir(os.path.join(main_experiment_path, experiment_name))
except:
print('Saving over experiment ' + experiment_name)
means = list()
variances = list()
for i in range(1, trials+1):
trial_path = os.path.join(main_experiment_path, experiment_name, experiment_name+'_'+str(i))
#Trial i folder
try:
os.mkdir(trial_path)
except:
print('Saving over trial ' + str(i) + ' of experiment ' + experiment_name)
df = pd.read_csv(inFile)
split_objects(df, os.path.join(trial_path, 'splitdf.txt'))
y_classes = train_val_split(os.path.join(trial_path, 'splitdf.txt'), outFile=os.path.join(trial_path, 'df'), categoryName=categoryName, numCat=0, val_prop=0.3)
train_path = os.path.join(trial_path, 'df_train.txt')
val_path = os.path.join(trial_path, 'df_val.txt')
#Train model i
train_regression(gpus=gpus,
experiment_path=trial_path,
train_df_path=train_path,
train_image_path=image_path,
val_df_path=val_path,
val_image_path=image_path,
y_name=categoryName,
augment_number=augment_number,
model_type=model_type,
model_save_name=experiment_name,
epoch_number=epoch_number,
learning_rate=learning_rate,
decay=decay,
regularization=regularization,
batch_size=batch_size,
img_height=img_height,
img_width=img_width
)
#Evaluate model i
i_stats = eval_regression(experiment_path=trial_path,
test_df_path=val_path,
test_image_path=image_path,
y_name=categoryName,
model_test_name=experiment_name,
height=img_height,
width=img_width
)
#Top 1 accuracy
means.append(i_stats['mean'])
variances.append(i_stats['variance'])
#Mean and variance of means and variances
meanofmeans = statistics.mean(means)
varianceofmeans = statistics.variance(means)
meanofvariances = statistics.mean(variances)
varianceofvariances = statistics.variance(variances)
#Write summary of test results
with open(os.path.join(main_experiment_path, experiment_name, experiment_name+':experiment_summary.txt'), 'w') as summary_writefile:
summary_writefile.write('Regressing on ' + categoryName + '\n')
summary_writefile.write('Mean error in years: ' + str(meanofmeans) + '\n')
summary_writefile.write('Mean error variance in years: ' + str(meanofvariances) + '\n')
summary_writefile.write('Variance of mean error in years: ' + str(varianceofmeans) + '\n')
summary_writefile.write('Variance of variances in years: ' + str(varianceofvariances) + '\n')
for i in range(1, len(means)+1):
summary_writefile.write('Model ' + str(i) + ': ' + 'mean: ' + str(means[i-1]) + ', v ariance: ' + str(variances[i-1]) + '\n')
#Return test results summary
return {'meanofmeans': meanofmeans,
'varianceofmeans': varianceofmeans,
'meansofvariances': meanofvariances,
'varianceofvariances': varianceofvariances}
params = {"batch_size": batch_sz, "shuffle": False, "num_workers": 4}
train_pct = 0.7
device = "cpu"
if __name__ == '__main__':
# Extract data for the ticker mentioned above
extractHistory(SYMBOL, start_date, end_date, pth)
# Get the inputs and outputs from the extracted ticker data
inputs, labels, dates = curateData(pth, "Close", "Date", n_lags)
N = len(inputs)
# Perform the train validation split
trainX, trainY, valX, valY = train_val_split(inputs, labels, train_pct)
# Standardize the data to bring the inputs on a uniform scale
trnX, SS_ = standardizeData(trainX, train=True)
valX, _ = standardizeData(valX, SS_)
# Create dataloaders for both training and validation datasets
training_generator = getDL(trnX, trainY, params)
validation_generator = getDL(valX, valY, params)
# Create the model
model = forecasterModel(n_lags, hidden_dim, rnn_layers, dropout).to(device)
# Define the loss function and the optimizer
loss_func = nn.MSELoss()
optim = torch.optim.Adam(model.parameters(), lr=learning_rate)
def eval_CNN_categorical(experiment_name, trials, inFile, image_path, categoryName, numCat, gpus, model_type, epoch_number, augment_number, learning_rate, decay, regularization, batch_size, img_height, img_width):
#Experiment folder
try:
os.mkdir(os.path.join(main_experiment_path, experiment_name))
except:
print('Saving over experiment ' + experiment_name)
top_1_accuracies = list()
for i in range(1, trials+1):
trial_path = os.path.join(main_experiment_path, experiment_name, experiment_name+'_'+str(i))
#Trial i folder
try:
os.mkdir(trial_path)
except:
print('Saving over trial ' + str(i) + ' of experiment ' + experiment_name)
df = pd.read_csv(inFile)
split_objects(df, os.path.join(trial_path, 'splitdf.txt'))
y_classes = train_val_split(os.path.join(trial_path, 'splitdf.txt'), outFile=os.path.join(trial_path, 'df'), categoryName=categoryName, numCat=numCat, val_prop=0.3)
train_path = os.path.join(trial_path, 'df_train.txt')
val_path = os.path.join(trial_path, 'df_val.txt')
#Train model i
train_categorical(gpus=gpus,
experiment_path=trial_path,
train_df_path=train_path,
train_image_path=image_path,
val_df_path=val_path,
val_image_path=image_path,
y_name=categoryName,
y_classes=y_classes,
augment_number=augment_number,
model_type=model_type,
model_save_name=experiment_name,
epoch_number=epoch_number,
learning_rate=learning_rate,
decay=decay,
regularization=regularization,
batch_size=batch_size,
img_height=img_height,
img_width=img_width
)
#Evaluate model i
k_accuracies = eval_categorical(experiment_path=trial_path,
test_df_path=val_path,
test_image_path=image_path,
y_name=categoryName,
model_test_name=experiment_name,
height=img_height,
width=img_width
)
#Top 1 accuracy
top_1_accuracies.append(k_accuracies[0])
#Mean and variance of top_1 accuracies
mean = statistics.mean(top_1_accuracies)
variance = statistics.variance(top_1_accuracies)
#Write summary of test results
with open(os.path.join(main_experiment_path, experiment_name, experiment_name+':experiment_summary.txt'), 'w') as summary_writefile:
summary_writefile.write('Predicting ' + categoryName + ', trained on top ' + str(numCat) + ' classes\n')
summary_writefile.write('Prediction error mean in years: ' + str(mean) + '\n')
summary_writefile.write('Prediction error variance in years: ' + str(variance) + '\n')
for i in range(1, len(top_1_accuracies)+1):
summary_writefile.write('Accuracy of model ' + str(i) + ': ' + str(top_1_accuracies[i-1]) + '\n')
#Return test results summary
return {'mean': mean,
'variance': variance}