Add files via upload

Author: meeeeee

prepare.py

@@ -19,18 +19,20 @@
 train_lst, eval_lst = list(), list()
 font = ImageFont.truetype(random.choice(font_list)+".ttf",14)
 for i in range(args.train_num):
-    img = Image.new("RGBA", (110,20),(255,255,255))
-    word = ''.join(random.choice(string.ascii_letters) for i in range(random.randrange(5,10)))
+    img = Image.new("RGBA", (20,20),(255,255,255))
+    word = ''.join(random.choice(string.ascii_letters))
     ImageDraw.Draw(img).text((5, 0), word, (0,0,0), font=font)
     img = np.array(img.convert("L"), dtype=np.float32)
     train_lst.append((img, img + img * np.random.normal(0,1,img.size).reshape(img.shape[0], img.shape[1]).astype('uint8')))
+    if i%1000 == 0: print("Train image range ", i)
 
 for i in range(args.eval_num):
     img = Image.new("RGBA", (110,20),(255,255,255))
     word = ''.join(random.choice(string.ascii_letters) for i in range(random.randrange(5,10)))
     ImageDraw.Draw(img).text((5, 0), word, (0,0,0), font=font)
     img = np.array(img.convert("L"), dtype=np.float32)
     eval_lst.append((img, img + img * np.random.normal(0,1,img.size).reshape(img.shape[0], img.shape[1]).astype('uint8')))
+    if i%1000 == 0: print("Test image range ", i)
 out1, out2 = open(args.train_path, "wb"), open(args.eval_path, "wb")
 pickle.dump(train_lst, out1), pickle.dump(eval_lst, out2)
 out1.close(), out2.close()

train.py

@@ -72,7 +72,7 @@ def tensor_to_image(tensor):
             for data in train_dataloader:
                 inputs, labels = data[0].to(device), data[1]
                 preds = model(inputs)
-                tensor_to_image(preds[0].detach().numpy().reshape((20, 110))).save("1.jpg")
+                tensor_to_image(preds[0].detach().numpy().reshape((20, 20))).save("1.jpg")
                 loss_f = nn.MSELoss()
                 loss = loss_f(preds, labels)
                 epoch_losses.update(loss.item(), len(inputs))
@@ -89,5 +89,6 @@ def tensor_to_image(tensor):
     for data in eval_dataloader:
         inputs, labels = data[0].to(device), data[1]
         with torch.no_grad(): preds = model(inputs).clamp(0.0, 255.0)
-        tensor_to_image(preds[0].reshape((20, 110))).save(args.outputs_dir + str(eval_counter) + ".jpg")
+        tensor_to_image(preds[0].reshape((20, 20))).save(args.outputs_dir + str(eval_counter) + ".jpg")
         eval_counter+=1
+# 0.000370268038*1000*200 \approx 37%