Owleye / codes /crt_train_models.py
owleyetracker's picture
Upload 45 files
b4c5d36 verified
"""This file is for creating and training the neural network models for eye movement prediction. Also, this is for creating and training
The in-out model which is for predicting whether the subject is looking inside of the screen or outside of the screen. To understand this
module, you should know about how to build neural network models with keras and tensorflow"""
from tensorflow.keras.layers import (Input, Conv2D, Flatten, MaxPooling2D,
Dense, Concatenate)
from tensorflow.keras.models import Model
import numpy as np
import os
from tensorflow.keras.callbacks import EarlyStopping
from tensorflow.keras.models import load_model
from sklearn.preprocessing import StandardScaler
from sklearn.utils import shuffle
from joblib import dump as j_dump
from joblib import load as j_load
import random
from codes.base import eyeing as ey
from openpyxl import Workbook
class Modeling():
@staticmethod
def create_io():
"""
creating in-out model (a CNN model) using tensorflow and keras
Parameters:
None
Returns:
None
"""
print("Starting to create an empty in_out model...")
inp1_shape = (ey.EYE_SIZE[0], ey.EYE_SIZE[1]*2, 1)
x2_chosen_features = (0, 1, 2, 3, 4, 5, 6, 7, 8, 9)
inp2_shape = (len(x2_chosen_features),)
inp1 = Input(inp1_shape)
layer = Conv2D(16, (11, 11), (1, 1), 'same', activation='relu')(inp1)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Conv2D(32, (7, 7), (1, 1), 'same', activation='relu')(layer)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Conv2D(64, (5, 5), (1, 1), 'same', activation='relu')(layer)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Conv2D(128, (3, 3), (1, 1), activation='relu')(layer)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Flatten()(layer)
inp2 = Input(inp2_shape)
layer = Concatenate()([layer, inp2])
layer = Dense(256, 'relu')(layer)
layer = Dense(128, 'relu')(layer)
layer = Dense(32, 'relu')(layer)
layer = Dense(8, 'relu')(layer)
output_layer = Dense(1, "sigmoid")(layer)
input_layers = [inp1, inp2]
model = Model(inputs=input_layers, outputs=output_layer)
model.compile(optimizer="adam", loss="binary_crossentropy", metrics="acc")
print(model.summary())
n_weights = np.sum([np.prod(v.get_shape()) for v in model.trainable_weights])
mdl_num = ey.find_max_mdl(ey.io_raw_dir) + 1
info = {"n_weights": n_weights,
"input1_shape": inp1_shape,
"input2_shape": inp2_shape,
"x2_chosen_features": x2_chosen_features}
mdl_name = ey.MDL + f"{mdl_num}"
mdl_dir = ey.io_raw_dir + mdl_name + ".h5"
model.save(mdl_dir)
ey.save([info], ey.io_raw_dir, [mdl_name])
print("\nEmpty in_out model created and saved to " + mdl_dir)
@staticmethod
def create_et():
"""
Creating eye tracking model (CNN model) using tensorflow and keras. You can change the structure in following, as you want.
Parameters:
None
Returns:
None
"""
print("Starting to create empty eye_tracking models...")
inp1_shape = (ey.EYE_SIZE[0], ey.EYE_SIZE[1]*2, 1)
x2_chosen_features = (0, 1, 2, 3, 4, 5, 6, 7, 8, 9)
inp2_shape = (len(x2_chosen_features),)
inp1 = Input(inp1_shape)
layer = Conv2D(16, (11, 11), (1, 1), 'same', activation='relu')(inp1)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Conv2D(32, (7, 7), (1, 1), 'same', activation='relu')(layer)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Conv2D(64, (5, 5), (1, 1), 'same', activation='relu')(layer)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Conv2D(128, (3, 3), (1, 1), activation='relu')(layer)
layer = MaxPooling2D((2, 2), (2, 2))(layer)
layer = Flatten()(layer)
inp2 = Input(inp2_shape)
layer = Concatenate()([layer, inp2])
layer = Dense(256, 'relu')(layer)
layer = Dense(128, 'relu')(layer)
layer = Dense(32, 'relu')(layer)
layer = Dense(8, 'relu')(layer)
out = Dense(1, 'linear')(layer)
input_layers = [inp1, inp2]
model = Model(inputs=input_layers, outputs=out)
model.compile(optimizer='adam', loss='mse')
print(model.summary())
n_weights = np.sum([np.prod(v.get_shape()) for v in model.trainable_weights])
mdl_num = ey.find_max_mdl(ey.et_raw_dir) + 1
info = {"n_weights": n_weights,
"input1_shape": inp1_shape,
"input2_shape": inp2_shape,
"x2_chosen_features": x2_chosen_features}
mdl_name = ey.MDL + f"{mdl_num}"
mdl_dir = ey.et_raw_dir + mdl_name + ".h5"
model.save(mdl_dir)
ey.save([info], ey.et_raw_dir, [mdl_name])
print("\nEmpty eye_tracking model created and saved to " + mdl_dir)
@staticmethod
def train_io(
subjects,
models_list,
min_max_brightness_ratio=[[0.65, 1.45]],
r_train_list=[0.85],
n_epochs_patience=[[160, 10]],
save_scaler=False,
show_model=False
):
"""
Training the io models. This method uses the dataset in the io folder of subject's number folder. The parameters should be lists.
So, you can train each model with several parameters and hyper parameters to see which one works better.
Parameters:
subjects: a list of subject numbers that you want to train the model with them.
models_list: You can train several models at a same time. So, you can enter a list of model numbers
min_max_brightness_ratio: To make the models robust to the brightness, the eyes images are multiplies into a number between two considered numbers
r_train_list: The ratio for train dataset
n_epochs_patience: The number of epochs and patience to intrupt training
save_scaler: To save the scaler
show_model: To show the model
Returns:
None
"""
print("Starting to train in_out model...")
# Loading all subjects
x1_load = []
x2_load = []
y_load = []
for sbj in subjects:
data_io_dir = ey.create_dir([ey.subjects_dir, f"{sbj}", ey.IO])
x1_load0, x2_load0, y_load0 = ey.load(data_io_dir, [ey.X1, ey.X2, ey.Y])
for (x10, x20, y10) in zip(x1_load0[0], x2_load0[0], y_load0[0]):
x1_load.append(x10)
x2_load.append(x20)
y_load.append(y10)
x1_load = np.array(x1_load)
x2_load = np.array(x2_load)
y_load = np.array(y_load)
n_smp = x1_load.shape[0]
print(f"\nNumber of samples : {n_smp}")
# Going through each brightness in min_max_brightness_ratio list
j = 1
for mbr in min_max_brightness_ratio:
x1_new = x1_load.copy()
for (i, _) in enumerate(x1_load):
r = random.uniform(mbr[0], mbr[1])
x1_new[i] = (x1_new[i] * r).astype(np.uint8)
# Going through each model
for raw_mdl_num in models_list:
info = ey.load(ey.io_raw_dir, [ey.MDL + f"{raw_mdl_num}"])[0]
x2_chosen_features = info["x2_chosen_features"]
x2_new = x2_load[:, x2_chosen_features]
x1_shf, x2_shf, y_shf = shuffle(x1_new, x2_new, y_load)
x1_scaler = ey.X1_SCALER
x1 = x1_shf / x1_scaler
x2_scaler = StandardScaler()
x2 = x2_scaler.fit_transform(x2_shf)
scalers = [x1_scaler, x2_scaler]
if save_scaler:
j_dump(scalers, ey.scalers_dir + f"scl_io_{len(x2_chosen_features)}.bin")
# Going through each training ratio in r_train_list
for rt in r_train_list:
n_train = int(rt * n_smp)
x1_train, x2_train = x1[:n_train], x2[:n_train]
x1_val, x2_val = x1[n_train:], x2[n_train:]
y_train = y_shf[:n_train]
y_val = y_shf[n_train:]
print("\nTrain and val data shape:")
print(x1_train.shape, x1_val.shape, x2_train.shape, x2_val.shape,
y_train.shape, y_val.shape)
x_train = [x1_train, x2_train]
x_val = [x1_val, x2_val]
# Going throught each epoch and patience in n_epochs_patience
for nep in n_epochs_patience:
# Training the models
info["min_max_brightness_ratio"] = mbr
info["r_train"] = rt
info["n_epochs_patience"] = nep
cb = EarlyStopping(patience=nep[1], verbose=1, restore_best_weights=True)
raw_model_dir = ey.io_raw_dir + ey.MDL + f"{raw_mdl_num}.h5"
print("\nLoading blink_in_out model from " + raw_model_dir)
model = load_model(raw_model_dir)
if show_model:
print(model.summary())
print(f"\n<<<<<<< {j}-model:{raw_mdl_num}-min_max_ratio:{mbr}-r_train:{rt}-epoch_patience:{nep} >>>>>>>>")
model.fit(x_train,
y_train,
validation_data=(x_val, y_val),
epochs=nep[0],
callbacks=cb)
train_loss = model.evaluate(x_train, y_train)
val_loss = model.evaluate(x_val, y_val)
info["train_loss"] = train_loss
info["val_loss"] = val_loss
trained_mdl_num = ey.find_max_mdl(ey.io_trained_dir) + 1
mdl_name = ey.MDL + f'{trained_mdl_num}'
ey.save([info], ey.io_trained_dir, [mdl_name])
mdl_tr_dir = ey.io_trained_dir + mdl_name + ".h5"
model.save(mdl_tr_dir)
print("\nSaving in_out model in " + mdl_tr_dir)
j += 1
@staticmethod
def train_et(
subjects,
models_list,
min_max_brightness_ratio=[[0.65, 1.45]],
r_train_list=[0.8],
n_epochs_patience=[[100, 15]],
shift_samples=None,
blinking_threshold="d",
save_scaler=False,
show_model=False
):
"""
Training the et (base) models. This method uses the dataset in the et folder of subject's number folder. The parameters should be lists.
So, you can train each model with several parameters and hyper parameters to see which one works better.
Parameters:
subjects: a list of subject numbers that you want to train the model with them.
models_list: You can train several models at a same time. So, you can enter a list of model numbers
min_max_brightness_ratio: To make the models robust to the brightness, the eyes images are multiplies into a number between two considered numbers
r_train_list: The ratio for train dataset
n_epochs_patience: The number of epochs and patience to intrupt training
shift_samples: To shift sample if there is a high delay
blinking_threshold: It can have three types --> d: default, ao: app offered, uo: user offered
save_scaler: To save the scaler
show_model: To show the model
Returns:
None
"""
print("Starting to train eye_tracking models...")
# Loading all subjects
x1_load = []
x2_load = []
y_load = []
kk = 0
for sbj in subjects:
sbj_dir = ey.create_dir([ey.subjects_dir, f"{sbj}"])
sbj_clb_dir = ey.create_dir([sbj_dir, ey.CLB])
(
sbj_x1_load,
sbj_x2_load,
sbj_y_load,
sbj_t_mat,
sbj_eyes_ratio
) = ey.load(sbj_clb_dir, [ey.X1, ey.X2, ey.Y, ey.T, ey.ER])
# If there is any shifting samples, doing that
if shift_samples:
if shift_samples[kk]:
ii = 0
for (x11, x21, y1, t1, eyr1) in zip(sbj_x1_load, sbj_x2_load, sbj_y_load, sbj_t_mat, sbj_eyes_ratio):
sbj_t_mat[ii] = t1[:-shift_samples[kk]]
sbj_x1_load[ii] = x11[shift_samples[kk]:]
sbj_x2_load[ii] = x21[shift_samples[kk]:]
sbj_y_load[ii] = y1[:-shift_samples[kk]]
sbj_eyes_ratio[ii] = eyr1[shift_samples[kk]:]
ii += 1
kk += 1
sbj_er_dir = ey.create_dir([sbj_dir, ey.ER])
# Removing the samples that are during blinking
sbj_blinking_threshold = ey.get_threshold(sbj_er_dir, blinking_threshold)
sbj_blinking = ey.get_blinking(sbj_t_mat, sbj_eyes_ratio, sbj_blinking_threshold)[1]
for (x11, x21, y1, b1) in zip(sbj_x1_load, sbj_x2_load, sbj_y_load, sbj_blinking):
for (x10, x20, y0, b0) in zip(x11, x21, y1, b1):
if not b0:
x1_load.append(x10)
x2_load.append(x20)
y_load.append(y0)
x1_load = np.array(x1_load)
x2_load = np.array(x2_load)
y_load = np.array(y_load)
n_smp = x1_load.shape[0]
print(f"\nNumber of samples : {n_smp}")
# Going through each brightness in min_max_brightness_ratio list
j = 1
for mbr in min_max_brightness_ratio:
x1_new = x1_load.copy()
for (i, _) in enumerate(x1_load):
r = random.uniform(mbr[0], mbr[1])
x1_new[i] = (x1_new[i] * r).astype(np.uint8)
# Going through each model
for raw_mdl_num in models_list:
info = ey.load(ey.et_raw_dir, [ey.MDL + f"{raw_mdl_num}"])[0]
x2_chosen_features = info["x2_chosen_features"]
x2_new = x2_load[:, x2_chosen_features]
x1_shf, x2_shf, y_hrz_shf, y_vrt_shf = shuffle(x1_new, x2_new, y_load[:, 0], y_load[:, 1])
x1_scaler = ey.X1_SCALER
x1 = x1_shf / x1_scaler
x2_scaler = StandardScaler()
x2 = x2_scaler.fit_transform(x2_shf)
y_scaler = ey.Y_SCALER
scalers = [x1_scaler, x2_scaler, y_scaler]
if save_scaler:
j_dump(scalers, ey.scalers_dir + f"scl_et_{len(x2_chosen_features)}.bin")
# Going through each training ratio in r_train_list
for rt in r_train_list:
n_train = int(rt * n_smp)
x1_train, x2_train = x1[:n_train], x2[:n_train]
x1_val, x2_val = x1[n_train:], x2[n_train:]
y_hrz_train, y_vrt_train = y_hrz_shf[:n_train], y_vrt_shf[:n_train]
y_hrz_val, y_vrt_val = y_hrz_shf[n_train:], y_vrt_shf[n_train:]
print("\nTrain and val data shape:")
print(x1_train.shape, x1_val.shape, x2_train.shape, x2_val.shape,
y_hrz_train.shape, y_hrz_val.shape, y_vrt_train.shape, y_vrt_val.shape)
x_train = [x1_train, x2_train]
x_val = [x1_val, x2_val]
# Going throught each epoch and patience in n_epochs_patience
for nep in n_epochs_patience:
# Training the models
info["min_max_brightness_ratio"] = mbr
info["r_train"] = rt
info["n_epochs_patience"] = nep
cb = EarlyStopping(patience=nep[1], verbose=1, restore_best_weights=True)
raw_model_dir = ey.et_raw_dir + ey.MDL + f"{raw_mdl_num}.h5"
print("\nLoading eye_tracking model from " + raw_model_dir)
model_hrz = load_model(raw_model_dir)
model_vrt = load_model(raw_model_dir)
if show_model:
print(model_hrz.summary())
trained_mdl_num = ey.find_max_mdl(ey.et_trained_dir, b=-7) + 1
print(f"\n<<<<<<< {j}-model-hrz:{raw_mdl_num}-min_max_ratio:{mbr}-r_train:{rt}-epoch_patience:{nep} >>>>>>>>")
model_hrz.fit(x_train,
y_hrz_train * y_scaler,
validation_data=(x_val, y_hrz_val * y_scaler),
epochs=nep[0],
callbacks=cb)
mdl_name = ey.MDL + f"{trained_mdl_num}"
mdl_hrz_tr_dir = ey.et_trained_dir + mdl_name + "-hrz.h5"
print("\nSaving horizontally eye_tracking model in " + mdl_hrz_tr_dir)
model_hrz.save(mdl_hrz_tr_dir)
hrz_train_loss = model_hrz.evaluate(x_train, y_hrz_train * y_scaler)
hrz_val_loss = model_hrz.evaluate(x_val, y_hrz_val * y_scaler)
info["hrz_train_loss"] = hrz_train_loss
info["hrz_val_loss"] = hrz_val_loss
print(f"\n<<<<<<< {j}-model-vrt:{raw_mdl_num}-min_max_ratio:{mbr}-r_train:{rt}-epoch_patience:{nep} >>>>>>>>")
model_vrt.fit(x_train,
y_vrt_train * y_scaler,
validation_data=(x_val, y_vrt_val * y_scaler),
epochs=nep[0],
callbacks=cb)
tr_model_vrt_dir = ey.et_trained_dir + mdl_name + f"-vrt.h5"
print("Saving vertically eye_tracking model in " + tr_model_vrt_dir)
model_vrt.save(tr_model_vrt_dir)
vrt_train_loss = model_vrt.evaluate(x_train, y_vrt_train * y_scaler)
vrt_val_loss = model_vrt.evaluate(x_val, y_vrt_val * y_scaler)
info["vrt_train_loss"] = vrt_train_loss
info["vrt_val_loss"] = vrt_val_loss
ey.save([info], ey.et_trained_dir, [mdl_name])
j += 1
@staticmethod
def get_models_information(io=True, raw=True, show_model=False):
"""
To write the models information in an excel file. It gets the information from attached pickle file for each model.
There are raw models and trained models in the io and the et.
Parameters:
io: If it's io or et
raw: If the model is trained or not
show_model: If you want to show the model
Returns:
None
"""
wb = Workbook()
ws = wb.active
ws['A1'] = "# of model"
ws['B1'] = "# of weights"
ws['C1'] = "input 1 shape"
ws['D1'] = "input 2 shape"
ws['E1'] = "x2 chosen features"
if io:
if raw:
files_name = os.listdir(ey.io_raw_dir)
if files_name:
for fn in files_name:
if fn[-7:] == ".pickle":
mdl_num = int(fn[3:-7])
mdl_name = ey.MDL + f"{mdl_num}"
if show_model:
mdl_dir = ey.io_raw_dir + mdl_name + ".h5"
mdl = load_model(mdl_dir)
print(f"<<<<<<<<<<<<<< {mdl_dir} >>>>>>>>>>>>>>")
print(mdl.summary())
info = ey.load(ey.io_raw_dir, [mdl_name])[0]
ws[f'A{mdl_num+1}'] = str(mdl_num)
ws[f'B{mdl_num+1}'] = str(info['n_weights'])
ws[f'C{mdl_num+1}'] = str(info['input1_shape'])
ws[f'D{mdl_num+1}'] = str(info['input2_shape'])
ws[f'E{mdl_num+1}'] = str(info['x2_chosen_features'])
else:
ws['F1'] = "min-Max brightness ratio"
ws['G1'] = "r_train"
ws['H1'] = "# of epochs and patience"
ws['I1'] = "train loss"
ws['J1'] = "val loss"
files_name = os.listdir(ey.io_trained_dir)
if files_name:
for fn in files_name:
if fn[-7:] == ".pickle":
mdl_num = int(fn[3:-7])
mdl_name = ey.MDL + f"{mdl_num}"
if show_model:
mdl_dir = ey.io_trained_dir + mdl_name + ".h5"
mdl = load_model(mdl_dir)
print(f"<<<<<<<<<<<<<< {mdl_dir} >>>>>>>>>>>>>>")
print(mdl.summary())
info = ey.load(ey.io_trained_dir, [mdl_name])[0]
ws[f'A{mdl_num+1}'] = str(mdl_num)
ws[f'B{mdl_num+1}'] = str(info['n_weights'])
ws[f'C{mdl_num+1}'] = str(info['input1_shape'])
ws[f'D{mdl_num+1}'] = str(info['input2_shape'])
ws[f'E{mdl_num+1}'] = str(info['x2_chosen_features'])
ws[f'F{mdl_num+1}'] = str(info['min_max_brightness_ratio'])
ws[f'G{mdl_num+1}'] = str(info['r_train'])
ws[f'H{mdl_num+1}'] = str(info['n_epochs_patience'])
ws[f'I{mdl_num+1}'] = str(info['train_loss'])
ws[f'J{mdl_num+1}'] = str(info['val_loss'])
else:
if raw:
files_name = os.listdir(ey.et_raw_dir)
if files_name:
for fn in files_name:
if fn[-7:] == ".pickle":
mdl_num = int(fn[3:-7])
mdl_name = ey.MDL + f"{mdl_num}"
if show_model:
mdl_dir = ey.et_raw_dir + mdl_name + ".h5"
mdl = load_model(mdl_dir)
print(f"<<<<<<<<<<<<<< {mdl_dir} >>>>>>>>>>>>>>")
print(mdl.summary())
info = ey.load(ey.et_raw_dir, [mdl_name])[0]
ws[f'A{mdl_num+1}'] = str(mdl_num)
ws[f'B{mdl_num+1}'] = str(info['n_weights'])
ws[f'C{mdl_num+1}'] = str(info['input1_shape'])
ws[f'D{mdl_num+1}'] = str(info['input2_shape'])
ws[f'E{mdl_num+1}'] = str(info['x2_chosen_features'])
else:
ws['F1'] = "min-Max brightness ratio"
ws['G1'] = "r_train"
ws['H1'] = "# of epochs and patience"
ws['I1'] = "model-hrz train loss"
ws['J1'] = "model-hrz val loss"
ws['K1'] = "model-vrt train loss"
ws['L1'] = "model-vrt val loss"
files_name = os.listdir(ey.et_trained_dir)
if files_name:
for fn in files_name:
if fn[-7:] == ".pickle":
mdl_num = int(fn[3:-7])
mdl_name = ey.MDL + f"{mdl_num}"
if show_model:
mdl_dir = ey.et_trained_dir + mdl_name + "-hrz.h5"
mdl = load_model(mdl_dir)
print(f"<<<<<<<<<<<<<< {mdl_dir} >>>>>>>>>>>>>>")
print(mdl.summary())
info = ey.load(ey.et_trained_dir, [mdl_name])[0]
ws[f'A{mdl_num+1}'] = str(mdl_num)
ws[f'B{mdl_num+1}'] = str(info['n_weights'])
ws[f'C{mdl_num+1}'] = str(info['input1_shape'])
ws[f'D{mdl_num+1}'] = str(info['input2_shape'])
ws[f'E{mdl_num+1}'] = str(info['x2_chosen_features'])
ws[f'F{mdl_num+1}'] = str(info['min_max_brightness_ratio'])
ws[f'G{mdl_num+1}'] = str(info['r_train'])
ws[f'H{mdl_num+1}'] = str(info['n_epochs_patience'])
ws[f'I{mdl_num+1}'] = str(info['hrz_train_loss'])
ws[f'J{mdl_num+1}'] = str(info['hrz_val_loss'])
ws[f'K{mdl_num+1}'] = str(info['vrt_train_loss'])
ws[f'L{mdl_num+1}'] = str(info['vrt_val_loss'])
if io and raw:
info_name = "info_io_raw"
elif io and not raw:
info_name = "info_io_trained"
elif not io and raw:
info_name = "info_et_raw"
else:
info_name = "info_et_trained"
wb.save(ey.files_dir + info_name + ".xlsx")