Related
I am trying to retrain resnet50 from scratch using a dataset that is similar to ImageNet. I wrote the following training loop:
def train_network(epochs , train_loader , val_loader , optimizer , network):
since = time.time ( )
train_acc_history = []
val_acc_history = []
best_model_weights = copy.deepcopy (network.state_dict ( ))
best_accuracy = 0.0
for epoch in range (epochs):
correct_train = 0
correct_val = 0
for x , t in train_loader:
x = x.to (device)
t = t.to (device)
optimizer.zero_grad ( )
z = network (x)
J = loss (z , t)
J.backward ( )
optimizer.step ( )
_ , y = torch.max (z , 1)
correct_train += torch.sum (y == t.data)
with torch.no_grad ( ):
network.eval ( )
for x_val , t_val in val_loader:
x_val = x_val.to (device)
t_val = t_val.to (device)
z_val = network (x_val)
_ , y_val = torch.max (z_val , 1)
correct_val += torch.sum (y_val == t_val.data)
network.train ( )
train_accuracy = correct_train.float ( ) / len (train_loader.dataset)
val_accuracy = correct_val.float ( ) / len (val_loader.dataset)
print (
F"Epoch: {epoch + 1} train_accuracy: {(train_accuracy.item ( ) * 100):.3f}% val_accuracy: {(val_accuracy.item ( ) * 100):.3f}%" ,
flush = True)
# time_elapsed_epoch = time.time() - since
# print ('Time taken for Epoch {} is {:.0f}m {:.0f}s'.format (epoch + 1, time_elapsed_epoch // 60 , time_elapsed_epoch % 60))
if val_accuracy > best_accuracy:
best_accuracy = val_accuracy
best_model_weights = copy.deepcopy (network.state_dict ( ))
train_acc_history.append (train_accuracy)
val_acc_history.append (val_accuracy)
print ( )
time_elapsed = time.time ( ) - since
print ('Training complete in {:.0f}m {:.0f}s'.format (time_elapsed // 60 , time_elapsed % 60))
print ('Best Validation Accuracy: {:3f}'.format (best_accuracy * 100))
network.load_state_dict (best_model_weights)
return network , train_acc_history , val_acc_history
But I am getting extremely poor training and validation accuracies as below:
> Epoch: 1 train_accuracy: 3.573% val_accuracy: 3.481%
> Epoch: 2 train_accuracy: 3.414% val_accuracy: 3.273%
> Epoch: 3 train_accuracy: 3.515% val_accuracy: 4.039%
> Epoch: 4 train_accuracy: 3.567% val_accuracy: 4.195%
Upon googling, I found that the accuracies of training from scratch are usually not so poor (in fact they start off from around 40% - 50%). I am finding it difficult to understand where the glitch might be. It would be great if someone could help me figure out where I might be going wrong.
Thanks
I tried your training loop without the weight checkpoint and got accuracy over 90% on fashionMNIST dataset using my own ResNet. So if you are using a good loss/optimizer I would suggest looking at the network architecture or creation of the data-loaders.
def train_network(epochs , train_loader , val_loader , optimizer , network):
#since = time.time ( )
train_acc_history = []
val_acc_history = []
loss = nn.CrossEntropyLoss()
#best_model_weights = copy.deepcopy (network.state_dict ( ))
#best_accuracy = 0.0
for epoch in range (epochs):
correct_train = 0
correct_val = 0
network.train ( )
for x , t in train_loader:
x = x.to (device)
t = t.to (device)
optimizer.zero_grad ( )
z = network (x)
J = loss (z , t)
J.backward ( )
optimizer.step ( )
_ , y = torch.max (z , 1)
correct_train += torch.sum (y == t.data)
with torch.no_grad ( ):
network.eval ( )
for x_val , t_val in val_loader:
x_val = x_val.to (device)
t_val = t_val.to (device)
z_val = network (x_val)
_ , y_val = torch.max (z_val , 1)
correct_val += torch.sum (y_val == t_val.data)
network.train ( )
train_accuracy = correct_train.float ( ) / len (train_loader.dataset)
val_accuracy = correct_val.float ( ) / len (val_loader.dataset)
print (
F"Epoch: {epoch + 1} train_accuracy: {(train_accuracy.item ( ) * 100):.3f}% val_accuracy: {(val_accuracy.item ( ) * 100):.3f}%" ,
flush = True)
'''
if val_accuracy > best_accuracy:
best_accuracy = val_accuracy
best_model_weights = copy.deepcopy (network.state_dict ( ))
train_acc_history.append (train_accuracy)
val_acc_history.append (val_accuracy)
#time_elapsed = time.time ( ) - since
#print ('Training complete in {:.0f}m {:.0f}s'.format (time_elapsed // 60 , time_elapsed % 60))
print ('Best Validation Accuracy: {:3f}'.format (best_accuracy * 100))
#network.load_state_dict (best_model_weights)
'''
return network , train_acc_history , val_acc_history
optimizer = optim.Adam(net.parameters(), lr = 0.01)
train_network(10,trainloader, testloader, optimizer, net)
Epoch: 1 train_accuracy: 83.703% val_accuracy: 86.820%
Epoch: 2 train_accuracy: 88.893% val_accuracy: 89.400%
Epoch: 3 train_accuracy: 90.297% val_accuracy: 89.700%
Epoch: 4 train_accuracy: 91.272% val_accuracy: 90.640%
Epoch: 5 train_accuracy: 91.948% val_accuracy: 91.250%
...
So, if you tested with the training loop I used (yours with small mods) and it still doesn't work I would check data loader and play around with network architecture.
Hi I'm brand new in coding, and I am getting stuck every new line of code I try to write but hey its a learning process.
I'm doing a strategy based on the MACD variables.
-MACD_Line is either positive or negative.
-Signal_Line is either positive or negative.
-Histogram is either positive or negative.
Based on the historical prices there are 6 possibilities of combined signals either: ---, -++, --+, +--, ++- or +++.
What I want to do is pre-set different position sizes depending on these 6 possible output signals.
So for example: if "---" then short 50% of equity,
if "+++" then long 100% of equity,
if "-++" then short 25% of equity.
Therefore the equity position would change after one of the initial 3 variables changes.
My attempt:
strategy("HIST", overlay= false, initial_capital = 1, default_qty_type= strategy.percent_of_equity, default_qty_value= 100 )
//time inputs
startDate = input(title="Start Date", type=input.integer, defval=1, minval=1, maxval=31)
startMonth = input(title="Start Month", type=input.integer, defval=1, minval=1, maxval=12)
startYear = input(title="Start Year", type=input.integer, defval=2014, minval=1800, maxval=2100)
endDate = input(title="End Date", type=input.integer, defval=29, minval=1, maxval=31)
endMonth = input(title="End Month", type=input.integer, defval=3, minval=1, maxval=12)
endYear = input(title="End Year", type=input.integer, defval=2021, minval=1800, maxval=2100)
inDateRange = (time >= timestamp(syminfo.timezone, startYear, startMonth, startDate, 0, 0)) and
(time < timestamp(syminfo.timezone, endYear, endMonth, endDate, 0, 0))
//variable
ema26= ema(close,26)
ema12= ema(close,12 )
macdl= ema12-ema26
signal= ema(macdl, 9)
hist= macdl-signal
enterLong = crossover(macdl,0)
enterShort = crossunder(macdl,0)
s000 = if (hist <= 0 and macdl <= 0 and signal <=0)
s001 = if (hist > 0 and macdl <= 0 and signal <= 0)
s011 = if (hist > 0 and macdl > 0 and signal <= 0)
s111 = if (hist > 0 and macdl > 0 and signal > 0)
s011 = if (hist <= 0 and macdl > 0 signal > 0)
s001 = if (hist <= 0 and macdl <= 0 signal > 0)
if (inDateRange and s111)
strategy.entry(id="+", long=true)
if (inDateRange and s000)
strategy.entry(id="-", long=false)
if (not inDateRange)
strategy.close_all()
This should get you started in the right direction. You'll need to finish coding all the conditions yourself. See comments in code:
//#version=4
strategy("HIST", overlay= false, initial_capital = 1, default_qty_type= strategy.percent_of_equity, default_qty_value= 100 )
//time inputs
startDate = input(title="Start Date", type=input.integer, defval=1, minval=1, maxval=31)
startMonth = input(title="Start Month", type=input.integer, defval=1, minval=1, maxval=12)
startYear = input(title="Start Year", type=input.integer, defval=2014, minval=1800, maxval=2100)
endDate = input(title="End Date", type=input.integer, defval=29, minval=1, maxval=31)
endMonth = input(title="End Month", type=input.integer, defval=3, minval=1, maxval=12)
endYear = input(title="End Year", type=input.integer, defval=2021, minval=1800, maxval=2100)
inDateRange = (time >= timestamp(syminfo.timezone, startYear, startMonth, startDate, 0, 0)) and
(time < timestamp(syminfo.timezone, endYear, endMonth, endDate, 0, 0))
//variable
ema26= ema(close,26)
ema12= ema(close,12 )
macdl= ema12-ema26
signal= ema(macdl, 9)
hist= macdl-signal
enterLong = crossover(macdl,0)
enterShort = crossunder(macdl,0)
// Two last var names clashed with others, so used "X" in them.
s000 = hist <= 0 and macdl <= 0 and signal <= 0
s001 = hist > 0 and macdl <= 0 and signal <= 0
s011 = hist > 0 and macdl > 0 and signal <= 0
s111 = hist > 0 and macdl > 0 and signal > 0
s01X = hist <= 0 and macdl > 0 and signal > 0
s00X = hist <= 0 and macdl <= 0 and signal > 0
// Detect changes in conditions.
f_changeIn(_cond) => _cond and not _cond[1]
c000 = f_changeIn(s000)
c001 = f_changeIn(s001)
c011 = f_changeIn(s011)
c111 = f_changeIn(s111)
c01X = f_changeIn(s01X)
c00X = f_changeIn(s00X)
// Functions calculates position size from a % (0 - 1.0).
f_positionSize(_percentEquity) => strategy.equity * _percentEquity / close
// Generate orders on trasitions into conditions.
float positionSize = na
if inDateRange
if c000
positionSize := f_positionSize(0.5)
strategy.entry(id="+", long=false, qty = positionSize)
else if c011
positionSize := f_positionSize(1.0)
strategy.entry(id="+", long=false, qty = positionSize)
else if c111
positionSize := f_positionSize(1.0)
strategy.entry(id="+", long=true, qty = positionSize)
else
strategy.close_all()
// For debugging.
plot(positionSize, "Position size", color.orange, 2, plot.style_circles)
I have a list of timestamp tuples of the form List((startTime,endTime)), which are basically denoting periods of time throughout the day.
For example:
(("2016-03-28 14:00:00","2016-03-28 15:00:00"),
("2016-03-28 17:00:00","2016-03-28 21:00:00"),
("2016-03-28 01:00:00","2016-03-28 13:00:00"))
I want to get a list of the durations which are not included in this list.
So the output should be:
(("2016-03-28 00:00:00","2016-03-28 01:00:00"),
("2016-03-28 13:00:00","2016-03-28 14:00:00"),
("2016-03-28 15:00:00","2016-03-28 17:00:00"),
("2016-03-28 17:00:00","2016-03-28 24:00:00"))
Can anyone suggest a good and efficient way of doing that in Scala?
The naive solution that I've tried so far is as follows:
import java.sql.Timestamp
import scala.collection.mutable.ListBuffer
def comparator(first: (Timestamp,Timestamp), second: (Timestamp, Timestamp)) = first._2.getTime <= second._1.getTime
val data = List((Timestamp.valueOf("2016-03-28 00:00:00"),Timestamp.valueOf("2016-03-28 10:00:00")),
(Timestamp.valueOf("2016-03-28 12:00:00"),Timestamp.valueOf("2016-03-28 15:00:00")),
(Timestamp.valueOf("2016-03-28 23:00:00"),Timestamp.valueOf("2016-03-28 23:59:59")),
(Timestamp.valueOf("2016-03-28 16:00:00"),Timestamp.valueOf("2016-03-28 21:00:00"))
).sortWith(comparator)
var emptySlots = new ListBuffer[(Timestamp,Timestamp)]()
var currTime = Timestamp.valueOf("2016-03-28 00:00:00")
var index = 0
var cond = 0
while(cond == 0){
if (currTime.compareTo(Timestamp.valueOf("2016-03-28 23:59:59")) < 0 && index >= data.size){
emptySlots += ((currTime,Timestamp.valueOf("2016-03-28 23:59:59") ))
cond = 1
}
else if(index >= data.size)
{
cond = 1
}
else if(currTime.compareTo(data(index)._1) < 0) {
emptySlots += ((currTime, data(index)._1))
currTime = data(index)._2
index += 1
}
else if(currTime.compareTo(data(index)._1) >= 0 && currTime.compareTo(data(index)._2) < 0 ) {
currTime = data(index)._2
index += 1
}
else if(currTime.compareTo(data(index)._1) > 0 && currTime.compareTo(data(index)._2) > 0 ) {
index += 1
}
}
emptySlots.toList
I am creating a Sudoku solver in C++ while implementing Lua scripting for the actual solving of the puzzle. I have created the following Lua code, but get a
PANIC: unprotected error in call to Lua API (attempt to call a nil value)
error whenever my C++ code reaches the first instance of lua_call.
When compiling the code in SciTE, I get the following error:
lua: SudokuSolver.lua:99: 'end' expected (to close 'for' at line 61)
near ''
Adding three 'end's to the end of the function that has the for loop at line 61 clears that error, but causes errors in the C++ program. Can someone please look at my Lua and see if there's any syntax errors or other issues which may be causing this? Thank you
CODE
-- Table Declaration
SudokuGrid = {}
function RecieveGrid ( _Pos, _Value )
-- Recives the cell value at _Pos position from C++
SudokuGrid[_Pos] = _Value
end
function SolveSudoku ( _Pos )
-- Recursive function which solves the sudoku puzzle
local iNewValue = 1
-- If Position is 82+, all cells are solved
if( _Pos >= 82 ) then
return true
end
-- If Position already has a value
if( SudokuGrid[_Pos] ~= 0) then
return SolveSudoku( _Pos + 1 )
else
while(true) do
SudokuGrid[_Pos] = iNewValue
iNewValue = iNewValue + 1
-- If the new value of the cell is higher than 9 its not valid
if( SudokuGrid[_Pos] > 9 ) then
--Reset value
SudokuGrid[_Pos] = 0
return false
end
if( IsValid( _Pos ) and SolveSudoku( _Pos + 1 ) ) then
return true
end
end
end
end
function IsValid ( _Pos )
-- Calculate Column and Row in Grid
x = _Pos % 9
if( x == 0 ) then
x = 9
end
y = math.ceil(_Pos / 9)
-- Check Rows
for i=1, 9 do
CheckVal = ((y - 1) * 9) + i
if( CheckVal == _Pos ) then
-- Do nothing
else if ( SudokuGrid[_Pos] == SudokuGrid[CheckVal]and SudokuGrid[_Pos] ~= 0 ) then
return false
else
-- Do nothing
end
end
-- Check Columns
for i=1, 9 do
CheckVal = ((i - 1) * 9) + x
if( CheckVal == _Pos ) then
-- Do nothing
else if ( SudokuGrid[_Pos] == SudokuGrid[CheckVal] and SudokuGrid[_Pos] ~= 0 ) then
return false
else
-- Do nothing
end
end
-- Check 3X3 Grid
SquareCol = math.ceil(x/3)
SquareRow = math.ceil(y/3)
StartVal = (SquareCol - 1) * 27 + (SquareRow * 3) -2
for j=0, 2 do
for i=0, 2 do
CheckVal = StartVal + i
if( CheckVal == _Pos ) then
-- Do nothing
else if ( SudokuGrid[_Pos] == SudokuGrid[CheckVal] and SudokuGrid[_Pos] ~= 0 ) then
return false
else
-- Do nothing
end
end
StartVal = StartVal + 9
end
return true
end
function SendGrid ( _Pos )
-- Sends the value at _Pos to C++
return SudokuGrid[_Pos]
end
The syntax error is in all lines containing else if:
else if ( SudokuGrid[_Pos] == SudokuGrid[CheckVal]and SudokuGrid[_Pos] ~= 0 ) then
In Lua, use elseif instead. Using else if would need more closing end.
elseif SudokuGrid[_Pos] == SudokuGrid[CheckVal] and SudokuGrid[_Pos] ~= 0 then
Just a basic Casaer Cipher. I've tested all of the sub functions, just encryptChar() does not particularly work. I get an infinite loop. It's supposed to be recursive. Here's the all code:
fun replace (str : string, index : int, newChar : char) : string = String.substring(str,0,index) ^ String.str(newChar) ^ String.substring(str,index+1,(size str) - index - 1;
fun encryptChar (msgStr : string, shiftAmnt : int, index : int) : string =
let val asciiCode = 0
in
if (not (String.sub(msgStr, index) = #" ")) then
(
asciiCode = ord( String.sub(msgStr, index) ) + shiftAmnt;
if (asciiCode < ord(#"A")) then asciiCode = asciiCode + 26
else if (asciiCode > ord(#"Z")) then asciiCode = asciiCode - 26
else asciiCode = asciiCode;
msgStr = replace(msgStr, index, chr(asciiCode))
)
else asciiCode = asciiCode;
index = index + 1;
if (index < (size msgStr - 1)) then encryptChar(msgStr, shiftAmnt, index)
else msgStr
end
;
fun encrypt(msgStr : string, shiftAmnt : int) : string = encryptChar (String.map Char.toUpper msgStr, shiftAmnt mod 26, 0);
The problem here is that you're misusing =. Outside of a variable definition, = is simply a boolean function which checks its arguments for equality. So if you do for example asciiCode = ord( String.sub(msgStr, index) ) + shiftAmnt;, it will simply return false (because asciiCode is not equal to ord( String.sub(msgStr, index) ) + shiftAmnt) and then throw that result away (because you have additional expressions after the ;). It will not reassign asciiCode.
Variables in SML are immutable. If you want to emulate mutable variables you can use refs and the := operator. However I would not recommend that approach as it is generally not good functional style and not necessary in this case. The preferable approach would be to rewrite the code in a way that each variable is only assigned once.
This is very basic indeed, and it's surprising that you ran into it in such a complicated situation.
Did you port this from some other language?
You need to forget everything you know about programming using assignments.
let val x = y in something
means more or less "within 'something', replace the identifier 'x' with the value of 'y'".
There is no way for you to change the value of x.
Do the substitution (this is not the actual evaluation order or anything, but it should give you an idea of what's going on):
encryptChar("THIS", amount, 0)
=>
let val asciiCode = 0
in
if (not (String.sub("THIS", 0) = #" ")) then
(
asciiCode = ord( String.sub("THIS", 0) ) + amount;
if (asciiCode < ord(#"A")) then asciiCode = asciiCode + 26
else if (asciiCode > ord(#"Z")) then asciiCode = asciiCode - 26
else asciiCode = asciiCode;
"THIS" = replace("THIS", 0, chr(asciiCode))
)
else asciiCode = asciiCode;
0 = 0 + 1;
if (0 < (size "THIS" - 1)) then encryptChar("THIS", amount, 0)
else str
end ;
=>
if (not (String.sub("THIS", 0) = #" ")) then
(
0 = ord( String.sub("THIS", 0) ) + amount;
if (0 < ord(#"A")) then 0 = 0 + 26
else if (0 > ord(#"Z")) then 0 = 0 - 26
else 0 = 0;
"THIS" = replace("THIS", 0, chr(0))
)
else 0 = 0;
0 = 0 + 1;
if (0 < (size "THIS" - 1)) then encryptChar("THIS", amount, 0)
else str
=>
if (not (String.sub("THIS", 0) = #" ")) then
(
0 = ord( String.sub("THIS", 0) ) + amount;
if true then false
else if false then false
else true;
false
)
else true;
false;
if (0 < (size "THIS" - 1)) then encryptChar("THIS", amount, 0)
else "this"
->
if (not false) then
(
false;
false;
false
)
else true;
false;
if true then encryptChar("THIS", amount, 0)
else "THIS"
=>
(
false;
false;
false
)
false;
encryptChar("THIS", amount, 0)
=>
encryptChar("THIS", amount, 0)
Which is where your infinite loop came from.
You would do well to get hold of an introductory text about ML programming.