From 2bd5bc1f4f92e5a6ef4de8b45698358ecb24a50d Mon Sep 17 00:00:00 2001
From: Moritz Ibing <ibing@cs.rwth-aachen.de>
Date: Thu, 24 Feb 2022 09:16:42 +0100
Subject: [PATCH] revert latent filtering
---
.../check_sequence_length_transform.py | 24 ++-----------------
modules/generative_head/composite_head_A.py | 14 ++---------
modules/generative_head/composite_head_B.py | 14 ++---------
modules/generative_head/composite_head_C.py | 14 ++---------
modules/generative_head/composite_head_D.py | 14 ++---------
.../double_substitution_head.py | 9 -------
modules/generative_head/substitution_head.py | 6 -----
.../double_substitution_embedding.py | 23 ++++--------------
.../token_embedding/substitution_embedding.py | 19 +++------------
9 files changed, 17 insertions(+), 120 deletions(-)
diff --git a/data/transform/check_sequence_length_transform.py b/data/transform/check_sequence_length_transform.py
index 79ff339..961610a 100644
--- a/data/transform/check_sequence_length_transform.py
+++ b/data/transform/check_sequence_length_transform.py
@@ -1,6 +1,3 @@
-import torch
-
-
class CheckSequenceLenghtTransform():
# TODO: make this maps actually properties of the embedding class or decouple them from this module
@@ -67,26 +64,9 @@ class CheckSequenceLenghtTransform():
for i in range(min(len(self.substitution_level), max_depth)):
sub_diff = self.substitution_level[i]
conv_fac = self.convolution_factor[i]
- dep_level = i + 1
-
- if sub_diff == 0:
- num_vectors = torch.sum(torch.from_numpy(dep) == dep_level) // conv_fac
- elif sub_diff == 1:
- val_1 = torch.from_numpy(val)[torch.from_numpy(dep) == (dep_level - 1)]
- num_vectors = (val_1.view(-1, conv_fac) == 2).max(dim=-1)[0].sum()
- elif sub_diff == 2:
- val_1 = torch.from_numpy(val)[torch.from_numpy(dep) == (dep_level - 1)]
- val_2 = torch.from_numpy(val)[torch.from_numpy(dep) == (dep_level - 2)]
- mask_1 = (val_1.view(-1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- mask_2 = mask_2.view(-1, conv_fac).max(dim=-1)[0]
- num_vectors = mask_2.sum()
- else:
- print("ERROR: substitution factors bigger than 2 are not implemented")
- return None
+ dep_level = i + 1 - sub_diff
- sum_sequence_length += num_vectors
+ sum_sequence_length += len(dep[dep == dep_level]) // conv_fac
if sum_sequence_length > self.num_positions:
return None
diff --git a/modules/generative_head/composite_head_A.py b/modules/generative_head/composite_head_A.py
index d10911e..09a4709 100644
--- a/modules/generative_head/composite_head_A.py
+++ b/modules/generative_head/composite_head_A.py
@@ -107,10 +107,7 @@ class CompositeHeadA(nn.Module):
layer_dep = torch.cat([dep[dep == (layer_depth - 1)], dep[dep == layer_depth]])
layer_pos = torch.cat([pos[dep == (layer_depth - 1)], pos[dep == layer_depth]])
# compute number of vectors in latent vector of current layer
- val_1 = val[dep == (layer_depth - 1)]
- num_vectors = (val_1.view(-1, self.reduction_factor[layer_depth]) == 2).max(dim=-1)[0].sum()
-
- # num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
elif layer_depth in (7, 8): # handle substitution
# get value, depth and position sequence of previous and current layer
layer_val = torch.cat(
@@ -135,14 +132,7 @@ class CompositeHeadA(nn.Module):
]
)
# compute number of vectors in latent vector of current layer
- val_1 = val[dep == (layer_depth - 1)]
- val_2 = val[dep == (layer_depth - 2)]
- mask_1 = (val_1.view(-1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- num_vectors = mask_2.view(-1, self.reduction_factor[layer_depth]).max(dim=-1)[0].sum()
-
- # num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
# filter latent vector of current layer
layer_vec = latent_vec[vector_idx:vector_idx + num_vectors]
diff --git a/modules/generative_head/composite_head_B.py b/modules/generative_head/composite_head_B.py
index 635ac11..14133ff 100644
--- a/modules/generative_head/composite_head_B.py
+++ b/modules/generative_head/composite_head_B.py
@@ -100,10 +100,7 @@ class CompositeHeadB(nn.Module):
layer_dep = torch.cat([dep[dep == (layer_depth - 1)], dep[dep == layer_depth]])
layer_pos = torch.cat([pos[dep == (layer_depth - 1)], pos[dep == layer_depth]])
# compute number of vectors in latent vector of current layer
- val_1 = val[dep == (layer_depth - 1)]
- num_vectors = (val_1.view(-1, self.reduction_factor[layer_depth]) == 2).max(dim=-1)[0].sum()
-
- # num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
elif layer_depth in (7, 8): # handle substitution
# get value, depth and position sequence of previous and current layer
layer_val = torch.cat(
@@ -128,14 +125,7 @@ class CompositeHeadB(nn.Module):
]
)
# compute number of vectors in latent vector of current layer
- val_1 = val[dep == (layer_depth - 1)]
- val_2 = val[dep == (layer_depth - 2)]
- mask_1 = (val_1.view(-1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- num_vectors = mask_2.view(-1, self.reduction_factor[layer_depth]).max(dim=-1)[0].sum()
-
- # num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
# filter latent vector of current layer
layer_vec = latent_vec[vector_idx:vector_idx + num_vectors]
diff --git a/modules/generative_head/composite_head_C.py b/modules/generative_head/composite_head_C.py
index cfd4e71..39f58f0 100644
--- a/modules/generative_head/composite_head_C.py
+++ b/modules/generative_head/composite_head_C.py
@@ -100,10 +100,7 @@ class CompositeHeadC(nn.Module):
layer_dep = torch.cat([dep[dep == (layer_depth - 1)], dep[dep == layer_depth]])
layer_pos = torch.cat([pos[dep == (layer_depth - 1)], pos[dep == layer_depth]])
# compute number of vectors in latent vector of current layer
- val_1 = val[dep == (layer_depth - 1)]
- num_vectors = (val_1.view(-1, self.reduction_factor[layer_depth]) == 2).max(dim=-1)[0].sum()
-
- # num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
elif layer_depth in (7, 8): # handle double substitution
# get value, depth and position sequence of previous and current layer
layer_val = torch.cat(
@@ -127,15 +124,8 @@ class CompositeHeadC(nn.Module):
pos[dep == layer_depth],
]
)
- val_1 = val[dep == (layer_depth - 1)]
- val_2 = val[dep == (layer_depth - 2)]
- mask_1 = (val_1.view(-1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- num_vectors = mask_2.view(-1, self.reduction_factor[layer_depth]).max(dim=-1)[0].sum()
-
# compute number of vectors in latent vector of current layer
- # num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
# filter latent vector of current layer
layer_vec = latent_vec[vector_idx:vector_idx + num_vectors]
diff --git a/modules/generative_head/composite_head_D.py b/modules/generative_head/composite_head_D.py
index ae09651..480e6a6 100644
--- a/modules/generative_head/composite_head_D.py
+++ b/modules/generative_head/composite_head_D.py
@@ -107,10 +107,7 @@ class CompositeHeadD(nn.Module):
layer_dep = torch.cat([dep[dep == (layer_depth - 1)], dep[dep == layer_depth]])
layer_pos = torch.cat([pos[dep == (layer_depth - 1)], pos[dep == layer_depth]])
# compute number of vectors in latent vector of current layer
- val_1 = val[dep == (layer_depth - 1)]
- num_vectors = (val_1.view(-1, self.reduction_factor[layer_depth]) == 2).max(dim=-1)[0].sum()
-
- # num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 1)) // self.reduction_factor[layer_depth]
elif layer_depth in (7, 8): # handle double substitution
# get value, depth and position sequence of previous and current layer
layer_val = torch.cat(
@@ -134,15 +131,8 @@ class CompositeHeadD(nn.Module):
pos[dep == layer_depth],
]
)
- val_1 = val[dep == (layer_depth - 1)]
- val_2 = val[dep == (layer_depth - 2)]
- mask_1 = (val_1.view(-1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- num_vectors = mask_2.view(-1, self.reduction_factor[layer_depth]).max(dim=-1)[0].sum()
-
# compute number of vectors in latent vector of current layer
- # num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
+ num_vectors = torch.sum(dep == (layer_depth - 2)) // self.reduction_factor[layer_depth]
# filter latent vector of current layer
layer_vec = latent_vec[vector_idx:vector_idx + num_vectors]
diff --git a/modules/generative_head/double_substitution_head.py b/modules/generative_head/double_substitution_head.py
index da18761..e1d41dd 100644
--- a/modules/generative_head/double_substitution_head.py
+++ b/modules/generative_head/double_substitution_head.py
@@ -20,7 +20,6 @@ class DoubleSubstitutionHead(nn.Module):
"""
super(DoubleSubstitutionHead, self).__init__()
self.head_dim = head_dim
- self.conv_size = conv_size
deconvolution_2 = [nn.GELU(), Deconvolution(embed_dim, head_dim, conv_size)]
for i in range(n_layer - 1):
@@ -115,14 +114,6 @@ class DoubleSubstitutionHead(nn.Module):
# substitute all mixed token embeddings of third to last layer, with token embeddings of penultimate layer
emb_2[val_2 == 2] = self.down_convolution_1(emb_1) # [N, T1, C]
- # filter out all tokens, that do not have any descendants in last layer
- mask_1 = (val_1.view(batch_size, -1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- mask_2 = mask_2.view(batch_size, -1, self.conv_size).max(dim=-1)[0]
- emb_2 = emb_2.view(batch_size, -1, self.conv_size, self.head_dim)[mask_2].view(batch_size, -1, self.head_dim)
- val_2 = val_2.view(batch_size, -1, self.conv_size)[mask_2].view(batch_size, -1)
-
emb_0 = self.convolution_0(emb_0[:, :mix_1 * 8])
emb_1 = self.convolution_1(emb_1)
emb_2 = self.convolution_2(emb_2)
diff --git a/modules/generative_head/substitution_head.py b/modules/generative_head/substitution_head.py
index 9fe20d4..b4a7fd6 100644
--- a/modules/generative_head/substitution_head.py
+++ b/modules/generative_head/substitution_head.py
@@ -20,7 +20,6 @@ class SubstitutionHead(nn.Module):
"""
super(SubstitutionHead, self).__init__()
self.head_dim = head_dim
- self.conv_size = conv_size
deconvolution_1 = [nn.GELU(), Deconvolution(embed_dim, head_dim, conv_size)]
for i in range(n_layer - 1):
@@ -98,11 +97,6 @@ class SubstitutionHead(nn.Module):
# substitite all mixed token embeddings of penultimate layer, with token embeddings of last layer
emb_1[val_1 == 2] = self.down_convolution(emb_0[:, :mix_1 * 8]) # [N, T1, C]
- # filter out all tokens, that do not have any descendants in last layer
- mask = (val_1.view(batch_size, -1, self.conv_size) == 2).max(dim=-1)[0]
- emb_1 = emb_1.view(batch_size, -1, self.conv_size, self.head_dim)[mask].view(batch_size, -1, self.head_dim)
- val_1 = val_1.view(batch_size, -1, self.conv_size)[mask].view(batch_size, -1)
-
emb_0 = self.convolution_0(emb_0[:, :mix_1 * 8])
emb_1 = self.convolution_1(emb_1)
diff --git a/modules/token_embedding/double_substitution_embedding.py b/modules/token_embedding/double_substitution_embedding.py
index bad4581..868a91f 100644
--- a/modules/token_embedding/double_substitution_embedding.py
+++ b/modules/token_embedding/double_substitution_embedding.py
@@ -97,6 +97,9 @@ class DoubleSubstitutionEmbedding(nn.Module):
dep_0[i, :len_0[i]] = depth[i, len_2[i] + len_1[i]:len_2[i] + len_1[i] + len_0[i]]
pos_0[i, :len_0[i]] = position[i, len_2[i] + len_1[i]:len_2[i] + len_1[i] + len_0[i]]
+ # precompute padding mask
+ self.mask = padding_mask(val_2[:, ::self.conv_size], device=value.device) # [N, S'_2, E]
+
# convolute embedded tokens of last layer
y_0 = self.convolution_0(x_0) # [N, S'_0, E // 4]
# substitite all mixed token embeddings of second-last layer, with token embeddings of last layer
@@ -108,25 +111,7 @@ class DoubleSubstitutionEmbedding(nn.Module):
x_2[val_2 == 2] = y_1[val_1[:, ::8] != 0] # [N, S_2, E // 2]
# convolute substituted tokens of second-last layer
- x_out = self.convolution_2(x_2.contiguous()) # [N, S'_2, E]
-
- # filter out all tokens, that do not have any descendants in last layer
- mask_1 = (val_1.view(batch_size, -1, 8) == 2).max(dim=-1)[0]
- mask_2 = torch.zeros_like(val_2, dtype=torch.bool)
- mask_2[val_2 == 2] = mask_1
- mask_2 = mask_2.view(batch_size, -1, self.conv_size).max(dim=-1)[0]
- len_out = torch.max(torch.sum(mask_2, dim=-1)).item()
-
- x_masked = torch.zeros(batch_size, len_out, embedding.shape[2], dtype=torch.float, device=value.device)
- val_masked = torch.zeros((batch_size, len_out), dtype=torch.long, device=value.device)
- for i in range(batch_size):
- x_masked[i] = x_out[i, mask_2[i].nonzero().squeeze(-1)]
- val_masked[i] = val_2[:, ::self.conv_size][i, mask_2[i].nonzero().squeeze(-1)]
-
- # precompute padding mask
- self.mask = padding_mask(val_masked, device=value.device) # [N, S'_2, E]
-
- return x_masked
+ return self.convolution_2(x_2.contiguous()) # [N, S'_2, E]
def forward(self, value, depth, position):
""" Transform sequences into embedding space for the encoder.
diff --git a/modules/token_embedding/substitution_embedding.py b/modules/token_embedding/substitution_embedding.py
index b41f181..24f15b3 100644
--- a/modules/token_embedding/substitution_embedding.py
+++ b/modules/token_embedding/substitution_embedding.py
@@ -82,6 +82,8 @@ class SubstitutionEmbedding(nn.Module):
dep_0[i, :len_0[i]] = depth[i, len_1[i]:len_1[i] + len_0[i]]
pos_0[i, :len_0[i]] = position[i, len_1[i]:len_1[i] + len_0[i]]
+ # precompute padding mask
+ self.mask = padding_mask(val_1[:, ::self.conv_size], device=value.device)
# convolve embedded tokens of last layer
y_0 = self.convolution_0(x_0) # [N, T2', C]
@@ -90,22 +92,7 @@ class SubstitutionEmbedding(nn.Module):
x_1[val_1 == 2] = y_0[val_0[:, ::8] != 0] # [N, T1, C]
# convolve substituted tokens of penultimate layer
- x_out = self.convolution_1(x_1.contiguous()) # [N, T1', E]
-
- # filter out all tokens, that do not have any descendants in last layer
- mask = (val_1.view(batch_size, -1, self.conv_size) == 2).max(dim=-1)[0]
- len_out = torch.max(torch.sum(mask, dim=-1)).item()
-
- x_masked = torch.zeros(batch_size, len_out, embedding.shape[2], dtype=torch.float, device=value.device)
- val_masked = torch.zeros((batch_size, len_out), dtype=torch.long, device=value.device)
- for i in range(batch_size):
- x_masked[i] = x_out[i, mask[i].nonzero().squeeze(-1)]
- val_masked[i] = val_1[i, ::self.conv_size][mask[i].nonzero().squeeze(-1)]
-
- # precompute padding mask
- self.mask = padding_mask(val_masked, device=value.device)
-
- return x_masked
+ return self.convolution_1(x_1.contiguous()) # [N, T1', E]
def forward(self, value, depth, position):
""" Transform sequences into embedding space for the encoder.
--
GitLab