Sequence learner Load error

[vaibhavch]

    b0 = BlankBlock(num_s=hgt.num_bits)
    sl = SequenceLearner(num_spc=10, num_dps=10,num_rpd=10, d_thresh=6, perm_thr=1, perm_inc=1, perm_dec=0) 

    sl.input.add_child(b0.output)
    #after some computations
    sl.save(file_str='mysequence.bin')

    sl.load(file_str='mysequence.bin')
    print(sl.get_historical_count())

when I load saved file, error happens.

Error: 0

[vaibhavch]

it seems there is an issue loading next available coincidence detector if (fread(sl->s_next_d, sl->num_s * sizeof(uint32_t), 1, fptr) == 0) { printf("Error:\n"); // TODO }

[ddigiorg]

I'm getting the same error. Looking into it now.

[vaibhavch]

@ddigiorg did you find the problem?

[ddigiorg]

Got it fixed and we're going through our process to update the repo. Stay tuned.

[vaibhavch]

@ddigiorg Good to hear. if the process takes time, can you show me quick fix for now?

[ddigiorg]

In src/bbcore/sequence_learner.c please update the save and load functions:

// =============================================================================
// Save
// =============================================================================
void sequence_learner_save(struct SequenceLearner* sl, const char* file) {
    FILE *fptr;

    // check if file can be opened
    if ((fptr = fopen(file,"wb")) == NULL) {
       printf("Error in sequence_learner_save(): cannot open file");
       exit(1);
    }

    // check if block has been initialized
    if (sl->init_flag == 0) {
        printf("Error in sequence_learner_save(): block not initialized\n");
    }

    struct CoincidenceSet* cs;

    // save hidden coincidence detector receptor addresses and permanences
    for (uint32_t d = 0; d < sl->num_d; d++) {
        cs = &sl->d_hidden[d];
        fwrite(cs->addrs, cs->num_r * sizeof(cs->addrs[0]), 1, fptr);
        fwrite(cs->perms, cs->num_r * sizeof(cs->perms[0]), 1, fptr);
    }

    // save output coincidence detector receptor addresses and permanences
    for (uint32_t d = 0; d < sl->num_d; d++) {
        cs = &sl->d_output[d];
        fwrite(cs->addrs, cs->num_r * sizeof(cs->addrs[0]), 1, fptr);
        fwrite(cs->perms, cs->num_r * sizeof(cs->perms[0]), 1, fptr);        
    }

    // save next available coincidence detector on each statelet
    fwrite(sl->s_next_d, sl->num_s * sizeof(sl->s_next_d[0]), 1, fptr);

    fclose(fptr);
}

// =============================================================================
// Load
// =============================================================================
void sequence_learner_load(struct SequenceLearner* sl, const char* file) {
    FILE *fptr;

    // check if file can be opened
    if ((fptr = fopen(file,"rb")) == NULL) {
       printf("Error in sequence_learner_load(): cannot open file\n");
       exit(1);
    }

    // check if block has been initialized
    if (sl->init_flag == 0) {
        printf("Error in sequence_learner_load(): block not initialized\n");
    }

    struct CoincidenceSet* cs;

    // load hidden coincidence detector receptor addresses and permanences
    for (uint32_t d = 0; d < sl->num_d; d++) {
        cs = &sl->d_hidden[d];
        fread(cs->addrs, cs->num_r * sizeof(cs->addrs[0]), 1, fptr);
        fread(cs->perms, cs->num_r * sizeof(cs->perms[0]), 1, fptr);
    }

    // load output coincidence detector receptor addresses and permanences
    for (uint32_t d = 0; d < sl->num_d; d++) {
        cs = &sl->d_output[d];
        fread(cs->addrs, cs->num_r * sizeof(cs->addrs[0]), 1, fptr);
        fread(cs->perms, cs->num_r * sizeof(cs->perms[0]), 1, fptr);        
    }

    // load next available coincidence detector on each statelet
    fread(sl->s_next_d, sl->num_s * sizeof(sl->s_next_d[0]), 1, fptr);

    fclose(fptr); 
}

[vaibhavch]

now I need to do sl.initialize() before sl.load()to make it work . it is weird that anomaly score is different every time with same reset sequence though learning disabled sl.compute(learn=False)

[ddigiorg]

I will update the source code to automatically initialize the block when sl.load() is called. Can you share the code in your script? I'd like to help debug it or fix any BrainBlocks issues if necessary.

[vaibhavch]

train.py

hgt = HyperGridTransform(num_grids=100, num_bins=32, num_subspace_dims=1)
b0 = BlankBlock(num_s=hgt.num_bits)
sl = SequenceLearner(num_spc=10, num_dps=10, num_rpd=10, d_thresh=6, perm_thr=1, perm_inc=1, perm_dec=0) 
sl.input.add_child(b0.output)
df = prepare_data() # this function returns a pandas data frame consisting 30 columns
idf =  idf.drop('signal', 1) # this column is a label so we drop it for input

for index in range(10,len(df)-10):
   edf = pd.DataFrame() 
   if int(df['signal'].iloc[index]) == 1:
      edf = edf.append(idf[index-10:index]) # collecting 10 row sequences based on signal column
   else:
    continue

   X = edf.to_numpy()
   hgt.fit(X)
   X_bits = hgt.transform(X)

   for k in range(len(X_bits)):
      X_array = X_bits[k, :].flatten()
      b0.output.bits = X_array
      sl.compute(learn=True)
   b0.clear()
   sl.clear()
sl.save(file_str='sl.bin')

test.py

hgt = HyperGridTransform(num_grids=100, num_bins=32, num_subspace_dims=1)
b0 = BlankBlock(num_s=hgt.num_bits)
sl = SequenceLearner(num_spc=10, num_dps=10, num_rpd=10, d_thresh=6, perm_thr=1, perm_inc=1, perm_dec=0) 
sl.input.add_child(b0.output)
sl.initialize()
sl.load(file_str='sl.bin')

edf = df[17970:17980]  # this is one of the sequence used in training
print(edf) 
X = edf.to_numpy()
X_bits = hgt.transform(X)

for k in range(len(X_bits)):
  X_array = X_bits[k, :].flatten()
  b0.output.bits = X_array
  sl.compute(learn=False)
  if k == len(X_bits) - 1:
    score = sl.get_score()   #predicting score on last row
    print(score)

if we run test.py, anomaly score is different every time. same sequence, different anomaly every run

[jacobeverist]

At least one source of your difficulty would be the HyperGridTransform being newly created for each script. Currently we don't have a feature to save its configuration. So whenever you create one, it's completely different.

One thing you can do is pass in a seed for the random number generator. For the constructor, add the following keyword argument:

random_state=0

This ensures that the randomly created configuration is the same for both train and test scripts.

Also, make sure you "fit()" the HGT on the same data for each case. That way you know it is configured identically.

[jacobeverist]

@vaibhavch Did making the recommended changes to the HGT result in more consistent performance?

[vaibhavch]

Thanks, passing the seed solved the issue.