x, y, l_infos, tags_oc = artificiel_occlusion(x, y, l_infos)
list_minibatchs_vl = split_data_to_minibatchs_eval(
{"x": x, "y": y}, ts_batch_size)
fold_exp = "../../exps/" + sys.argv[1]
with open(fold_exp+"/model.pkl", 'r') as f:
stuff = pkl.load(f)
layers_infos, params_vl = stuff["layers_infos"], stuff["params_vl"]
print layers_infos
tag = stuff["tag"]
dropout = stuff["dropout"]
rng = np.random.RandomState(23455)
input = T.fmatrix("x_input")
for l in layers_infos:
l["W"], l["b"], l["rng"] = None, None, rng
model = ModelMLP(layers_infos, input, dropout=dropout)
model.set_params_vals(fold_exp+"/model.pkl")
eval_fn = get_eval_fn(model)
# Unit test
unit_test(fold_exp+"/unit_imgs/", w, h, path_mean_shap, eval_fn, ds)
# Perf mean shape.
# TRAIN
tr_path = "../../inout/data/face/" + faceset + "_data/ch_tr_1800_0_0_0.pkl"
print "TRAIN EVAL:"
with open(tr_path, 'r') as f:
tr_data = pkl.load(f)
list_minibatchs_train = split_data_to_minibatchs_eval(
{"x": tr_data["x"], "y": tr_data['y']}, ts_batch_size)
cdf_ms, cdf0_1_ms, auc_ms = evaluate_mean_shape(path_mean_shap, l_infos,
w, h, y=tr_data['y'],
}
layer3 = {
"rng": rng,
"n_in": nhid_l2,
"n_out": 68*2,
"W": dae_l3.hidden.W_prime,
"b": dae_l3.hidden.b_prime,
"activation": NonLinearity.TANH
}
layers = [layer0, layer1, layer2, layer3]
# dropout = [float(sys.argv[1]), float(sys.argv[2]), float(sys.argv[3]),
# float(sys.argv[4])]
dropout = [0.0, 0.0, 0.0, 0.0]
# number of the hidden layer just before the output ae. Default: None
id_code = None
model = ModelMLP(layers, input, l1_reg=0., l2_reg=0., reg_bias=False,
dropout=dropout, id_code=id_code)
aes_in = []
aes_out = []
if id_code is not None:
assert aes_out != []
# Train
# Data
tr_batch_size = 10
vl_batch_size = 8000
with open(path_valid, 'r') as f:
l_samples_vl = pkl.load(f)
list_minibatchs_vl = split_data_to_minibatchs_eval(
l_samples_vl, vl_batch_size)
max_epochs = int(1000)
lr_vl = 1e-3
