def lrs2charlm_checker():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = LRS2CharLM()
model.load_state_dict(
torch.load(args["TRAINED_LM_FILE"], map_location=device))
model.to(device)
inp = torch.tensor(args["CHAR_TO_INDEX"][" "] - 1)
initStateBatch = None
string = list()
for i in range(100):
inputBatch = inp.reshape(1, 1)
inputBatch = inputBatch.to(device)
model.eval()
with torch.no_grad():
outputBatch, finalStateBatch = model(inputBatch, initStateBatch)
outputBatch = torch.exp(outputBatch)
out = outputBatch.squeeze()
probs = out.tolist()
ix = np.random.choice(np.arange(len(probs)), p=probs / np.sum(probs))
char = args["INDEX_TO_CHAR"][ix + 1]
string.append(char)
inp = torch.tensor(ix)
initStateBatch = finalStateBatch
print("".join(string))
return
def main():
np.random.seed(args["SEED"])
torch.manual_seed(args["SEED"])
gpuAvailable = torch.cuda.is_available()
device = torch.device("cuda:1" if gpuAvailable else "cpu")
kwargs = {"num_workers":args["NUM_WORKERS"], "pin_memory":True} if gpuAvailable else {}
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
#declaring the test dataset and test dataloader
videoParams = {"videoFPS":args["VIDEO_FPS"]}
testData = LRS2Main("test", args["DATA_DIRECTORY"], args["MAIN_REQ_INPUT_LENGTH"], args["CHAR_TO_INDEX"], args["STEP_SIZE"],
videoParams)
testLoader = DataLoader(testData, batch_size=args["BATCH_SIZE"], collate_fn=collate_fn, shuffle=True, **kwargs)
if args["TRAINED_MODEL_FILE"] is not None:
print("\nTrained Model File: %s" %(args["TRAINED_MODEL_FILE"]))
#declaring the model, loss function and loading the trained model weights
model = VideoNet(args["TX_NUM_FEATURES"], args["TX_ATTENTION_HEADS"], args["TX_NUM_LAYERS"], args["PE_MAX_LENGTH"],
args["TX_FEEDFORWARD_DIM"], args["TX_DROPOUT"], args["NUM_CLASSES"])
model.load_state_dict(torch.load(args["CODE_DIRECTORY"] + args["TRAINED_MODEL_FILE"], map_location=device))
model.to(device)
loss_function = nn.CTCLoss(blank=0, zero_infinity=False)
#declaring the language model
lm = LRS2CharLM()
lm.load_state_dict(torch.load(args["TRAINED_LM_FILE"], map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
print("\nTesting the trained model .... \n")
beamSearchParams = {"beamWidth":args["BEAM_WIDTH"], "alpha":args["LM_WEIGHT_ALPHA"], "beta":args["LENGTH_PENALTY_BETA"],
"threshProb":args["THRESH_PROBABILITY"]}
testParams = {"decodeScheme":args["TEST_DEMO_DECODING"], "beamSearchParams":beamSearchParams, "spaceIx":args["CHAR_TO_INDEX"][" "],
"eosIx":args["CHAR_TO_INDEX"]["<EOS>"], "lm":lm}
#evaluating the model over the test set
testLoss, testCER, testWER = evaluate(model, testLoader, loss_function, device, testParams)
#printing the test set loss, CER and WER
print("Test Loss: %.6f || Test CER: %.3f || Test WER: %.3f" %(testLoss, testCER, testWER))
print("\nTesting Done.\n")
else:
print("Path to the trained model file not specified.\n")
return
def ctc_search_decode_checker():
outputs = [
"TTTEEEST-IINNNN-G- CC-TTCCC- -DEEE-CO-DD---E -FUU-NCCC--TAA-B-FA--E",
"ONNE SSSTEEEP ISSS OOOOVVEERA- FDDA-S A FD-AASDF - AD-AFA DF-ADF SF-ADF",
"EVERYTHING ALRIGHT CHECK DONE SH-SG-GAD-G HS- RA-R H J- J-AM GA-AM GA-GA-AD",
"SSEEEE-E-- -EEE-VE-NNN ---DDDOOOO-ODDE-E --O-OOOOTTTY AAAASS-SSAAM WORK",
"---------------------------------------------------------------------------"
]
inpLens = [64, 32, 29, 75, 56]
outputProbs = 0.01 * torch.ones(
(len(outputs[0]), len(inpLens), args["NUM_CLASSES"]))
inpLens = torch.tensor(inpLens)
for n in range(len(outputs)):
for i in range(len(outputs[n])):
char = outputs[n][i]
if char == "-":
ix = 0
else:
ix = args["CHAR_TO_INDEX"][char]
outputProbs[i, n, ix] = 1.5
outputLogProbs = torch.log(outputProbs)
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
lm = LRS2CharLM()
lm.load_state_dict(torch.load(args["TRAINED_LM_FILE"],
map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
predictions, predictionLens = ctc_search_decode(
outputLogProbs, inpLens, beamSearchParams, args["CHAR_TO_INDEX"][" "],
args["CHAR_TO_INDEX"]["<EOS>"], lm)
predictions = [
args["INDEX_TO_CHAR"][ix] for ix in predictions.tolist()
if ix != args["CHAR_TO_INDEX"]["<EOS>"]
]
predictedSequences = list()
s = 0
for ln in predictionLens.tolist():
predictedSequences.append("".join(predictions[s:s + ln - 1]))
s = s + ln - 1
print(predictedSequences)
return
print("\nTrained Model File: %s" % (args["TRAINED_MODEL_FILE"]))
#declaring the model,loss function and loading the trained model weights
model = AVNet(args["TX_NUM_FEATURES"], args["TX_ATTENTION_HEADS"],
args["TX_NUM_LAYERS"], args["PE_MAX_LENGTH"],
args["AUDIO_FEATURE_SIZE"], args["TX_FEEDFORWARD_DIM"],
args["TX_DROPOUT"], args["NUM_CLASSES"])
model.load_state_dict(
torch.load(args["CODE_DIRECTORY"] + args["TRAINED_MODEL_FILE"],
map_location=device))
model.to(device)
loss_function = nn.CTCLoss(blank=0, zero_infinity=False)
#declaring the language model
lm = LRS2CharLM()
lm.load_state_dict(torch.load(args["TRAINED_LM_FILE"],
map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
print("\nTesting the trained model .... \n")
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
if args["TEST_DEMO_MODE"] == "AO":
def main():
np.random.seed(args["SEED"])
torch.manual_seed(args["SEED"])
gpuAvailable = torch.cuda.is_available()
device = torch.device("cuda" if gpuAvailable else "cpu")
if args["TRAINED_MODEL_FILE"] is not None:
print("\nTrained Model File: %s" % (args["TRAINED_MODEL_FILE"]))
print("\nDemo Directory: %s" % (args["DEMO_DIRECTORY"]))
#declaring the model and loading the trained weights
model = VideoNet(args["TX_NUM_FEATURES"], args["TX_ATTENTION_HEADS"],
args["TX_NUM_LAYERS"], args["PE_MAX_LENGTH"],
args["TX_FEEDFORWARD_DIM"], args["TX_DROPOUT"],
args["NUM_CLASSES"])
model.load_state_dict(
torch.load(args["CODE_DIRECTORY"] + args["TRAINED_MODEL_FILE"],
map_location=device))
model.to(device)
#declaring the visual frontend module
vf = VisualFrontend()
vf.load_state_dict(
torch.load(args["TRAINED_FRONTEND_FILE"], map_location=device))
vf.to(device)
#declaring the language model
lm = LRS2CharLM()
lm.load_state_dict(
torch.load(args["TRAINED_LM_FILE"], map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
print("\n\nRunning Demo .... \n")
#walking through the demo directory and running the model on all video files in it
for root, dirs, files in os.walk(args["DEMO_DIRECTORY"]):
for file in files:
if file.endswith(".mp4"):
sampleFile = os.path.join(root, file[:-4])
#preprocessing the sample
params = {
"roiSize": args["ROI_SIZE"],
"normMean": args["NORMALIZATION_MEAN"],
"normStd": args["NORMALIZATION_STD"],
"vf": vf
}
preprocess_sample(sampleFile, params)
#converting the data sample into appropriate tensors for input to the model
visualFeaturesFile = os.path.join(root, file[:-4]) + ".npy"
videoParams = {"videoFPS": args["VIDEO_FPS"]}
inp, _, inpLen, _ = prepare_main_input(
visualFeaturesFile, None,
args["MAIN_REQ_INPUT_LENGTH"], args["CHAR_TO_INDEX"],
videoParams)
inputBatch, _, inputLenBatch, _ = collate_fn([
(inp, None, inpLen, None)
])
#running the model
inputBatch = (inputBatch.float()).to(device)
inputLenBatch = (inputLenBatch.int()).to(device)
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(
outputBatch, inputLenBatch,
args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
predictionBatch, predictionLenBatch = ctc_search_decode(
outputBatch, inputLenBatch, beamSearchParams,
args["CHAR_TO_INDEX"][" "],
args["CHAR_TO_INDEX"]["<EOS>"], lm)
else:
print("Invalid Decode Scheme")
exit()
#converting character indices back to characters
pred = predictionBatch[:][:-1]
pred = "".join(
[args["INDEX_TO_CHAR"][ix] for ix in pred.tolist()])
#printing the predictions
print("File: %s" % (file))
print("Prediction: %s" % (pred))
print("\n")
print("Demo Completed.\n")
else:
print("\nPath to trained model file not specified.\n")
return
def main():
np.random.seed(args["SEED"])
torch.manual_seed(args["SEED"])
gpuAvailable = torch.cuda.is_available()
device = torch.device("cuda" if gpuAvailable else "cpu")
if args["TRAINED_MODEL_FILE"] is not None:
print("\nTrained Model File: %s" % (args["TRAINED_MODEL_FILE"]))
print("\nDemo Directory: %s" % (args["DEMO_DIRECTORY"]))
#declaring the model and loading the trained weights
model = AVNet(args["TX_NUM_FEATURES"], args["TX_ATTENTION_HEADS"],
args["TX_NUM_LAYERS"], args["PE_MAX_LENGTH"],
args["AUDIO_FEATURE_SIZE"], args["TX_FEEDFORWARD_DIM"],
args["TX_DROPOUT"], args["NUM_CLASSES"])
model.load_state_dict(
torch.load(args["CODE_DIRECTORY"] + args["TRAINED_MODEL_FILE"],
map_location=device))
model.to(device)
#declaring the visual frontend module
vf = VisualFrontend()
vf.load_state_dict(
torch.load(args["TRAINED_FRONTEND_FILE"], map_location=device))
vf.to(device)
#declaring the language model
lm = LRS2CharLM()
lm.load_state_dict(
torch.load(args["TRAINED_LM_FILE"], map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
#reading the noise file
if args["TEST_DEMO_NOISY"]:
_, noise = wavfile.read(args["DATA_DIRECTORY"] + "/noise.wav")
else:
noise = None
print("\n\nRunning Demo .... \n")
rows = []
print(args['TEST_DEMO_MODE'])
#walking through the demo directory and running the model on all video files in it
for root, dirs, files in os.walk(args["DEMO_DIRECTORY"]):
for file in files:
if file.endswith(".mp4"):
sNum = file[7]
if file[13] <= "9" and file[13] >= "0":
cNum = file[12:14]
else:
cNum = file[12]
if file[-6] == "l":
cType = "jumble"
elif file[-6] == "s":
cType = "base"
else:
cType = file[-8:-4]
sampleFile = os.path.join(root, file[:-4])
#preprocessing the sample
params = {
"roiSize": args["ROI_SIZE"],
"normMean": args["NORMALIZATION_MEAN"],
"normStd": args["NORMALIZATION_STD"],
"vf": vf
}
preprocess_sample(sampleFile, params)
#converting the data sample into appropriate tensors for input to the model
audioFile = os.path.join(root, file[:-4]) + ".wav"
visualFeaturesFile = os.path.join(root, file[:-4]) + ".npy"
audioParams = {
"stftWindow": args["STFT_WINDOW"],
"stftWinLen": args["STFT_WIN_LENGTH"],
"stftOverlap": args["STFT_OVERLAP"]
}
videoParams = {"videoFPS": args["VIDEO_FPS"]}
inp, _, inpLen, _ = prepare_main_input(
audioFile, visualFeaturesFile, None, noise,
args["MAIN_REQ_INPUT_LENGTH"], args["CHAR_TO_INDEX"],
args["NOISE_SNR_DB"], audioParams, videoParams)
inputBatch, _, inputLenBatch, _ = collate_fn([
(inp, None, inpLen, None)
])
#running the model
inputBatch = ((inputBatch[0].float()).to(device),
(inputBatch[1].float()).to(device))
inputLenBatch = (inputLenBatch.int()).to(device)
if args["TEST_DEMO_MODE"] == "AO":
inputBatch = (inputBatch[0], None)
elif args["TEST_DEMO_MODE"] == "VO":
inputBatch = (None, inputBatch[1])
elif args["TEST_DEMO_MODE"] == "AV":
pass
else:
print("Invalid Operation Mode.")
exit()
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(
outputBatch, inputLenBatch,
args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
predictionBatch, predictionLenBatch = ctc_search_decode(
outputBatch, inputLenBatch, beamSearchParams,
args["CHAR_TO_INDEX"][" "],
args["CHAR_TO_INDEX"]["<EOS>"], lm)
else:
print("Invalid Decode Scheme")
exit()
#converting character indices back to characters
pred = predictionBatch[:][:-1]
pred = "".join(
[args["INDEX_TO_CHAR"][ix] for ix in pred.tolist()])
#printing the predictions
print("File: %s" % (file))
print("Speaker: " + sNum + " Clip: " + cNum +
" Clip Type: " + cType)
print("Prediction: %s" % (pred))
print("\n")
row = [sNum, cNum, cType, pred]
rows.append(row)
print("Demo Completed.\n")
with open("predictions.csv", "w", newline="") as file:
writer = csv.writer(file)
for x in range(len(rows)):
writer.writerow(rows[x])
else:
print("\nPath to trained model file not specified.\n")
return