def train_test(batch_data, reward):
place = fluid.CPUPlace()
feed_dict = {}
for name in batch_data.conf.user_slot_names + batch_data.conf.item_slot_names:
ft = batch_data.tensor_dict[name]
feed_dict[name] = create_tensor(ft.values, lod=ft.lod, place=place)
decode_len = batch_data.decode_len().reshape([-1, 1]).astype('int64')
lod = [seq_len_2_lod([1] * len(decode_len))]
feed_dict['decode_len'] = create_tensor(decode_len,
lod=lod,
place=place)
reward = reward.reshape([-1, 1]).astype('float32')
lod = [seq_len_2_lod(batch_data.decode_len())]
feed_dict['reward'] = create_tensor(reward, lod=lod, place=place)
return feed_dict
def infer_onestep(batch_data, prev_hidden, selected_items, last_click_id):
"""
batch_data: the complete data
item_feature: (b*seq_len,)
prev_hidden: (b, dim)
last_click_id: (b,)
selected_items: (b,)
"""
batch_size = batch_data.batch_size()
seq_len = batch_data.seq_lens()[0]
batch_offset = batch_data.offset()
place = fluid.CPUPlace()
feed_dict = {}
lod = [seq_len_2_lod([1] * batch_size)]
offset_selected_items = np.array(selected_items) + batch_offset # (b,)
for name in batch_data.conf.item_slot_names:
if name == 'last_click_id':
v = last_click_id.reshape([-1, 1])
else:
v = batch_data.tensor_dict[name].values[offset_selected_items]
feed_dict[name] = create_tensor(v, lod=lod, place=place)
# prev_hidden
feed_dict['prev_hidden'] = create_tensor(prev_hidden,
lod=lod,
place=place)
return feed_dict
def softmax_sampling(batch_data, eta):
place = fluid.CPUPlace()
feed_dict = {}
for name in batch_data.conf.user_slot_names + batch_data.conf.item_slot_names:
ft = batch_data.tensor_dict[name]
feed_dict[name] = create_tensor(ft.values, lod=ft.lod, place=place)
decode_len = batch_data.decode_len().reshape([-1, 1]).astype('int64')
lod = [seq_len_2_lod([1] * len(decode_len))]
feed_dict['decode_len'] = create_tensor(decode_len,
lod=lod,
place=place)
eta = np.array([eta]).astype('float32')
feed_dict['eta'] = create_tensor(eta, lod=[], place=place)
return feed_dict
def infer_onestep(batch_data,
prev_hidden,
candidate_items,
last_click_id,
conf=None):
"""
batch_data: the complete data
item_feature: (b*seq_len,)
prev_hidden: (b, dim)
last_click_id: (b,)
candidate_items: 2d list
"""
if conf is None:
conf = batch_data.conf
batch_size = batch_data.batch_size()
seq_len = batch_data.seq_lens()[0]
cand_len = len(candidate_items[0])
batch_offset = batch_data.offset()
# expand last_click_id
last_click_id = np.repeat(last_click_id, cand_len,
axis=0).reshape([-1, 1]) # (b*cand_len, 1)
place = fluid.CPUPlace()
feed_dict = {}
lod = [seq_len_2_lod([1] * (batch_size * cand_len))]
offset_candidate_items = np.array(candidate_items).flatten(
) + np.repeat(batch_offset, cand_len, axis=0) # (b*cand_len)
for name in conf.item_slot_names:
if name == 'last_click_id':
v = last_click_id
else:
v = batch_data.tensor_dict[name].values[offset_candidate_items]
feed_dict[name] = create_tensor(v, lod=lod, place=place)
# expand prev_hidden
prev_hidden = np.repeat(prev_hidden, cand_len,
axis=0) # (b*cand_len, dim)
feed_dict['prev_hidden'] = create_tensor(prev_hidden,
lod=lod,
place=place)
return feed_dict
def apply_masks(batch_data, list_item_masks, conf=None):
"""
list_item_masks: (n_masks, seq_len), a list of 1d item_mask
Apply mask on item_level_slot_names except last_click_id
"""
if conf is None:
conf = batch_data.conf
batch_size = batch_data.batch_size()
seq_len = batch_data.seq_lens()[0]
AssertEqual(len(list_item_masks[0]), seq_len)
batch_data.add_last_click_id()
n_masks = len(list_item_masks)
batch_item_masks = np.tile(
np.array(list_item_masks).flatten(),
[batch_size]) # (batch_size * n_masks * seq_len)
place = fluid.CPUPlace()
feed_dict = {}
for name in conf.recent_slot_names + \
conf.item_slot_names:
ft = batch_data.tensor_dict[name]
v = ft.values
extra_shape = list(v.shape[1:])
v = v.reshape([batch_size, -1] +
extra_shape) # (batch_size, seq_len/recent_len, ...)
v = np.repeat(
v, n_masks,
axis=0) # (batch_size * n_masks, seq_len/recent_len, ...)
seq_lens = [v.shape[1]] * (batch_size * n_masks)
v = v.reshape([-1] + extra_shape
) # (batch_size * n_masks * seq_len/recent_len, ...)
if name in conf.item_slot_names and name != 'last_click_id':
v = v * batch_item_masks.reshape([-1] + [1] *
(len(v.shape) - 1))
feed_dict[name] = create_tensor(v,
lod=[seq_len_2_lod(seq_lens)],
place=place)
return feed_dict
def infer_onestep(batch_data,
prev_hidden,
candidate_items,
last_click_id,
last_item,
step,
conf=None):
"""
batch_data: the complete data
item_feature: (b*seq_len,)
prev_hidden: (b, dim)
last_click_id: (b,)
last_item: (b,)
step: int
candidate_items: 2d list
Different from other model, ddpg don't need to expand.
"""
if conf is None:
conf = batch_data.conf
batch_size = batch_data.batch_size()
seq_len = batch_data.seq_lens()[0]
cand_len = len(candidate_items[0])
batch_offset = batch_data.offset()
place = fluid.CPUPlace()
feed_dict = {}
item_slot_names = list(conf.item_slot_names)
item_slot_names.remove('last_click_id')
last_item_slot_names = ['last_' + name for name in item_slot_names]
### candidates
lod = [seq_len_2_lod([cand_len] * batch_size)]
offset_candidate_items = np.array(candidate_items).flatten(
) + np.repeat(batch_offset, cand_len, axis=0) # (b*cand_len)
for name in item_slot_names:
v = batch_data.tensor_dict[name].values[offset_candidate_items]
feed_dict[name] = create_tensor(v, lod=lod, place=place)
### last item
lod = [seq_len_2_lod([1] * batch_size)]
offset_last_item = np.array(last_item).flatten() + batch_offset # (b,)
for last_name in last_item_slot_names:
name = last_name[len('last_'):]
v = batch_data.tensor_dict[name].values[offset_last_item]
feed_dict[last_name] = create_tensor(v, lod=lod, place=place)
### last click
lod = [seq_len_2_lod([1] * batch_size)]
feed_dict['last_click_id'] = create_tensor(last_click_id.reshape(
[-1, 1]),
lod=lod,
place=place) # (b, 1)
### prev_hidden
lod = [seq_len_2_lod([1] * batch_size)]
feed_dict['prev_hidden'] = create_tensor(prev_hidden,
lod=lod,
place=place) # (b, dim)
### first_step_mask
lod = [seq_len_2_lod([1] * batch_size)]
first_step_mask = np.full([batch_size, 1],
float(step > 0)).astype('float32')
feed_dict['first_step_mask'] = create_tensor(first_step_mask,
lod=lod,
place=place) # (b, 1)
return feed_dict