Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Modify EmbeddingDropout and adapt AWD_LSTM to trigger hooks in Embedding layers #2906

Open
wants to merge 3 commits into
base: master
Choose a base branch
from
Open

Modify EmbeddingDropout and adapt AWD_LSTM to trigger hooks in Embedding layers #2906

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4ec4ebb
Modify EmbeddingDropout and adapt AWD_LSTM
arnaujc91 Oct 25, 2020
8d22796
added encoder_dp layer to AWD_LSTM in order to pass the tests
arnaujc91 Nov 3, 2020
5f6f754
tests fixed for the new EmbeddingDropout class
arnaujc91 Nov 5, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
32 changes: 16 additions & 16 deletions fastai/text/models/awdlstm.py
View file
Open in desktop
Expand Up @@ -61,21 +61,21 @@ def reset(self):
def _do_nothing(self): pass

# Cell
class EmbeddingDropout(Module):
"Apply dropout with probability `embed_p` to an embedding layer `emb`."

def __init__(self, emb, embed_p):
self.emb,self.embed_p = emb,embed_p
class EmbeddingDropout(nn.Embedding):
"Apply dropout with probability `embed_p` to an embedding layer."
def __init__(self, *args, embed_p, **kwargs):
super().__init__(*args, **kwargs)
self.embed_p = embed_p

def forward(self, words, scale=None):
if self.training and self.embed_p != 0:
size = (self.emb.weight.size(0),1)
mask = dropout_mask(self.emb.weight.data, size, self.embed_p)
masked_embed = self.emb.weight * mask
else: masked_embed = self.emb.weight
size = (self.weight.size(0),1)
mask = dropout_mask(self.weight.data, size, self.embed_p)
masked_embed = self.weight * mask
else: masked_embed = self.weight
if scale: masked_embed.mul_(scale)
return F.embedding(words, masked_embed, ifnone(self.emb.padding_idx, -1), self.emb.max_norm,
self.emb.norm_type, self.emb.scale_grad_by_freq, self.emb.sparse)
return F.embedding(words, masked_embed, ifnone(self.padding_idx, -1), self.max_norm,
self.norm_type, self.scale_grad_by_freq, self.sparse)

# Cell
class AWD_LSTM(Module):
Expand All @@ -87,11 +87,11 @@ def __init__(self, vocab_sz, emb_sz, n_hid, n_layers, pad_token=1, hidden_p=0.2,
store_attr('emb_sz,n_hid,n_layers,pad_token')
self.bs = 1
self.n_dir = 2 if bidir else 1
self.encoder = nn.Embedding(vocab_sz, emb_sz, padding_idx=pad_token)
self.encoder_dp = EmbeddingDropout(self.encoder, embed_p)
self.encoder = EmbeddingDropout(vocab_sz, emb_sz, embed_p=embed_p, padding_idx=pad_token)
self.encoder_dp = self.encoder
self.encoder.weight.data.uniform_(-self.initrange, self.initrange)
self.rnns = nn.ModuleList([self._one_rnn(emb_sz if l == 0 else n_hid, (n_hid if l != n_layers - 1 else emb_sz)//self.n_dir,
bidir, weight_p, l) for l in range(n_layers)])
self.encoder.weight.data.uniform_(-self.initrange, self.initrange)
self.input_dp = RNNDropout(input_p)
self.hidden_dps = nn.ModuleList([RNNDropout(hidden_p) for l in range(n_layers)])
self.reset()
Expand Down Expand Up @@ -139,7 +139,7 @@ def reset(self):
def awd_lstm_lm_split(model):
"Split a RNN `model` in groups for differential learning rates."
groups = [nn.Sequential(rnn, dp) for rnn, dp in zip(model[0].rnns, model[0].hidden_dps)]
groups = L(groups + [nn.Sequential(model[0].encoder, model[0].encoder_dp, model[1])])
groups = L(groups + [nn.Sequential(model[0].encoder, model[0].encoder, model[1])])
return groups.map(params)

# Cell
Expand All @@ -149,7 +149,7 @@ def awd_lstm_lm_split(model):
# Cell
def awd_lstm_clas_split(model):
"Split a RNN `model` in groups for differential learning rates."
groups = [nn.Sequential(model[0].module.encoder, model[0].module.encoder_dp)]
groups = [nn.Sequential(model[0].module.encoder, model[0].module.encoder)]
groups += [nn.Sequential(rnn, dp) for rnn, dp in zip(model[0].module.rnns, model[0].module.hidden_dps)]
groups = L(groups + [model[1]])
return groups.map(params)
Expand Down
37 changes: 18 additions & 19 deletions nbs/32_text.models.awdlstm.ipynb
View file
Open in desktop
Expand Up @@ -210,21 +210,21 @@
"outputs": [],
"source": [
"#export\n",
"class EmbeddingDropout(Module):\n",
" \"Apply dropout with probability `embed_p` to an embedding layer `emb`.\"\n",
"\n",
" def __init__(self, emb, embed_p):\n",
" self.emb,self.embed_p = emb,embed_p\n",
"class EmbeddingDropout(nn.Embedding):\n",
" \"Apply dropout with probability `embed_p` to an embedding layer.\"\n",
" def __init__(self, *args, embed_p, **kwargs):\n",
" super().__init__(*args, **kwargs)\n",
" self.embed_p = embed_p\n",
"\n",
" def forward(self, words, scale=None):\n",
" if self.training and self.embed_p != 0:\n",
" size = (self.emb.weight.size(0),1)\n",
" mask = dropout_mask(self.emb.weight.data, size, self.embed_p)\n",
" masked_embed = self.emb.weight * mask\n",
" else: masked_embed = self.emb.weight\n",
" size = (self.weight.size(0),1)\n",
" mask = dropout_mask(self.weight.data, size, self.embed_p)\n",
" masked_embed = self.weight * mask\n",
" else: masked_embed = self.weight\n",
" if scale: masked_embed.mul_(scale)\n",
" return F.embedding(words, masked_embed, ifnone(self.emb.padding_idx, -1), self.emb.max_norm,\n",
" self.emb.norm_type, self.emb.scale_grad_by_freq, self.emb.sparse)"
" return F.embedding(words, masked_embed, ifnone(self.padding_idx, -1), self.max_norm,\n",
" self.norm_type, self.scale_grad_by_freq, self.sparse)"
]
},
{
Expand All @@ -233,10 +233,9 @@
"metadata": {},
"outputs": [],
"source": [
"enc = nn.Embedding(10, 7, padding_idx=1)\n",
"enc_dp = EmbeddingDropout(enc, 0.5)\n",
"enc = EmbeddingDropout(10, 7, embed_p=0.5, padding_idx=1)\n",
"tst_inp = torch.randint(0,10,(8,))\n",
"tst_out = enc_dp(tst_inp)\n",
"tst_out = enc(tst_inp)\n",
"for i in range(8):\n",
" assert (tst_out[i]==0).all() or torch.allclose(tst_out[i], 2*enc.weight[tst_inp[i]])"
]
Expand All @@ -257,11 +256,11 @@
" store_attr('emb_sz,n_hid,n_layers,pad_token')\n",
" self.bs = 1\n",
" self.n_dir = 2 if bidir else 1\n",
" self.encoder = nn.Embedding(vocab_sz, emb_sz, padding_idx=pad_token)\n",
" self.encoder_dp = EmbeddingDropout(self.encoder, embed_p)\n",
" self.encoder = EmbeddingDropout(vocab_sz, emb_sz, embed_p=embed_p, padding_idx=pad_token)\n",
" self.encoder_dp = self.encoder\n",
" self.encoder.weight.data.uniform_(-self.initrange, self.initrange)\n",
" self.rnns = nn.ModuleList([self._one_rnn(emb_sz if l == 0 else n_hid, (n_hid if l != n_layers - 1 else emb_sz)//self.n_dir,\n",
" bidir, weight_p, l) for l in range(n_layers)])\n",
" self.encoder.weight.data.uniform_(-self.initrange, self.initrange)\n",
" self.input_dp = RNNDropout(input_p)\n",
" self.hidden_dps = nn.ModuleList([RNNDropout(hidden_p) for l in range(n_layers)])\n",
" self.reset()\n",
Expand Down Expand Up @@ -382,7 +381,7 @@
"def awd_lstm_lm_split(model):\n",
" \"Split a RNN `model` in groups for differential learning rates.\"\n",
" groups = [nn.Sequential(rnn, dp) for rnn, dp in zip(model[0].rnns, model[0].hidden_dps)]\n",
" groups = L(groups + [nn.Sequential(model[0].encoder, model[0].encoder_dp, model[1])])\n",
" groups = L(groups + [nn.Sequential(model[0].encoder, model[0].encoder, model[1])])\n",
" return groups.map(params)"
]
},
Expand All @@ -406,7 +405,7 @@
"#export\n",
"def awd_lstm_clas_split(model):\n",
" \"Split a RNN `model` in groups for differential learning rates.\"\n",
" groups = [nn.Sequential(model[0].module.encoder, model[0].module.encoder_dp)]\n",
" groups = [nn.Sequential(model[0].module.encoder, model[0].module.encoder)]\n",
" groups += [nn.Sequential(rnn, dp) for rnn, dp in zip(model[0].module.rnns, model[0].module.hidden_dps)]\n",
" groups = L(groups + [model[1]])\n",
" return groups.map(params)"
Expand Down