def main(data_file, vocab_path):
"""Build and evaluate Naive Bayes classifiers for the federalist papers"""
authors, essays, essay_ids = parse_federalist_papers(data_file)
function_words = load_function_words(vocab_path)
# load the attributed essays into a feature matrix
# label mapping is for me to track
# make them into two classifiers, zero and one.
# the distribution of the zero (ham) was higher？
# the distribution of one (man) was higher?
# output: two classes zero and one
X = load_features(essays, function_words)
# TODO: load the author names into a vector y, mapped to 0 and 1, using functions from util.py
labels_map = labels_to_key(authors)
print(labels_map)
# y output, a list of zeros and ones, 相对应，第几篇文章里面是什么
# y is the golden standard, it is used for both training, and evaluation
y = np.asarray(labels_to_y(authors, labels_map))
# numerical
print(f"Numpy array has shape {X.shape} and dtype {X.dtype}")
# TODO shuffle, then split the data
# if split has already had a shuffle function embedded in it, no need for importing
train, test = split_data(X, y, 0.25)
# TODO: train a multinomial NB model, evaluate on validation split
nbm = MultinomialNB()
# to see what is the definition of nbm, what it requires as in the parameter
# train is array, two tuples with [] in it, the first one is a array, teh second one is target
# rows of X and the len of y are not identical.
# y 的长度要大于X， 不能直接用y, 需要用剪裁过在train 里面的
nbm.fit(train[0], train[1]) # change
preds_nbm = nbm.predict(test[0])
test_y = test[1]
accuracy = calculate_accuracy(preds_nbm, test_y)
print(f" the accuracy for multinomial NB model is {accuracy}")
# TODO: train a Bernoulli NB model, evaluate on validation split
nbb = BernoulliNB()
nbb.fit(train[0], train[1])
preds_nbb = nbb.predict(test[0])
accuracy = calculate_accuracy(preds_nbb, test_y)
print(f" the accuracy for Bernoulli NB model is {accuracy}")
# TODO: fit the zero rule
train_y = train[1]
most_frequent_class = find_zero_rule_class(train_y)
print(f"the most frequent class is {most_frequent_class}")
test_predictions = apply_zero_rule(test[0], most_frequent_class)
test_accuracy = calculate_accuracy(test_predictions, test_y)
print(f" the accuracy for the baseline is {test_accuracy}")
def main():
parser = argparse.ArgumentParser()
parser.add_argument('model_path',
type=str,
metavar='MODEL',
help='Path to Keras model')
parser.add_argument('x_path', type=str, metavar='X', help='Path to inputs')
parser.add_argument('y_path',
type=str,
metavar='Y',
help='Path to targets')
parser.add_argument('-f',
'--output_file',
type=str,
default=None,
metavar='FILE',
help='File to print output to')
args = parser.parse_args()
model = load_model(args.model_path)
x = np.load(args.x_path)
y = np.load(args.y_path)
y_pred = model.predict(x)
acc = calculate_accuracy(y, y_pred)
cm = generate_confusion_matrix(y, y_pred)
cm_norm = cm.astype(np.float) / cm.sum(axis=1)[:, np.newaxis]
stats = calculate_recall_precision_f1_mcc(cm)
print('Accuracy: {:.2f}%'.format(acc * 100.0))
print('Confusion matrix (rows are actual, columns are predicted):')
print(cm)
print('Normalized confusion matrix:')
print(cm_norm)
print('Recall: {}'.format(stats['recall']))
print('Precision: {}'.format(stats['precision']))
print('F1 score: {}'.format(stats['f1']))
print('Matthews correlation coefficient: {}'.format(stats['mcc']))
if args.output_file is not None:
with open(args.output_file, 'w') as f:
f.write('Accuracy: {:.2f}%\n'.format(acc * 100.0))
f.write(
'Confusion matrix (rows are actual, columns are predicted):\n')
f.write(str(cm) + '\n')
f.write('Normalized confusion matrix:\n')
f.write(str(cm_norm) + '\n')
f.write('Recall: {}\n'.format(stats['recall']))
f.write('Precision: {}\n'.format(stats['precision']))
f.write('F1 score: {}\n'.format(stats['f1']))
f.write('Matthews correlation coefficient: {}\n'.format(
stats['mcc']))
def quantize(model_name, out_path):
keras_model = tf.keras.Sequential(
[hub.KerasLayer(tf_hub_links[model_name], output_shape=[1001])])
if "inception_v3" in model_name:
keras_model._set_inputs(
preprocess_cv_dataset_299[0]) # Batch input shape.
else:
keras_model._set_inputs(
preprocess_cv_dataset_224[0]) # Batch input shape.
tflite_convertor = TFLiteConvertor(\
saved_model_dir=keras_model,
base_path=out_path,
model_name=model_name)
tflite_convertor.fp32()
tflite_convertor.weight_int8_act_fp32()
if "inception_v3" in model_name:
tflite_convertor.full_integer_except_io(representative_data_gen_299)
else:
tflite_convertor.full_integer_except_io(representative_data_gen_224)
im_height = 299 if "inception_v3" in model_name else 224
im_width = im_height
tflite_model_path = out_path + "/" + model_name + "_fp32.tflite"
top1, top5 = calculate_accuracy(test_dataset, tflite_model_path, im_height,
im_width, preprocess,
postprocess[model_name], 10)
print("{:15} {:20} {:10} {:10}".format(model_name, "fp32", top1, top5))
tflite_model_path = out_path + "/" + model_name + "_full_integer_except_io.tflite"
top1, top5 = calculate_accuracy(test_dataset, tflite_model_path, im_height,
im_width, preprocess,
postprocess[model_name], 10)
print("{:15} {:20} {:10} {:10}".format(model_name, "full_integer", top1,
top5))
def main():
# bank marketing dataset
bank_marketing = fetch_openml(data_id=1461)
# only using 5k of the available 45211 instances
data = bank_marketing.data[:5000]
labels = bank_marketing.target[:5000]
# converting nominal features to strings
nominal_features = {1, 2, 3, 4, 6, 7, 8, 10, 15}
converted_data = []
num_rows = data.shape[0]
num_cols = data.shape[1]
for row in range(num_rows):
new_row = []
for col in range(num_cols):
if col in nominal_features:
new_row.append(str(data[row, col]))
else:
new_row.append(data[row, col])
new_row.append(labels[row])
converted_data.append(new_row)
train_n = int(0.8 * num_rows)
random.shuffle(converted_data)
train = converted_data[:train_n]
test = converted_data[train_n:]
print("Number of training rows: {}".format(train_n))
print("Number of testing rows: {}".format(len(test)))
train_X = [row[:-1] for row in train]
train_y = [row[-1] for row in train]
test_X = [row[:-1] for row in test]
test_y = [row[-1] for row in test]
dt = DecisionTree()
t1 = time.time()
dt.fit(train_X, train_y)
t2 = time.time()
diff = t2 - t1
print("Time to fit the decision tree: {:.2f} seconds".format(diff))
predictions = [dt.predict(record) for record in test_X]
accuracy = calculate_accuracy(predictions, test_y)
print("Accuracy on the test set: {:.2f}%".format(accuracy * 100))
def load_and_train(model_path,
x_train,
x_val,
x_test,
y_train,
y_val,
y_test,
max_epochs=1000,
batch_size=32,
patience=20):
"""
Load model and resume training.
"""
model = load_model(model_path)
checkpointer = ModelCheckpoint(model_path, save_best_only=True)
model.fit(x_train,
y_train,
epochs=max_epochs,
batch_size=batch_size,
verbose=1,
validation_data=(x_val, y_val),
callbacks=[EarlyStopping(patience=patience), checkpointer])
model = load_model(model_path)
y_train_pred = model.predict(x_train)
print('Train accuracy: {:.2f}%'.format(
calculate_accuracy(y_train, y_train_pred) * 100.0))
# print('Test accuracy: {:.2f}%'.format(calculate_accuracy(y_test, y_test_pred) * 100.0))
y_test_pred = []
for i in range(3):
y_test_pred.append(model.predict(x_test[i]))
print('Test accuracy: {:.2f}%'.format(
calculate_accuracy(y_test[i], y_test_pred[i]) * 100.0))
return model
def main(data_file):
print(data_file)
# load the data
authors, essays, essay_ids = parse_federalist_papers(data_file)
num_essays = len(essays)
print(f"Working with {num_essays} reviews")
# create a key that links author id string -> integer
author_key = labels_to_key(authors)
print(len(author_key))
print(author_key)
# convert all the labels using the key
y = labels_to_y(authors, author_key)
assert y.size == len(
authors
), f"Size of label array (y.size) must equal number of labels {len(authors)}"
# shuffle and split the data
train, test = split_data(essays, y, 0.3)
data_size_after = len(train[1]) + len(test[1])
assert data_size_after == y.size, f"Number of datapoints after split {data_size_after} must match size before {y.size}"
print(f"{len(train[0])} in train; {len(test[0])} in test")
# learn zero rule on train
train_y = train[1]
most_frequent_class = find_zero_rule_class(train_y)
print(most_frequent_class)
# lookup label string from class #
reverse_author_key = {v: k for k, v in author_key.items()}
print(
f"The most frequent class is {reverse_author_key[most_frequent_class]}"
)
# apply zero rule to test reviews
test_predictions = apply_zero_rule(test[0], most_frequent_class)
print(f"Zero rule predictions on held-out data: {test_predictions}")
# score accuracy
test_y = test[1]
test_accuracy = calculate_accuracy(test_predictions, test_y)
print(f"Accuracy of zero rule: {test_accuracy:0.03f}")
def val_epoch(epoch, data_loader, model, criterion, opt, logger):
print('validation at epoch {}'.format(epoch))
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
for i, (inputs, targets) in enumerate(data_loader):
data_time.update(time.time() - end_time)
if not opt.no_cuda:
targets = targets.cuda(async=True)
inputs = Variable(inputs, volatile=True)
targets = Variable(targets, volatile=True)
outputs = model(inputs)
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
losses.update(loss.data[0], inputs.size(0))
accuracies.update(acc, inputs.size(0))
batch_time.update(time.time() - end_time)
end_time = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
logger.log({'epoch': epoch, 'loss': losses.avg, 'acc': accuracies.avg})
return losses.avg
def val_epoch(epoch, data_loader, model, criterion, opt, vis,vallogwindow):
print('validation at epoch {}'.format(epoch))
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
mmap = meter.mAPMeter()
AP = meter.APMeter()
top = meter.ClassErrorMeter(topk=[1, 3, 5], accuracy=True)
mmap.reset()
AP.reset()
top.reset()
end_time = time.time()
for i, (inputs, targets) in enumerate(data_loader):
data_time.update(time.time() - end_time)
if type(inputs) is list:
inputs = [Variable(inputs[ii].cuda()) for ii in range(len(inputs))]
else:
inputs = Variable(inputs.cuda())
targets = targets.cuda()
with torch.no_grad():
#inputs = Variable(inputs)
targets = Variable(targets)
outputs ,context= model(inputs)
#if i %5==0:
#for jj in range(num):
# org_img = inverse_normalize(inputs[0,jj,:,:,:].detach().cpu().numpy())
# show_keypoint(org_img, context[0].detach().cpu(),vis=vis,title = str(jj+1))
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
losses.update(loss.data.item(), targets.detach().size(0))
accuracies.update(acc, targets.detach().size(0))
one_hot = torch.zeros_like(outputs).cuda().scatter_(1, targets.view(-1, 1), 1)
mmap.add(outputs.detach(), one_hot.detach())
top.add(outputs.detach(), targets.detach())
AP.add(outputs.detach(), one_hot.detach())
batch_time.update(time.time() - end_time)
end_time = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})\t'
'mmap {mmap}\t'
'top1 3 5: {top}\t'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies,
mmap=mmap.value(),
top=top.value() ))
vis.text("gpu:{}, epoch: {},loss: {},accu:{},mAP:{}, top135 {}\nAP:{}".format(torch.cuda.current_device(),epoch,losses.avg,accuracies.avg,mmap.value(),top.value(),AP.value())
,win=vallogwindow,append=True)
#exit()
#if epoch==10:
# exit()
return losses.avg, mmap.value()
def train_epoch(epoch, data_loader, model, criterion, optimizer, opt, vis,
trainlogwindow):
print('train at epoch {}'.format(epoch))
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
mmap = meter.mAPMeter()
top = meter.ClassErrorMeter(topk=[1, 3, 5], accuracy=True)
mmap.reset()
top.reset()
end_time = time.time()
for i, (inputs, targets) in enumerate(data_loader):
data_time.update(time.time() - end_time)
targets = targets.cuda()
if type(inputs) is list:
inputs = [Variable(inputs[ii]).cuda() for ii in range(len(inputs))]
else:
inputs = inputs.cuda()
#inputs, targets_a, targets_b, lam = mixup_data(inputs, targets, opt.DATASET.ALPHA, True)
#inputs, targets_a, targets_b = Variable(inputs), Variable(targets_a), Variable(targets_b)
inputs = Variable(inputs)
#print(targets)
targets = Variable(targets)
outputs, context = model(inputs)
#loss_func = mixup_criterion(targets_a, targets_b, lam)
#loss = loss_func(criterion, outputs)
loss = criterion(outputs, targets)
#print(outputs.shape)
#print(targets)
acc = calculate_accuracy(outputs, targets)
one_hot = torch.zeros_like(outputs).cuda().scatter_(
1, targets.view(-1, 1), 1)
mmap.add(outputs.detach(), one_hot.detach())
top.add(outputs.detach(), targets.detach())
losses.update(loss.data.item(), targets.detach().size(0))
accuracies.update(acc, targets.detach().size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end_time)
end_time = time.time()
vis.text(
"gpu{}, epoch: {},batch:{},iter: {},loss: {},acc:{},lr: {}\n".format(torch.cuda.current_device(),epoch, i + 1,(epoch - 1) * len(data_loader) + (i + 1),losses.val, \
accuracies.val,optimizer.param_groups[0]['lr'])
,win=trainlogwindow,append=True)
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})\t'
'mmap {mmap}\t'
'top1 3 5: {top}\t'.format(epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies,
mmap=mmap.value(),
top=top.value()))
vis.text(
"total:\n gpu:{} epoch: {},loss: {},lr: {}, accu:{},mAP:{}, top135 {}\n"
.format(torch.cuda.current_device(), epoch, losses.avg,
optimizer.param_groups[0]['lr'], accuracies.avg, mmap.value(),
top.value()),
win=trainlogwindow,
append=True)
if torch.cuda.current_device() == 0:
print("saveing ckp ########################################")
if epoch % opt.MODEL.CKP_DURING == 0:
save_file_path = os.path.join(opt.MODEL.RESULT, opt.MODEL.NAME,
'save_{}.pth'.format(epoch))
if not os.path.exists(
os.path.join(opt.MODEL.RESULT, opt.MODEL.NAME)):
os.makedirs(os.path.join(opt.MODEL.RESULT, opt.MODEL.NAME))
states = {
'epoch': epoch + 1,
'arch': opt.MODEL.NAME,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
torch.save(states, save_file_path)
return losses.avg, mmap.value()
def main(data_file, vocab_path):
"""Build and evaluate Naive Bayes classifiers for the federalist papers"""
authors, essays, essay_ids = parse_federalist_papers(data_file)
function_words = load_function_words(vocab_path)
# load the attributed essays into a feature matrix
X = load_features(essays, function_words)
# TODO: load the author names into a vector y, mapped to 0 and 1, using functions from util.py
labels_map = labels_to_key(authors)
y = np.asarray(labels_to_y(authors, labels_map))
print(f"X has shape {X.shape} and dtype {X.dtype}")
print(f"y has shape {y.shape} and dtype {y.dtype}")
# TODO shuffle, then split the data
train, test = split_data(X, y, 0.25)
data_size_after = len(train[1]) + len(test[1])
assert data_size_after == y.size, f"Number of datapoints after split {data_size_after} must match size before {y.size}"
print(f"{len(train[0])} in train; {len(test[0])} in test")
# TODO: train a multinomial NB model, evaluate on validation split
nb_mul = MultinomialNB()
nb_mul.fit(train[0], train[1])
pred_mul = nb_mul.predict(test[0])
acc_mul = metrics.accuracy_score(test[1], pred_mul)
print(f"Accuracy of Multinomial NB method: {acc_mul:0.03f}")
# TODO: train a Bernoulli NB model, evaluate on validation split
### make a binary feature vector
train_bi = np.copy(train[0])
for i in range(len(train_bi)):
train_bi[i] = np.where(train_bi[i] > 0, 1, 0)
test_bi = np.copy(test[0])
for i in range(len(test_bi)):
test_bi[i] = np.where(test_bi[i] > 0, 1, 0)
nb_ber = BernoulliNB()
nb_ber.fit(train_bi, train[1])
pred_ber = nb_ber.predict(test_bi)
acc_ber = metrics.accuracy_score(test[1], pred_ber)
print(f"Accuracy of Bernoulli NB method: {acc_ber:0.03f}")
# TODO: fit the zero rule
# learn zero rule on train
most_frequent_class = find_zero_rule_class(train[1])
# apply zero rule to test reviews
test_predictions = apply_zero_rule(test[0], most_frequent_class)
# score accuracy
test_accuracy = calculate_accuracy(test_predictions, test[1])
print(f"Accuracy of Zero rule: {test_accuracy:0.03f}")
# lookup label string from class #
author_key = labels_to_key(authors)
reverse_author_key = {v: k for k, v in author_key.items()}
print(
f"The author predicted by the Zero rule is {reverse_author_key[most_frequent_class]}"
)
def bgru(x_train,
x_val,
x_test,
y_train,
y_val,
y_test,
out_dir,
name='bgru_model',
hidden_units=10,
layers=1,
max_epochs=1000,
batch_size=32,
patience=20,
dropout=0.0,
recurrent_dropout=0.0):
"""
Bidirectional GRU model for protein secondary structure prediction.
"""
num_samples = x_train.shape[0]
max_seq_len = x_train.shape[1]
num_features = x_train.shape[2]
num_classes = y_train.shape[2]
# Build Keras model
model = Sequential()
model.add(Masking(mask_value=0, input_shape=(max_seq_len, num_features)))
model.add(
Bidirectional(
GRU(hidden_units,
return_sequences=True,
input_shape=(max_seq_len, num_features),
dropout=dropout,
recurrent_dropout=recurrent_dropout)))
if layers > 1:
for _ in range(layers - 1):
model.add(
Bidirectional(
GRU(hidden_units,
return_sequences=True,
dropout=dropout,
recurrent_dropout=recurrent_dropout)))
model.add(TimeDistributed(Dense(num_classes)))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
print(model.summary())
# Train model. Use early-stopping on validation data to determine when to stop training.
model_path = os.path.join(out_dir, name + '.h5')
checkpointer = ModelCheckpoint(model_path, save_best_only=True)
model.fit(x_train,
y_train,
epochs=max_epochs,
batch_size=batch_size,
verbose=1,
validation_data=(x_val, y_val),
callbacks=[EarlyStopping(patience=patience), checkpointer])
model = load_model(
model_path
) # Best model is not necessarily current model instance b/c patience != 0
y_train_pred = model.predict(x_train)
print('Train accuracy: {:.2f}%'.format(
calculate_accuracy(y_train, y_train_pred) * 100.0))
# Test set accuracy
y_test_pred = []
for i in range(3):
y_test_pred.append(model.predict(x_test[i]))
print('Test accuracy: {:.2f}%'.format(
calculate_accuracy(y_test[i], y_test_pred[i]) * 100.0))
return model
def train_epoch(epoch, data_loader, model, criterion, optimizer, opt,
epoch_logger, batch_logger):
print('train at epoch {}'.format(epoch))
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
for i, (inputs, targets) in enumerate(data_loader):
data_time.update(time.time() - end_time)
if not opt.no_cuda:
targets = targets.cuda(async=True)
inputs = Variable(inputs)
targets = Variable(targets)
outputs = model(inputs)
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
losses.update(loss.data[0], inputs.size(0))
accuracies.update(acc, inputs.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end_time)
end_time = time.time()
batch_logger.log({
'epoch': epoch,
'batch': i + 1,
'iter': (epoch - 1) * len(data_loader) + (i + 1),
'loss': losses.val,
'acc': accuracies.val,
'lr': optimizer.param_groups[0]['lr']
})
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
epoch_logger.log({
'epoch': epoch,
'loss': losses.avg,
'acc': accuracies.avg,
'lr': optimizer.param_groups[0]['lr']
})
if epoch % opt.checkpoint == 0:
save_file_path = os.path.join(opt.result_path,
'save_{}.pth'.format(epoch))
states = {
'epoch': epoch + 1,
'arch': opt.arch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
torch.save(states, save_file_path)