recompile python

keon · Jul 14, 2019 · 8855eda · 8855eda
1 parent 49739df
commit 8855eda
Show file tree

Hide file tree

Showing 10 changed files with 50 additions and 710 deletions.
diff --git a/07-순차적인_데이터를_처리하는_RNN/02-sequence-to-sequence-lstm.ipynb b/07-순차적인_데이터를_처리하는_RNN/02-sequence-to-sequence-lstm.ipynb
diff --git a/07-순차적인_데이터를_처리하는_RNN/02-sequence-to-sequence-lstm.py b/07-순차적인_데이터를_처리하는_RNN/02-sequence-to-sequence-lstm.py
diff --git a/07-순차적인_데이터를_처리하는_RNN/02-sequence-to-sequence.py b/07-순차적인_데이터를_처리하는_RNN/02-sequence-to-sequence.py
@@ -9,6 +9,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+import random
 
 get_ipython().run_line_magic('matplotlib', 'inline')
 
@@ -51,18 +52,15 @@ def forward(self, inputs, targets):
 
         # 디코더 (Decoder)
         outputs = []
-        for i in range(targets.size()[0]): 
+
+        for i in range(targets.size()[0]):
             decoder_input = self.embedding(decoder_input)
             decoder_output, decoder_state = self.decoder(decoder_input, decoder_state)
+            projection = self.project(decoder_output)
+            outputs.append(projection)
 
-            # 디코더의 출력값으로 다음 글자 예측하기
-            projection = self.project(decoder_output.view(1, -1))  # batch x vocab_size
-            prediction = F.softmax(projection, dim=1)  # batch x vocab_size
-            outputs.append(prediction)
-
-            # 디코더 입력 갱신
-            _, top_i = prediction.data.topk(1)  # 1 x 1
-            decoder_input = top_i
+            #티처 포싱(Teacher Forcing) 사용
+            decoder_input = torch.LongTensor([[targets[i]]])
 
         outputs = torch.stack(outputs).squeeze()
         return outputs

diff --git a/07-순차적인_데이터를_처리하는_RNN/03-seq2seq_gru.py b/07-순차적인_데이터를_처리하는_RNN/03-seq2seq_gru.py
diff --git a/07-순차적인_데이터를_처리하는_RNN/assets/encoder_decoder.png b/07-순차적인_데이터를_처리하는_RNN/assets/encoder_decoder.png
diff --git a/07-순차적인_데이터를_처리하는_RNN/assets/pics b/07-순차적인_데이터를_처리하는_RNN/assets/pics
diff --git a/08-딥러닝_해킹하기/01-fgsm-attack.py b/08-딥러닝_해킹하기/01-fgsm-attack.py
@@ -31,7 +31,6 @@
 
 CLASSES = json.load(open('./imagenet_samples/imagenet_classes.json'))
 idx2class = [CLASSES[str(i)] for i in range(1000)]
-class2idx = {v:i for i,v in enumerate(idx2class)}
 
 
 # ## 이미지 불러오기
@@ -53,8 +52,8 @@
 
 
 # 시각화를 위해 넘파이 행렬 변환
-original_img_view = img_tensor.squeeze(0)  # [1, 3, 244, 244] -> [3, 244, 244]
-original_img_view = original_img_view.transpose(0,2).transpose(0,1).detach().numpy()
+original_img_view = img_tensor.squeeze(0).detach()  # [1, 3, 244, 244] -> [3, 244, 244]
+original_img_view = original_img_view.transpose(0,2).transpose(0,1).numpy()
 
 # 텐서 시각화
 plt.imshow(original_img_view)
@@ -86,7 +85,7 @@ def fgsm_attack(image, epsilon, gradient):
 # ## 적대적 예제 생성
 
 # 이미지의 기울기값을 구하도록 설정
-img_tensor.requires_grad = True
+img_tensor.requires_grad_(True)
 
 # 이미지를 모델에 통과시킴
 output = model(img_tensor)
@@ -121,7 +120,8 @@ def fgsm_attack(image, epsilon, gradient):
 
 
 # 시각화를 위해 넘파이 행렬 변환
-perturbed_data_view = perturbed_data.squeeze(0).transpose(0,2).transpose(0,1).detach().numpy()
+perturbed_data_view = perturbed_data.squeeze(0).detach()
+perturbed_data_view = perturbed_data_view.transpose(0,2).transpose(0,1).numpy()
 
 plt.imshow(perturbed_data_view)
 
@@ -130,11 +130,11 @@ def fgsm_attack(image, epsilon, gradient):
 
 f, a = plt.subplots(1, 2, figsize=(10, 10))
 
-# 원본 사진
+# 원본
 a[0].set_title(prediction_name)
 a[0].imshow(original_img_view)
 
-# 복원된 사진
+# 적대적 예제
 a[1].set_title(perturbed_prediction_name)
 a[1].imshow(perturbed_data_view)