def step(self, batch_index, mode):
'''
This function is one step in an epoch and will run a training or testing step depending on the parameter.
Args:
batch_index (int): step number for the epoch
mode (str): 'train' or 'test' based on the mode of
Returns:
Dictionary of predictions, answers, loss, number skipped, and the normal and gradient parameters
'''
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn # Theano function set
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
input_mask = input_masks[batch_index]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
gradient_value = self.get_gradient_fn(inp, q, ans, input_mask) # Get and calculate the gradient function
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
print("==> gradient is nan at index %d." % batch_index)
print("==> skipping")
skipped = 1
if skipped == 0:
ret = theano_fn(inp, q, ans, input_mask) # Run the theano function
else:
ret = [-1, -1]
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"skipped": skipped,
"log": "pn: %.3f \t gn: %.3f" % (param_norm, grad_norm)
}
def conjugate_gradient(params, func, init_values, stop_condition=1e-2):
# * PRP
values = Matrix(init_values)
lam = Symbol('lam')
beta = 0
previous_d = 0
previous_g = 0
step = 0
while True:
g = get_grad(params, func)
g = g.subs(dict(zip(params, list(values))))
if get_norm(g) <= stop_condition:
return list(values), func.subs(dict(zip(params, list(values))))
if previous_g != 0:
beta = (g.T * (g - previous_g)) / (get_norm(previous_g)**2)
d = -g + beta[0] * previous_d
else:
d = -g
lam_func = func.subs(dict(zip(params, list(values + lam * d))))
lam_value = get_stagnation(lam_func)
values = values + lam_value * d
previous_d = d
previous_g = g
f_value = func.subs(dict(zip(params, list(values))))
print('step: {} params: {} f: {}'.format(step, list(values),
f_value))
step += 1
def step(self, batch_index, mode):
'''
This function is one step in an epoch and will run a training or testing step depending on the parameter.
Args:
batch_index (int): step number for the epoch
mode (str): 'train' or 'test' based on the mode of
Returns:
Dictionary of predictions, answers, loss, number skipped, and the normal and gradient parameters
'''
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn # Theano function set
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
input_mask = input_masks[batch_index]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
gradient_value = self.get_gradient_fn(inp, q, ans, input_mask) # Get and calculate the gradient function
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
print "==> gradient is nan at index %d." % batch_index
print "==> skipping"
skipped = 1
if skipped == 0:
ret = theano_fn(inp, q, ans, input_mask) # Run the theano function
else:
ret = [-1, -1]
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"skipped": skipped,
"log": "pn: %.3f \t gn: %.3f" % (param_norm, grad_norm)
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
sgates = self.train_gates
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
sgates = self.test_gates
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
input_mask = input_masks[batch_index]
sgate = sgates[batch_index]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
gradient_value = self.get_gradient_fn(inp, q, ans, input_mask,
sgate)
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
#print "==> gradient is nan at index %d." % batch_index
#print "==> skipping"
skipped = 1
if skipped == 0:
ret = theano_fn(inp, q, ans, input_mask, sgate)
else:
ret = [-1, -1, -1, -1, -1]
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"log": "pn: %.3f" % param_norm,
"skipped": skipped
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
fact_counts = self.train_fact_count
input_masks = self.train_input_mask
if mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
fact_counts = self.test_fact_count
input_masks = self.test_input_mask
start_index = batch_index * self.batch_size
inp = inputs[start_index:start_index+self.batch_size]
q = qs[start_index:start_index+self.batch_size]
ans = answers[start_index:start_index+self.batch_size]
fact_count = fact_counts[start_index:start_index+self.batch_size]
input_mask = input_masks[start_index:start_index+self.batch_size]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
gradient_value = self.get_gradient_fn(inp, q, ans, fact_count, input_mask)
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
print "==> gradient is nan at index %d." % batch_index
print "==> skipping"
skipped = 1
if skipped == 0:
ret = theano_fn(inp, q, ans, fact_count, input_mask)
else:
ret = [float('NaN'), float('NaN')]
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": ret[0],
"answers": ans,
"current_loss": ret[1],
"skipped": skipped,
"grad_norm": grad_norm,
"param_norm": param_norm,
"log": "",
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
sgates = self.train_gates
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
sgates = self.test_gates
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
input_mask = input_masks[batch_index]
sgate = sgates[batch_index]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
gradient_value = self.get_gradient_fn(inp, q, ans, input_mask,sgate)
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
#print "==> gradient is nan at index %d." % batch_index
#print "==> skipping"
skipped = 1
if skipped==0:
ret = theano_fn(inp, q, ans, input_mask, sgate)
else:
ret=[-1,-1,-1,-1,-1]
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"log": "pn: %.3f" % param_norm,
"skipped": skipped
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
input_mask = input_masks[batch_index]
ret = theano_fn(inp, q, ans, input_mask)
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
if mode == "test":
theano_fn = self.test_fn
inp, q, ans, ans_inp, ans_mask = self._process_batch_sind(
batch_index, mode)
ret = theano_fn(inp, q, ans, ans_mask, ans_inp)
#theano_fn.profile.print_summary()
#sys.exit()
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": ret[0],
"answers": ans,
"current_loss": ret[1],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
fact_counts = self.train_fact_count
input_masks = self.train_input_mask
if mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
fact_counts = self.test_fact_count
input_masks = self.test_input_mask
start_index = batch_index * self.batch_size
inp = inputs[start_index:start_index+self.batch_size]
q = qs[start_index:start_index+self.batch_size]
ans = answers[start_index:start_index+self.batch_size]
fact_count = fact_counts[start_index:start_index+self.batch_size]
input_mask = input_masks[start_index:start_index+self.batch_size]
ret = theano_fn(inp, q, ans, fact_count, input_mask)
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": ret[0],
"answers": ans,
"current_loss": ret[1],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
input_mask = input_masks[batch_index]
ret = theano_fn(inp, q, ans, input_mask)
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
}
def quasi_newton(params, func, init_values, stop_condition=1e-5):
# * BFGS
values = Matrix(init_values)
lam = Symbol('lam')
next_g = 0
next_values = 0
h = eye(len(params))
step = 0
while True:
g = get_grad(params, func)
g = g.subs(dict(zip(params, list(values))))
d = -h**(-1) * g
lam_func = func.subs(dict(zip(params, list(values + lam * d))))
lam_value = get_stagnation(lam_func)
next_values = values + lam_value * d
if get_norm(g) <= stop_condition:
return list(values), func.subs(dict(zip(params,
list(next_values))))
else:
next_g = get_grad(params, func)
next_g = next_g.subs(dict(zip(params, list(next_values))))
s = next_values - values
y = next_g - g
h = (eye(len(params)) - (s * y.T) /
(s.T * y)[0]) * h * (eye(len(params)) - (s * y.T) /
(s.T * y)[0]).T + (s * s.T) / (s.T *
y)[0]
values = next_values
f_value = func.subs(dict(zip(params, list(values))))
print('step: {} params: {} f: {}'.format(step, list(values),
f_value))
step += 1
def measure_velocity_sensor(self, poiList, rover): min_dist = self.min_sensor_dist_sqr max_dist = self.max_sensor_dist sum = np.zeros(4) for poi in poiList: # get quadrant of POI vect = utils.vect_sub(poi.pos, rover.pos) dist = utils.get_norm(vect) angle = utils.get_angle(vect) % (2 * math.pi) # Between 0 to 2pi relative_angle = (angle - rover.heading + math.pi / 2) % (2 * math.pi) q = utils.get_quadrant(relative_angle) - 1 # get relative velocity of POI to agent. rel_vel_vect = poi.vel_lin rel_pos_vect = utils.vect_sub(rover.pos, poi.pos) rel_pos_norm = utils.get_norm(rel_pos_vect) rel_pos_unit = [rel_pos_vect[0]/rel_pos_norm, rel_pos_vect[1]/rel_pos_norm] dot = np.dot(rel_pos_unit, rel_vel_vect) normalized_dot = poi.value * dot / rel_pos_norm**2 sum[q] += normalized_dot return list(sum)
def newton(params, func, init_values, stop_condition=1e-2):
values = Matrix(init_values)
step = 0
while True:
g = get_grad(params, func)
g = g.subs(dict(zip(params, list(values))))
if get_norm(g) <= stop_condition:
return list(values), func.subs(dict(zip(params, list(values))))
h = get_hessian(params, func)
h = h.subs(dict(zip(params, list(values))))
values = values - h**(-1) * g
f_value = func.subs(dict(zip(params, list(values))))
print('step: {} params: {} f: {}'.format(step, list(values),
f_value))
step += 1
def return_sensor_rover(self, roverList, quadrant, max_dist=500):
min_dist = 10
sum = 0
for rover in roverList:
vect = utils.vect_sub(rover.pos, self.pos)
dist = utils.get_norm(vect)
angle = utils.get_angle(vect) % (2 * math.pi) # Between 0 to 2pi
relative_angle = (angle - self.heading + math.pi / 2) % (2 *
math.pi)
# print 'Vect: ', vect
# print 'Angle: ', angle*360/2/math.pi, relative_angle*360/2/math.pi
if dist < max_dist and utils.check_quadrant(
relative_angle, quadrant):
# print 'I SEE YOU', quadrant
sum += 1 / max(dist**2, min_dist**2)
return sum
def steepest_descent(params, func, init_values, stop_condition=1e-10):
values = Matrix(init_values)
lam = Symbol('lam')
step = 0
while True:
g = get_grad(params, func)
g = g.subs(dict(zip(params, list(values))))
if get_norm(g) <= stop_condition:
return list(values), func.subs(dict(zip(params, list(values))))
lam_func = func.subs(dict(zip(params, list(values - lam * g))))
lam_value = get_stagnation(lam_func)
values = values - lam_value * g
f_value = func.subs(dict(zip(params, list(values))))
print('step: {} params: {} f: {}'.format(step, list(values),
f_value))
step += 1
def learn_lr_classifier(training_corpus):
D = get_vocabulary_size()
labels = get_labels()
w = [0] * (D + 1)
norm = 1.0
num_iters = 0
while norm > convergence_threshold:
num_iters += 1
if num_iters > max_iters:
break
old_w = list(w)
shuffled = list(training_corpus)
shuffle(shuffled)
for vector in shuffled:
label = 1.0 if float(vector[0]) == labels[0] else 0.0
prediction = get_prediction(vector[1:], w)
delta = label - prediction
update_weights(vector[1:], w, delta)
norm = get_norm(w,old_w)
return w
def main():
"""Fuction to compute 1-dimensional correlation matrix using kernel density
estimation method
"""
data = np.load(INPUT_PATH)
data = get_norm(data)
num_atoms = data.shape[1]
corr_matrix = np.zeros((num_atoms, num_atoms))
for row in range(num_atoms):
# Compute only inferior diagonal matrix
for col in range(row):
corr_matrix[row, col] = mi_kde(data, row, col)
print(row, col, corr_matrix[row, col])
corr_matrix = gen_corr_coef(corr_matrix, dim=1)
np.save(file=OUTPUT_PATH, arr=corr_matrix)
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
if mode == "test" or mode == 'val':
theano_fn = self.test_fn
q, ans, ans_inp, ans_mask, img_ids = self._process_batch_sind(
batch_index, mode)
ret = theano_fn(q, ans, ans_mask, ans_inp)
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": ret[0],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
fact_counts = self.train_fact_count
input_masks = self.train_input_mask
if mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
fact_counts = self.test_fact_count
input_masks = self.test_input_mask
start_index = batch_index * self.batch_size
inp = inputs[start_index:start_index + self.batch_size]
q = qs[start_index:start_index + self.batch_size]
ans = answers[start_index:start_index + self.batch_size]
fact_count = fact_counts[start_index:start_index + self.batch_size]
input_mask = input_masks[start_index:start_index + self.batch_size]
inp, q, ans, fact_count, input_mask = self._process_batch(
inp, q, ans, fact_count, input_mask)
ret = theano_fn(inp, q, ans, fact_count, input_mask)
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": ret[0],
"answers": ans,
"current_loss": ret[1],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
}
def step(self,batch_index,mode):
if mode== "train" and self.mode== "test":
raise Exception("Cannot train during test mode")
start_index=batch_index * self.batch_size
inputs, qs, answers, fact_counts, input_masks,img_feats =self.process_vqa_data(mode,start_index,start_index+self.batch_size)
if mode=="train":
theano_fn=self.train_fn
# inputs = self.process_vqa_data(self.h5file['cap_train'][start_index:start_index+self.batch_size])
# qs=self.process_vqa_data(self.h5file['ques_train' ][ start_index:start_index+self.batch_size] )
# answers=self.process_vqa_data(self.h5file['answers'][start_index:start_index+self.batch_size] )
# fact_counts=np.zeros(self.batch_size,dtype="int")
# fact_counts.fill(20)
# input_masks= process_masks( inputs ) # figure it out
if mode=="test":
theano_fn=self.test_fn
# inputs=self.process_vqa_data( self.h5file['cap_test'][start_index:start_index+self.batch_size ] )
# qs=self.process_vqa_data( self.h5file['ques_test'][start_index:start_index+self.batch_size ] )
# answers=self.process_vqa_data( self.h5file['ans_test'][start_index:start_index+self.batch_size ] )
# fact_counts=np.zeros(self.batch_size,dtype="int")
# fact_counts.fill(20)
# input_masks= process_masks( inputs ) # figure it out
inp,q,ans,fact_count,input_mask,img_feat=self._process_batch(inputs,qs,answers,fact_counts,input_masks,img_feats )
img_feat=img_feat.reshape((self.batch_size*self.img_seq_len,self.img_vector_size))
ret = theano_fn( inp,q,ans,fact_count,input_mask,img_feat,self.lr)
param_norm=np.max( [ utils.get_norm( x.get_value()) for x in self.params])
return { "prediction":ret[0],
"answers":ans,
"current_loss":ret[1],
"skipped":0,
"log":"pn: %.3f" % param_norm
}
'W_mem_upd_in', 'W_mem_upd_hid', 'b_mem_upd', 'W_mem_hid_in',
'W_mem_hid_hid', 'b_mem_hid', 'W_b', 'W_1', 'W_2', 'b_1', 'b_2', 'W_a'
]
fig, ax = plt.subplots(figsize=(9, 4))
with open(file_name, 'r') as load_file:
dict = pickle.load(load_file)
loaded_params = dict['params']
if flag:
for i in xrange(len(params)):
if params[i] in to_write:
out_obj[params[i]] = loaded_params[i]
with open(sys.argv[2], 'w') as save_file:
pickle.dump(obj=out_obj, file=save_file, protocol=-1)
print "finish dumping file to " + sys.argv[2]
for (x, y) in zip(params, loaded_params):
n = y.shape
if len(n) == 1:
n = n[0]
else:
n = n[0] * n[1]
norm = utils.get_norm(y) / n**0.5
print x, ' shape: ', y.shape, ', norm: ', norm, ', max: ', np.max(
np.abs(y))
if len(y.shape) > 1:
ax.imshow(y, cmap='Blues', interpolation='none')
plt.title('Train. ' + x + ', norm ' + str(norm))
fig.show()
input_str = raw_input("Press ENTER to continue.")
def train(dataset, alpha, A_type, normalize_type, model_pretrained_params,
model_type, batch_size, test_batch_size, negative_nums, item_emb_dim,
hid_dim1, hid_dim2, hid_dim3, lr_emb, l2_emb, lr_gcn, l2_gcn, lr_cnn,
l2_cnn, epochs, params_file_name):
# init
if dataset == 'LastFM':
# use LastFM dataset
data_obj = LastfmData()
elif dataset == 'Diginetica':
# use Diginetica dataset
data_obj = DigineticaData()
else:
# use yoochoose1_64 dataset
data_obj = YoochooseData(dataset=dataset)
# gpu device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# init A
# A: type=scipy.sparse
A = data_obj.get_decay_adj(
data_obj.d, tail=None,
alpha=alpha) if A_type == 'decay' else data_obj.get_gcn_adj(data_obj.d)
# normalize the adj, type = 'ramdom_walk'(row 1) or type = 'symmetric'
if normalize_type == 'random_walk':
print('----------------------------------')
print('Normalize_type is random_walk:')
A = spmx_1_normalize(A)
print('----------------------------------')
else:
print('----------------------------------')
print('Normalize_type is symmetric:')
A = spmx_sym_normalize(A)
print('----------------------------------')
# transform the adj to a sparse cpu tensor
A = spmx2torch_sparse_tensor(A)
# get cpu tensor: labels
labels = data_obj.get_labels(data_obj.d)
# get cpu sparse tensor: session adj
SI = data_obj.get_session_adj(data_obj.d, alpha=alpha)
# load model pretrained params
if model_pretrained_params == 'True':
print('----------------------------------')
if dataset == 'LastFM':
# use LastFM params
print('Use LastFM model pretraned params: ' + params_file_name +
'.pkl')
pretrained_state_dict = torch.load('./lastfm_pretrained_params/' +
params_file_name + '.pkl')
elif dataset == 'Diginetica':
# use Diginetica params
print('Use Diginetica model pretraned params: ' +
params_file_name + '.pkl')
pretrained_state_dict = torch.load('./dig_pretrained_params/' +
params_file_name + '.pkl')
else:
# use yoochoose1_64 params
print('Use yoochoose1_64 model pretraned params: ' +
params_file_name + '.pkl')
pretrained_state_dict = torch.load('./yoo1_64_pretrained_params/' +
params_file_name + '.pkl')
print('----------------------------------')
else:
pretrained_state_dict = None
# transform all tensor to cuda
A = A.to(device)
labels = labels.to(device)
SI = SI.to(device)
# define the evalution object
evalution5 = Evaluation(k=5)
evalution10 = Evaluation(k=10)
evalution15 = Evaluation(k=15)
evalution20 = Evaluation(k=20)
# define yoochoose data object
trainloader = SessionDataloader(train_size=data_obj.train_size,
test_size=data_obj.test_size,
item_size=data_obj.item_size,
labels=labels,
batch_size=batch_size,
train=True,
negative_nums=negative_nums,
shuffle=True)
testloader = SessionDataloader(train_size=data_obj.train_size,
test_size=data_obj.test_size,
item_size=data_obj.item_size,
labels=labels,
batch_size=test_batch_size *
data_obj.item_size,
train=False,
negative_nums=negative_nums,
shuffle=False)
# define model, then transform to cuda
if model_type == 'sgncf1_cnn':
# use sgncf1_cnn model:
model = sgncf1_cnn(dataset_nums=data_obj.train_size +
data_obj.test_size,
item_nums=data_obj.item_size,
item_emb_dim=item_emb_dim,
hid_dim1=hid_dim1)
else:
# use sgncf2_cnn model:
model = sgncf2_cnn(dataset_nums=data_obj.train_size +
data_obj.test_size,
item_nums=data_obj.item_size,
item_emb_dim=item_emb_dim,
hid_dim1=hid_dim1,
hid_dim2=hid_dim2)
model.to(device)
# update model_state_dict
if pretrained_state_dict is not None:
model_state_dict = model.state_dict()
pretrained_state_dict = {
k: v
for k, v in pretrained_state_dict.items() if k in model_state_dict
}
model_state_dict.update(pretrained_state_dict)
model.load_state_dict(model_state_dict)
# define loss and optim
criterion = nn.BCEWithLogitsLoss()
if model_type == 'sgncf1_cnn':
# use sgncf1 model parameters:
optim_emb = optim.Adagrad([{
'params': model.item_emb.parameters()
}],
lr=lr_emb,
weight_decay=l2_emb)
optim_gcn = optim.Adam([{
'params': model.gconv1.parameters()
}],
lr=lr_gcn,
weight_decay=l2_gcn)
optim_cnn = optim.Adam([{
'params': model.cnn_1d.parameters()
}, {
'params': model.fc.parameters()
}],
lr=lr_cnn,
weight_decay=l2_cnn)
else:
# use sgncf2 model parameters:
optim_emb = optim.Adagrad([{
'params': model.item_emb.parameters()
}],
lr=lr_emb,
weight_decay=l2_emb)
optim_gcn = optim.Adam([{
'params': model.gconv1.parameters()
}, {
'params': model.gconv2.parameters()
}],
lr=lr_gcn,
weight_decay=l2_gcn)
optim_cnn = optim.Adam([{
'params': model.cnn_1d.parameters()
}, {
'params': model.fc.parameters()
}],
lr=lr_cnn,
weight_decay=l2_cnn)
# figure recall mrr norm
fig_recalls = []
fig_mrrs = []
fig_emb_norms = []
fig_gcn_norms = []
fig_cnn_norms = []
fig_epochs = []
# train epochs
for epoch in range(epochs):
# model training
start = time.time()
# test evalution dict
r = {'5': [], '10': [], '15': [], '20': []}
m = {'5': [], '10': [], '15': [], '20': []}
# loss list
losses = []
model.train()
for i, data in enumerate(trainloader):
# zero optim
optim_emb.zero_grad()
optim_gcn.zero_grad()
optim_cnn.zero_grad()
# batch inputs
batch_sidxes, batch_iidxes, batch_labels = data[:, 0].long().to(
device), data[:,
1].long().to(device), data[:,
2].float().to(device)
# predicting
outs = model(batch_sidxes, batch_iidxes, A, SI)
# loss
loss = criterion(outs, batch_labels)
# backward
loss.backward()
# optim step
optim_emb.step()
optim_gcn.step()
optim_cnn.step()
# losses
losses.append(loss.item())
# print loss, recall, mrr
if i % 20 == 19:
print('[{0: 2d}, {1:5d}, {2: 7d}], loss:{3:.4f}'.format(
epoch + 1, int(i * (batch_size / (negative_nums + 1))),
data_obj.train_size, np.mean(losses)))
# print gcn_norm, emb_norm
emb_norm = get_norm(model, 'emb')
gcn_norm = get_norm(model, 'gcn')
cnn_norm = get_norm(model, 'cnn')
fig_emb_norms.append(emb_norm)
fig_gcn_norms.append(gcn_norm)
fig_cnn_norms.append(gcn_norm)
print('[gcn_norm]:{0:.4f} [emb_norm]:{1:.4f} [cnn_norm]:{2:.4f}'.
format(gcn_norm, emb_norm, cnn_norm))
# epoch time
print('[epoch time]:{0:.4f}'.format(time.time() - start))
# model eval
model.eval()
with torch.no_grad():
for j, d in enumerate(testloader):
# test batch inputs
b_sidxes, b_iidxes, b_labels = d[0][:, 0].long().to(
device), d[0][:, 1].long().to(device), d[1].to(device)
# predicting
o = model(b_sidxes, b_iidxes, A, SI)
o = o.view(-1, data_obj.item_size)
# evalution, k=5, 10, 15, 20
r['5'].append(evalution5.evaluate(o, b_labels)[0])
r['10'].append(evalution10.evaluate(o, b_labels)[0])
r['15'].append(evalution15.evaluate(o, b_labels)[0])
r['20'].append(evalution20.evaluate(o, b_labels)[0])
m['5'].append(evalution5.evaluate(o, b_labels)[1])
m['10'].append(evalution10.evaluate(o, b_labels)[1])
m['15'].append(evalution15.evaluate(o, b_labels)[1])
m['20'].append(evalution20.evaluate(o, b_labels)[1])
# print test inf
# print('[{0: 2d}, {1: 5d}, {2: 7d}]'.format(epoch+1,
# j * test_batch_size,
# data_obj.test_size))
# print test recall mrr
print('[{0: 2d}]'.format(epoch + 1))
print('[recall@5 ]:{0:.4f} [mrr@5 ]:{1:.4f}'.format(
np.sum(r['5']) / data_obj.test_size,
np.sum(m['5']) / data_obj.test_size))
print('[recall@10]:{0:.4f} [mrr@10]:{1:.4f}'.format(
np.sum(r['10']) / data_obj.test_size,
np.sum(m['10']) / data_obj.test_size))
print('[recall@15]:{0:.4f} [mrr@15]:{1:.4f}'.format(
np.sum(r['15']) / data_obj.test_size,
np.sum(m['15']) / data_obj.test_size))
print('[recall@20]:{0:.4f} [mrr@20]:{1:.4f}'.format(
np.sum(r['20']) / data_obj.test_size,
np.sum(m['20']) / data_obj.test_size))
# plt recall and mrr and norm
fig_epochs.append(epoch)
fig_recalls.append(np.sum(r['20']) / data_obj.test_size)
fig_mrrs.append(np.sum(m['20']) / data_obj.test_size)
plt_evalution(fig_epochs,
fig_recalls,
fig_mrrs,
k=20,
alpha=alpha,
lr_emb=lr_emb,
l2_emb=l2_emb,
lr_gcn=lr_gcn,
l2_gcn=l2_gcn,
model_type=model_type,
lr_cnn=lr_cnn,
l2_cnn=l2_cnn)
plt_norm(fig_epochs,
fig_emb_norms,
fig_gcn_norms,
fig_cnn_norms,
alpha=alpha,
lr_emb=lr_emb,
l2_emb=l2_emb,
lr_gcn=lr_gcn,
l2_gcn=l2_gcn,
model_type=model_type,
lr_cnn=lr_cnn,
l2_cnn=l2_cnn)
'W_mem_res_in', 'W_mem_res_hid', 'b_mem_res',
'W_mem_upd_in', 'W_mem_upd_hid', 'b_mem_upd',
'W_mem_hid_in', 'W_mem_hid_hid', 'b_mem_hid', 'W_b',
'W_1', 'W_2', 'b_1', 'b_2', 'W_a']
fig, ax = plt.subplots(figsize=(9,4))
with open(file_name, 'r') as load_file:
dict = pickle.load(load_file)
loaded_params = dict['params']
if flag:
for i in xrange(len(params)):
if params[i] in to_write:
out_obj[params[i]] = loaded_params[i]
with open(sys.argv[2], 'w') as save_file:
pickle.dump(obj = out_obj,file = save_file,protocol = -1)
print "finish dumping file to "+sys.argv[2]
for (x, y) in zip(params, loaded_params):
n = y.shape
if len(n)==1:
n=n[0]
else:
n=n[0]*n[1]
norm = utils.get_norm(y)/n**0.5
print x,' shape: ',y.shape,', norm: ',norm,', max: ',np.max(np.abs(y))
if len(y.shape)>1:
ax.imshow(y,cmap = 'Blues',interpolation='none')
plt.title('Train. '+x+', norm '+str(norm))
fig.show()
input_str = raw_input("Press ENTER to continue.")
def train(dataset, alpha, A_type, normalize_type, session_type,
pretrained_item_emb, model_type, batch_size, shuffle, item_emb_dim,
hid_dim1, hid_dim2, hid_dim3, lr_emb, lr_gcn, l2_emb, l2_gcn,
epochs):
# init
if dataset == 'LastFM':
# use LastFM dataset
data_obj = LastfmData()
elif dataset == 'Diginetica':
# use Diginetica dataset
data_obj = DigineticaData()
else:
# use yoochoose1_64 dataset
data_obj = YoochooseData(dataset=dataset)
# gpu device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# init A
# A: type=scipy.sparse
A = data_obj.get_decay_adj(
data_obj.d, tail=None,
alpha=alpha) if A_type == 'decay' else data_obj.get_gcn_adj(data_obj.d)
# normalize the adj, type = 'ramdom_walk'(row 1) or type = 'symmetric'
if normalize_type == 'random_walk':
print('----------------------------------')
print('Normalize_type is random_walk:')
A = spmx_1_normalize(A)
print('----------------------------------')
else:
print('----------------------------------')
print('Normalize_type is symmetric:')
A = spmx_sym_normalize(A)
print('----------------------------------')
# transform the adj to a sparse cpu tensor
A = spmx2torch_sparse_tensor(A)
# get cpu tensor: labels
labels = data_obj.get_labels(data_obj.d)
# get cpu tensor: item_idxes
_, _, item_idxes = data_obj.get_indexes()
if session_type == 'session_hot_items':
# get cpu sparse tensor: session adj
session_adj = data_obj.get_session_adj(data_obj.d, alpha=alpha)
else:
# if not use session adj, then session_adj = None
session_adj = None
if session_type == 'session_last_item':
# get cpu LongTensor: session_last_item
session_last_item = data_obj.get_session_last_item(data_obj.d).long()
else:
# if not use session_last_item, then session_last_item = None
session_last_item = None
# get pretrained_item_emb
if pretrained_item_emb == 'True' and alpha != 0.0:
print('----------------------------------')
if dataset == 'yoochoose1_64':
print('Use yoochoose1_64 pretrained item embedding: ' +
'pretrained_emb' + str(alpha) + '.pkl')
pretrained_item_emb = torch.load(
'./yoo1_64_pretrained_item_emb/pretrained_emb' + str(alpha) +
'.pkl')['item_emb.weight']
elif dataset == 'yoochoose1_8':
print('Use yoochoose1_8 pretrained item embedding: ' +
'pretrained_emb' + str(alpha) + '.pkl')
pretrained_item_emb = torch.load(
'./yoo1_8_pretrained_item_emb/pretrained_emb' + str(alpha) +
'.pkl')['item_emb.weight']
elif dataset == 'LastFM':
print('Use LastFM pretrained item embedding: ' + 'pretrained_emb' +
str(alpha) + '.pkl')
pretrained_item_emb = torch.load(
'./lastfm_pretrained_item_emb/pretrained_emb' + str(alpha) +
'.pkl')['item_emb.weight']
else:
print('Use Diginetica pretrained item embedding: ' +
'pretrained_emb' + str(alpha) + '.pkl')
pretrained_item_emb = torch.load(
'./dig_pretrained_item_emb/pretrained_emb' + str(alpha) +
'.pkl')['item_emb.weight']
print('----------------------------------')
else:
print('----------------------------------')
print('Not use pretrained item embedding:')
pretrained_item_emb = None
print('----------------------------------')
# get cpu LongTensor: item_emb_idxes
item_emb_idxes = torch.arange(data_obj.item_size).long()
# transform all tensor to cuda
A = A.to(device)
labels = labels.to(device)
item_idxes = item_idxes.to(device)
item_emb_idxes = item_emb_idxes.to(device)
if session_last_item is not None:
session_last_item = session_last_item.to(device)
if session_adj is not None:
session_adj = session_adj.to(device)
# define the evalution object
evalution5 = Evaluation(k=5)
evalution10 = Evaluation(k=10)
evalution15 = Evaluation(k=15)
evalution20 = Evaluation(k=20)
# define yoochoose data object
trainset = SessionDataset(train_size=data_obj.train_size,
test_size=data_obj.test_size,
train=True,
labels=labels)
trainloader = DataLoader(dataset=trainset,
batch_size=batch_size,
shuffle=shuffle)
testset = SessionDataset(train_size=data_obj.train_size,
test_size=data_obj.test_size,
train=False,
labels=labels)
testloader = DataLoader(dataset=testset,
batch_size=batch_size,
shuffle=False)
# define model, then transform to cuda
if model_type == 'ngcf1_session_hot_items':
# use ngcf1_session_hot_items model:
model = ngcf1_session_hot_items(
item_nums=data_obj.item_size,
item_emb_dim=item_emb_dim,
hid_dim1=hid_dim1,
pretrained_item_emb=pretrained_item_emb)
elif model_type == 'ngcf2_session_hot_items':
# use ngcf2_session_hot_items model:
model = ngcf2_session_hot_items(
item_nums=data_obj.item_size,
item_emb_dim=item_emb_dim,
hid_dim1=hid_dim1,
hid_dim2=hid_dim2,
pretrained_item_emb=pretrained_item_emb)
elif model_type == 'ngcf3_session_hot_items':
# use ngcf3_session_hot_items model:
model = ngcf3_session_hot_items(
item_nums=data_obj.item_size,
item_emb_dim=item_emb_dim,
hid_dim1=hid_dim1,
hid_dim2=hid_dim2,
hid_dim3=hid_dim3,
pretrained_item_emb=pretrained_item_emb)
else:
# use ngcf2_session_last_item model:
model = ngcf2_session_last_item(
item_nums=data_obj.item_size,
item_emb_dim=item_emb_dim,
hid_dim1=hid_dim1,
hid_dim2=hid_dim2,
pretrained_item_emb=pretrained_item_emb)
model.to(device)
# define loss and optim
criterion = nn.CrossEntropyLoss()
if model_type == 'ngcf1_session_hot_items':
# use ngcf1_session_hot_items model parameters:
optim_emb = optim.Adagrad([{
'params': model.item_emb.parameters()
}],
lr=lr_emb,
weight_decay=l2_emb)
optim_gcn = optim.Adam([{
'params': model.gconv1.parameters()
}],
lr=lr_gcn,
weight_decay=l2_gcn)
elif model_type == 'ngcf2_session_hot_items':
# use ngcf2_session_hot_items model parameters:
optim_emb = optim.Adagrad([{
'params': model.item_emb.parameters()
}],
lr=lr_emb,
weight_decay=l2_emb)
optim_gcn = optim.Adam([{
'params': model.gconv1.parameters()
}, {
'params': model.gconv2.parameters()
}],
lr=lr_gcn,
weight_decay=l2_gcn)
elif model_type == 'ngcf3_session_hot_items':
# use ngcf3_session_hot_items model parameters:
optim_emb = optim.Adagrad([{
'params': model.item_emb.parameters()
}],
lr=lr_emb,
weight_decay=l2_emb)
optim_gcn = optim.Adam([{
'params': model.gconv1.parameters()
}, {
'params': model.gconv2.parameters()
}, {
'params': model.gconv3.parameters()
}],
lr=lr_gcn,
weight_decay=l2_gcn)
else:
# use ngcf2_session_last_item model parameters:
optim_emb = optim.Adagrad([{
'params': model.item_emb.parameters()
}],
lr=lr_emb,
weight_decay=l2_emb)
optim_gcn = optim.Adam([{
'params': model.gconv1.parameters()
}, {
'params': model.gconv2.parameters()
}],
lr=lr_gcn,
weight_decay=l2_gcn)
# figure recall mrr norm
fig_recalls = []
fig_mrrs = []
fig_emb_norms = []
fig_gcn_norms = []
fig_epochs = []
# train epochs
for epoch in range(epochs):
# model training
start = time.time()
# train evalution dict
recall = {'5': [], '10': [], '15': [], '20': []}
mrr = {'5': [], '10': [], '15': [], '20': []}
# test evalution dict
r = {'5': [], '10': [], '15': [], '20': []}
m = {'5': [], '10': [], '15': [], '20': []}
# loss list
losses = []
model.train()
for i, data in enumerate(trainloader):
# zero optim
optim_emb.zero_grad()
optim_gcn.zero_grad()
# batch inputs
batch_idxes, batch_labels = data[0].long().to(
device), data[1].long().to(device)
# predicting
if model_type == 'ngcf1_session_hot_items':
# use ngcf1_session_hot_items model to predict
outs = model(batch_idxes, A, item_idxes, session_adj,
item_emb_idxes)
elif model_type == 'ngcf2_session_hot_items':
# use ngcf2_session_hot_items model to predict
outs = model(batch_idxes, A, item_idxes, session_adj,
item_emb_idxes)
elif model_type == 'ngcf3_session_hot_items':
# use ngcf3_session_hot_items model to predict
outs = model(batch_idxes, A, item_idxes, session_adj,
item_emb_idxes)
else:
# use ngcf2_session_last_item model to predict
outs = model(batch_idxes, A, item_idxes, session_last_item,
item_emb_idxes)
# loss
loss = criterion(outs, batch_labels)
# backward
loss.backward()
# optim step
optim_emb.step()
optim_gcn.step()
# evalution, k=5, 10, 15, 20
recall['5'].append(evalution5.evaluate(outs, batch_labels)[0])
recall['10'].append(evalution10.evaluate(outs, batch_labels)[0])
recall['15'].append(evalution15.evaluate(outs, batch_labels)[0])
recall['20'].append(evalution20.evaluate(outs, batch_labels)[0])
mrr['5'].append(evalution5.evaluate(outs, batch_labels)[1])
mrr['10'].append(evalution10.evaluate(outs, batch_labels)[1])
mrr['15'].append(evalution15.evaluate(outs, batch_labels)[1])
mrr['20'].append(evalution20.evaluate(outs, batch_labels)[1])
# losses
losses.append(loss.item())
# print loss, recall, mrr
if i % 50 == 49:
print('[{0: 2d}, {1:5d}] loss:{2:.4f}'.format(
epoch + 1, i + 1, np.mean(losses)))
print('[recall@5 ]:{0:.4f} [mrr@5 ]:{1:.4f}'.format(
np.mean(recall['5']), np.mean(mrr['5'])))
print('[recall@10]:{0:.4f} [mrr@10]:{1:.4f}'.format(
np.mean(recall['10']), np.mean(mrr['10'])))
print('[recall@15]:{0:.4f} [mrr@15]:{1:.4f}'.format(
np.mean(recall['15']), np.mean(mrr['15'])))
print('[recall@20]:{0:.4f} [mrr@20]:{1:.4f}'.format(
np.mean(recall['20']), np.mean(mrr['20'])))
# print gcn_norm, emb_norm
emb_norm = get_norm(model, 'emb')
gcn_norm = get_norm(model, 'gcn')
fig_emb_norms.append(emb_norm)
fig_gcn_norms.append(gcn_norm)
print('[gcn_norm]:{0:.4f} [emb_norm]:{1:.4f}'.format(
gcn_norm, emb_norm))
# epoch time
print('[epoch time]:{0:.4f}'.format(time.time() - start))
# save model
if epoch % 10 == 9:
torch.save(
model.state_dict(),
'params' + model_type + '-Alpha' + str(alpha) + '_' +
'_lr_emb' + str(lr_emb) + '_l2_emb' + str(l2_emb) + '_lr_gcn' +
str(lr_gcn) + '_l2_gcn' + str(l2_gcn) + '.pkl')
# model eval
model.eval()
with torch.no_grad():
for j, d in enumerate(testloader):
# test batch inputs
b_idxes, b_labels = d[0].long().to(device), d[1].long().to(
device)
# predicting
if model_type == 'ngcf1_session_hot_items':
# use ngcf1_session_hot_items model to predict
o = model(b_idxes, A, item_idxes, session_adj,
item_emb_idxes)
elif model_type == 'ngcf2_session_hot_items':
# use ngcf2_session_hot_items model to predict
o = model(b_idxes, A, item_idxes, session_adj,
item_emb_idxes)
elif model_type == 'ngcf3_session_hot_items':
# use ngcf3_session_hot_items model to predict
o = model(b_idxes, A, item_idxes, session_adj,
item_emb_idxes)
else:
# use ngcf2_session_last_item model to predict
o = model(b_idxes, A, item_idxes, session_last_item,
item_emb_idxes)
# evalution, k=5, 10, 15, 20
r['5'].append(evalution5.evaluate(o, b_labels)[0])
r['10'].append(evalution10.evaluate(o, b_labels)[0])
r['15'].append(evalution15.evaluate(o, b_labels)[0])
r['20'].append(evalution20.evaluate(o, b_labels)[0])
m['5'].append(evalution5.evaluate(o, b_labels)[1])
m['10'].append(evalution10.evaluate(o, b_labels)[1])
m['15'].append(evalution15.evaluate(o, b_labels)[1])
m['20'].append(evalution20.evaluate(o, b_labels)[1])
# print test recall mrr
print('[{0: 2d}]'.format(epoch + 1))
print('[recall@5 ]:{0:.4f} [mrr@5 ]:{1:.4f}'.format(
np.mean(r['5']), np.mean(m['5'])))
print('[recall@10]:{0:.4f} [mrr@10]:{1:.4f}'.format(
np.mean(r['10']), np.mean(m['10'])))
print('[recall@15]:{0:.4f} [mrr@15]:{1:.4f}'.format(
np.mean(r['15']), np.mean(m['15'])))
print('[recall@20]:{0:.4f} [mrr@20]:{1:.4f}'.format(
np.mean(r['20']), np.mean(m['20'])))
# plt recall and mrr and norm
fig_epochs.append(epoch)
fig_recalls.append(np.mean(r['20']))
fig_mrrs.append(np.mean(m['20']))
plt_evalution(fig_epochs,
fig_recalls,
fig_mrrs,
k=20,
alpha=alpha,
lr_emb=lr_emb,
l2_emb=l2_emb,
lr_gcn=lr_gcn,
l2_gcn=l2_gcn,
model_type=model_type)
plt_norm(fig_epochs,
fig_emb_norms,
fig_gcn_norms,
alpha=alpha,
lr_emb=lr_emb,
l2_emb=l2_emb,
lr_gcn=lr_gcn,
l2_gcn=l2_gcn,
model_type=model_type)
def step(self, batch_idx, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
input_masks = self.train_input_mask
qinfo = self.train_qinfo
elif mode == "train_val":
theano_fn = self.test_fn
inputs = self.train_val_input
qs = self.train_val_q
answers = self.train_val_answer
input_masks = self.test_input_mask
qinfo = self.train_val_qinfo
elif mode == 'test':
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
input_masks = self.test_input_mask
qinfo = self.test_qinfo
else:
raise Exception("Invalid mode")
num_ma_opts = answers.shape[1]
p_q = np.zeros((len(batch_idx), 300),
dtype='float32') # question input vector
target = np.zeros((len(batch_idx))) # answer (as a single number)
p_inp = np.zeros(
(len(batch_idx), self.max_sent_len, self.sent_vector_size),
dtype='float32') # story statements
p_ans = np.zeros((len(batch_idx), num_ma_opts, 300),
dtype='float32') # multiple choice answers
#b_qinfo = []
input_mask = input_masks
for b, bi in enumerate(batch_idx):
inp = inputs[qinfo[bi]['qid']]
q = qs[bi]
ans = answers[bi]
target[b] = qinfo[bi]['correct_option']
for i in range(len(inp)):
p_inp[b][i] = inp[i]
for j in range(len(ans)):
p_ans[b][j] = self.pos_encodings(ans[j])
p_q[b] = self.pos_encodings(q)
#b_qinfo.append(qinfo[bi])
ret = theano_fn(p_inp, p_q, p_ans, target)
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": np.array(ret[0]),
"answers": np.array(target),
"current_loss": ret[1],
"skipped": 0,
"log": "pn: %.3f" % param_norm,
"inp": np.array([inp]),
"q": np.array([q]),
"probabilities": np.array([ret[0]]),
"attentions": np.array([ret[2]]),
}
def get_vel_lin(self): return utils.get_norm(self.vel_lin)
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
ca = self.train_choices
cb = self.train_choices
#cc = self.train_choices
#cd = self.train_choices
input_masks = self.train_input_mask
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
ca = self.test_choices
cb = self.test_choices
#cc = self.test_choices
#cd = self.test_choices
input_masks = self.test_input_mask
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
ca = ca[batch_index][0]
cb = cb[batch_index][1]
#cc = cc[batch_index][2]
#cd = cd[batch_index][3]
input_mask = input_masks[batch_index]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
#gradient_value = self.get_gradient_fn(inp, q, ans, ca, cb, cc, cd, input_mask)
gradient_value = self.get_gradient_fn(inp, q, ans, ca, cb,
input_mask)
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
print "==> gradient is nan at index %d." % batch_index
print "==> skipping"
skipped = 1
if skipped == 0:
#ret = theano_fn(inp, q, ans, ca, cb, cc, cd, input_mask)
ret = theano_fn(inp, q, ans, ca, cb, input_mask)
else:
ret = [float('NaN'), float('NaN')]
param_norm = np.max(
[utils.get_norm(x.get_value()) for x in self.params])
return {
"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"skipped": skipped,
"log": "pn: %.3f \t gn: %.3f" % (param_norm, grad_norm)
}
def step(self, batch_index, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
inputs = self.train_input
qs = self.train_q
answers = self.train_answer
choices = self.train_choices
input_masks = self.train_input_mask
elif mode == "test":
theano_fn = self.test_fn
inputs = self.test_input
qs = self.test_q
answers = self.test_answer
choices = self.test_choices
input_masks = self.test_input_mask
elif mode == "dev":
theano_fn = self.test_fn
inputs = self.dev_input
qs = self.dev_q
answers = self.dev_answer
choices = self.dev_choices
input_masks = self.dev_input_mask
else:
raise Exception("Invalid mode")
inp = inputs[batch_index]
q = qs[batch_index]
ans = answers[batch_index]
ca = choices[batch_index][0]
cb = choices[batch_index][1]
cc = choices[batch_index][2]
cd = choices[batch_index][3]
input_mask = input_masks[batch_index]
skipped = 0
grad_norm = float('NaN')
if mode == 'train':
gradient_value = self.get_gradient_fn(inp, q, ans, ca, cb, cc, cd, input_mask)
grad_norm = np.max([utils.get_norm(x) for x in gradient_value])
if (np.isnan(grad_norm)):
print "==> gradient is nan at index %d." % batch_index
print "==> skipping"
skipped = 1
if skipped == 0:
ret = theano_fn(inp, q, ans, ca, cb, cc, cd, input_mask)
else:
ret = [float('NaN'), float('NaN')]
param_norm = np.max([utils.get_norm(x.get_value()) for x in self.params])
return {"prediction": np.array([ret[0]]),
"answers": np.array([ans]),
"current_loss": ret[1],
"skipped": skipped,
"log": "pn: %.3f \t gn: %.3f" % (param_norm, grad_norm)
}