def __init__(self, args, vocab, transition_system):
super(Parser, self).__init__()
self.args = args
self.vocab = vocab
self.transition_system = transition_system
self.grammar = self.transition_system.grammar
# Embedding layers
self.src_embed = nn.Embedding(len(vocab.source), args.embed_size)
self.production_embed = nn.Embedding(
len(transition_system.grammar) + 1, args.action_embed_size)
self.primitive_embed = nn.Embedding(len(vocab.primitive),
args.action_embed_size)
self.field_embed = nn.Embedding(len(transition_system.grammar.fields),
args.field_embed_size)
self.type_embed = nn.Embedding(len(transition_system.grammar.types),
args.type_embed_size)
nn.init.xavier_normal(self.src_embed.weight.data)
nn.init.xavier_normal(self.production_embed.weight.data)
nn.init.xavier_normal(self.primitive_embed.weight.data)
nn.init.xavier_normal(self.field_embed.weight.data)
nn.init.xavier_normal(self.type_embed.weight.data)
# LSTMs
if args.lstm == 'lstm':
self.encoder_lstm = nn.LSTM(args.embed_size,
args.hidden_size // 2,
bidirectional=True)
self.decoder_lstm = nn.LSTMCell(
args.action_embed_size + # previous action
args.action_embed_size + args.field_embed_size +
args.type_embed_size + # frontier info
args.hidden_size + # parent hidden state
args.hidden_size, # input feeding
args.hidden_size)
else:
from .lstm import LSTM, LSTMCell
self.encoder_lstm = LSTM(args.embed_size,
args.hidden_size // 2,
bidirectional=True,
dropout=args.dropout)
self.decoder_lstm = LSTMCell(
args.action_embed_size + # previous action
args.action_embed_size + args.field_embed_size +
args.type_embed_size + # frontier info
args.hidden_size + args.hidden_size, # parent hidden state
args.hidden_size,
dropout=args.dropout)
# pointer net
self.src_pointer_net = PointerNet(args.hidden_size, args.hidden_size)
self.primitive_predictor = nn.Linear(args.hidden_size, 2)
# initialize the decoder's state and cells with encoder hidden states
self.decoder_cell_init = nn.Linear(args.hidden_size, args.hidden_size)
# attention: dot product attention
# project source encoding to decoder rnn's h space
self.att_src_linear = nn.Linear(args.hidden_size,
args.hidden_size,
bias=False)
# transformation of decoder hidden states and context vectors before reading out target words
# this produces the `attentional vector` in (Luong et al., 2015)
self.att_vec_linear = nn.Linear(args.hidden_size + args.hidden_size,
args.hidden_size,
bias=False)
# embedding layers
self.query_vec_to_embed = nn.Linear(args.hidden_size,
args.embed_size,
bias=False)
self.production_readout_b = nn.Parameter(
torch.FloatTensor(len(transition_system.grammar) + 1).zero_())
self.tgt_token_readout_b = nn.Parameter(
torch.FloatTensor(len(vocab.primitive)).zero_())
self.production_readout = self.production_readout_func
self.tgt_token_readout = self.tgt_token_readout_func
# self.production_readout = nn.Linear(args.hidden_size, len(transition_system.grammar) + 1)
# self.tgt_token_readout = nn.Linear(args.hidden_size, len(vocab.primitive))
# dropout layer
self.dropout = nn.Dropout(args.dropout)
if args.cuda:
self.new_long_tensor = torch.cuda.LongTensor
self.new_tensor = torch.cuda.FloatTensor
else:
self.new_long_tensor = torch.LongTensor
self.new_tensor = torch.FloatTensor
def __init__(self, input_size, hidden_size, noisin=0):
super().__init__()
self.cell = nn.LSTMCell(input_size=input_size, hidden_size=hidden_size)
self.reset_parameters()
def __init__(self):
super(Sequence, self).__init__()
self.lstm1 = nn.LSTMCell(1, 51)
self.lstm2 = nn.LSTMCell(51, 51)
self.linear = nn.Linear(51, 1)
def __init__(self, args):
print("Encoder model --- LSTMCELL")
super(Encoder_WordLstm, self).__init__()
self.args = args
# random
self.char_embed = nn.Embedding(
self.args.embed_char_num,
self.args.embed_char_dim,
sparse=False,
padding_idx=self.args.create_alphabet.char_PaddingID)
# for index in range(self.args.embed_char_dim):
# self.char_embed.weight.data[self.args.create_alphabet.char_PaddingID][index] = 0
self.char_embed.weight.requires_grad = True
self.bichar_embed = nn.Embedding(
self.args.embed_bichar_num,
self.args.embed_bichar_dim,
sparse=False,
padding_idx=self.args.create_alphabet.bichar_PaddingID)
# for index in range(self.args.embed_bichar_dim):
# self.bichar_embed.weight.data[self.args.create_alphabet.bichar_PaddingID][index] = 0
self.bichar_embed.weight.requires_grad = True
# fix the word embedding
self.static_char_embed = nn.Embedding(
self.args.static_embed_char_num,
self.args.embed_char_dim,
sparse=False,
padding_idx=self.args.create_static_alphabet.char_PaddingID)
init.uniform(self.static_char_embed.weight,
a=-np.sqrt(3 / self.args.embed_char_dim),
b=np.sqrt(3 / self.args.embed_char_dim))
self.static_bichar_embed = nn.Embedding(
self.args.static_embed_bichar_num,
self.args.embed_bichar_dim,
sparse=False,
padding_idx=self.args.create_static_alphabet.bichar_PaddingID)
init.uniform(self.static_bichar_embed.weight,
a=-np.sqrt(3 / self.args.embed_bichar_dim),
b=np.sqrt(3 / self.args.embed_bichar_dim))
# load external word embedding
if args.char_Embedding is True:
print("char_Embedding")
pretrained_char_weight = np.array(args.pre_char_word_vecs)
self.static_char_embed.weight.data.copy_(
torch.from_numpy(pretrained_char_weight))
for index in range(self.args.embed_char_dim):
self.static_char_embed.weight.data[
self.args.create_static_alphabet.char_PaddingID][index] = 0
self.static_char_embed.weight.requires_grad = False
if args.bichar_Embedding is True:
print("bichar_Embedding")
pretrained_bichar_weight = np.array(args.pre_bichar_word_vecs)
self.static_bichar_embed.weight.data.copy_(
torch.from_numpy(pretrained_bichar_weight))
# print(self.static_bichar_embed.weight.data[self.args.create_static_alphabet.bichar_PaddingID])
# print(self.static_bichar_embed.weight.data[self.args.create_static_alphabet.bichar_UnkID])
for index in range(self.args.embed_bichar_dim):
self.static_bichar_embed.weight.data[
self.args.create_static_alphabet.
bichar_PaddingID][index] = 0
self.static_bichar_embed.weight.requires_grad = False
self.lstm_left = nn.LSTMCell(input_size=self.args.hidden_size,
hidden_size=self.args.rnn_hidden_dim,
bias=True)
self.lstm_right = nn.LSTMCell(input_size=self.args.hidden_size,
hidden_size=self.args.rnn_hidden_dim,
bias=True)
# init lstm weight and bias
init.xavier_uniform(self.lstm_left.weight_ih)
init.xavier_uniform(self.lstm_left.weight_hh)
init.xavier_uniform(self.lstm_right.weight_ih)
init.xavier_uniform(self.lstm_right.weight_hh)
value = np.sqrt(6 / (self.args.rnn_hidden_dim + 1))
self.lstm_left.bias_hh.data.uniform_(-value, value)
self.lstm_left.bias_ih.data.uniform_(-value, value)
self.lstm_right.bias_hh.data.uniform_(-value, value)
self.lstm_right.bias_ih.data.uniform_(-value, value)
self.hidden_l = self.init_hidden_cell(self.args.batch_size)
self.hidden_r = self.init_hidden_cell(self.args.batch_size)
self.dropout = nn.Dropout(self.args.dropout)
self.dropout_embed = nn.Dropout(self.args.dropout_embed)
self.input_dim = (self.args.embed_char_dim +
self.args.embed_bichar_dim) * 2
if self.args.use_cuda is True:
self.liner = nn.Linear(in_features=self.input_dim,
out_features=self.args.hidden_size,
bias=True).cuda()
else:
self.liner = nn.Linear(in_features=self.input_dim,
out_features=self.args.hidden_size,
bias=True)
# init linear
init.xavier_uniform(self.liner.weight)
init_linear_value = np.sqrt(6 / (self.args.hidden_size + 1))
self.liner.bias.data.uniform_(-init_linear_value, init_linear_value)
def __init__(self, embed_size, hidden_size, vocab, dropout_rate=0.2):
""" Init NMT Model.
@param embed_size (int): Embedding size (dimensionality)
@param hidden_size (int): Hidden Size (dimensionality)
@param vocab (Vocab): Vocabulary object containing src and tgt languages
See vocab.py for documentation.
@param dropout_rate (float): Dropout probability, for attention
"""
super(NMT, self).__init__()
self.model_embeddings = ModelEmbeddings(embed_size, vocab)
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.vocab = vocab
# default values
self.encoder = None
self.decoder = None
self.h_projection = None
self.c_projection = None
self.att_projection = None
self.combined_output_projection = None
self.target_vocab_projection = None
self.dropout = None
### TODO - Initialize the following variables:
### self.encoder (Bidirectional LSTM with bias)
### self.decoder (LSTM Cell with bias)
### self.h_projection (Linear Layer with no bias), called W_{h} in the PDF.
### self.c_projection (Linear Layer with no bias), called W_{c} in the PDF.
### self.att_projection (Linear Layer with no bias), called W_{attProj} in the PDF.
### self.combined_output_projection (Linear Layer with no bias), called W_{u} in the PDF.
### self.target_vocab_projection (Linear Layer with no bias), called W_{vocab} in the PDF.
### self.dropout (Dropout Layer)
# encoder will be feed the word embeddings for the source sentence, and yield hidden states and cell states for both the forwards and backwards LSTMs
self.encoder = nn.LSTM(input_size=embed_size,
hidden_size=hidden_size,
bidirectional=True,
bias=True)
# decoder is initialized with a linear projection of the engcoder's final hidden state and final cell state, and feed the matching target sentence word embeddings
self.decoder = nn.LSTMCell(input_size=embed_size + hidden_size,
hidden_size=hidden_size,
bias=True)
self.h_projection = nn.Linear(in_features=hidden_size * 2,
out_features=hidden_size,
bias=False)
self.c_projection = nn.Linear(in_features=hidden_size * 2,
out_features=hidden_size,
bias=False)
self.att_projection = nn.Linear(in_features=hidden_size * 2,
out_features=hidden_size,
bias=False)
# transformation of decoder hidden states and context vectors before reading out target words
# this produces the `attentional vector` in (Luong et al., 2015)
self.combined_output_projection = nn.Linear(
in_features=hidden_size * 2 + hidden_size,
out_features=hidden_size,
bias=False)
# prediction layer of the target vocabulary
self.target_vocab_projection = nn.Linear(in_features=hidden_size,
out_features=len(vocab.tgt),
bias=False)
self.dropout = nn.Dropout(dropout_rate)
def __init__(self):
super(LstmCell, self).__init__()
self.LstmCell = nn.LSTMCell(input_size=256, hidden_size=256)
def __init__(self, input_dim, process_step=4):
super(QueryEncoder, self).__init__()
self.input_dim = input_dim
self.process_step = process_step
# self.batch_size = batch_size
self.process = nn.LSTMCell(input_dim, 2*input_dim)
def __init__(self, in_channels, frame_channels, r, attn_type, attn_win,
attn_norm, prenet_type, prenet_dropout, forward_attn,
trans_agent, forward_attn_mask, location_attn, attn_K,
separate_stopnet):
super(Decoder, self).__init__()
self.frame_channels = frame_channels
self.r_init = r
self.r = r
self.encoder_embedding_dim = in_channels
self.separate_stopnet = separate_stopnet
self.max_decoder_steps = 1000
self.stop_threshold = 0.5
# model dimensions
self.query_dim = 1024
self.decoder_rnn_dim = 1024
self.prenet_dim = 256
self.attn_dim = 128
self.p_attention_dropout = 0.1
self.p_decoder_dropout = 0.1
# memory -> |Prenet| -> processed_memory
prenet_dim = self.frame_channels
self.prenet = Prenet(prenet_dim,
prenet_type,
prenet_dropout,
out_features=[self.prenet_dim, self.prenet_dim],
bias=False)
self.attention_rnn = nn.LSTMCell(self.prenet_dim + in_channels,
self.query_dim,
bias=True)
self.attention = init_attn(attn_type=attn_type,
query_dim=self.query_dim,
embedding_dim=in_channels,
attention_dim=128,
location_attention=location_attn,
attention_location_n_filters=32,
attention_location_kernel_size=31,
windowing=attn_win,
norm=attn_norm,
forward_attn=forward_attn,
trans_agent=trans_agent,
forward_attn_mask=forward_attn_mask,
attn_K=attn_K)
self.decoder_rnn = nn.LSTMCell(self.query_dim + in_channels,
self.decoder_rnn_dim,
bias=True)
self.linear_projection = Linear(self.decoder_rnn_dim + in_channels,
self.frame_channels * self.r_init)
self.stopnet = nn.Sequential(
nn.Dropout(0.1),
Linear(self.decoder_rnn_dim + self.frame_channels * self.r_init,
1,
bias=True,
init_gain='sigmoid'))
self.memory_truncated = None
def __init__(self,
lstm_settings_dict,
feature_size_dict={
'acous': 0,
'visual': 0
},
batch_size=32,
seq_length=200,
prediction_length=60,
embedding_info=[],
seq_wind=10):
super(LSTMPredictor, self).__init__()
# General model settings
self.batch_size = batch_size
self.seq_length = seq_length
self.feature_size_dict = feature_size_dict
self.prediction_length = prediction_length
# lstm_settings_dict
self.lstm_settings_dict = lstm_settings_dict
self.feature_size_dict['master'] = 0
if self.lstm_settings_dict['no_subnets']:
for act_mod in self.lstm_settings_dict['active_modalities']:
self.feature_size_dict['master'] += self.feature_size_dict[
act_mod]
else:
for act_mod in self.lstm_settings_dict['active_modalities']:
self.feature_size_dict['master'] += self.lstm_settings_dict[
'hidden_dims'][act_mod]
self.num_layers = lstm_settings_dict['layers']
# embedding settings
self.embedding_info = embedding_info
self.embeddings = {'acous': [], 'visual': []}
self.embedding_indices = {'acous': [], 'visual': []}
self.embed_delete_index_list = {'acous': [], 'visual': []}
self.embed_data_types = {'acous': [], 'visual': []}
self.len_output_of_embeddings = {'acous': 0, 'visual': 0}
self.embedding_flags = {}
# attention
self.attn = nn.Linear(
self.feature_size_dict['visual'] +
self.lstm_settings_dict['hidden_dims']['acous'] * 2,
self.lstm_settings_dict['hidden_dims']['acous']).type(dtype)
self.seq_wind = seq_wind
# m = nn.Linear(20, 30); input = torch.randn(128, 20); output = m(input)
#####################################
for modality in self.embedding_info.keys():
self.embedding_flags[modality] = bool(
len(self.embedding_info[modality]))
if self.embedding_flags[modality]:
for embedding in self.embedding_info[modality]:
self.len_output_of_embeddings[
modality] += 2 * embedding['embedding_out_dim']
for emb_func_indx in range(len(self.embedding_info[modality])):
if self.embedding_info[modality][emb_func_indx][
'embedding_use_func']:
self.embeddings[modality].append(
nn.Embedding(
self.embedding_info[modality][emb_func_indx]
['embedding_num'],
self.embedding_info[modality][emb_func_indx]
['embedding_out_dim']).type(dtype))
self.embedding_func = self.embeddings[modality][-1]
self.embed_data_types[modality].append(dtype_long)
elif self.embedding_info[modality][emb_func_indx][
'use_glove']:
embed_tab_path = self.embedding_info[modality][
emb_func_indx]['glove_embed_table']
glove_embed_table = pickle.load(
open(embed_tab_path, 'rb'))
glove_embed_table[0] = np.random.normal(
0, 1e5, 300) # need this to deal with BCE error
self.embeddings[modality].append(
nn.Embedding.from_pretrained(
torch.FloatTensor(glove_embed_table).type(
dtype),
freeze=self.lstm_settings_dict['freeze_glove'])
)
self.embedding_func = self.embeddings[modality][-1]
self.embed_data_types[modality].append(dtype_long)
print('using glove embeddings')
else:
self.embeddings[modality].append(
nn.Linear(self.embedding_info[modality]
[emb_func_indx]['embedding_num'],
self.embedding_info[modality]
[emb_func_indx]['embedding_out_dim'],
bias=True).type(dtype))
self.embedding_linear = self.embeddings[modality][-1]
self.embed_data_types[modality].append(dtype)
self.embedding_indices[modality].append(
self.embedding_info[modality][emb_func_indx]
['emb_indices']) # two tuples for start and end
for emb_func_indx in range(len(self.embedding_info[modality])):
self.embed_delete_index_list[modality] += list(
range(
self.embedding_indices[modality][emb_func_indx][0]
[0], self.embedding_indices[modality]
[emb_func_indx][0][1]))
self.embed_delete_index_list[modality] += list(
range(
self.embedding_indices[modality][emb_func_indx][1]
[0], self.embedding_indices[modality]
[emb_func_indx][1][1]))
# Initialize LSTMs
self.lstm_dict = {}
if self.lstm_settings_dict['no_subnets']:
if not (len(self.lstm_settings_dict['active_modalities']) == 1):
raise ValueError('Can only have one modality if no subnets')
else:
self.lstm_settings_dict['is_irregular'][
'master'] = self.lstm_settings_dict['is_irregular'][
self.lstm_settings_dict['active_modalities'][0]]
if self.lstm_settings_dict['is_irregular']['master']:
# self.lstm_dict['master'] = nn.LSTMCell(self.feature_size_dict['master'],
# self.lstm_settings_dict['hidden_dims']['master']).type(dtype)
self.lstm_dict['master'] = nn.LSTMCell(
self.feature_size_dict['master'],
self.lstm_settings_dict['hidden_dims']['master']).type(
dtype)
self.lstm_master = self.lstm_dict['master']
else:
self.lstm_dict['master'] = nn.LSTM(
self.feature_size_dict['master'],
self.lstm_settings_dict['hidden_dims']['master']).type(
dtype)
self.lstm_master = self.lstm_dict['master']
else: # Two subnets
self.lstm_settings_dict['is_irregular']['master'] = False
self.lstm_dict['master'] = nn.LSTM(
self.feature_size_dict['master'],
self.lstm_settings_dict['hidden_dims']['master']).type(dtype)
self.lstm_master = self.lstm_dict['master']
for lstm in self.lstm_settings_dict['active_modalities']:
if self.lstm_settings_dict['is_irregular'][lstm]:
# self.lstm_dict[lstm] = nn.LSTMCell(self.feature_size_dict[lstm],
# self.lstm_settings_dict['hidden_dims'][lstm]).type(dtype)
self.lstm_dict[lstm] = nn.LSTMCell(
self.feature_size_dict[lstm],
self.lstm_settings_dict['hidden_dims'][lstm]).type(
dtype)
if lstm == 'acous':
self.lstm_cell_acous = self.lstm_dict[lstm]
else:
self.lstm_cell_visual = self.lstm_dict[lstm]
else:
self.lstm_dict[lstm] = nn.LSTM(
self.feature_size_dict[lstm],
self.lstm_settings_dict['hidden_dims'][lstm]).type(
dtype)
if lstm == 'acous':
self.lstm_acous = self.lstm_dict[lstm]
else:
self.lstm_visual = self.lstm_dict[lstm]
if self.lstm_settings_dict['visual_as_id']['visual']:
# self.lstm_cell_visual = self.lstm_dict['acous'].copy()
self.lstm_visual = self.lstm_dict['acous']
# init dropout layers
self.dropout_dict = {}
for drop_key, drop_val in self.lstm_settings_dict['dropout'].items():
self.dropout_dict[drop_key] = nn.Dropout(drop_val)
setattr(self, 'dropout_' + str(drop_key),
self.dropout_dict[drop_key])
self.out = nn.Linear(self.lstm_settings_dict['hidden_dims']['master'],
prediction_length).type(dtype)
self.init_hidden()
def __init__(self):
super(Decoder, self).__init__()
self.lstm = nn.LSTMCell(conf('emb-size'), conf('dec-hidden-size'))
self.y_concat = nn.Linear(2 * conf('enc-hidden-size') + conf('emb-size'), conf('emb-size'))
trainloader = DataLoader(traindataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=4)
testdataset = MyDataset('test.csv')
testloader = DataLoader(testdataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=4)
mlp1 = MLP(EMB_SIZE).to(device)
mlp2 = MLP(2 * HIDDEN_SIZE).to(device)
mlp3 = MLP(2 * HIDDEN_SIZE).to(device)
r = Pred(HIDDEN_SIZE).to(device)
lstm = nn.LSTMCell(HIDDEN_SIZE, HIDDEN_SIZE).to(device)
embed = torch.nn.functional.one_hot
optimizer_mlp1 = torch.optim.Adam(mlp1.parameters(),
lr=2e-4,
weight_decay=1e-4)
optimizer_mlp2 = torch.optim.Adam(mlp2.parameters(),
lr=2e-4,
weight_decay=1e-4)
optimizer_mlp3 = torch.optim.Adam(mlp3.parameters(),
lr=2e-4,
weight_decay=1e-4)
optimizer_r = torch.optim.Adam(r.parameters(), lr=2e-4, weight_decay=1e-4)
optimizer_lstm = torch.optim.Adam(lstm.parameters(),
lr=2e-4,
weight_decay=1e-4)
def __init__(self, pc_embedder, address_embedder, cache_line_embedder,
positional_embedder, lstm_hidden_size, max_attention_history,
loss_fns=None, cache_pc_embedder=None):
"""Constructs a model to predict evictions from a EvictionEntries history.
At each timestep t, receives:
- pc_t: program counter of t-th memory access.
- a_t: (cache-aligned) address of t-th memory access.
- [l^0_t, ..., l^N_t]: the cache lines present in the cache set accessed
by a_t. Each cache line consists of the cache-aligned address and the pc
of the last access to that address.
Computes:
c_0, h_0 = zeros(lstm_hidden_size)
c_{t + 1}, h_{t + 1} = LSTM([e(pc_t)]; e(a_t)], c_t, h_t)
h^i = attention([h_{t - K}, ..., h_t], query=e(l^i_t)) for i = 1, ..., N
eviction_score s^i = softmax(f(h^i))
The line with the highest eviction score is evicted.
Args:
pc_embedder (embed.Embedder): embeds the program counter.
address_embedder (embed.Embedder): embeds the address.
cache_line_embedder (embed.Embedder): embed the cache line.
positional_embedder (embed.Embedder): embeds positions of the access
history.
lstm_hidden_size (int): dimension of output of LSTM (h and c).
max_attention_history (int): maximum number of past hidden states to
attend over (K in the equation above).
loss_fns (dict): maps a name (str) to a loss function (LossFunction).
The name is used in the loss method. Defaults to top_1_log_likelihood.
cache_pc_embedder (embed.Embedder | None): embeds the pc of each cache
line, if provided. Otherwise cache line pcs are not embedded.
"""
super(EvictionPolicyModel, self).__init__()
self._pc_embedder = pc_embedder
self._address_embedder = address_embedder
self._cache_line_embedder = cache_line_embedder
self._cache_pc_embedder = cache_pc_embedder
self._lstm_cell = nn.LSTMCell(
pc_embedder.embed_dim + address_embedder.embed_dim, lstm_hidden_size)
self._positional_embedder = positional_embedder
query_dim = cache_line_embedder.embed_dim
if cache_pc_embedder is not None:
query_dim += cache_pc_embedder.embed_dim
self._history_attention = attention.MultiQueryAttention(
attention.GeneralAttention(query_dim, lstm_hidden_size))
# f(h, e(l))
self._cache_line_scorer = nn.Linear(
lstm_hidden_size + self._positional_embedder.embed_dim, 1)
self._reuse_distance_estimator = nn.Linear(
lstm_hidden_size + self._positional_embedder.embed_dim, 1)
# Needs to be capped because of limited GPU memory
self._max_attention_history = max_attention_history
if loss_fns is None:
loss_fns = {"log_likelihood": LogProbLoss()}
self._loss_fns = loss_fns
def __init__(self,
num_steps,
x_size,
window_size,
z_what_size,
rnn_hidden_size,
encoder_net=[],
decoder_net=[],
predict_net=[],
embed_net=None,
bl_predict_net=[],
non_linearity='ReLU',
decoder_output_bias=None,
decoder_output_use_sigmoid=False,
use_masking=True,
use_baselines=True,
baseline_scalar=None,
scale_prior_mean=3.0,
scale_prior_sd=0.1,
pos_prior_mean=0.0,
pos_prior_sd=1.0,
likelihood_sd=0.3,
use_cuda=False):
super().__init__()
self.num_steps = num_steps
self.x_size = x_size
self.window_size = window_size
self.z_what_size = z_what_size
self.rnn_hidden_size = rnn_hidden_size
self.use_masking = use_masking
self.use_baselines = use_baselines
self.baseline_scalar = baseline_scalar
self.likelihood_sd = likelihood_sd
self.use_cuda = use_cuda
prototype = torch.tensor(0.).cuda() if use_cuda else torch.tensor(0.)
self.options = dict(dtype=prototype.dtype, device=prototype.device)
self.z_pres_size = 1
self.z_where_size = 3
# By making these parameters they will be moved to the gpu
# when necessary. (They are not registered with pyro for
# optimization.)
self.z_where_loc_prior = nn.Parameter(
torch.FloatTensor([scale_prior_mean, pos_prior_mean, pos_prior_mean]),
requires_grad=False)
self.z_where_scale_prior = nn.Parameter(
torch.FloatTensor([scale_prior_sd, pos_prior_sd, pos_prior_sd]),
requires_grad=False)
# Create nn modules.
rnn_input_size = x_size ** 2 if embed_net is None else embed_net[-1]
rnn_input_size += self.z_where_size + z_what_size + self.z_pres_size
nl = getattr(nn, non_linearity)
self.rnn = nn.LSTMCell(rnn_input_size, rnn_hidden_size)
self.encode = Encoder(window_size ** 2, encoder_net, z_what_size, nl)
self.decode = Decoder(window_size ** 2, decoder_net, z_what_size,
decoder_output_bias, decoder_output_use_sigmoid, nl)
self.predict = Predict(rnn_hidden_size, predict_net, self.z_pres_size, self.z_where_size, nl)
self.embed = Identity() if embed_net is None else MLP(x_size ** 2, embed_net, nl, True)
self.bl_rnn = nn.LSTMCell(rnn_input_size, rnn_hidden_size)
self.bl_predict = MLP(rnn_hidden_size, bl_predict_net + [1], nl)
self.bl_embed = Identity() if embed_net is None else MLP(x_size ** 2, embed_net, nl, True)
# Create parameters.
self.h_init = nn.Parameter(torch.zeros(1, rnn_hidden_size))
self.c_init = nn.Parameter(torch.zeros(1, rnn_hidden_size))
self.bl_h_init = nn.Parameter(torch.zeros(1, rnn_hidden_size))
self.bl_c_init = nn.Parameter(torch.zeros(1, rnn_hidden_size))
self.z_where_init = nn.Parameter(torch.zeros(1, self.z_where_size))
self.z_what_init = nn.Parameter(torch.zeros(1, self.z_what_size))
if use_cuda:
self.cuda()
def __init__(self, args):
super(JointNS, self).__init__(args)
self.image_feature_size = 512
self.object_feature_size = 512
self.hidden_size = 512
self.num_layers = 3
self.loss_function = args.loss
self.number_of_cp = args.number_of_cp
self.environment = args.instance_environment
self.sequence_length = args.sequence_length
self.gpu_ids = args.gpu_ids
self.all_obj_names = args.object_list
self.use_gt_cp = args.use_gt_cp
self.clean_force = True
# configs w.r.t. two losses
self.joint_two_losses = args.joint_two_losses
self.loss1_or_loss2 = None
if args.loss1_w < 0.00001:
self.loss1_or_loss2 = False # update loss2 only
elif args.loss2_w < 0.00001:
self.loss1_or_loss2 = True # update loss1 only
self.loss1_optim, self.loss2_optim, self.joint_optim = None, None, None
# neural force simulator
self.use_image_feature = True
if not self.use_image_feature:
self.one_ns_layer = MLPNS(hidden_size=64, layer_norm=False)
else:
self.one_ns_layer = NSWithImageFeature(hidden_size=64,
layer_norm=False,
image_feature_dim=512)
# self.ns_layer = {obj_name: MLPNS(hidden_size=64, layer_norm=False) for obj_name in self.all_obj_names}
# force predictor networks.
self.feature_extractor = resnet18(pretrained=args.pretrain)
del self.feature_extractor.fc
self.feature_extractor.eval()
self.input_feature_size = self.object_feature_size
self.cp_feature_size = self.number_of_cp * 3
self.image_embed = combine_block_w_do(512, 64, args.dropout_ratio)
self.contact_point_image_embed = combine_block_w_do(
512, 64, args.dropout_ratio)
input_object_embed_size = torch.Tensor(
[3 + 4, 100, self.object_feature_size])
self.input_object_embed = input_embedding_net(
input_object_embed_size.long().tolist(),
dropout=args.dropout_ratio)
self.contact_point_input_object_embed = input_embedding_net(
input_object_embed_size.long().tolist(),
dropout=args.dropout_ratio)
state_embed_size = torch.Tensor([
NoGradEnvState.total_size + self.cp_feature_size, 100,
self.object_feature_size
])
self.state_embed = input_embedding_net(
state_embed_size.long().tolist(), dropout=args.dropout_ratio)
self.lstm_encoder = nn.LSTM(input_size=self.hidden_size + 64 * 7 * 7,
hidden_size=self.hidden_size,
batch_first=True,
num_layers=self.num_layers)
self.contact_point_encoder = nn.LSTM(input_size=self.hidden_size +
64 * 7 * 7,
hidden_size=self.hidden_size,
batch_first=True,
num_layers=self.num_layers)
contact_point_decoder_size = torch.Tensor(
[self.hidden_size, 100, (3) * self.number_of_cp])
self.contact_point_decoder = input_embedding_net(
contact_point_decoder_size.long().tolist(),
dropout=args.dropout_ratio)
self.lstm_decoder = nn.LSTMCell(input_size=self.hidden_size * 2,
hidden_size=self.hidden_size)
forces_directions_decoder_size = torch.Tensor(
[self.hidden_size, 100, (3) * self.number_of_cp])
self.forces_directions_decoder = input_embedding_net(
forces_directions_decoder_size.long().tolist(),
dropout=args.dropout_ratio)
assert args.batch_size == 1, 'have not been implemented yet, because of the environment'
assert self.number_of_cp == 5 # for five fingers
self.all_objects_keypoint_tensor = get_all_objects_keypoint_tensors(
args.data)
if args.gpu_ids != -1:
for obj, val in self.all_objects_keypoint_tensor.items():
self.all_objects_keypoint_tensor[obj] = val.cuda()
self.force_predictor_modules = [
self.feature_extractor, self.image_embed,
self.contact_point_image_embed, self.input_object_embed,
self.contact_point_input_object_embed, self.state_embed,
self.lstm_encoder, self.contact_point_encoder,
self.contact_point_decoder, self.forces_directions_decoder
]
# see gradients for debugging
self.vis_grad = args.vis_grad
self.grad_vis = None
self.train_res = args.train_res or self.vis_grad
def __init__(self,
embed_size,
hidden_size,
vocab,
dropout_rate=0.2,
input_feed=True,
label_smoothing=0.):
super(NMT, self).__init__()
self.embed_size = embed_size
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.vocab = vocab
self.input_feed = True
self.src_code_embed = nn.Embedding(len(vocab.src_code),
embed_size,
padding_idx=vocab.src_code['<pad>'])
self.src_nl_embed = nn.Embedding(len(vocab.src_nl),
embed_size,
padding_idx=vocab.src_nl['<pad>'])
self.tgt_embed = nn.Embedding(len(vocab.tgt),
embed_size,
padding_idx=vocab.tgt['<pad>'])
self.code_encoder_lstm = nn.LSTM(embed_size,
hidden_size,
bidirectional=True)
self.nl_encoder_lstm = nn.LSTM(embed_size,
hidden_size,
bidirectional=True)
decoder_lstm_input = embed_size + (
4 * hidden_size) if self.input_feed else embed_size
self.decoder_lstm = nn.LSTMCell(decoder_lstm_input, hidden_size)
# attention: dot product attention
# project source encoding to decoder rnn's state spacexxxx
self.att_src_code_linear = nn.Linear(hidden_size * 2,
hidden_size,
bias=False)
self.att_src_nl_linear = nn.Linear(hidden_size * 2,
hidden_size,
bias=False)
# transformation of decoder hidden states and context vectors before reading out target words
# this produces the `attentional vector` in (Luong et al., 2015)
self.att_vec_linear = nn.Linear(hidden_size * 2 * 2 + hidden_size,
hidden_size,
bias=False)
# prediction layer of the target vocabulary
self.readout = nn.Linear(hidden_size, len(vocab.tgt), bias=False)
# dropout layer
self.dropout = nn.Dropout(self.dropout_rate)
# initialize the decoder's state and cells with encoder hidden states
self.decoder_cell_init = nn.Linear(hidden_size * 2, hidden_size)
# copy related layers
self.p_gen_linear = nn.Linear(hidden_size * 9 + embed_size, 3)
self.label_smoothing = label_smoothing
if label_smoothing > 0.:
self.label_smoothing_loss = LabelSmoothingLoss(
label_smoothing,
tgt_vocab_size=len(vocab.tgt),
padding_idx=vocab.tgt['<pad>'])
def __init__(self, agent_params, **kwargs):
# call the super-class init
super(ActorCritic, self).__init__()
self.gamma = agent_params['gamma'] # discount factor
self.input_dims = agent_params['input_dims']
self.action_dims = agent_params['action_dims']
if 'rfsize' not in agent_params.keys():
self.rfsize = kwargs.get('rfsize', 4)
else:
self.rfsize = agent_params['rfsize']
if 'padding' not in agent_params.keys():
self.padding = kwargs.get('padding', 1)
else:
self.padding = agent_params['padding']
if 'dilation' not in agent_params.keys():
self.dilation = 1
else:
self.dilation = kwargs.get('dilation', 1)
if 'stride' not in agent_params.keys():
self.stride = kwargs.get('stride', 1)
else:
self.stride = agent_params['stride']
if 'batch_size' not in agent_params.keys():
self.batch_size = kwargs.get('batch_size', 1)
else:
self.batch_size = agent_params['batch_size']
self.use_SR = kwargs.get('use_SR', True)
if 'hidden_types' in agent_params.keys():
if len(agent_params['hidden_dims']) != len(
agent_params['hidden_types']):
raise Exception(
'Incorrect specification of hidden layer dimensions')
hidden_types = agent_params['hidden_types']
# create lists for tracking hidden layers
self.hidden = nn.ModuleList()
self.hidden_dims = agent_params['hidden_dims']
self.hx = []
self.cx = []
# calculate dimensions for each layer
for ind, htype in enumerate(hidden_types):
if htype not in ['linear', 'lstm', 'gru', 'conv', 'pool']:
raise Exception(
f'Unrecognized type for hidden layer {ind}')
if ind == 0:
input_d = self.input_dims
else:
if hidden_types[ind - 1] in [
'conv', 'pool'
] and not htype in ['conv', 'pool']:
input_d = int(np.prod(self.hidden_dims[ind - 1]))
else:
input_d = self.hidden_dims[ind - 1]
if htype in ['conv', 'pool']:
output_d = tuple(self.conv_output(input_d))
self.hidden_dims[ind] = output_d
else:
output_d = self.hidden_dims[ind]
# construct the layer
if htype is 'linear':
self.hidden.append(nn.Linear(input_d, output_d))
self.hx.append(None)
self.cx.append(None)
elif htype is 'lstm':
self.hidden.append(nn.LSTMCell(input_d, output_d))
self.hx.append(
Variable(torch.zeros(self.batch_size, output_d)))
self.cx.append(
Variable(torch.zeros(self.batch_size, output_d)))
elif htype is 'gru':
self.hidden.append(nn.GRUCell(input_d, output_d))
self.hx.append(
Variable(torch.zeros(self.batch_size, output_d)))
self.cx.append(None)
elif htype is 'conv':
in_channels = input_d[0]
out_channels = output_d[0]
self.hidden.append(
nn.Conv2d(in_channels,
out_channels,
kernel_size=self.rfsize,
padding=self.padding,
stride=self.stride,
dilation=self.dilation))
self.hx.append(None)
self.cx.append(None)
elif htype is 'pool':
self.hidden.append(
nn.MaxPool2d(kernel_size=self.rfsize,
padding=self.padding,
stride=self.stride,
dilation=self.dilation))
self.hx.append(None)
self.cx.append(None)
# create the actor and critic layers
self.layers = [self.input_dims
] + self.hidden_dims + [self.action_dims]
self.output = nn.ModuleList([
nn.Linear(output_d, self.action_dims), #actor
nn.Linear(output_d, 1) #critic
])
if self.use_SR:
self.SR = nn.Linear(output_d, output_d) # psi
else:
self.layers = [self.input_dims, self.action_dims]
self.output = nn.ModuleList([
nn.Linear(input_dimensions, action_dimensions), # ACTOR
nn.Linear(input_dimensions, 1)
]) # CRITIC
self.output_d = self.hidden_dims[-1]
self.saved_actions = []
self.saved_rewards = []
self.saved_phi = []
self.saved_psi = []
'''
main_params = []
SR_params = []
for name, para in self.named_parameters():
if name[0:2] == 'SR':
SR_params.append(para)
else:
main_params.append(para)
self.SR_opt = opt([{'params': SR_params,
'lr': 0.01 * agent_params.eta}]) # opt([{'params': freeze, 'lr': 0.0}, {'params': unfreeze, 'lr': agent_params['eta']}], lr=0.0)
self.optimizer = opt(main_params, lr=agent_params.eta)
'''
self.optimizer = optim.Adam(self.parameters(), lr=agent_params['eta'])
def LSTMCell(input_size, hidden_size, **kwargs):
m = nn.LSTMCell(input_size, hidden_size, **kwargs)
for name, param in m.named_parameters():
if 'weight' in name or 'bias' in name:
param.data.uniform_(-0.1, 0.1)
return m
def test_lstm_cell(self):
model = nn.LSTMCell(RNN_INPUT_SIZE, RNN_HIDDEN_SIZE)
input = torch.randn(BATCH_SIZE, RNN_INPUT_SIZE)
h0 = torch.randn(BATCH_SIZE, RNN_HIDDEN_SIZE)
c0 = torch.randn(BATCH_SIZE, RNN_HIDDEN_SIZE)
self.run_model_test(model, train=False, batch_size=BATCH_SIZE, input=(input, (h0, c0)), use_gpu=False)
def __init__(self, num_inputs, action_space):
super(ActorCritic, self).__init__()
self.conv1 = nn.Conv2d(num_inputs, 32, 3, stride=2, padding=1)
self.conv2 = nn.Conv2d(32, 32, 3, stride=2, padding=1)
self.conv3 = nn.Conv2d(32, 32, 3, stride=2, padding=1)
self.conv4 = nn.Conv2d(32, 32, 3, stride=2, padding=1)
self.lstm = nn.LSTMCell(32 * 3 * 3, 256)
num_outputs = action_space.n
self.critic_linear = nn.Linear(256, 1)
self.actor_linear = nn.Linear(256, num_outputs)
################################################################
self.icm_conv1 = nn.Conv2d(num_inputs, 32, 3, stride=2, padding=1)
self.icm_conv2 = nn.Conv2d(32, 32, 3, stride=2, padding=1)
self.icm_conv3 = nn.Conv2d(32, 32, 3, stride=2, padding=1)
self.icm_conv4 = nn.Conv2d(32, 32, 3, stride=2, padding=1)
# self.icm_lstm = nn.LSTMCell(32 * 3 * 3, 256)
self.inverse_linear1 = nn.Linear(288 + 288, 256)
self.inverse_linear2 = nn.Linear(256, num_outputs)
self.forward_linear1 = nn.Linear(288 + num_outputs, 256)
self.forward_linear2 = nn.Linear(256, 288)
# self.inverse_linear1 = nn.Linear(256 + 256, 256)
# self.inverse_linear2 = nn.Linear(256, num_outputs)
# self.forward_linear1 = nn.Linear(256 + num_outputs, 256)
# self.forward_linear2 = nn.Linear(256, 256)
################################################################
self.apply(weights_init)
self.inverse_linear1.weight.data = normalized_columns_initializer(
self.inverse_linear1.weight.data, 0.01)
self.inverse_linear1.bias.data.fill_(0)
self.inverse_linear2.weight.data = normalized_columns_initializer(
self.inverse_linear2.weight.data, 1.0)
self.inverse_linear2.bias.data.fill_(0)
self.forward_linear1.weight.data = normalized_columns_initializer(
self.forward_linear1.weight.data, 0.01)
self.forward_linear1.bias.data.fill_(0)
self.forward_linear2.weight.data = normalized_columns_initializer(
self.forward_linear2.weight.data, 1.0)
self.forward_linear2.bias.data.fill_(0)
'''
self.icm_lstm.bias_ih.data.fill_(0)
self.icm_lstm.bias_hh.data.fill_(0)
'''
################################################################
self.actor_linear.weight.data = normalized_columns_initializer(
self.actor_linear.weight.data, 0.01)
self.actor_linear.bias.data.fill_(0)
self.critic_linear.weight.data = normalized_columns_initializer(
self.critic_linear.weight.data, 1.0)
self.critic_linear.bias.data.fill_(0)
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
self.train()
def test_lstm_cell_is_half(self):
cell = nn.LSTMCell(self.h, self.h)
self.run_cell_test(cell, state_tuple=True)
def __init__(self, args):
super(Graph_MFN, self).__init__()
# print("Graph_MFN initialization ....")
# print(args)
self.d_l, self.d_a, self.d_v = args.feature_dims
self.dh_l, self.dh_a, self.dh_v = args.hidden_dims_l, args.hidden_dims_a, args.hidden_dims_v
total_h_dim = self.dh_l + self.dh_a + self.dh_v
self.mem_dim = args.memsize
self.inner_node_dim = args.inner_node_dim
self.singleton_l_size = args.hidden_dims_l
self.singleton_a_size = args.hidden_dims_a
self.singleton_v_size = args.hidden_dims_v
# Here Changed! (rm window_dim)
# window_dim = args.windowsize
output_dim = args.num_classes
# Here Changed! (rm attInShape, use inner_node_dim instead)
# attInShape = total_h_dim * window_dim
# gammaInShape = attInShape + self.mem_dim
gammaInShape = self.inner_node_dim + self.mem_dim # Todo : we need get inner_node_dim from args.
final_out = total_h_dim + self.mem_dim
# h_att1 = args.NN1Config_shapes
h_att2 = args.NNConfig_shapes
h_gamma1 = args.gamma1Config_shapes
h_gamma2 = args.gamma2Config_shapes
h_out = args.outConfig_shapes
# att1_dropout = args.NN1Config_drop
att2_dropout = args.NNConfig_drop
gamma1_dropout = args.gamma1Config_drop
gamma2_dropout = args.gamma2Config_drop
out_dropout = args.outConfig_drop
self.lstm_l = nn.LSTMCell(self.d_l, self.dh_l)
self.lstm_a = nn.LSTMCell(self.d_a, self.dh_a)
self.lstm_v = nn.LSTMCell(self.d_v, self.dh_v)
# Here Changed! Todo : add Arg param singleton_l singleton_a singleton_v
self.l_transform = nn.Linear(self.dh_l * 2, self.singleton_l_size)
self.a_transform = nn.Linear(self.dh_a * 2, self.singleton_a_size)
self.v_transform = nn.Linear(self.dh_v * 2, self.singleton_v_size)
# Here Changed! (initialize the DFG part) Todo : add Arg param inner node dimension.
pattern_model = nn.Sequential(nn.Linear(100, self.inner_node_dim)).to(
args.device)
efficacy_model = nn.Sequential(nn.Linear(100, self.inner_node_dim)).to(
args.device
) # Note : actually here inner_node_dim can change arbitrarily
self.graph_mfn = DynamicFusionGraph(pattern_model, [
self.singleton_l_size, self.singleton_a_size, self.singleton_v_size
], self.inner_node_dim, efficacy_model, args.device).to(args.device)
# Here Changed! (delete att1 )
# self.att1_fc1 = nn.Linear(attInShape, h_att1)
# self.att1_fc2 = nn.Linear(h_att1, attInShape)
# self.att1_dropout = nn.Dropout(att1_dropout)
# Here Changed! (alter the dim param.)
self.att2_fc1 = nn.Linear(
self.inner_node_dim, h_att2
) # Note: might (inner_node_dim = self.mem_dim) is a common choice.
self.att2_fc2 = nn.Linear(h_att2, self.mem_dim)
self.att2_dropout = nn.Dropout(att2_dropout)
self.gamma1_fc1 = nn.Linear(gammaInShape, h_gamma1)
self.gamma1_fc2 = nn.Linear(h_gamma1, self.mem_dim)
self.gamma1_dropout = nn.Dropout(gamma1_dropout)
self.gamma2_fc1 = nn.Linear(gammaInShape, h_gamma2)
self.gamma2_fc2 = nn.Linear(h_gamma2, self.mem_dim)
self.gamma2_dropout = nn.Dropout(gamma2_dropout)
self.out_fc1 = nn.Linear(final_out, h_out)
self.out_fc2 = nn.Linear(h_out, output_dim)
self.out_dropout = nn.Dropout(out_dropout)
def __init__(self, ins = 2, es = 8, hs = 16):
super(EncoderRNN, self).__init__()
self.hs = hs
self.linear1 = nn.Linear(ins, es)
self.lstm1 = nn.LSTMCell(es, hs)
self.gru1 = nn.GRUCell(es, hs)
def __init__(self,
obs_space,
action_space,
use_memory=False,
use_text=False):
super().__init__()
# Decide which components are enabled
self.use_text = use_text
self.use_memory = use_memory
self.recurrent = use_memory
# Define image embedding
image_chans = obs_space["image"][2]
self.image_conv = nn.Sequential(
nn.Conv2d(image_chans, 16, (2, 2)),
nn.ReLU(),
nn.MaxPool2d((2, 2)),
nn.Conv2d(16, 32, (2, 2)),
nn.ReLU() #,
# nn.Conv2d(32, 64, (2, 2)),
# nn.ReLU()
)
n = obs_space["image"][0]
m = obs_space["image"][1]
# self.image_embedding_size = ((n-1)//2-2)*((m-1)//2-2)*64 # original. image_embedding_size is basically the number of elements at output of self.image_conv(x), not accounting for batch size.
self.image_embedding_size = 32 * 6 # 32 is outchan, 6 is h*w
# Define memory
if self.use_memory:
self.memory_rnn = nn.LSTMCell(self.image_embedding_size,
self.semi_memory_size)
# Define text embedding
if self.use_text:
self.word_embedding_size = 32
self.word_embedding = nn.Embedding(obs_space["text"],
self.word_embedding_size)
self.text_embedding_size = 128
self.text_rnn = nn.GRU(self.word_embedding_size,
self.text_embedding_size,
batch_first=True)
# Resize image embedding
self.embedding_size = self.semi_memory_size
if self.use_text:
self.embedding_size += self.text_embedding_size
# Define actor's model
if isinstance(action_space, gym.spaces.Discrete):
self.actor = nn.Sequential(nn.Linear(self.embedding_size, 16),
nn.Tanh(),
nn.Linear(16, action_space.n))
else:
raise ValueError("Unknown action space: " + str(action_space))
# Define critic's model
self.critic = nn.Sequential(nn.Linear(self.embedding_size, 16),
nn.Tanh(), nn.Linear(16, 1))
# Initialize parameters correctly
self.apply(initialize_parameters)
def __init__(self, input_size, hidden_size):
super(AdaptiveLSTMCell, self).__init__()
self.lstm_cell = nn.LSTMCell(input_size, hidden_size)
self.x_gate = nn.Linear(input_size, hidden_size)
self.h_gate = nn.Linear(hidden_size, hidden_size)
def __init__(self, embed_size, hidden_size, vocab, dropout_rate=0.2):
""" Init NMT Model.
@param embed_size (int): Embedding size (dimensionality)
@param hidden_size (int): Hidden Size (dimensionality)
@param vocab (Vocab): Vocabulary object containing src and tgt languages
See vocab.py for documentation.
@param dropout_rate (float): Dropout probability, for attention
"""
super(NMT, self).__init__()
self.model_embeddings = ModelEmbeddings(embed_size, vocab)
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.vocab = vocab
# default values
self.encoder = None
self.decoder = None
self.h_projection = None
self.c_projection = None
self.att_projection = None
self.combined_output_projection = None
self.target_vocab_projection = None
self.dropout = None
### YOUR CODE HERE (~8 Lines)
### TODO - Initialize the following variables:
### self.encoder (Bidirectional LSTM with bias)
### self.decoder (LSTM Cell with bias)
### self.h_projection (Linear Layer with no bias), called W_{h} in the PDF.
### self.c_projection (Linear Layer with no bias), called W_{c} in the PDF.
### self.att_projection (Linear Layer with no bias), called W_{attProj} in the PDF.
### self.combined_output_projection (Linear Layer with no bias), called W_{u} in the PDF.
### self.target_vocab_projection (Linear Layer with no bias), called W_{vocab} in the PDF.
### self.dropout (Dropout Layer)
###
### Use the following docs to properly initialize these variables:
### LSTM:
### https://pytorch.org/docs/stable/nn.html#torch.nn.LSTM
### LSTM Cell:
### https://pytorself.ch.org/docs/stable/nn.html#torch.nn.LSTMCell
### Linear Layer:
### https://pytorch.org/docs/stable/nn.html#torch.nn.Linear
### Dropout Layer:
### https://pytorch.org/docs/stable/nn.html#torch.nn.Dropout
self.encoder = nn.LSTM(embed_size,
hidden_size,
bias=True,
dropout=self.dropout_rate,
bidirectional=True)
self.decoder = nn.LSTMCell(
embed_size + hidden_size, hidden_size,
bias=True) # why add embed+hidden for input size?
self.h_projection = nn.Linear(
hidden_size * 2, hidden_size,
bias=False) # prj output of last h_state of encode (R^2h) to R^h
self.c_projection = nn.Linear(hidden_size * 2, hidden_size, bias=False)
self.att_projection = nn.Linear(hidden_size * 2,
hidden_size,
bias=False)
self.combined_output_projection = nn.Linear(
hidden_size * 3, hidden_size,
bias=False) # use after combined attention output and h_decode
self.target_vocab_projection = nn.Linear(
hidden_size, len(vocab.tgt), bias=False) # for softmax of last
self.dropout = nn.Dropout(self.dropout_rate)
def __init__(self, embed_size, hidden_size, vocab, dropout_rate=0.2):
""" Init NMT Model.
@param embed_size (int): Embedding size (dimensionality)
@param hidden_size (int): Hidden Size (dimensionality)
@param vocab (Vocab): Vocabulary object containing src and tgt languages
See vocab.py for documentation.
@param dropout_rate (float): Dropout probability, for attention
"""
super(NMT, self).__init__()
self.model_embeddings = ModelEmbeddings(embed_size, vocab)
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.vocab = vocab
# default values
self.encoder = None
self.decoder = None
self.h_projection = None
self.c_projection = None
self.att_projection = None
self.combined_output_projection = None
self.target_vocab_projection = None
self.dropout = None
### YOUR CODE HERE (~8 Lines)
### TODO - Initialize the following variables:
### self.encoder (Bidirectional LSTM with bias)
### self.decoder (LSTM Cell with bias)
### self.h_projection (Linear Layer with no bias), called W_{h} in the PDF.
### self.c_projection (Linear Layer with no bias), called W_{c} in the PDF.
### self.att_projection (Linear Layer with no bias), called W_{attProj} in the PDF.
### self.combined_output_projection (Linear Layer with no bias), called W_{u} in the PDF.
### self.target_vocab_projection (Linear Layer with no bias), called W_{vocab} in the PDF.
### self.dropout (Dropout Layer)
###
### Use the following docs to properly initialize these variables:
### LSTM:
### https://pytorch.org/docs/stable/nn.html#torch.nn.LSTM
### LSTM Cell:
### https://pytorch.org/docs/stable/nn.html#torch.nn.LSTMCell
### Linear Layer:
### https://pytorch.org/docs/stable/nn.html#torch.nn.Linear
### Dropout Layer:
### https://pytorch.org/docs/stable/nn.html#torch.nn.Dropout
self.encoder = nn.LSTM(embed_size,
self.hidden_size,
num_layers=1,
bias=True,
bidirectional=True)
self.decoder = nn.LSTMCell(
embed_size + self.hidden_size,
self.hidden_size,
bias=True,
)
self.h_projection = nn.Linear(2 * embed_size, embed_size, False)
self.c_projection = nn.Linear(2 * embed_size, embed_size, False)
self.att_projection = nn.Linear(2 * embed_size, embed_size, False)
self.combined_output_projection = nn.Linear(3 * embed_size, embed_size,
False)
self.target_vocab_projection = nn.Linear(
embed_size, len(self.vocab.tgt), False
) # all these layers are called projection because project input dim vector to output dim vector
self.dropout = nn.Dropout(dropout_rate)
def __init__(self, latents, actions, hiddens, gaussians):
super().__init__(latents, actions, hiddens, gaussians)
self.rnn = nn.LSTMCell(latents + actions, hiddens)
def __init__(self,
cell='gru',
use_baseline=True,
n_actions=10,
n_units=64,
fusion_dim=128,
n_input=76,
n_hidden=128,
demo_dim=17,
n_output=1,
dropout=0.0,
lamda=0.5,
device='cpu'):
super(Agent, self).__init__()
self.cell = cell
self.use_baseline = use_baseline
self.n_actions = n_actions
self.n_units = n_units
self.n_input = n_input
self.n_hidden = n_hidden
self.n_output = n_output
self.dropout = dropout
self.lamda = lamda
self.fusion_dim = fusion_dim
self.demo_dim = demo_dim
self.device = device
self.agent1_action = []
self.agent1_prob = []
self.agent1_entropy = []
self.agent1_baseline = []
self.agent2_action = []
self.agent2_prob = []
self.agent2_entropy = []
self.agent2_baseline = []
self.agent1_fc1 = nn.Linear(self.n_hidden + self.demo_dim,
self.n_units)
self.agent2_fc1 = nn.Linear(self.n_input + self.demo_dim, self.n_units)
self.agent1_fc2 = nn.Linear(self.n_units, self.n_actions)
self.agent2_fc2 = nn.Linear(self.n_units, self.n_actions)
if use_baseline == True:
self.agent1_value = nn.Linear(self.n_units, 1)
self.agent2_value = nn.Linear(self.n_units, 1)
if self.cell == 'lstm':
self.rnn = nn.LSTMCell(self.n_input, self.n_hidden)
else:
self.rnn = nn.GRUCell(self.n_input, self.n_hidden)
for name, param in self.rnn.named_parameters():
if 'bias' in name:
nn.init.constant(param, 0.0)
elif 'weight' in name:
nn.init.orthogonal_(param)
if dropout > 0.0:
self.nn_dropout = nn.Dropout(p=dropout)
self.init_h = nn.Linear(self.demo_dim, self.n_hidden)
self.init_c = nn.Linear(self.demo_dim, self.n_hidden)
self.fusion = nn.Linear(self.n_hidden + self.demo_dim, self.fusion_dim)
self.output = nn.Linear(self.fusion_dim, self.n_output)
self.sigmoid = nn.Sigmoid()
self.softmax = nn.Softmax()
self.tanh = nn.Tanh()
self.relu = nn.ReLU()
def _set_cell(self):
return nn.LSTMCell(input_size=self.in_dims + self.pb_dims,
hidden_size=self.unit_nums)
def __init__(self, embed_size, hidden_size, vocab, dropout_rate=0.2):
""" Init NMT Model.
@param embed_size (int): Embedding size (dimensionality)
@param hidden_size (int): Hidden Size, the size of hidden states (dimensionality)
@param vocab (Vocab): Vocabulary object containing src and tgt languages
See vocab.py for documentation.
@param dropout_rate (float): Dropout probability, for attention
"""
super(NMT, self).__init__()
self.model_embeddings = ModelEmbeddings(embed_size, vocab)
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.vocab = vocab
# default values
self.encoder = None
self.decoder = None
self.h_projection = None
self.c_projection = None
self.att_projection = None
self.combined_output_projection = None
self.target_vocab_projection = None
self.dropout = None
# For sanity check only, not relevant to implementation
self.gen_sanity_check = False
self.counter = 0
### YOUR CODE HERE (~8 Lines)
### TODO - Initialize the following variables:
### self.encoder (Bidirectional LSTM with bias)
### self.decoder (LSTM Cell with bias)
### self.h_projection (Linear Layer with no bias), called W_{h} in the PDF.
### self.c_projection (Linear Layer with no bias), called W_{c} in the PDF.
### self.att_projection (Linear Layer with no bias), called W_{attProj} in the PDF.
### self.combined_output_projection (Linear Layer with no bias), called W_{u} in the PDF.
### self.target_vocab_projection (Linear Layer with no bias), called W_{vocab} in the PDF.
### self.dropout (Dropout Layer)
###
### Use the following docs to properly initialize these variables:
### LSTM:
### https://pytorch.org/docs/stable/nn.html#torch.nn.LSTM
### LSTM Cell:
### https://pytorch.org/docs/stable/nn.html#torch.nn.LSTMCell
### Linear Layer:
### https://pytorch.org/docs/stable/nn.html#torch.nn.Linear
### Dropout Layer:
### https://pytorch.org/docs/stable/nn.html#torch.nn.Dropout
self.encoder = nn.LSTM(input_size=embed_size,
hidden_size=self.hidden_size,
bias=True,
bidirectional=True)
self.decoder = nn.LSTMCell(input_size=embed_size + self.hidden_size,
hidden_size=self.hidden_size,
bias=True)
self.h_projection = nn.Linear(self.hidden_size * 2,
self.hidden_size,
bias=False) # W_{h}
self.c_projection = nn.Linear(self.hidden_size * 2,
self.hidden_size,
bias=False) # W_{c}
self.att_projection = nn.Linear(self.hidden_size * 2,
self.hidden_size,
bias=False) # W_{attProj}
self.combined_output_projection = nn.Linear(self.hidden_size * 3,
self.hidden_size,
bias=False) # W_{u}
self.target_vocab_projection = nn.Linear(self.hidden_size,
len(self.vocab.tgt),
bias=False) # W_vocab
self.dropout = nn.Dropout(p=self.dropout_rate)