def rate(img_path):
"""
Returns: Scores, mean, std
"""
model_setenv()
device = model_device()
model = get_model()
model_name = 'models/ImageNima.pth'
model_load(model, model_name)
model = model.to(device)
model.eval()
image_filenames = sorted(glob.glob(img_path))
transform = transforms.Compose([
transforms.Resize(224),
transforms.CenterCrop(224),
transforms.ToTensor(),
])
weighted_votes = torch.arange(10, dtype=torch.float) + 1
weighted_votes = weighted_votes.to(device)
for index, filename in enumerate(image_filenames):
img = Image.open(filename).convert('RGB')
img = transform(img).to(device)
with torch.no_grad():
scores = model(img.view(1, 3, 224, 224))
mean = torch.matmul(scores, weighted_votes)
std = torch.sqrt((scores * torch.pow(
(weighted_votes - mean.view(-1, 1)), 2)).sum(dim=1))
print("{:.4f} {:.4f}--- {}".format(mean.item(), std.item(), filename))
def train():
"""
basic predict function for the API
the 'mode' give you the ability to toggle between a test version and a production verion of training
"""
regressor = request.args.get('regressor')
if regressor is None:
print(
"WARNING API (train): received request, but no regressor specified, assuming 'randomforest'"
)
regressor = "randomforest"
print("... training model")
data_dir = os.path.join(THIS_DIR, "cs-train")
try:
model_train(data_dir, test=True, regressor=regressor)
print("... training complete")
# reload models and data after re-train
print("... reloading models in cache")
global_data, global_models = model_load(training=False)
return (jsonify(True))
except Exception as e:
print("ERROR API (train): model_train returned: {}".format(str(e)))
return jsonify([]), 400
def main():
model_train(test=False)
model = model_load()
print("model training complete.")
def predict():
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'country' not in request.json:
print(
"ERROR API (predict): received request, but no 'country' found within"
)
return jsonify([])
if 'year' not in request.json:
print(
"ERROR API (predict): received request, but no 'year' found within"
)
return jsonify([])
if 'month' not in request.json:
print(
"ERROR API (predict): received request, but no 'month' found within"
)
return jsonify([])
if 'day' not in request.json:
print(
"ERROR API (predict): received request, but no 'day' found within")
return jsonify([])
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
## extract the query
country = request.json['country']
year = request.json['year']
month = request.json['month']
day = request.json['day']
## load model
(all_data, all_models) = model_load()
if not all_models:
print("ERROR: models are not available")
return jsonify([])
_result = model_predict(country,
year,
month,
day,
all_models,
all_data,
test=test)
result = {}
## convert numpy objects to ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
return (jsonify(result))
def predict():
"""
basic predict function for the API
"""
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'country' not in request.json:
print(
"ERROR API (predict): received request, but no 'country' found within"
)
return jsonify([])
if 'date' not in request.json:
print(
"WARNING API (predict): received request, but no 'date' was found within"
)
## set the test flag
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
## extract the country
country_input = request.json['country']
date_input = request.json['date']
## load model
model = model_load(test=test)
if not model:
print("ERROR: model is not available")
return jsonify([])
_result = model_predict(date=date_input,
country=country_input,
df=None,
model=model,
test=test)
result = {}
## convert numpy objects to ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
indexes = pd.Series([])
for i in range(len(_result["predicted"].index)):
indexes[i] = str(_result["predicted"].index[i])
_result["predicted"].index = indexes
_result["predicted"] = _result["predicted"].to_json()
return (jsonify(_result))
def main():
# train the model
model_train(test=False)
# load the model
model = model_load()
print("model training complete.")
def predict():
"""
basic predict function for the API
"""
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'country' not in request.json:
print(
"ERROR API (predict): received request, but no 'country' found within"
)
return jsonify([])
#if 'type' not in request.json:
# print("WARNING API (predict): received request, but no 'type' was found assuming 'numpy'")
# query_type = 'numpy'
## set the test flag
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
## extract the query
country = request.json['country']
year = request.json['year']
month = request.json['month']
day = request.json['day']
#if request.json['type'] == 'dict':
# pass
#else:
# print("ERROR API (predict): only dict data types have been implemented")
# return jsonify([])
## load model
data_dir = os.path.join("data", "cs-train")
all_data, all_models = model_load(country, data_dir=data_dir)
model = all_models[country]
if not model:
print("ERROR: model is not available")
return jsonify([])
_result = model_predict(country, year, month, day, test=test)
result = {}
## convert numpy objects to ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
return (jsonify(result))
def main():
## train the model
model_train()
## load the model
model = model_load()
print("model training complete.")
def main(data_dir):
## train the model
model_train(data_dir, test=False)
## load the model
model = model_load()
print("model training complete.")
def main():
# train the model
data_dir = os.path.join("..", "cs-train")
model_train(data_dir)
# load the model
model = model_load()
print("model training complete.")
def predict():
"""
basic predict function for the API
"""
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'query' not in request.json:
print(
"ERROR API (predict): received request, but no 'query' found within"
)
return jsonify([])
if 'type' not in request.json:
print(
"WARNING API (predict): received request, but no 'type' was found assuming 'numpy'"
)
query_type = 'numpy'
## set the test flag
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
## extract the query
query = request.json['query']
print(query)
if request.json['type'] == 'dict':
pass
else:
print(
"ERROR API (predict): only dict data types have been implemented")
return jsonify([])
## load model
model = model_load()
if not model:
print("ERROR: model is not available")
return jsonify([])
_result = model_predict(query, model, test=test)
result = {}
## convert numpy objects to ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
return (jsonify(result))
def test_02_load(self):
"""
test the train functionality
"""
## load the model
all_data, all_models = model_load()
self.assertTrue(all_data)
self.assertTrue(all_models)
def main():
## train the model
data_dir = os.path.join(".", "data", "cs-train")
model_train(data_dir, test=False)
## load the model
all_data, all_models = model_load()
print("model training complete.")
def predict():
"""
basic predict function for the API
"""
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'query' not in request.json:
print(
"ERROR API (predict): received request, but no 'query' found within"
)
return jsonify([])
if 'type' not in request.json:
print(
"WARNING API (predict): received request, but no 'type' was found assuming 'numpy'"
)
query_type = 'numpy'
query = request.json['query']
idx = query['idx']
query = np.array(query['data'])
## load model
model = model_load()
if not model:
print("ERROR: model is not navailable")
return jsonify([])
_result = {}
if idx == 1:
_result[idx] = (model_predict(query, model))
else:
for i, j in list(zip(query, idx)):
_result[j] = (model_predict(i, model))
result = {}
result['y_pred'] = {}
## convert numpy objects so ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result['y_pred'][key] = item.tolist()
else:
result['y_pred'][key] = item
return (jsonify(result))
def main():
data_dir = os.path.join(DATA_DIR, "cs-train")
## train the model
model_train(data_dir, test=False)
## load the model
model = model_load()
print("model training complete.")
def main():
## train the model
model_train(data_dir=DATA_DIR, prefix='sl', test=False)
## load the model
all_data, all_models = model_load(country='all',
prefix='sl',
data_dir=DATA_DIR,
training=False)
print("... models loaded: ", ",".join(all_models.keys()))
print("model training complete.")
def main():
## train the model
print("TRAINING MODELS")
data_dir = os.path.join(".", "data", "cs-train")
model_train(data_dir, test=True)
## load the model
print("LOADING MODELS")
all_data, all_models = model_load()
print("... models loaded: ", ",".join(all_models.keys()))
print("model training complete.")
def test_02_load(self):
"""
test the train functionality
"""
# Load the model
model_data, models = model_load(country='united_kingdom',
prefix='test',
data_dir=data_dir,
training=False)
model = list(models.values())[0]
self.assertTrue('predict' in dir(model))
self.assertTrue('fit' in dir(model))
def predict():
"""
basic predict function for the API
"""
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
print(request.json)
if 'country' not in request.json:
print(
"ERROR API (predict): received request, but no 'country' found within"
)
return jsonify(False)
## set the test flag
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
## extract the query parameters
country = request.json['country']
year = request.json['year']
month = request.json['month']
day = request.json['day']
print(country, year, month, day)
## load model
data_dir = os.path.join("data", "cs-train")
all_data, all_models = model_load(data_dir=data_dir)
model = all_models[country]
if not model:
print("ERROR: model is not available")
return jsonify([])
## predict
_result = model_predict(country, year, month, day, test=test)
result = {}
## convert numpy objects to ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
return (jsonify(result))
def predict():
"""
basic predict function for the API
"""
print(f"Request: {request.json}")
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'query' not in request.json:
print(
"ERROR API (predict): received request, but no 'query' found within"
)
return jsonify([])
# if 'type' not in request.json:
# print("WARNING API (predict): received request, but no 'type' was found assuming 'numpy'")
# query_type = 'numpy'
## set the test flag
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
## extract the query
query = request.json['query']
## load model
data_dir = os.path.join(".", "data", "cs-production")
all_data, all_models = model_load(data_dir=data_dir, training=False)
if not all_models:
print("ERROR: model is not available")
return jsonify([])
_result = model_predict(**query, test=test)
result = {}
## convert numpy objects to ensure they are serializable
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
print(f"Result: {result}")
print(f"JSON: {jsonify(result)}")
return (jsonify(result))
def predict():
"""
basic predict function for the API
"""
print(request.json)
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
if 'query' not in request.json:
print("ERROR API (predict): received request, but no 'query' found within")
return jsonify([])
if 'type' not in request.json:
print("WARNING API (predict): received request, but no 'type' was found assuming 'numpy'")
query_type = 'numpy'
query = request.json['query']
if request.json['type'] == 'numpy':
query = np.array(query)
else:
print("ERROR API (predict): only numpy data types have been implemented")
return jsonify([])
## load model
model = model_load()
if not model:
print("ERROR: model is not available")
return jsonify([])
_result = model_predict(query,model)
result = {}
## convert numpy objects so ensure they are serializable
for key,item in _result.items():
if isinstance(item,np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
return(jsonify(result))
def setting(model_config, checkpoint_path):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
_, test_dataloader, _ = data_loader.load_path_loss_with_detail_dataset(
input_dir=model_config['input_dir'],
model_type=model_config['model'],
num_workers=model_config['num_workers'],
batch_size=model_config['batch_size'],
shuffle=model_config['shuffle'],
input_size=model_config['input_size']
if model_config['model'] == 'DNN' else model_config['sequence_length'])
nn_model = model.model_load(model_config)
checkpoint = torch.load(checkpoint_path)
nn_model.load_state_dict(checkpoint['model_state_dict'])
test(model_config=model_config,
nn_model=nn_model,
dataloader=test_dataloader,
device=device)
def predict():
"""
basic predict function for the API
"""
## input checking
if not request.json:
print("ERROR: API (predict): did not receive request data")
return jsonify([])
country = request.json['country']
year = request.json['year']
month = request.json['month']
day = request.json['day']
model = model_load(country)
if not model:
print("ERROR: model is not available")
return jsonify([])
print("... predicting")
result = model_predict(country, year, month, day, model)
print("... prediction complete")
return (jsonify(result))
def main():
print("LOADING MODELS")
production_data_dir = os.path.join("data", "cs-production")
all_data, all_models = model_load(data_dir=production_data_dir)
print("... models loaded: ", ",".join(all_models.keys()))
count = 0
for country in all_data.keys():
if all_data[country]['X'].shape[0] > 0:
for date in all_data[country]['dates']:
dt = datetime.strptime(date, '%Y-%m-%d')
query = {
'country': country,
'year': str(dt.year),
'month': str(dt.month),
'day': str(dt.day)
}
# input checks
if country not in all_models.keys():
result = "ERROR (model_predict) - model for country '{}' could not be found".format(
country)
else:
result = model_predict(query,
data=all_data[country],
model=all_models[country],
test=True)
count += 1
print('result[', count, ']: ', result)
print("model test predict complete.")
def predict():
if not request.json:
print("No request data")
return jsonify([])
if 'country' not in request.json:
print("Please provide the country name")
return jsonify([])
if 'day' not in request.json:
print("Please provide the day")
return jsonify([])
if 'month' not in request.json:
print("Please provide the month")
return jsonify([])
if 'year' not in request.json:
print("Please provide the year")
return jsonify([])
test = False
if 'mode' in request.json and request.json['mode'] == 'test':
test = True
country = request.json['country']
day = request.json['day']
month = request.json['month']
year = request.json['year']
data_dir = os.path.join(".", "data", "cs-train")
all_data, all_models = model_load(data_dir=data_dir)
model = all_models[country]
if not model:
print("Mo models avaliable")
return jsonify([])
_result = model_predict(country, year, month, day, test=test)
result = {}
for key, item in _result.items():
if isinstance(item, np.ndarray):
result[key] = item.tolist()
else:
result[key] = item
return (jsonify(result))
parser.add_argument('--checkpoint',
type=str,
default="models/VideoZoom.pth",
help="checkpint file")
parser.add_argument('--input',
type=str,
default="dataset/predict/input",
help="input folder")
parser.add_argument('--output',
type=str,
default="dataset/predict/output",
help="output folder")
args = parser.parse_args()
model = get_model()
model_load(model, args.checkpoint)
device = model_device()
model.to(device)
model.eval()
enable_amp(model)
totensor = transforms.ToTensor()
toimage = transforms.ToPILImage()
video = Video()
video.reset(args.input)
progress_bar = tqdm(total=len(video))
h = video.height
w = video.width
def model_load(downloaded_path):
# Load "vggnet_model_acc_0.69.hdf5"
return model.model_load(downloaded_path)
def test_and_train(model_name='speech2speech', retrain=True):
"""
Test and/or train on given dataset
@param model_name name of model to save.
@param retrain True if retrain, False if load from pretrained model
"""
(X_train, y_train), (X_val, y_val), (X_test, y_test) = load_dataset(
'raw',
nfft=NFFT,
hop_len=HOP_LENGTH,
fs=FS,
stacked_frames=STACKED_FRAMES,
chunk=CHUNK)
model = None
X_train_norm = normalize_sample(X_train)
X_val_norm = normalize_sample(X_val)
X_test_norm = normalize_sample(X_test)
y_train_norm = normalize_sample(y_train)
y_val_norm = normalize_sample(y_val)
y_test_norm = normalize_sample(y_test)
print("X shape:", X_train_norm.shape, "y shape:", y_train_norm.shape)
# Xtn_strided = stride_over(X_train_norm)
# Xvn_strided = stride_over(X_val_norm)
# Xten_strided = stride_over(X_test_norm)
# Xtn_reshape = Xtn_strided
# Xvn_reshape = Xvn_strided
# Xten_reshape = Xten_strided
# ytn_reshape = y_train_norm.reshape(-1, NFFT//2 + 1, 1, 1)
# yvn_reshape = y_val_norm.reshape(-1, NFFT//2 + 1, 1, 1)
# yten_reshape = y_test_norm.reshape(-1, NFFT//2 + 1, 1, 1)
# train_dataset = tf.data.Dataset.from_tensor_slices((Xtn_reshape,
# ytn_reshape)).batch(X_train_norm.shape[1]).shuffle(X_train.shape[0]).repeat()
# val_dataset = tf.data.Dataset.from_tensor_slices((Xvn_reshape, yvn_reshape)).batch(X_val_norm.shape[1]).repeat(1)
# train_dataset = tf.data.Dataset.from_tensor_slices((X_train_norm, y_train_norm)).batch(BATCH_SIZE).shuffle(BATCH_SIZE).repeat()
# val_dataset = tf.data.Dataset.from_tensor_slices((X_val_norm, y_val_norm)).batch(BATCH_SIZE).repeat(1)
# print(list(train_dataset.as_numpy_iterator())[0])
# Scale the sample X and get the scaler
# scaler = scale_sample(X)
# Check if model already exists and retrain is not being called again
if (os.path.isfile(os.path.join(MODEL_DIR, model_name, 'model.json'))
and not retrain):
model = model_load(model_name)
# Compile the model
model.compile(loss=LOSS, optimizer=OPTIMIZER, metrics=METRICS)
else:
if not os.path.isdir(os.path.join(MODEL_DIR, model_name)):
create_model_directory(model_name)
baseline_val_loss = None
model = None
# model = gen_model(tuple(Xtn_reshape.shape[1:]))
model = gen_model(tuple(X_train_norm.shape[1:]))
print('Created Model...')
model.compile(loss=LOSS, optimizer=OPTIMIZER, metrics=METRICS)
print('Metrics for Model...')
# print(list(train_dataset.as_numpy_iterator())[0])
tf.keras.utils.plot_model(model,
show_shapes=True,
dpi=96,
to_file=os.path.join(MODEL_DIR, model_name,
'model.png'))
print(model.metrics_names)
early_stopping_callback = tf.keras.callbacks.EarlyStopping(
monitor='val_loss', patience=5, restore_best_weights=True)
if (os.path.isfile(path(MODEL_DIR, model_name))):
model.load_weights(path(MODEL_DIR, model_name))
baseline_val_loss = model.evaluate(X_val_norm, y_val_norm)[0]
print(baseline_val_loss)
early_stopping_callback = tf.keras.callbacks.EarlyStopping(
monitor='val_loss',
patience=5,
restore_best_weights=True,
baseline=baseline_val_loss)
log_dir = os.path.join(
LOGS_DIR, 'files',
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=log_dir, update_freq='batch')
model_checkpoint_callback = ModelCheckpoint(monitor='val_loss',
filepath=os.path.join(
MODEL_DIR, model_name,
'model.h5'),
save_best_only=True,
save_weights_only=False,
mode='min')
# fit the keras model on the dataset
cbs = [early_stopping_callback,
model_checkpoint_callback] # tensorboard_callback
model.fit(X_train_norm,
y_train_norm,
epochs=EPOCHS,
validation_data=(X_val_norm, y_val_norm),
verbose=1,
callbacks=cbs,
batch_size=BATCH_SIZE)
print('Model Fit...')
model = save_model(model, model_name)
# model = model_load(model_name)
[loss, mse, accuracy,
rmse] = model.evaluate(X_test_norm, y_test_norm,
verbose=0) # _, mse, accuracy =
print(
'Testing accuracy: {}, Testing MSE: {}, Testing Loss: {}, Testing RMSE: {}'
.format(accuracy * 100, mse, loss, rmse))
# # Randomly pick 1 test
idx = 32
print(idx)
X = X_test_norm[idx]
y = y_test_norm[idx]
# y = y_test_norm[idx].reshape(-1, NFFT//2 + 1)
# min_y, max_y = np.min(y_test_norm[idx]), np.max(y_test_norm[idx])
# min_x, max_x = np.min(y_test_norm[idx]), np.max(y_test_norm[idx])
# print("MinY: {}\tMaxY{}".format(min_y, max_y))
# print("MinX: {}\tMaxX{}".format(min_x, max_x))
X = np.expand_dims(X, axis=0)
# X = stride_over(X)
# mean = np.mean(X)
# std = np.std(X)
# X = (X-mean) / std
print(X.shape)
# y_pred = model.predict(X)
y_pred = np.squeeze(model.predict(X), axis=0)
# y_pred = y_pred.reshape(-1, NFFT//2 + 1)
print(y.shape)
print(y_pred.shape)
y = y.T
y_pred = y_pred.T
X_test_norm = X_test_norm[idx].T
# GriffinLim Vocoder
output_sound = librosa.core.griffinlim(y_pred)
input_sound = librosa.core.griffinlim(X_test_norm)
target_sound = librosa.core.griffinlim(y)
# Play and plot all
play_sound(input_sound, output_sound, target_sound, FS)
if not os.path.isdir(os.path.join(MODEL_DIR, model_name, 'audio_output')):
create_model_directory(os.path.join(model_name, 'audio_output'))
librosa.output.write_wav(path(MODEL_DIR, model_name, 'audio_output',
'input.wav'),
input_sound,
sr=FS,
norm=True)
librosa.output.write_wav(path(MODEL_DIR, model_name, 'audio_output',
'target.wav'),
target_sound,
sr=FS,
norm=True)
librosa.output.write_wav(path(MODEL_DIR, model_name, 'audio_output',
'predicted.wav'),
output_sound,
sr=FS,
norm=True)
return
def startup():
global global_data, global_models
print(".. loading models")
global_data, global_models = model_load(training=False)
print(".. all models loaded")
default="output/ImageColor_D.pth", help="checkpoint D file")
parser.add_argument('--bs', type=int, default=16, help="batch size")
parser.add_argument('--lr', type=float, default=1e-4, help="learning rate")
parser.add_argument('--epochs', type=int, default=1000)
args = parser.parse_args()
# Create directory to store weights
if not os.path.exists(args.outputdir):
os.makedirs(args.outputdir)
# get model
model = get_model(trainning=True)
device = model_device()
model.set_optimizer(args.lr)
model_load(model.net_G, args.checkpoint_g)
if model.use_D:
model_load(model.net_D, args.checkpoint_d)
model.to(device)
lr_scheduler_G = optim.lr_scheduler.StepLR(
model.optimizer_G, step_size=100, gamma=0.1)
if model.use_D:
lr_scheduler_D = optim.lr_scheduler.StepLR(
model.optimizer_D, step_size=100, gamma=0.1)
# get data loader
train_dl, valid_dl = get_data(trainning=True, bs=args.bs)
for epoch in range(args.epochs):